Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Understanding, modeling and predicting trophic interactions between marine species Valls, Audrey Emilie 2014

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2015_february_valls_audrey.pdf [ 4.6MB ]
Metadata
JSON: 24-1.0167272.json
JSON-LD: 24-1.0167272-ld.json
RDF/XML (Pretty): 24-1.0167272-rdf.xml
RDF/JSON: 24-1.0167272-rdf.json
Turtle: 24-1.0167272-turtle.txt
N-Triples: 24-1.0167272-rdf-ntriples.txt
Original Record: 24-1.0167272-source.json
Full Text
24-1.0167272-fulltext.txt
Citation
24-1.0167272.ris

Full Text

 UNDERSTANDING, MODELING AND PREDICTING TROPHIC INTERACTIONS BETWEEN MARINE SPECIES  by  Audrey Emilie Valls  A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF  THE REQUIREMENTS FOR THE DEGREE OF  DOCTOR OF PHILOSOPHY  in  The Faculty of Graduate and Postdoctoral Studies (Zoology)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)  December 2014  © Audrey Emilie Valls, 2014 ii  Abstract In this thesis, I explore and propose new methodologies and intend to provide new insights for understanding, modeling and predicting trophic interactions between marine species. I introduce the importance of trophic interactions in marine ecosystems and the common ecosystem modeling approaches applied to marine food webs. First, I present an attempt to synthetize information from all published Ecopath with Ecosim (EwE) models worldwide, gathered in the EcoBase digital repository. Through the compilation and standardization of a suite of metadata, I describe and discuss the usage of the EwE modeling approach and its evolution over time, since its very first application. I also present a meta-analysis of Ecopath models based on these metadata, where I select the models of potential interest, using a scoring method, and focus on one particular aspect of food web modeling, relating to the identification of keystone species. I propose a comprehensive and critical review of the ill-defined concept of keystone species and argue for a restored, exclusive and operational definition of the concept. The proposed definition is placed in a larger framework that considers different categories of ecologically important species. Then, a new functional index of keystoneness is derived from the EwE modeling approach, so as to identify potential keystone species in marine food webs. The proposed index addresses some of the biases observed in previously applied indices. Finally, I present an attempt to predict diet composition for predatory fish species. The intention is to build on existing large datasets, provided in the FishBase biodiversity information system, to identify predictors of fish feeding selectivity. The determination of clear or consistent pattern between biological and ecological species straits and diet composition is challenged by data restrictions, but some recommendations for future studies are provided. In conclusion, data availability may be a critical issue when considering some aspects of trophic interactions, especially for modeling and predictions at the species level. Data sharing within the scientific community, notably through the use of digital and open-access information repositories, is critical for the development of global meta-analyses in marine ecology.   iii  Preface With the exception of Chapters 1 and 6, all of the chapters in this dissertation have been prepared as stand-alone, peer-reviewed publications. Chapters 2 and 3 are under review at refereed journals, while Chapter 4 is in press. Chapter 5 is in preparation for submission. I am the senior author on all of the papers, and I led the design, implementation, analysis and writing of all co-authored chapters. Details of co-authorship contributions are stated below.   Chapter 2 is co-authored by Mathieu Colléter, Jérôme Guitton, Didier Gascuel, Daniel Pauly, and Villy Christensen. Mathieu Colléter and Didier Gascuel initiated the project of EcoBase. I initiated this study, following my collaboration with Mathieu Colléter on the production of a scientific report introducing EcoBase. I share first authorship with Mathieu Colléter on this study, to which we equally contributed regarding the design of the study, collection of the information, analysis of the data and writing of the manuscript. Jérôme Guitton assisted with technical issues with the development and use of the EcoBase database. Didier Gascuel, Daniel Pauly, and Villy Christensen provided edits, comments, and suggestions on the manuscript. A version of this chapter is under review.  Chapter 3 is co-authored by Marta Coll and Villy Christensen. Marta Coll suggested the idea of a literature review. I expanded on the original idea and conducted a more in-depth review than originally envisaged. I reviewed the literature and wrote the manuscript. Marta Coll contributed to this chapter through stimulating discussions and suggestions on drafts of the manuscript. Villy Christensen provided guidance throughout the development of the manuscript. A version of this chapter is under review.  Chapter 4 is co-authored by Marta Coll and Villy Christensen. The original idea and its conceptualization for this chapter are mine. I developed the methods, carried out all analyses and prepared the manuscript. Marta Coll provided data, methodological guidance on statistical analyses, and contributed to this chapter through stimulating discussions and suggestions on drafts of the manuscript. Villy Christensen provided guidance throughout the development of the manuscript. A version of this chapter is in press in Ecological Monographs.  iv   Chapter 5 is co-authored by Deng Palomares and Villy Christensen. I designed the study, conducted the analyses and prepared the manuscript. Deng Palomares assisted with the use of the FishBase database, and contributed to this chapter through stimulating discussions and comments on drafts of the manuscript. Villy Christensen suggested the original idea of this chapter and provided guidance throughout the development of the manuscript. A version of this chapter is in preparation for submission.    v  Table of Contents Abstract .......................................................................................................................................... iiPreface ........................................................................................................................................... iiiTable of Contents .......................................................................................................................... vList of Tables ................................................................................................................................ ixList of Figures ................................................................................................................................ xAcknowledgements .................................................................................................................... xiiiDedication .................................................................................................................................... xv1. Introduction ........................................................................................................................... 11.1. Context ............................................................................................................................. 11.1.1. Drivers of changes in marine ecosystems ................................................................. 21.1.2. Impacts of fishing on marine ecosystems ................................................................. 31.1.3. Effects of climate change on marine ecosystems ..................................................... 41.1.4. Consequences on marine ecosystems functions and services ................................... 61.2. Concepts and definitions .................................................................................................. 71.2.1. Marine biodiversity ................................................................................................... 71.2.2. Trophic interactions .................................................................................................. 81.2.3. Diet composition ..................................................................................................... 101.3. Methods and approaches ................................................................................................ 121.3.1. Ecosystem-based management ............................................................................... 121.3.2. Ecosystem models ................................................................................................... 131.3.3. Ecopath with Ecosim .............................................................................................. 151.3.4. EcoBase................................................................................................................... 181.4. Research questions and objectives ................................................................................. 192. Global overview of the applications of the Ecopath with Ecosim modeling approach using the EcoBase information repository ................................................................................ 222.1. Introduction .................................................................................................................... 22vi  2.2. Methods .......................................................................................................................... 252.2.1. Compilation of the models and publications .......................................................... 252.2.2. Compilation of the metadata ................................................................................... 262.3. Results ............................................................................................................................ 292.3.1. General description of the EwE applications published worldwide ....................... 292.3.2. Evolution of the EwE applications over the last three decades .............................. 312.4. Discussion ...................................................................................................................... 332.4.1. Challenges and potentials of the compilation and meta-analysis of EwE models .. 332.4.2. Conclusions on the EwE modeling approach: past, present and future .................. 353. Proposal for a restored definition of the concept of ‘keystone species’ .......................... 443.1. Introduction .................................................................................................................... 443.2. Methods .......................................................................................................................... 463.3. Results ............................................................................................................................ 473.3.1. Evolution of the keystone species concept ............................................................. 473.3.2. Different categories of critical species, including keystone species ....................... 503.3.3. Examples of sensu stricto keystone versus pseudo-keystone species .................... 553.3.4. A restrictive definition for the keystone species concept ....................................... 563.4. Discussion ...................................................................................................................... 573.4.1. A restored concept of keystone species .................................................................. 573.4.2. Context-dependency issues with the keystone species concept .............................. 603.4.3. ‘Keystone species’: more than an ecological concept? ........................................... 614. Keystone species: toward an operational concept for marine biodiversity conservation .    ............................................................................................................................................... 724.1. Introduction .................................................................................................................... 724.2. Methods .......................................................................................................................... 754.2.1. Derivation of indices of keystoneness from Ecopath food web models ................. 754.2.2. Analysis of the contributions of the IC and BC components to the keystoneness estimation............................................................................................................................... 814.2.3. Selection of the most promising KS index ............................................................. 824.2.4. Application to a case study ..................................................................................... 85vii  4.3. Results ............................................................................................................................ 854.3.1. Pool of selected Ecopath models ............................................................................ 854.3.2. Balance in the contributions of IC and BC to the KS indices ................................. 874.3.3. Selection of the most promising KS index ............................................................. 884.3.4. Identification of potential keystone groups with the selected KS index and the classification tree ................................................................................................................... 884.3.5. Case study of the Prince William Sound (1994-1996) food web ........................... 904.4. Discussion ...................................................................................................................... 904.4.1. A balanced index of keystoneness for marine food webs ....................................... 904.4.2. Methodological considerations on the derived functional KS index ...................... 924.4.3. Perspectives for future directions in keystoneness analyses ................................... 934.4.4. Toward an index of keystoneness applicable to biodiversity conservation ............ 955. Selectivity of fish feeding: potentials, challenges and recommendations ..................... 1075.1. Introduction .................................................................................................................. 1075.2. Methods ........................................................................................................................ 1095.2.1. Definition of the species pool ............................................................................... 1095.2.2. Compilation of the diet composition from FishBase ............................................ 1105.2.3. Compilation of the species parameters from FishBase ......................................... 1115.2.4. Statistical analyses ................................................................................................ 1135.3. Results .......................................................................................................................... 1135.3.1. Pool of predator and prey species ......................................................................... 1135.3.2. Diet composition matrix ....................................................................................... 1145.3.3. Biological and ecological parameters ................................................................... 1145.3.4. Identification of predator groups .......................................................................... 1155.3.5. Discrimination of prey and predator based on their species traits ........................ 1165.4. Discussion .................................................................................................................... 1175.4.1. Methodological considerations ............................................................................. 1175.4.2. Potentials and limitations of the use of FishBase ................................................. 1195.4.3. Challenges in predicting fish feeding selectivity .................................................. 1205.4.4. Recommendations for future studies .................................................................... 122viii  6. Conclusion .......................................................................................................................... 139Bibliography .............................................................................................................................. 148Appendices ................................................................................................................................. 167Appendix A: Detailed list of the 433 unique EwE models documented in EcoBase ............. 167Appendix B: Example of the Ecopath model of the Prince William Sound (1994-1996) ...... 202Appendix C: Overview of the 101 selected Ecopath models .................................................. 205Appendix D: Weighted averaged diet percent calculated based on FishBase diet data .......... 227Appendix E: Selection of biological and ecological parameters from FishBase .................... 234   ix  List of Tables Table 2.1. List of the metadata used for the analyses. Metadata are listed in alphabetical order, with their corresponding fields in EcoBase and definitions. ............................................ 37Table 2.2. Number and percentage of articles presenting EwE-based studies, detailed by journal............................................................................................................................................ 38Table 3.1. Comparative framework describing 12 categories of species of potential interest, based on a literature review. ............................................................................................. 63Table 3.2. List of examples of species labeled as ‘keystone predator’ or ‘keystone species’ (Sensu Paine 1966, 1969).. ................................................................................................ 66Table 3.3. List of examples of ‘pseudo-keystone’ species, i.e. species labeled as ‘keystone’ in the literature but whose description would rather correspond to other categories of species of potential interest.. .............................................................................................................. 67Table 4.1. Equations of the 12 indices of keystoneness (KS), obtained by combining each impact components (IC) with each biomass components (BC).. .................................................. 97Table 4.2. Results from the comparison between groups ordered by keystoneness, and by trophic impact or biomass, for each KS index, over all selected models.. .................................... 98Table 4.3. Results from the rank correlation tests (column 2), the classification tree (column 3 – see Figure 4.1) and the status allocation (columns 4 to 8 – see Figure 4.2), for each KS index.. ................................................................................................................................ 99Table 5.1. Nomenclature and taxonomy classification for the 12 predator and 31 prey species selected from FishBase.. ................................................................................................. 123Table 5.2. Diet composition matrix for selected predator-prey pairs. (a)  values summed to one and expressed in percentage. (b)  values expressed in rank.. ................... 125Table 5.3. List of biological and ecological parameters selected for predator and prey species, with associated acronyms and classes or categories.. ..................................................... 127Table 5.4. Parameters selected for the 12 selected predator species.. ........................................ 128Table 5.5. Parameters selected for the 31 selected prey species.. ............................................... 129  x  List of Figures Figure 2.1. World maps of the EwE models documented in EcoBase representing (a) the centroids of the geographic extent of the models and the number of models (b) by FAO area and (c) by LME.. ....................................................................................................... 39Figure 2.2. Classification of the 433 unique EwE models documented in EcoBase by: (a) type of ecosystem represented by the models; (b) research questions answered with the models; (c) number of functional groups of the models; (d) decade of the first year of the models; (e) time period (in number of years) represented by the models; and (f) area (in km2) covered by the models.. .................................................................................................... 40Figure 2.3. World maps of the number of EwE models documented in EcoBase by FAO area, for each of the last three decades: (a) 1984-1993; (b) 1994-2003; and (c) 2004-2014. ......... 41Figure 2.4. Comparison of the classified EwE models between the last three decades: 1984-1993; 1994-2003; and 2004-2014; based on the year of publication of the models. Classification of the models by: (a) category of ecosystem represented by the models; (b) climatic zone of the modeled ecosystems; (c) research questions answered with the models; (d) number of functional groups of the models; (e); time period (in number of years) represented by the models; and (f) area (in km2) covered by the models.. ............................................... 42Figure 2.5. Number of publications presenting EwE-based studies over the last thirty years.. ... 43Figure 3.1. Evolutionary tree representing the evolution of the keystone species concept over time, based on a literature review.. ................................................................................... 68Figure 3.2. Comparative framework of the 12 categories of species of potential interest, classified by category and represented based on their potential interests. ........................ 69Figure 3.3. Comparative framework of the 12 categories of species of potential interest, distributed on trophic level versus biomass scales.. ......................................................... 70Figure 3.4. Conceptual diagram representing the restored definition of the keystone species concept.. ............................................................................................................................ 71Figure 4.1. Identification of keystone species with a classification tree. (a) Classification tree for the identification of keystone species among the functional groups of each modeled ecosystem. (b) Simplified graphical representation of the classification tree on a scatter xi  plot, with the biomass () on the x-axis, and the trophic impact (in squared values) () on the y-axis, both in log values.. ................................................................................... 100Figure 4.2. Rules for status allocation to each pair of ‘Ecopath model – KS index’.. ................ 101Figure 4.3. Boxplots of (a) spatial extent (Area, thousands of km2), (b) total biomass (		, t·km-2), and (c) total catch (		, t·km-2·year-1), for all selected Ecopath models, categorized by type of modeled ecosystem. ....................................................................................... 102Figure 4.4. Boxplots of (a) number of living groups, (b) number of fishing fleets, and (c) mean trophic level of the catch (), for all selected Ecopath models.. ................................ 103Figure 4.5. Bar plots of the groups identified as keystone with (a) the selected  index, and (b) the classification tree....................................................................................................... 104Figure 4.6. Application of the classification tree (see Figure 4.1) to the Prince William Sound (1994-1996) food web (Okey and Pauly 1999).. ............................................................ 105Figure 4.7. Application of the (a) , (b) , and (c)  indices (see Table 4.1) to the Prince William Sound (1994-1996) food web (Okey and Pauly 1999).. ................................... 106Figure 5.1. Results from the PCA on the diet matrix expressed in rank, with the predator species as individuals: (a) scree plot of Eigen values and (b) scatter diagram of the coordinates (or factor map) of the individuals. .................................................................................. 132Figure 5.2. Results from the PCA: factor map of the individuals with the number of prey (PreyNo) projected as a supplementary variable on the plan formed by (a) axes 1 and 2 and (b) axes 1 and 3. ....................................................................................................... 133Figure 5.3. Results from the PCA: factor map of the individuals with the maximum length (ML) projected as a supplementary variable on the plan formed by (a) axes 1 and 2 and (b) axes 2 and 3. ............................................................................................................................ 134Figure 5.4. Results from the PCA: factor map of the individuals with the average aspect ratio (AAR) projected as a supplementary variable on the plan formed by (a) axes 1 and 3 and (b) axes 2 and 3. .............................................................................................................. 135Figure 5.5. Results from the CA on the prey parameters, with the prey species as individuals: factor map of the individuals with the different levels of each parameter projected on the plan formed by the axes 1 and 2.. ................................................................................... 136xii  Figure 5.6. Results from the CA on the prey parameters: factor map of the individuals with the group of the predator (PredatorGroup) with the highest diet composition value projected as a supplementary variable on the plan formed by the axes 1 and 2. ............................ 137Figure 5.7. Results from the CA on the predator parameters, with the predator species as individuals: factor map of the individuals with the different levels of each parameter projected on the plan formed by the axes 1 and 2.. ......................................................... 138  xiii  Acknowledgements First and foremost, I wish to thank my supervisor, Dr. Villy Christensen, for his guidance and support. Thank you for giving me the opportunity of doing a PhD and allowing me to make it the most valuable learning experience, especially by encouraging me to attend to scientific meetings and participate to international conferences. Then, I would like to thank the members of my committee, Dr. Daniel Pauly, Dr. William Cheung, and Dr. Evgeny Pakhomov for their guidance throughout the years of my studies. Thank you for your time and advice on my work.  Thanks are also due to my co-authors on the different chapters of my thesis. I am grateful to Dr. Marta Coll and Dr. Deng Palomares, for their precious time, contribution and support. Thank you for believing in my ideas and helping me to turn them into research projects. Special thanks to Dr. Marta Coll for hosting me at the Institut de Ciències del Mar (ICM-CSIC) in Barcelona for six weeks. I would like to also thank Mathieu Colléter for a fruitful collaboration on the EcoBase project. Thank you for welcoming me into the project at first and for your contribution to our collaborative paper. Thanks also to Dr. Didier Gascuel and Dr. Daniel Pauly for constructive comments on earlier drafts of this paper. Many thanks to Jérôme Guitton for his assistance and advice on the technical aspects of the EcoBase project.  I wish to acknowledge financial support, which made my research work possible. The first year of my PhD studies was funded by the Netherlands Environmental Assessment Agency (PBL). The following years were funded by the Nereus Program, a scientific cooperation between the Nippon Foundation and The University of British Columbia with six additional partners. As a junior fellow of the Nereus Program, I had the opportunity to present my work to the other members of the program during annual meetings. I thank them all, and particularly the junior and senior fellows, for thought-provoking discussions and inspiration to think ‘outside the box’. Special thanks to Dr. Yoshitaka Ota and the Nippon Foundation for giving us, the fellows, the opportunity to travel to Japan and have a very unique experience while meeting inspiring people in Northern Japan. Additional financial support was provided by the International Tuition Award, the Faculty of Science Graduate Award, and the Graduate Student International Research Mobility Award from The University of British Columbia. I would like to also thank Dr. Sylvie Guénette and Dr. Didier Gascuel, for encouraging me to pursue my graduate studies and helping to make things possible by kindly referring me.   xiv  Many thanks to the colleagues and friends who made my journey at UBC more enjoyable: Chiara, Sherman, Dalai, Léa, Cécile, Marian, Marcela, Leah, Jeroen, and Shannon. Thank you all for the shared lunches, fun times and good memories!  Thank you to my family members for their constant love and support, with an extra thank you to my parents for their patience and understanding. And last, but certainly not least, thank you Arnaud, for being there for me, since the very beginning, at every moment, and every time I needed it the most.    xv  Dedication     To my grandparents: Jean & Amparo and Jean & Jacqueline  1  1. Introduction 1.1. Context It is a general perception that the oceans are the last frontier and that we have but scratched the surface when it comes to understanding how marine ecosystems function as well as what life they shelter. This is in many respects true, but we also have gathered considerable information about life in the world’s oceans. About 226,000 marine species have been described, and 58,000 to 72,000 species have been sampled but not yet described, for a predicted total of 0.7 to 1.0 million marine species (Appeltans et al. 2012). Previous higher estimates are now considered unlikely (Appeltans et al. 2012, Costello et al. 2013). Moreover, the number of described species have been increasing over the last decade, so that most species could be discovered by the end of the century (Appeltans et al. 2012, Costello et al. 2013). In comparison, the total number of species on the planet was estimated to be about five millions, with a three millions error margin, and of which 1.5 million were valid described species (Costello et al. 2013).  The world’s oceans are dynamic systems, which undergo different types of changes, from species extinctions, population declines or habitat degradations to species invasions (Sala and Knowlton 2006). In a recent review (including about 80 species or groups of species from various marine environments), the percent of population decline ranged from 39% for crustaceans to 91% for oysters, with most of the mammals, birds, and reptiles severely depleted by 1900 and even further by 1950 (Jackson 2008). Other studies demonstrated that the biomass of large predatory fish has declined by two-thirds from 1950 to 1999 in the North-Atlantic (Christensen et al. 2003b), and globally by two-thirds over the last hundred years (Christensen et al. 2014a). Habitat-forming species in the coastal zones, such as corals, mangroves or seagrasses, were estimated to be critically endangered on a global scale, with a reduction percentage ranging from 50 to more than 90% (Jackson 2008).  2  1.1.1. Drivers of changes in marine ecosystems There are several natural factors causing changes in marine ecosystems. First, natural variability may be caused by modifications in physical processes, either due to external factors such as solar radiation, or due to internal factors, such as ocean-atmosphere interactions. Natural variability may also be induced by biological processes, such as the combined effect of bottom-up and top-down controls. Bottom-up or resource control corresponds to the regulation of ecosystems by variability in primary production or nutrient inputs, while top-down control corresponds to the regulation by predation (Pace et al. 1999).  Anthropogenic factors also represent strong drivers of change in marine ecosystems and food webs (Sala and Knowlton 2006, Halpern et al. 2008). Indeed, marine species within a given habitat are adapted to natural pulse disturbances, but not necessarily to human-made chronic disturbances (Gray 1997, Sala and Knowlton 2006). Halpern et al. (2008) showed that no area of the world’s oceans is unaffected by anthropogenic impact, and that 41% is strongly affected by multiple human activities. Moreover, synergistic effects of combined human activities make human impacts on marine ecosystems stronger, while potential thresholds effects make them unpredictable (Jackson et al. 2001, Folke et al. 2004, Sala and Knowlton 2006). ‘Regime shifts’, i.e. changes from one rather stable condition or ‘state’ to another, have been observed in marine ecosystems and were associated with different human-induced causes (Scheffer et al. 2001, Knowlton 2004, Daskalov et al. 2007). According to demographic projections, global human population will increase up to nine billions by 2050 (ten Brink 2010). Thus, under current conditions, human pressure on marine ecosystems is likely to increase as well, especially in coastal areas where human population is concentrated (Martínez et al. 2007).  The two main globally-distributed anthropogenic drivers impacting marine ecosystems are commercial fishing and human-induced climate change (Halpern et al. 2008). Commercial fishing may have significant impacts on marine ecosystems if it occurs above sustainable levels and leads to overfishing. For instance, overfishing was identified as the main cause of disturbance and collapse in coastal ecosystems (Jackson et al. 2001). Human-induced climate change designates the perturbation of natural variability in oceanic physical, biological and chemical conditions, driven by increasing greenhouse gases (notably carbon dioxide) emissions from human activities (Doney 2010). In addition to overexploitation and anthropogenic climate 3  change, species introductions and water pollution (contamination and eutrophication), caused by human activities (e.g., shipping, fertilizer use) also represent important threats to marine biodiversity (Sala and Knowlton 2006, Costello et al. 2010).  1.1.2. Impacts of fishing on marine ecosystems Fishing activities exert significant impacts on marine ecosystems and biodiversity worldwide (Pauly et al. 2002). Indeed, overfishing causes serious depletions worldwide (Pitcher and Cheung 2013). Overall, global capture fish production in marine waters have remained steady since the 1980s (FAO 2014). However, when considering the increase in fishing power, the efficiency of the world’s fisheries (i.e. landings/effort) actually shows a reduction by half since the 1950s (Watson et al. 2013). While well-assessed fisheries may be moving toward sustainable exploitation patterns, unassessed stocks show continued decline due to overfishing (Costello et al. 2012). The fraction of overfished stocks (i.e. stocks exploited at biologically unsustainable levels) has increased over the last decades, up to nearly 30 percent in 2011, while the proportion of fully fished stocks accounted for about 60 percent of the total number of stocks assessed (FAO 2014).  Fishing activities are responsible for population declines of target species, but also of non-targeted species through by-catch, i.e. incidental catch or discards (Hall et al. 2000). Marine mammals, sea birds, and sea turtles are the most severely impacted species due to incidental mortality (Hall et al. 2000). Shrimp trawls and bottom trawls are the fishing gears that have the highest amount of by-catch (Pitcher and Cheung 2013). Moreover, when used on sensitive habitats, such as seagrass or coral beds, bottom trawls represent one the most destructive fishing practices (Pitcher and Cheung 2013). Thus, in benthic and deep-sea marine ecosystems, chronic disturbance from trawl fishing leads to the removal of seabed species, lower productivity and fragmentation of the habitat they form (Kaiser et al. 2002). Severe and continuous declines may result in the extinction or collapse of overfished populations (Casey and Myers 1998, Sadovy and Cheung 2003). Large, long-living and slowly maturing species, such as sharks (Stevens et al. 2000), are more particularly  threatened with extinction (Sadovy 2001). Of the 133 local, regional, and global extinctions of marine populations documented worldwide, 55% were caused by unsustainable exploitation, and 37% 4  driven by habitat loss (Dulvy et al. 2003). Even when not spatially extinct, a species may become functionally or ecologically extinct, when it becomes so rare that it no longer fulfills its function in the ecosystem (Carlton et al. 1999, Sala and Knowlton 2006). Ecological extinction may be due to over-exploitation, such as for sea otters in the Northeast Pacific, baleen whales in the Southern oceans, or the Newfoundland cod (Carlton et al. 1999, Jackson 2008).  Fishing activities initially targeted valuable, large predatory fish species, located at the upper end of marine food webs. However, as the populations of large predators progressively became depleted by overfishing, targeted species were gradually replaced by less valuable, smaller foraging fish species, a phenomenon known as ‘fishing down marine food webs’ (Christensen 1996, Pauly et al. 1998). This shift in fish stocks composition was revealed by the decline of the mean trophic level of global fisheries catch (Pauly et al. 1998, Pauly and Palomares 2005, Pauly and Watson 2005), which has been observed in many regions of the world (Stergiou and Christensen 2011). More generally, the loss of large apex consumers due to human influence on nature, also called ‘trophic downgrading’, has been observed worldwide, both in aquatic and terrestrial ecosystems (Estes et al. 2011). The direct consequence of human-induced removal of top predator species is a simplification of the food web, which may lead to major shifts in the ecosystem structure and dynamics (Pauly et al. 2002, Estes et al. 2011). In addition, by targeting the larger individuals, fishing selects for individuals that grow slower and reproduce earlier and at smaller sizes (Conover and Munch 2002). Thus, due to phenotypic plasticity, fishing induces changes in the size distribution and the life-history traits of exploited fish stocks, and may thus impact their evolutionary characteristics and lead to genetic diversity loss in the long term (Gray 1997, Law 2007). Overall, the disturbance of marine ecosystems caused by fishing activities may amplify the impacts of environmental variability (Pauly et al. 2002, Sala and Knowlton 2006, Estes et al. 2011).  1.1.3. Effects of climate change on marine ecosystems Climate change modifies the ocean biogeochemistry by inducing warming, acidification and deoxygenation (the loss of dissolved oxygen from the ocean) (Doney 2010, Gruber 2011). These three processes interact, creating synergistic effects, and occur at the global scale, but with distinct regional patterns (Gruber 2011). Changes in environmental conditions impact marine 5  organisms and ecosystems in various and complex ways (Hoegh-Guldberg and Bruno 2010, Doney et al. 2012). New environmental conditions may be tolerable if they do not alter species’ physiological performance. Tolerance corresponds to the persistence in situ of a marine species in response to climate change (Dawson et al. 2011), and involves either acclimatization, i.e. adjustment of individuals physiology, or adaptation, i.e. natural selection of tolerant genotypes over generations (Parmesan 2006, Doney et al. 2012). When climate change creates intolerable conditions, species may show different modes of response: change in geographic range (either depth range, i.e. by habitat shift, or latitudinal range, i.e. migration), change in phenology (timing of annual events), or local extinction (Parmesan 2006, Dawson et al. 2011, Doney et al. 2012).  The best-documented biological impacts are shifts in the abundance and distribution of populations due to water temperature increase (e.g., Perry et al. 2005, Nye et al. 2009, Raitsos et al. 2010). Projections of global marine species distributions under climate change scenarios predicted pole-ward migration of species, leading to local climate-driven extinctions or invasions (Cheung et al. 2009). Indeed, fish have the tendency to remain in the same temperature ranges, and so respond to climate change by moving so as to stay where the temperature is tolerable (Pauly 2010). Nevertheless, changes in ocean chemistry, and notably ocean acidification, may be more important than ocean warming for the physiological performance of many organisms (Harley et al. 2006). For instance, the ability of corals to maintain their external calcium carbonate skeletons is directly affected by seawater CO2 chemistry (Orr et al. 2005, Fabry et al. 2008). Global ocean warming and acidification lead to mass coral bleaching and mortality threatening coral reef communities and ecosystems (Hoegh-Guldberg et al. 2007). Moreover, while higher water temperature or acidity increases oxygen demand for growth in fish (and other heterotrophic organisms) (Pörtner and Knust 2007, Pörtner 2008), ocean deoxygenation reduces oxygen supply and creates oxygen minimum zones (Keeling et al. 2010). As a result, climate change may increase aerobic stress for fish, which in turn may modify their feeding behavior and affect their growth and reproduction (Pörtner and Knust 2007, Pauly 2010). Thus, assemblage-averaged maximum body weight of fishes is projected to decrease by 14 to 24% globally from 2000 to 2050 under climate change scenarios (Cheung et al. 2012).  In response to warming, acidifying, and stratifying oceans, marine ecosystems also show: decreasing primary production with potential consequences on nutrient input (Polovina et al. 2008, Steinacher et al. 2010); changes in ocean circulation driving population dynamics through 6  larval dispersal (Harley et al. 2006); and increasing sea-level rise causing coastal habitat loss (Hoegh-Guldberg and Bruno 2010). All these climate-induced changes in marine species, populations, or communities may have consequences on ecosystem properties, such as biodiversity or food-web structure, notably through the disturbance of material and energy flows and biogeochemical cycles (Doney et al. 2012). For instance, the ‘match-mismatch’ hypothesis corresponds to a trophic asynchrony in the climate-induced phenological changes between interacting predator and prey populations (Cushing 1990, Edwards and Richardson 2004, Durant et al. 2007). As a result, global drivers of climate change have significant impacts on marine ecosystem functions and services (Doney et al. 2012). For instance, large-scale redistribution of global fisheries catch potential for exploited marine species was predicted under climate change scenarios (Cheung et al. 2008, Cheung et al. 2010, Cheung et al. 2011), with potential impacts on fisheries economics worldwide (Sumaila et al. 2011).  1.1.4. Consequences on marine ecosystems functions and services Ecosystem functions designate the ecological processes regulating the fluxes of energy, nutrients and organic matter in the ecosystem, while ecosystem services describe the suite of benefits that ecosystems provide to humanity (Cardinale et al. 2012). Examples of marine ecosystem services notably encompass seafood supply, protection against erosion or extreme events, recycling of pollutants (Sala and Knowlton 2006). Biodiversity controls basic ecosystem functions, directly influences certain provisioning and regulating ecosystem services, and may be considered as an ecosystem service (Cardinale et al. 2012, Mace et al. 2012). For instance, greater fish diversity strongly correlated with higher fisheries yields and productivity (Bracken et al. 2007), and was associated with greater stability of fisheries yields (Cardinale et al. 2012). However, anthropogenic drivers of changes in the oceans have caused a rapid decline in global marine biodiversity, with subsequent consequences on marine ecosystem functions and services (Sala and Knowlton 2006). Globally, indicators of the state of biodiversity show declines, with no significant recent reductions in rate, whereas indicators of pressures show increases (Butchart et al. 2010). Moreover, biodiversity loss may significantly alter ecosystem functions and services globally, so that it could be a major driver of ecosystem change (Cardinale et al. 2012, Hooper et 7  al. 2012). Thus, while seafood demand is increasing worldwide and food security issues are rising, marine biodiversity loss is a serious concern to human societies (Rice and Garcia 2011).  In addition, marine ecosystems provide multiple services, which may involve trade-offs between the supplies of different services. Indeed, simplifying ecosystems may optimize certain provisioning services locally and temporary, but may also lead to biodiversity loss, and thus reduce regulating services on larger scales and in longer terms (Pereira et al. 2010, Cardinale et al. 2012). The loss of individual species may have indirect ecosystem-wide impacts, especially for species playing important functional roles in marine ecosystems (Sala and Knowlton 2006). Thus, the removal of top predators caused by overexploitation in marine ecosystems impacts functional diversity and has effects on a wide range of ecosystem processes (Estes et al. 2011, Cardinale et al. 2012). Another example is the predicted loss of coral reefs and its consequences on coastal protection, fisheries and tourism activities (Hoegh-Guldberg et al. 2007).  1.2. Concepts and definitions 1.2.1. Marine biodiversity Biodiversity is usually shorthanded for biological diversity, defined by the Convention on Biological Diversity as “the variability among living organisms from all sources […] and the ecological complexes of which they are a part”. In other words, biodiversity encompasses the variety of life forms including plants, animals and microorganisms, the genes that they contain and the ecosystems that they form. Biodiversity can be subdivided into terrestrial, freshwater or marine biodiversity, by distinguishing the main types of ecosystems occurring on Earth. Thus, marine biodiversity represents the variety of life in the oceans (Sala and Knowlton 2006). Biodiversity is a multi-faceted concept, comprising four main aspects: compositional diversity, which measures the number of entities (or richness); structural diversity, which measures the distribution of abundance of the entities (or evenness); functional diversity, which measures the functional roles the entities play in ecosystems; and what we could call ‘differential’ diversity, i.e. a measure of the differences among species (e.g., morphological disparity or genetic divergence) (Sala and Knowlton 2006). Entities may consist of genes or 8  individuals, species, communities or ecosystems, depending on the chosen resolution of the biodiversity concept. A species may comprise one or more populations, i.e. groups of individuals that can reproduce and interchange genetic material, while a community corresponds to a group of species co-occurring under the same environmental conditions (Gray 1997). The most commonly used metric of biodiversity is the species and its most common measure is species richness, i.e. the number of species found in a given area (Gotelli and Colwell 2001).  Changes in marine biodiversity can be considered over different temporal scales: the evolutionary timescale (i.e. over many generations); or the ecological timescale (i.e. over few generations) (Gray 1997, Sala and Knowlton 2006). Over evolutionary timescales, marine biodiversity has been increasing, apart from some mass extinction events, and its recovery capacity has not been reduced (Sala and Knowlton 2006). Over ecological timescales, human activities have caused more rapid changes to marine biodiversity than the natural changes caused by ecological succession (Sala and Knowlton 2006). Anthropogenic impacts might be outcompeting natural processes in such a way that the term ‘Anthropocene’ has recently been popularized to describe the current geological epoch (Crutzen and Steffen 2003, Steffen et al. 2007, Rockstrom et al. 2009). In addition, marine biodiversity can be measured at different spatial scales (Gray 1997). At local scales, within-habitat (or alpha) diversity is defined for a given sampled ecosystem or community, and between-habitat (or beta) diversity when sampling covers more than one ecosystem or community. From regional to global scales, landscape (or gamma) diversity is defined as the mosaic of ecosystems or communities over larger samples. Within-species (or point) diversity can also be defined when considering a smaller sample consisting of a single species (Sala and Knowlton 2006).  1.2.2. Trophic interactions Trophic interactions designate feeding relationships between organisms and may be described using food webs, trophic pyramids, flow diagrams, or trophic spectra. Food webs are maps representing the linkages between predator and prey species in the community (Pimm et al. 1991, Menge 1995). Trophic pyramids are diagrams where producers and consumers are arranged according to their size, number and production rate (Lindeman 1942). Based on trophic pyramids, Lindeman (1942) introduced the concept of trophic level, which is the position of an 9  organism in the food web, determined by its feeding preferences. Seeing that most marine organisms function at several trophic levels, fractional trophic levels were then defined (Odum and Heald 1975). Thus, using trophic levels, functional biodiversity can be characterized in two principal dimensions, horizontal and vertical: horizontal diversity considers diversity within a single trophic level, while vertical diversity considers diversity across trophic levels (which may be measured by the degree of omnivory or the food-chain length) (Duffy et al. 2007). Flow diagrams are graphical representations of trophic interactions in food-web models. Organisms are arranged based on their trophic level and interconnected boxes or nodes are used to represent their biomass and trophic flows (e.g., Christensen et al. 2008). An ecosystem may also be represented by the distribution of one of its descriptive parameter (such as biomass) across trophic levels, called the (biomass) trophic spectrum (e.g., Gascuel et al. 2011). Trophic interactions may have direct or indirect effects, through changes in species abundance (Menge 1995). Direct effects are defined as the influence of a species on another species resulting from their direct interaction, while indirect effects correspond to the influence of a species on another species with which it does not interact directly (Menge 1995). Direct effects caused by trophic interactions correspond to predation, while several types of indirect effects may be due to trophic interactions, such as: (i) ‘keystone predation’, when a predator indirectly increases the abundance of competitors of its prey via consumption of the highly competitive prey; (ii) ‘trophic cascade’, when an increase in abundance is caused by the control of predator by a predator of the predator, i.e. it is an effect over three trophic levels; (iii) ‘exploitation competition’, when the reduction of a consumer is caused by reduction of its prey by another consumer; (iv) ‘apparent competition’, when a reduction of a prey is caused by the increase of another prey which enhances predation by a common predator (Holt 1977); and (v) ‘habitat facilitation’, when a species indirectly improves the habitat of another species by reducing the abundance of a third species which has a negative effect on this habitat (Menge 1995). Keystone predation was identified as the most common type of indirect effect in rocky intertidal habitats, while apparent competition and trophic cascades were also common (Menge 1995). Trophic cascades were originally assumed to be more common in pelagic communities, whereas keystone predation would be more common in benthic communities (Menge 1995). Subsequent studies suggested that trophic cascades were not restricted by the type of ecosystems, but were less likely under certain biotic or abiotic conditions (Pace et al. 1999). In fact, trophic 10  cascades have now been documented in both benthic and pelagic marine ecosystems (Verity and Smetacek 1996, Pinnegar et al. 2000), as well as in terrestrial and freshwater ecosystems (Estes et al. 2011).  The importance of a trophic interaction, called ‘interaction strength’, is defined by the magnitude of its effect on the community (Paine 1980, Menge 1995). A strong interactor species has a large effect on the species with which it interacts, while a weak interactor species has a small effect on the other species (Paine 1980, Menge and Freidenburg 2001). A reduction in the abundance of a strong interactor may induce community-wide changes and affect many other species, through the disturbance of trophic interactions (Paine 1980, Sala and Knowlton 2006). In contrast, changes caused by the depletion of weak interactors may be unpredictable (Sala and Knowlton 2006). Due to their strong influence on the food-web structure of the community, strongly interacting species may be defined as critical species or ‘ecologically important species’ (Perry 2010). Several types of critical species may be discriminated based on their role and position in the food web. At the top of the food web, ‘keystone’ species correspond to strongly interacting species of high trophic level whose trophic impacts are disproportionately large relative to their abundance (Paine 1966, 1969a). Lower in the food web, ‘key-industry’ species correspond to species of intermediate trophic level that act as trophic links between primary producers and top predators (Elton 1927). At the bottom of the food web, ‘foundation’ species, such as abundant primary producers, also contribute to the maintenance of the food-web structure of the community (Dayton 1972, Ellison et al. 2005).  1.2.3. Diet composition In quantitative food-web analyses, trophic interactions are defined using the diet composition (DC) of species. DC corresponds to the proportion of each prey in the diet of a predator (Christensen and Walters 2004). DC should ideally be expressed in weight or volume, so that trophic relationships may be determined in energetic terms (Stobberup et al. 2009). Fish species’ diet composition is commonly derived from stomach contents data, which are complex, time-consuming, and expensive to collect (Link 2004, Albouy et al. 2011). The main difficulty in stomach contents analysis is the taxonomical identification of all species remains. Thus, diet composition data is often very scarce in the literature, either available for well-known fish 11  species only, or lacking taxonomic details. When no data is available, DC is either derived from data on similar species in the literature, or arbitrary defined based on expert knowledge. DC is sometimes assumed to be equal to prey relative abundance, although the relative abundance of prey is not a good predictor of diet composition for fish species (Link 2004). As a consequence, DC is usually one of the most poorly estimated parameters in quantitative food-web analyses (Link 2004).  Most fish are generalist feeders, that feed on a broad range of prey items (Gerking 1994). Besides, fish commonly are opportunists when they feed, i.e. they take advantage of available sources of food in their environment (Gerking 1994, Link 2004). Thus, in absence of data, predicting fish diet composition may be challenging. Nevertheless, when several prey items are available, fish select certain prey items among all available based on some constraints. The challenge is to identify the main constraints that influence fish feeding selectivity. Size is known as a key parameter to determine predator-prey interactions (Cohen et al. 1993, Thiebaux and Dickie 1993, Gill 2003), but characteristics other than size may influence prey preference and utilization for fish species (Gerking 1994, Link 2004). Prey preference may be defined as the favored choice of the predator for the most suitable prey items, whereas prey utilization (or diet composition) corresponds to the realized preference, coupled with prey availability (Link 2004).  In a few cases, fish diet compositions were predicted based on a suite of characteristics describing predator and prey species (e.g., Sibbing and Nagelkerke 2001, Link 2004). Yet, there is no strong consensus in the literature on good predictors of fish diet composition. Based on the ecomorphological hypothesis of a correlation between species morphology and ecology, the morphology of the feeding apparatus could be a good predictor of prey utilization for fish (Motta et al. 1995, Norton 1995). However, there is contrasting evidence from the literature to verify the hypothesis of dietary-morphological relationships in fish. In many cases, the ecomorphological traits could be used to predict some general aspects of fish feeding only (i.e. trophic guilds, e.g, Boyle and Horn 2006, Ibañez et al. 2007, Oliveira et al. 2010, Albouy et al. 2011). Fish morphology may predict the way fish are feeding (prey preference or feeding mode), but not the prey they feed on (prey utilization or diet composition) (Motta 1988, Motta et al. 1995, Clifton and Motta 1998, Barnett et al. 2006).  12  1.3. Methods and approaches 1.3.1. Ecosystem-based management Ecosystem-based management (EBM) has become one of the key policy objectives in the strategic plans of international organizations (Link and Browman 2014). Ecosystems were defined by the Convention on Biological Diversity as “a dynamic complex of plant, animal and micro-organism communities and their non-living environment interacting as a functional unit”. According to this definition, ecosystem-based management should consider the whole food web, as well as the abiotic factors that influence it (Link 2002b). A more recent definition described ecosystems as a geographically specified system including biotic and abiotic factors, but also humans (Francis et al. 2007).  In fact, EBM aims at considering and addressing the multiple, simultaneous, and sometimes cumulative, natural and anthropogenic pressures faced by the important components of ecosystems (Link and Browman 2014). The implementation of EBM for marine ecosystems requires integrated assessments, which can be applied to different ocean-use sectors and adopted at different levels: (i) EBM integrate all sectors of activities over all levels of management; (ii) ecosystem-based fisheries management (EBFM) uses integrated ecosystem assessments with a focus on the fisheries sector; and (iii) ecosystem approach to fisheries (EAF) focuses on fisheries stocks with some ecosystem considerations (Link and Browman 2014).  Interest in ecosystem-based management in the fisheries sector notably rose from issues related to conflicting stakeholders, debate over the most important ecosystem processes, and limitations of traditional single-species management (Link 2002b).  Indeed, fisheries management occur in a broadly interdisciplinary context, involving biological, ecological, oceanographic, social, and economic aspects (Link 2002b). Moreover, there has been a recognition that ecosystems provide services beyond the scope of fisheries management (i.e. the targeted stock) and that fishing impacts these services via habitat degradation, incidental mortality of non-targeted species, changes in targeted populations, and disturbance of food webs (Link 2002b, Pikitch et al. 2004). Thus, facing the inherent complexity of natural ecosystems, a more effective and holistic approach was needed to prioritize and maintain the ability of ecosystems to produce these services (Link 2002b, Pikitch et al. 2004). The objective of EBFM 13  is to manage the conditions in the ecosystem that maintain critical processes for the persistence of the desirable ecosystem state (Link 2002b), by “reversing the order of management priorities to start with the ecosystem rather than the target species” (Pikitch et al. 2004). In the context of EBFM, ecosystem models (EM) were one of the tools used to evaluate ecosystem properties and give insights on the potential effects that changes in management practices would have on ecosystems (Christensen and Walters 2005, Christensen and Walters 2011, Espinoza-Tenorio et al. 2012).  1.3.2. Ecosystem models Models are coherent and purposeful representations of systems and of the processes therein, and may consist of words, graphs or equations, which describe the system elements and their relationships with the surrounding environment (Starfield et al. 1990, Pauly and Christensen 2002). Mathematical models representing marine ecosystems may be classified into different types, based on their characteristics and objectives. First, we can discriminate statistical or empirical models from process-based or mechanistic models (Whipple et al. 2000). Statistical models encompass exploratory models and analytical models, the latter type corresponding to (single-species, structured or not) stock assessment models (SAM) (Whipple et al. 2000). Another category described as extended stock assessment models (ESAM, taking a few additional species interactions into account) may be included in statistical models (Plagányi 2007, Hunsicker et al. 2011). Statistical models are used to detect potential patterns using frequentist or Bayesian approaches to probability, and provide algorithms and parameters used in process-based models (Whipple et al. 2000). Process-based models may be sub-categorized into qualitative and quantitative models. Qualitative process-based models describe qualitative interactions between the elements of the system and include: (i) food web models, i.e. unsigned directed graph models; and (ii) loop analysis models, i.e. signed diagraph models (Whipple et al. 2000, Fulton 2010). Although substantial insights into ecosystem functioning may be gained from qualitative models, whose implementation is rapid and flexible, the development and implementation of process-based models within the framework of EBFM has mainly focused on quantitative modeling approaches (Fulton 2010). 14  Quantitative process-based models correspond to the main models used in fisheries and ecosystem modelling today, ranging from minimal realistic models to whole system models (Hunsicker et al. 2011). They comprise multi-species models, ‘ecosystem models’, and ‘end-to-end models’. Multi-species models (MSM) correspond to the multi-species versions of stock assessment models (Hollowed et al. 2000, Whipple et al. 2000, Pauly and Christensen 2002), such as multi-species virtual population analysis (MSVPA) (Magnusson 1995), or models based on the Lotka-Volterra equations (May et al. 1979). MSMs focus on a limited number of species, the ones that have important interactions with the target species (Plagányi 2007). In contrast, ecosystem models (EM) may be defined as models that attempt to represent all biological components of the ecosystem (i.e. the entire food web) (Whipple et al. 2000, Plagányi 2007). EMs may be static if developed for descriptive purpose, or dynamic and predictive. Components of EMs may be defined as individuals, species, or groups of species. Species-based ecosystem models include trophic flow models (e.g., Prognostic Bulk Biomass (PROBUB) and Dynamical Numerical Marine Ecosystem Simulation (DYNUMES) models for the Bering Sea (Laevastu and Larkins 1981), Ecopath with Ecosim (EwE) models (Christensen and Pauly 1992, Pauly et al. 2000, Christensen and Walters 2004)) and ecological network analyses (e.g., Netwrk (Ulanowicz and Kay 1991, Ulanowicz 2004)). Individual-based models (IBM) use decision algorithms to simulate the behaviour of individuals so as to understand system properties and dynamics (Grimm and Railsback 2005). EMs may use specific metrics, such as size or trophic level. Size spectra models represent the ecosystem as a continuous distribution of biomass per body mass class, with biological rates and feeding interactions determined by size ratios (Jennings et al. 2008). Similarly, trophic spectra models represent the ecosystem as a continuous distribution of biomass per trophic level class (e.g., EcoTroph (Gascuel 2005, Gascuel and Pauly 2009, Gasche and Gascuel 2013)).  End-to-end models (E2EMs) (or whole system models) are the last generation of marine ecosystem models, that attempt to represent the entire ecosystem, including biological, physical and human components, allowing for two-way coupling between them, and integrating both bottom-up and top-down controls (Travers et al. 2007, Fulton 2010, Rose et al. 2010). E2EMs may be inappropriate tools for tactical management, but useful for answering system-level ‘what-if’ management questions or testing impact scenarios (Fulton 2010). E2EMs may correspond to network-based models (e.g., Ecopath with Ecosim (EwE), see Section 1.3.3) or 15  Agent-Based Models (ABM) (individual-based models allowing different types of individuals, such as InVitro (Fulton 2010)). Some E2EMs, such as the NEMURO-FISH (Rose et al. 2007) or ATLANTIS (Fulton et al. 2011) models, are coupled with a biogeochemical model. Another sub-category of E2EMs consist of hybrid models, i.e. models created by coupling different types of models (Fulton 2010). Examples of hybrid E2EMs under on-going development include: Spatial Ecosystem And POpulation DYnamics Models (SEAPODYM) (Lehodey et al. 2008) or Object-oriented Simulator of Marine ecOSystem Exploitation (OSMOSE) (Travers-Trolet et al. 2014).  1.3.3. Ecopath with Ecosim  Ecopath with Ecosim (EwE) is commonly described as an ecosystem model, and more precisely corresponds to a quantitative, process-based and species-based model, representing trophic flows in the ecosystem. The EwE modeling approach was primarily developed as a tool-box to help fisheries management and answer ‘what if’ questions about policy that could not be addressed with single-species assessment models (Pauly et al. 2000, Christensen and Walters 2004, Christensen and Walters 2011). In fact, EwE has been widely applied to inform ecosystem-based management (e.g., Jarre-Teichmann 1998, Plagányi and Butterworth 2004, Christensen and Walters 2011, Coll and Libralato 2012). Also, the EwE software is user-friendly, free (under the terms of the General Public License) and downloadable (www.ecopath.org). Thus, EwE has been applied to hundreds of ecosystems worldwide.  EwE is based on the ECOPATH model, proposed and applied to estimate the main predator-prey pathways in the marine ecosystem of Northwestern Hawaiian Islands (Polovina 1984). The original ECOPATH model was then developed further (Christensen and Pauly 1992), and turned into the EwE modeling framework, which can be seen as a tool-box offering a large collection of methods to analyze various ecological phenomena (Pauly et al. 2000). EwE consists of a suite of three sub-models: (i) Ecopath, static and descriptive; (ii) Ecosim, dynamic and predictive; and (iii) Ecospace, spatially explicit, dynamic and predictive (Christensen and Walters 2004). Ecopath is a descriptive model, representing a static snapshot of the food-web structure, and used as a parameter estimation methodology in the EwE modeling complex (Walters et al. 1997). Ecosim uses the outputs of Ecopath to produce time-dynamic simulations of changing trophic interactions with changes in the ecosystem due to anthropogenic or 16  environmental disturbances (Walters et al. 1997). Ecosim may be used to fit model predictions to time-series data, and to explore alternative fisheries management or climate change scenarios (Christensen and Walters 2004). Ecospace replicates outputs of Ecosim over a spatial grid to explicitly account for the spatial aspects of trophic structure, and was primarily designed for exploring the possible consequences of alternative marine protected areas policies (Walters et al. 1999). Several modules were also developed and added to the EwE package, such as the ‘Network Analysis’ plugin, a tool available with the Ecopath model (Christensen et al. 2008). The Network Analysis is based on concepts from theoretical ecology, and notably network analysis theory (Ulanowicz 1986, Ulanowicz and Norden 1990, Ulanowicz and Puccia 1990, Ulanowicz 1995).  Ecopath describes the feeding relationships between all species occurring in the modeled food web. Species are aggregated into functional (or trophic) groups, corresponding to an age/size class of a single species, a single species, or a group of species sharing similar ecological traits (such as size, diet, predators, habitat and life cycle), and thus assumed to play similar functional roles in the food web (Christensen et al. 2008). Using functional groups is required to handle diet composition parameters, and also to run the spatially explicit Ecospace model. The main assumption in Ecopath is of mass balance over a given time period (typically one year) (Christensen and Walters 2004). For each functional group, Ecopath assumes that the energy input and output are balanced in the ecosystem, and that the system has the same biomass state at the end of the period as it had at the beginning (Walters et al. 1997). The mass-balance constraint serves as a filter for mutually incompatible estimates in the model. All available information collected about the components of the ecosystem pass through the ‘mass-balance filter’, which determines what parameters must be to support the current trophic structure and be consistent with observations (Walters et al. 1997, Christensen and Walters 2004).  The parameterization in Ecopath is based on two master equations (Equation 1.1 and Equation 1.2), to which the mass-balance constraint is applied (Christensen and Walters 2004). The first master equation describes the energy balance for each group, so that:  	  	  	  		   Equation 1.1 In the second master equation, describing the conservation of biomass, production is split in five components: the biomass removed by natural causes of mortality other than predation (diseases, old age…), by predation, and by fishing, plus the net migration and biomass accumulation: 17  !  "#! $ ! "%! $ !  &! $ !  '!  !     Equation 1.2 Where, for each functional group ,  is the production rate (t·km-2·year-1),  the biomass (t·km-2), "# the other mortality rate (year-1), "% the predation mortality rate (year-1), & the fishing mortality rate (year-1), ' the net migration rate (emigration – immigration) (t·km-2·year-1) and  the biomass accumulation rate (t·km-2·year-1). "#, the catch-all rate including all mortalities not elsewhere included, may be expressed through the ecotrophic efficiency ('', dimensionless), which is defined as the part of  not directed to detritus, but used by the trophic chain, exported, accumulated or fished: "#!  ()$*+,--)./)          Equation 1.3 "%, the mortality rate caused by predation, may be expressed as the sum, over all the predator groups 0 which feed partly on the group , of the product of the consumption/biomass ratios (12, year-1), the biomasses (3) and the diet compositions (43): "%!  ∑ *6 /.⁄ 898 $/8$:;8)/)         Equation 1.4 Therefore, the first master equation may also be written, for each functional group : * ⁄ .! $ !  * ⁄ .! $ ! $ * < ''!.  ∑ *1 .⁄ 3=3 $ 3 $ 43!  !  '!  !             Equation 1.5 Where, for each functional group ,  is expressed as the product of the production/biomass ratio (2, year-1) and !; and the catch rate (, t·km-2·year-1) is the product of & and , and corresponds to the total of all extractions by all the fishing fleets defined in the modeled ecosystem (if exploited).  To estimate the missing parameters, Ecopath solves a system of linear equations (with as many equations as there are groups in the model) by using a generalized method for matrix inversion (Christensen and Walters 2004). For each group, one of the six main parameters (biomass, production/biomass ratio, consumption/biomass ratio, ecotrophic efficiency, net migration rate and biomass accumulation rate) is estimated (ideally the ''), and the others must be entered, along with the two remaining ones (diet compositions and catch rates) (Christensen and Walters 2004). Ecopath provides a snapshot of the trophic web consisting of ‘instantaneous’ estimates of biomasses, trophic flows, and mortality rates, for the reference year or time period (Christensen and Walters 2004).  18  1.3.4. EcoBase  Building an EwE model requires the collection, compilation and harmonization of various types of information: descriptive data on species abundance, diet composition and catch; computed data on species production, consumption and ecosystem properties; and simulation data on species biomass trends, after applying alternate scenarios (Christensen et al. 2008). Then, EwE models help understanding the structure and functioning of the modeled ecosystems, by summarizing all available knowledge on the ecosystems and deriving various system properties (Walters et al. 1997). Thus, EwE-based studies may be seen as important sources of information, and several meta-analyses, based on a selection of EwE models, have been performed (Colléter et al. 2013b). Some meta-analyses studied ecological and trophic concepts (e.g., Christensen and Pauly 1993a, Christensen 1995a, Libralato et al. 2006, Gascuel et al. 2008, Arreguín-Sánchez 2011), while others focused on ecosystems and species of particular interest (e.g., Christensen et al. 2003a, Christensen et al. 2003b, Pauly et al. 2009).  However, only few meta-analyses have been using large collections (more than 50) of EwE models (e.g., Coll et al. 2012, Pikitch et al. 2012, Heymans et al. 2014). These meta-analyses were based on individual datasets since no comprehensive, open-access and digital collection of EwE models was available. Indeed, contrary to other historical “big” sciences, such as oceanography, meteorology or astronomy, extensive data-sharing is not the norm in life sciences, such as ecology, and faces sociological and technological challenges (Pauly 1995, Reichman et al. 2011, Thessen and Patterson 2011). Nevertheless, ecology is following the same path and becoming a data-intensive science (Kelling et al. 2009, Michener and Jones 2012). Ecological studies are more and more based on data-driven methodologies, relying on pre-existing large datasets and allowing for new insights on complex or underlying phenomena on global scale (e.g., Christensen et al. 2009). Thus, new practices are needed to make data sharing fully part of the culture in ecology (Zeller et al. 2005, Dalgleish et al. 2012, Hampton et al. 2013). Notably, open-access and digital repositories that “provide standardization, atomization and quality control services facilitate the [discovery, storage and] reuse of data and will play a stronger role in data-intensive science” (Thessen and Patterson 2011).  The EcoBase project was initiated to support and stimulate data sharing and global meta-analyses, using the EwE modeling approach. EcoBase is an online information repository of 19  EwE models published in the scientific literature, developed with the intention of making the models discoverable, accessible, and reusable to the scientific community. The main goals of EcoBase are to (i) provide a comprehensive and up-to-date list of published EwE models and EwE-based applications; (ii) compile and provide open-access and digital information on the referenced models; (iii) enable global meta-analyses based on EwE models. The EcoBase repository may be accessed online (sirs.agrocampus-ouest.fr/EcoBase/) via different (more or less restrictive) user types and details on the structure, usage, and capabilities of EcoBase can be found in the report introducing it and available online (Colléter et al. 2013b). In the long term, EcoBase is meant to be used by ecosystem modelers worldwide as a platform where to (i) look for published EwE models; (ii) select and access models of interest to their research work; (iii) download models as well as upload their own models.  1.4. Research questions and objectives The main objectives of this thesis are to propose new methods and provide new insights for understanding, modeling and potentially predicting marine species trophic interactions (MSTI). More precisely, I tackle the following research questions:  • How modeling approaches, such as EwE, are used to model MSTI? What are the potentials and challenges of using the EcoBase information repository to help answering this question?  • Why are MSTI important for the structure and functioning of marine ecosystems?  • Can modeling approaches, such as EwE, help identifying important MSTI? How are MSTI impacted by anthropogenic drivers, such as fishing?  • Can we predict MSTI? What are the potentials and challenges of using the FishBase information repository to help answering this question?   The thesis is structured into six chapters, including the Introduction and Conclusion chapters. In the Introduction, I presented current knowledge on the status, trends, and drivers of changes in the oceans. Loss of biodiversity, and notably loss of trophic diversity due to fishing, was identified as an important driver of changes in marine ecosystems. Concepts and definitions 20  related to biodiversity, trophic interactions and diet composition were briefly introduced, in a marine context. Then, I introduced the Ecopath with Ecosim (EwE) ecosystem modeling approach, within the framework of Ecosystem Based Management (EBM), and by comparison with other types of models. I also introduced the EcoBase repository of published EwE models worldwide.  In Chapter 2, we use EcoBase to propose a global overview of the applications of the EwE modeling approach. For the 433 unique models documented in EcoBase, we compile, standardize, and analyze all available metadata on the general characteristics of the modeled ecosystems, the objectives, complexity and scope of the models, as well as the associated publication(s). Based on the year of publication of the models, we also analyze the evolution of the EwE applications over the last thirty years. Thus, we intend to provide new insights on the interest for and usage of one of the most widely applied approaches in ecosystem modeling. Moreover, the standardized metadata provided in Chapter 2 are reused as selection criteria to perform a meta-analysis based on EwE models in Chapter 4. Indeed, by applying a scoring method on these criteria, a list of Ecopath models of potential interest is obtained.  Both Chapters 3 and 4 focus on understanding the importance of trophic interactions, using the EwE modeling approach. I briefly mentioned in the introduction different types of species that are critical to the ecosystem, due to their trophic roles. In Chapter 3, I propose a comparative and conceptual framework describing the different categories of critical species, as well as other species of potential managerial interest. I focus on the over-used and much debated concept of keystone species and I intend to clearly defining what is and is not a keystone species. I present a thorough literature review and analyze the origin and evolution of the multiple definitions, terms, and examples associated with the (sometimes inappropriate) applications of the concept in the peer-reviewed literature. Chapter 3 intends to be a proposal for a restored, restrictive and criteria-based definition of the concept of keystone species.  In Chapter 4, I apply the restored definition proposed in Chapter 3 to identify potential keystone species in marine food webs. I perform a meta-analysis, based on 101 published Ecopath models, selected from EcoBase so as to be representative of the variety of marine ecosystems worldwide. Based on the meta-analysis, I derive a functional index of keystoneness (KS) from the EwE modeling approach, directly applicable to marine food webs. Thus, I intend to make the concept of keystone species operational for marine biodiversity conservation. The 21  proposed KS index is calculated based on estimates of the trophic impact of species in the food web, itself obtained from estimates of the species diet composition (DC). Thus, DC is a critical parameter when assessing the relative importance of species in the food web, based on their trophic interactions. However, as mentioned in the Introduction, DC is one of the least addressed parameter in quantitative trophic flow models, and several factors may prey selection and utilization, notably for fish species.  In Chapter 5, I analyze fish diet composition and intend to identify good predictors of fish feeding selectivity. Multiple approaches have been proposed in the literature to predict fish diet composition, with more or less success. I choose to tackle this issue by using one of the most important open-access, digital and cross-disciplinary information repositories on marine biodiversity: FishBase (www.fishbase.org). I select and extract diet data and a suite of biological and ecological parameters from FishBase for a defined pool of predator and prey species. Predator species are then discriminated based on their diets, and parameters of potential significance to predict predator types are identified.  Finally, I summarize the major results and main conclusions from each chapter in the Conclusion. For each chapter, I also emphasize the applications of the proposed methodologies and the implications of the findings, draw some recommendations, discuss the limitations of the analyses and propose possible perspectives for future improvements.    22  2. Global overview of the applications of the Ecopath with Ecosim modeling approach using the EcoBase information repository 2.1. Introduction The Life Sciences have reached a new era, that of the ‘Big New Biology’ (Thessen and Patterson 2011). Ecology is following a similar path, and has turned into a ‘data-intensive science’ (Kelling et al. 2009, Michener and Jones 2012). Ecological studies are more and more based on data-driven methodologies, relying on large pre-existing datasets and allowing for new insights on complex or underlying phenomena on global scale (e.g., Christensen et al. 2009). A popular example of open-access, digital and cross-disciplinary database in aquatic ecology is FishBase, the online encyclopedia of fishes (www.fishbase.org). However, extensive data sharing is still rare in Life Sciences, and ecology has not yet joined the other historical “big” sciences, such as oceanography, meteorology or astronomy, where massive data-sharing is the norm (Pauly 1995, Edwards 2010, Hampton et al. 2013). Although incentives for digitization of non-digital materials have been growing, existing information repositories, such as FishBase, were estimated to represent less than 1% of the data in ecology (Reichman et al. 2011, Thessen and Patterson 2011).  Data sharing is a required principle for independent verification and reuse (Vision 2010), and published papers which make their data available are cited more frequently (Piwowar et al. 2007). Yet, the open-access principle of sharing information online for free has been increasingly applied to publications, but much less to data, mainly due to issues with recognition and sense of data ownership (Vision 2010, Thessen and Patterson 2011). Data sharing is not a tradition in ecology and faces multiple sociological and technological obstacles (Reichman et al. 2011, Thessen and Patterson 2011). New practices are needed to make data sharing fully part of the culture in Life Sciences (Pauly 1988, Reichman et al. 2011, Thessen and Patterson 2011, Dalgleish et al. 2012, Hampton et al. 2013). The two critical stages at which practices have to be improved to allow for data sharing are the very first, i.e., the collection of the data, and the very 23  last, i.e., their publication. Most of the time, biological data are not being collected with reuse in mind and are then published in a narrative or summarized style in scientific articles (Vision 2010, Thessen and Patterson 2011). The actual data are meant to be provided in online supplements or upon individual requests sent to the authors, but these options often remain unreliable (Vision 2010). Also, extensive data sharing may solve – at least in part – the problem of data loss, such as hard-copies or computer files in outmoded format (Zeller et al. 2005).   While facing the challenges related to open-access data, ecology is more and more relying on modeling-based approaches to inform management. In aquatic ecology, the Ecopath with Ecosim (EwE) modeling approach has been widely applied to inform ecosystem-based management (e.g., Jarre-Teichmann 1998, Plagányi and Butterworth 2004, Christensen and Walters 2011, Coll and Libralato 2012), since its original development in the early 1980s (Polovina 1984). The EwE modeling approach was primarily developed as a tool-box to help fisheries management and answer ‘what if’ questions about policy that could not be addressed with single-species assessment models (Pauly et al. 2000, Christensen and Walters 2004, Christensen and Walters 2011). Details on the core principles and equations of EwE can be found in the EwE user guide available online (Christensen et al. 2008). The EwE software is user-friendly, free (under the terms of the GNU General Public License) and downloadable online (www.ecopath.org). Thus, hundreds of EwE models representing aquatic (but also some terrestrial) ecosystems have been developed and published worldwide.  Building an EwE model require the collection, compilation and harmonization of various types of information: descriptive data on species abundance, diet composition and catch; computed data on species production, consumption and ecosystem properties; and simulation data on species biomass trends, after applying alternate scenarios (Christensen et al. 2008). By summarizing all available knowledge on the modeled ecosystems and deriving various system properties, EwE models help understanding the structure and functioning of ecosystems (Walters et al. 1997). Thus, EwE-based studies may be seen as important sources of information. On top of detailed information on the modeled species and food web, each EwE-based study also provides a general description of the modeled ecosystem, which represents essential information to reuse the model for performing meta-analyses. Several meta-analyses, based on a selection of EwE models, have been performed, focusing either on ecological and trophic concepts (e.g., 24  Christensen and Pauly 1993a, Christensen 1995a, Gascuel et al. 2008, Arreguín-Sánchez 2011), or on ecosystems and species of particular interest (e.g., Christensen et al. 2003a, Christensen et al. 2003b, Libralato et al. 2006, Pauly et al. 2009). However, only few meta-analyses have been using large collections of EwE models (more than 50) (e.g., Coll et al. 2012, Pikitch et al. 2012, Heymans et al. 2014). These meta-analyses were based on individual datasets since no comprehensive, open-access, and digital collection of EwE models was available, and this is why EcoBase was created.   The EcoBase project was initiated to support and stimulate data sharing and global meta-analyses using the EwE modeling approach. EcoBase is an online information repository of EwE models published in the scientific literature, developed with the intention of making the models discoverable, accessible, and reusable to the scientific community. The EcoBase repository may be accessed online (sirs.agrocampus-ouest.fr/EcoBase/) via different (more or less restrictive) user types and details on the structure, usage, and capabilities of EcoBase can be found in the report introducing it and available online (Colléter et al. 2013b). The main goals of EcoBase are to (i) provide a comprehensive and up-to-date list of published EwE models and EwE-based applications; (ii) compile and present information from the referenced EwE models; (iii) enable global meta-analyses based on EwE models. In the long term, EcoBase is meant to be used by modelers as a platform where to (i) look for published EwE models; (ii) select and access models of interest to their research work; (iii) download others’ models as well as upload their own models.  In this study, we used EcoBase to compile available critical metadata on all the EwE models referenced in the models repository. We analyzed the metadata to propose a global overview of the applications of the EwE modeling approach in the scientific literature. We focused on the objectives of the EwE-based studies, the complexity and scope of the models, the general characteristics of the modeled ecosystems, and the associated publication(s). We first presented a general description of the EwE applications published worldwide. Based on the year of publication of the models, we then analyzed the evolution of the EwE applications over the last three decades. We also analyzed the contribution of EwE-based studies in the scientific literature. We intended to provide new insights on past and recent usage of and interest for the EwE modeling approach in the scientific community.  25  2.2. Methods 2.2.1. Compilation of the models and publications  2.2.1.1  Compilation of 433 EwE models In EcoBase, we completed an inventory of all EwE models published from 1984 to 2014 (Colléter et al. 2013b). Our inventory intended to be as exhaustive as possible, although some models may be missing, especially if they were published after October 2013 (date of the publication of the EcoBase repository). The EwE modeling approach and software consists of a suite of three sub-models (or routines): (i) Ecopath, a static and descriptive model, representing a mass-balanced snapshot of the food web; (ii) Ecosim, a dynamic and predictive model, producing time-dynamic simulations for exploring alternative scenarios, based on Ecopath; and (iii) Ecospace, a spatially-explicit version of Ecosim (Christensen and Pauly 1992, Walters et al. 1997, Walters et al. 1999, Christensen and Walters 2004). In EcoBase, ‘EwE models’ primarily designate Ecopath models, based on which Ecosim or Ecospace models may have been developed later on.  Since three existing databases of EwE models were merged into the new EcoBase models repository (Colléter et al. 2013b), some duplicates were first obtained and we then extracted a list of 433 ‘unique’ EwE models from the 571 models recorded in EcoBase. Unique models were defined by the ecosystem they represented (i.e. the model area), the time period they covered (i.e. the model start and end years), and the author of the model (i.e. the first author on the reference of the model, commonly considered as the modeler who developed the model). Thus, if several models represented the same ecosystem but, either over older or more recent time periods, or over smaller or larger areas, or were developed by different modelers, then they were considered as unique. In addition, models representing the same ecosystem over different time periods were specifically identified as ‘replicates’. Most of the time, replicates were built by the same author and comprised the same number of groups, although there were some exceptions. The analyses presented in this study are based on the 433 unique EwE models documented in EcoBase (see 26  Appendix A and the EcoBase website for detailed references of the models). For each of the 433 models, we compiled, standardized and encoded in EcoBase all available metadata.  2.2.1.2  Compilation of 397 EwE-base publications In parallel, we conducted a comprehensive literature review and compiled 397 publications presenting EwE-based studies. The detailed references of these publications were encoded in EcoBase. Some of the information comprised in the references was used as metadata: the name of the first author; the year of publication; and the type of publication. We considered five publication types: journal articles, reports (entire report and report sections), book sections, theses, and conference papers. For journal articles, we also analyzed the number of publications by scientific journal, based on the title of the journals. Each of the 433 EwE models was associated to its corresponding publications. One model could have been described in more than one publication. In that case, we classified the publications as primary, secondary or tertiary reference, based on the year of publication and the level of details on the Ecopath model provided in the publications. When several references were available for one model, the primary one was used to define the year of publication. In the end, each model was associated to at least one (primary) reference. Then, the models were aggregated, based on their year of publication, into three groups corresponding to the last three decades (1984-1993; 1994-2003; 2004-2014), so as to analyze the evolution of EwE applications over time. Note that the same publication could describe several EwE models, so that we obtained only 397 publications for 433 models.  2.2.2. Compilation of the metadata For each of the 433 models, we analyzed the metadata describing: (i) the characteristics of the modeled ecosystems; (ii) the research objectives of the modeling approach; and (iii) the structure, scales and units of the models (see Table 2.1 for the detailed list of metadata). All data used in this study may be downloaded on the EcoBase website.  27  2.2.2.1  Characterization of the modeled ecosystems The 433 models were classified based on the characteristics of the ecosystems they represented and using information provided in the associated publications. First, using expert knowledge, we defined 13 ecosystem types: bay/fjord, beach, channel/strait, coastal lagoon, continental shelf (down to 200 m depth), coral reef, estuary, lake, ocean (deeper than 200 m), reservoir, river, terrestrial, and upwelling. The 13 types were aggregated into three broad categories: marine (including bay/fjord, beach, channel/strait, coastal lagoon, continental shelf, coral reef, ocean, and upwelling), freshwater (comprising estuary, lake, reservoir, and river), and terrestrial.  Then, we documented the geographic extent of the models, so as to precisely locate them on the world map. To do so, a spatial shape was defined for each model, based on the coordinates or the map provided by the modelers in the associated publications. When no precise indication was given, the spatial shape was defined as a point. The centroid of the spatial shape defined for each model was used to determine the climatic zone of the model. Three climatic zones were defined based on latitude: tropical-subtropical (less than 30°N/30°S), temperate (greater than or equal to 30°N-58°N/30°S-58°S), and high latitude (greater than 58°N/58°S) (Pikitch et al. 2012). The spatial shape was also used to classify the models representing marine ecosystems by FAO area and Large Marine Ecosystem (LME). FAO areas correspond to the 19 major marine fishing areas internationally established for statistical purposes by the Food and Agriculture Organization of the United Nations (www.fao.org/fishery/cwp/handbook/h/en), while LMEs correspond to 66 well-defined marine areas, usually of 200,000 km² or more, adjacent to the continents (www.lme.noaa.gov/). Therefore, only the 365 models representing marine ecosystems were classified into FAO areas, while the subset of 345 models representing coastal marine ecosystems was classified into LMEs.  2.2.2.2  Identification of the research objectives of the modeling approach We classified the research objectives of the 433 EwE models into seven topics, i.e., (1) aquaculture; (2) ecosystem structure and functioning; (3) environmental variability and climate change; (4) fisheries; (5) Marine Protected Areas (MPA); (6) pollution; and (7) (group of) 28  species of particular interest. The research objectives of each model were identified by reviewing the title, abstract and keywords of the associated publications. Note that the topics are not mutually exclusive, since a same model may have been developed to answer several questions. Although the 433 EwE models were defined based on the implementation of the Ecopath routine, we compiled metadata indicating if the Ecosim or Ecospace routines had also been used. Ecosim and Ecospace models add complexity to the Ecopath base models and enable to explore alternative management scenarios by representing the temporal and spatial variability of the ecosystem.  Some of the models were used to perform specific analyses, using some of the modules (or plug-ins) available in the EwE software. In this study, we considered three plug-ins of potential interest: Ecotracer, EcoTroph, and Keystoneness. Ecotracer uses Ecosim to predict movement and accumulation of contaminants and tracers in food webs (Coombs 2004, Booth and Zeller 2005). More precisely, changes in concentrations of chemicals (e.g., persistent organic pollutants (POP) and isotope tracers) are predicted using flow rates from Ecosim along with ancillary information, such as isotope decay rate and physical exchange rates (Christensen, 2008). EcoTroph is based on synthetic representation of ecosystems using trophic spectra, i.e. the continuous distribution of biomass (or production, catch, fishing mortality) as a function of continuous trophic levels (Gascuel and Pauly 2009, Gascuel et al. 2011). It enables the simulation of various fisheries changes and their impacts on computed trophic spectra, and provides indicators of the ecosystem impacts of fishing (Tremblay-Boyer et al. 2011, Colléter et al. 2013a, Gasche and Gascuel 2013). The Keystoneness tool in the Network Analyses plug-in may be used to estimate the extent to which any functional group defined in the model functions as keystone species (Christensen et al. 2008). The methodology was implemented based on two indices of keystoneness proposed in the literature: KS#1 (Libralato et al. 2006) and KS#2 (adapted from Power et al. 1996), both based on a consensus definition of keystone species (Christensen et al. 2008).  2.2.2.3  Description of the structure, scales and units of the models In EwE, species are aggregated into functional groups, corresponding to a single species or a group of species sharing similar ecological traits, such as size, diet, predator and life cycle 29  (Christensen et al. 2008). The number of functional groups defined in the 433 EwE models was recorded and used as a basic descriptor of model structure. Models comprising a large number of groups have a low trophic aggregation, and thus are more complex, than models including few groups. Also, the inclusion of multi-stanza groups was considered as a supplementary descriptor of model structure. Multi-stanza groups consist of a set of groups representing different life history stages (or stanzas) that are linked together, and may be used for a species with a complex trophic ontogeny. Thus, species defined using multi-stanza groups add levels of complexity in the models.  When specified in the associated publications, the temporal and spatial scales of the 433 EwE models were collected. Typically, Ecopath models represent an average situation for a period of one year, but some models may correspond to longer time periods. Thus, we compiled as first and last years, the records of the start and end years of the modeled periods, as stated by the modelers in the publications. The first year was used to determine the decade covered by the models, and the time period (in number of years) represented by the models was calculated. The area (in km2) covered by the EwE models was also documented. Lastly, currency and time units of the models were documented. 2.3. Results 2.3.1. General description of the EwE applications published worldwide Globally, EwE models have been mainly developed to study marine ecosystems in the tropical or temperate zones, with some regions better covered than others. 84% of the models (i.e. 365 models) represent marine ecosystems, whereas 15% (63 models) represent freshwater ecosystems and 1% (5 models) terrestrial ones (Figure 2.1a). All FAO areas comprise at least one model, but five areas concentrate about 40 models each: the Northeast Atlantic and the Eastern Central Atlantic comprise 10% of the models each; and the Western Central Atlantic, the Northwest Atlantic, and the Mediterranean and Black Sea comprise 9% of the models each (Figure 2.1b). The Humboldt Current, the Gulf of Alaska, the Mediterranean and the Guinea Current are the LMEs comprising the highest number of models (at least 5% each) (Figure 2.1c). 30  Four LMEs did not comprise any EwE models: the Somali Coastal Current, the Oyashio Current, the East Siberian Sea, and the Laptev Sea. Overall, the Northern and Central Atlantic Ocean is the region with the highest proportion of EwE models, whereas fewer EwE models have been developed for the Indian and Antarctic Oceans. Among the 13 ecosystem types we defined, the most represented ones in the marine category are: continental shelf (32% of the models), bay/fjord (14%), and ocean (13%) (Figure 2.2a). In the freshwater category, lake is the top represented ecosystem type, with 8% of the models. Based on the climatic zone we defined, 49% of the models are located in the tropics, while 44% are located in temperate areas and only 7% in high latitudes. EwE models were used to tackle a wide range of ecological issues (Figure 2.2b). 87% of the models were developed to answer questions regarding the structure and functioning of the ecosystem, 64% to analyze fisheries, 35% to focus on particular species of interest, and 11% to consider environmental variability. Less than 10% of the models tackled issues related to MPA, pollution or aquaculture. Time dynamic simulations in Ecosim were performed for 41% of the models, and spatially-explicit versions were developed in Ecospace for 7% of the models. The Keystoneness tool has been used in 11% of the EwE models (47 models). The EcoTroph and Ecotracer plug-ins have been rarely applied (i.e., in 2% and less than 1% of the models, respectively).  Overall, the majority of the models represented ecosystems between 1980 and 2009, over a time period of one year and an area ranging from 10,000 to 1,000,000 km2, and comprised from 10 to 40 functional groups. About two thirds of the 433 EwE models comprise between 10 and 40 functional groups, with 33% (141 models) including 10 to 20 groups (Figure 2.2c). The numbers of groups range from 7 to 171 groups, but only 5 models include between 75 and 100 groups and 2 models more than 100 groups. Besides, 30% of the models comprise groups corresponding to stanzas. About two third of the models refer to a time period comprised between 1980 and 2009, with 37% (159 models) corresponding to the 1990s (Figure 2.2d). Of the 433 unique models, about 30% (132 models) correspond to replicates of the same ecosystems. In particular, models developed to represent periods before 1950 are almost all (93%) replicates of models representing more recent periods. About two thirds of the models cover a time period lasting from one to five years, with 44% (191 models) corresponding to one year, which is the classical temporal scale of Ecopath models (Figure 2.2e). The longest time 31  period represented by a model is 40 years. The spatial extent covered by the models varies widely, from 0.005 km2 to 34,640,000 km2 (Figure 2.2f). Though, model area does not exceed 1,000,000 km2 for most models, and half of the models cover an area ranging from 10,000 to 1,000,000 km2. Twenty-two models were developed without specifying any temporal scale and 75 models without indicating any spatial scale. Finally, 89% of the 433 EwE models use wet weight as currency unit (of which 88% express it in t·km-²), 5% carbon, and 4% dry weight. Only three models use calories, one model joule and one model nitrogen. Almost all models use year as time unit, while only ten models use day, month, or season.  The 397 EwE-based studies were mainly published in the peer-reviewed literature, with 213 research articles published in 60 different scientific journals. However, only a few of the journals published more than three articles (Table 2.2). Ecological Modelling published the highest number of EwE-based studies (76 articles, i.e. 36%), followed by Estuarine, Coastal and Shelf Science (6%), and Progress in Oceanography (5%).  2.3.2. Evolution of the EwE applications over the last three decades Over the last three decades, EwE applications have progressively been developed in all regions of the world’s oceans. From 1984 to 1993, only eight FAO areas were applied at least one model, and the Western Central Atlantic area has the highest number of models (Figure 2.3a). Then, from 1994 to 2003, more models have been developed, and the Northeast Pacific and Northeast Atlantic areas show the highest number of models, while only two FAO areas located in the Antarctic do not show any model (Figure 2.3b). From 2004 to 2014, all FAO areas have been applied at least one model, and the highest number of models is observed in the Eastern Central Atlantic, while the Mediterranean and Black Sea and the Northwest Atlantic areas also show a high number of models (Figure 2.3c).  During the first decade of its development (1984-1993), the EwE modeling approach essentially consisted of Ecopath models representing tropical marine systems and using a simple trophic structure. In contrast, during the last two decades (1994-2014), EwE models were applied to a wider variety of ecosystems, including high latitudes and terrestrial systems. Besides, the numbers of models representing freshwater ecosystems has doubled over the last two decades (Figure 2.4a). Nonetheless, freshwater applications represent less than 20% of the models 32  published since 1994. Regarding climatic zones, 80% of the models represented tropical systems in the first decade (1984-1993), while the remaining 20% represent temperate systems (Figure 2.4b). Since 1994, the proportions of both tropical and temperate systems have stabilized between 40 to 50%. From 1994 to 2014, 32 models representing ecosystems at high latitudes have been published.  EwE models were also progressively used to answer a wider range of research questions, such as issues related to pollution or aquaculture (Figure 2.4c). Since 1984, the proportion of models developed to analyze ecosystem structure and functioning has remained the highest. However, the proportion of models looking at fisheries-related issues has significantly increased (from 10% in 1984-1993 to 30% in 1994-2003). The number of models built to study environmental variability (including climate change) and MPAs, respectively, has gradually increased since 1984, although their respective proportions have remained lower than 10%. The proportion of models focusing on a particular species (or taxa) of interest has stayed more or less constant over time (at about 15%).  Over the last 30 years, modeling practices have evolved toward Ecopath models with more complex trophic structure, shorter temporal scales, and larger spatial scales. During the first decade of the development of the EwE modeling approach, the total number of groups defined in the models range from 7 to 27 (Figure 2.4d). Over time, the range of the number of groups has expanded toward more groups, up to 67 groups in the last decade (excluding the few outlier models). Besides, the median is around 15 groups in 1984-1993, while it is around 30 groups in 2004-2014. Thus, recently developed models tend to be less aggregated and thus more complex, although highly aggregated models are still being proposed in recent times.  In contrast, the time period represented by the models tends to decrease over the last 30 years (Figure 2.4e). The number of years represented by the models ranges from 1 to 14 in the first decade (1984-1993), while it ranges from 1 to 8 in the last decade (2004-2014) (excluding the few outlier models). The median also shows a decrease over time, from 3 years in 1984-1993 to 1 year in 2004-2014. Besides, the proportion of models corresponding to replicates of the same ecosystem over different (anterior or posterior) time periods has been increasing through time: from 24% of replicates published in 1984-1993, to 27% in 1994-2003, and 33% in 2004-2014. The areas covered by the models have always varied over a wide range (Figure 2.4f). Nevertheless, in the last two decades, the range of the models area has expanded toward very 33  large areas, and the median has shifted accordingly, from about 1,000 km2 in 1984-1993 to about 100,000 km2 in 1994-2014. Over the last 30 years, the numbers of EwE-based publications has steadily increased (Figure 2.5). The annual number of publications reached a peak in 1993, increased slowly from 1994 to 2000 and more rapidly from 2000 to 2004, then stabilized at around 25 references per year. Thus, of the 397 EwE-based studies published over the last three decades, 9% were published in 1984-1993, 29% in 1994-2003, and 62% in 2004-2014 (respectively 35, 114, and 248 publications). Likewise, the cumulated number of EwE users (first authors on publications) has significantly increased over time, from 30 users in 1993, to 98 in 2003, and 230 in 2014 (Figure 2.5). Moreover, both the number and proportion of peer-reviewed publications presenting EwE applications have been growing over the last three decades (Figure 2.5). From 1984 to 1993, most publications corresponded to conference proceedings, mainly from the “Trophic Models of Aquatic Ecosystems” Theme session at the Statutory Meeting of ICES, in Copenhagen, in October 1990 (Christensen and Pauly 1993b). Between 1994 and 2003, EwE-based studies were principally published as reports (with the publications of numerous Fisheries Centre Research Reports) or journal articles. Over the most recent period (2004-2014), publications were predominantly journal articles.  2.4. Discussion 2.4.1. Challenges and potentials of the compilation and meta-analysis of EwE models All the metadata analyzed in this study were collected or derived from the publications presenting the EwE models. However, detailed information on some metadata was sometimes missing for many models, which prevented us from using the whole collection of metadata initially envisaged. Lack of information was sometimes observed for metadata of potentially high relevance to the EwE modeling approach. For instance, we obtained 5% of missing data regarding the time period represented by the models, and 17% for the area covered by the models. Though, it is critical to clearly define and indicate the temporal and spatial scales when 34  developing a model. When the geographic extent of the models was poorly described, no spatial shape could be defined for these models. Likewise, several metadata describing the physical characteristics of the modeled ecosystems were considered in EcoBase (e.g., temperature, depth, salinity, oxygen, primary production), but information on abiotic conditions was lacking for most models. Remarkably, information on minimum and maximum depths of the area covered by the models, which is critical to determine the type of ecosystem represented by the models, was sometimes not provided in sufficient details. Moreover, the compilation of the metadata was sometimes challenging due to some ambiguity in the description of the model. Some models included functional groups labeled as stanzas, but which were not always properly defined as multi-stanzas groups. Thus, we were not able to analyze in many details the usage of multi-stanzas groups.  For some metadata, the proportion of models with non-available information was too high for the metadata to be representative. This was notably the case with the version of the model used by the modeler. Since the EwE software evolved with time and upgraded versions were successively released, we intended to analyze the evolution of the use of the different versions. The first version of EwE, ECOPATH, was used only in the early 1980s (Polovina 1984), and the development of a user-friendly software in the early 1990s (version 2, Christensen and Pauly 1992) rapidly led to a broader use of the model. Versions 3 and 4 only had limited use and were rapidly replaced by version 5 (Christensen et al. 2005), which is now itself replaced by version 6 (Christensen et al. 2008). However, the EwE software versions were only specified by modelers for half of the models, so that we could not explore this aspect of the modeling approach much further. Likewise, the Pedigree index, corresponding to an estimation of the data quality of the models (Pauly et al. 2000), could not be analyzed since it was provided for only very few models.  The challenges we faced during the collection and compilation of the metadata emphasized the need for defining and enforcing best modeling practices (FAO 2008). Precisely, we would strongly recommend EwE modelers to provide basic information about their models in sufficient details in the associated publication(s). As a template, we would suggest the following metadata to be considered as essential: location, objectives, area, time period, units, software version, Pedigree index, trophic structure, and environmental conditions. Requiring such information to be systematically provided would facilitate the reuse of the models, and thus 35  increase the scientific value of EwE-based publications. Indeed, the metadata compiled in this study could serve as valuable criteria based on which EwE models would be selected for future studies. By applying a scoring method on these selection criteria, a pool of models of potential interest would be defined and the selected models could then be reused to perform EwE-based meta-analyses. Thus, EcoBase offers a framework where the metadata could be stored, in a standardized and granular fashion, so that they could be easily reused.  2.4.2. Conclusions on the EwE modeling approach: past, present and future The metadata compiled for this study were used to give a global overview of the EwE models and EwE-based publications. We also analyzed the evolution of the EwE applications over time, based on the year of publication of the models. The initial emphasis on the tropics was due to the development of EwE initially being centered at the International Center for Living Aquatic Resources Management (ICLARM, now WorldFish), which was based in the Philippines, and which was focused on developing methodologies for managing tropical ecosystems (Christensen and Pauly 1993b). EwE was also early applied to freshwater ecosystems, likely due to the ease for defining the boundaries and components of systems such as lakes or rivers. Over the last three decades, some regions have been widely analyzed using the EwE modeling approach, while others have remained poorly-studied. The focus of EwE applications on some regions, such as the North and Central Atlantic Ocean, were driven by research programs and funding opportunities, but also by data availability. In contrast, future modeling efforts could be concentrated in the Indian and Antarctic Oceans, seeing the smaller number of applications in these regions. Terrestrial systems have been very poorly studied using EwE until now. However, interests have recently been growing for applying the EwE modeling approach to terrestrial case studies (Fretzer 2014).  Over the last three decades, the research questions addressed using EwE models have progressively become more diversified. Particularly, the usefulness of EwE models for fisheries management has been demonstrated in several cases, such as upwelling systems (Jarre-Teichmann 1998) or the Mediterranean Sea (Coll and Libralato 2012). More generally, the main strengths of EwE relate to the improved understanding of the trophic structure of the modeled ecosystem and the possible comparison between modeled ecosystems (Jarre-Teichmann 1998, 36  Plagányi and Butterworth 2004). In parallel, the complexity of the EwE models has been increasing, notably with the inclusion of more functional groups and the evolution toward a partitioning approach for the representation the food web in some cases. However, a significant proportion of recent EwE applications focused on some species (or taxa) of particular interest, so that the representation of the food web was centered on these species, using a more selective approach.  Despite the development of the Ecosim routine allowing for time-dynamic simulations (Walters et al. 1997, Walters et al. 2000, Walters et al. 2006, Walters and Christensen 2007), the static Ecopath routine has also been used to analyze changes in ecosystems over time, through the comparison of replicate models. Indeed, using replicates of Ecopath models may constitute an easier (and complementary) approach than performing simulations in Ecosim, which are more complex and data-demanding. Despite its complementarity with Ecosim, the Ecospace routine has been little used to date (7%). In fact, Ecospace has been proposed more recently (Walters et al. 1999) and the latest improvements made to the routine have not been widely disseminated yet (Walters et al. 2010, Christensen et al. 2014b). Likewise, EcoTroph is a relatively recent plug-in (Gascuel and Pauly 2009), which has thus been rarely applied, but the recent implementation of the plug-in in R (Colléter et al. 2013a) may allow for a wider application. With the release of EwE version 6, more flexibility was implemented in the software and users were given access to the source code of the model. Thus, EwE modelers will be more and more inclined to develop their own plug-ins (e.g., Coll and Steenbeek 2014) and take full advantage of the multiple potentials of the EwE modeling approach, including the EcoBase repository.   37  Table 2.1. List of the metadata used for the analyses. Metadata are listed in alphabetical order, with their corresponding fields in EcoBase and definitions.  Metadata EcoBase field Definition Aquaculture aquaculture Was the model developed to analyze aquaculture? (TRUE/FALSE) Area area Model area (in km2) Climatic zone classification Climatic zone of the modeled ecosystem Currency unit currency_units Currency unit of the model, e.g., wet weight, in t·km-2 Ecopath ecopath Is the Ecopath routine used? (TRUE/FALSE) Ecosim ecosim Is the Ecosim routine used? (TRUE/FALSE) Ecospace ecospace Is the Ecospace routine used? (TRUE/FALSE) Ecosystem ecosyst_functioning Was the model developed to analyze ecosystem functioning/structure? (TRUE/FALSE) Ecosystem category ecosystem_category Category of the modeled ecosystem Ecosystem type ecosystem_type Type of the modeled ecosystem Ecotracer ecotracer Is the Ecotracer plug-in used? (TRUE/FALSE) EcoTroph ecotroph Is the EcoTroph plug-in used? (TRUE/FALSE) Environment environment_variability Was the model developed to analyze climate variability? (TRUE/FALSE) First Year model_year_start Start year of the model time period Fisheries fisheries Was the model developed to analyze fisheries? (TRUE/FALSE) Geographic extent geographic_extent Spatial polygon of the model (in decimal degrees) Keystoneness keystone_analysis Is the Keystoneness tool in the Network Analysis plug-in used? (TRUE/FALSE) Last Year model_year_end End year of the model time period Model model_number Model number (3-digits code) as recorded in EcoBase MPA mpa Is the modeled ecosystem a Marine Protected Area? (TRUE/FALSE) Number of groups number_groups Number of functional groups included in the model Pollution pollution Was the model developed to analyze pollution? (TRUE/FALSE) Publication id_ref Publication number (3-digits code) as recorded in EcoBase Publication xml_ref Reference in xml format (including author, year of publication, and type of publication) Reference referenced Is the reference of the model recorded in EcoBase? (TRUE/FALSE) Reference ref_importance Importance of the reference (1 by default) Replicates  Model number(s) of the replicate(s) Replicates overlapping_model Are there other models representing the same ecosystem but over a different time period? (TRUE/FALSE) Species species_of_interest Was the model developed to study a particular species/group? (TRUE/FALSE) Stanzas stanza_groups Are there any stanzas in the model? (TRUE/FALSE) Time unit time_units Time unit of the model Version ewe_version_original Version of EwE used by the modeler 38  Table 2.2. Number and percentage of articles presenting EwE-based studies, detailed by journal.  Journal Number of articles Percentage Ecological Modelling 76 36 Estuarine, Coastal and Shelf Science 12 6 Progress in Oceanography 10 5 Journal of Marine Systems 9 4 Marine Ecology Progress Series 9 4 Canadian Journal of Fisheries and Aquatic Sciences 8 4 African Journal of Marine Science 5 2 Bulletin of Marine Science 5 2 Fisheries Research 5 2 ICES Journal of Marine Science 4 2 Journal of Experimental Marine Biology and Ecology 4 2 Others 66 31  39   Figure 2.1. World maps of the EwE models documented in EcoBase representing (a) the centroids of the geographic extent of the models and the number of models (b) by FAO area and (c) by LME. Only the 365 models representing marine ecosystems are represented on the FAO area map, and the 345 models representing coastal marine ecosystems on the LME map. 40   Figure 2.2. Classification of the 433 unique EwE models documented in EcoBase by: (a) type of ecosystem represented by the models; (b) research questions answered with the models; (c) number of functional groups of the models; (d) decade of the first year of the models; (e) time period (in number of years) represented by the models; and (f) area (in km2) covered by the models. On each graph, values are displayed in percentage on the x-axis, and in number on the top of each bar. Missing data are displayed as ‘NA’ for Non Available.  12345678910> 10NAPercentage of models0 10 20 30 40 5019147402315141644273022Before 19501950s1960s1970s1980s1990s2000sNAPercentage of models0 10 20 30 4014181952761597322(0,10](10,20](20,30](30,40](40,50](50,75](75,100](100,200]Percentage of models0 10 20 30 40271419691422952AquacultureEcosystemEnvironmentFisheriesMPAPollutionSpeciesPercentage of models0 20 40 60 80 1001537648276305150bay/fjordbeachchannel/straitcoastal lagooncontinental shelfcoral reefestuarylakeoceanreservoirriverterrestrialupwellingPercentage of models0 10 20 30 4059417361402219335783530(0,100](100,1e+04](1e+04,1e+06](1e+06,1e+08]NAPercentage of models0 10 20 30 40 50501041842075 41   Figure 2.3. World maps of the number of EwE models documented in EcoBase by FAO area, for each of the last three decades: (a) 1984-1993; (b) 1994-2003; and (c) 2004-2014.  42   Figure 2.4. Comparison of the classified EwE models between the last three decades: 1984-1993; 1994-2003; and 2004-2014; based on the year of publication of the models. Classification of the models by: (a) category of ecosystem represented by the models; (b) climatic zone of the modeled ecosystems; (c) research questions answered with the models; (d) number of functional groups of the models; (e); time period (in number of years) represented by the models; and (f) area (in km2) covered by the models. On the bar plots, values are displayed in percentage on the y-axis, and in number on the top of each bar. On the boxplots, the horizontal line in the box represents the median number of models.  1984-1993 1994-2003 2004-2014MPAAquaculturePollutionEnvironmentSpeciesFisheriesEcosystemPublication periodPercentage of models0204060801001 0 0585375 7 184580111248 435971912281984-1993 1994-2003 2004-2014FreshwaterMarineTerrestrialPublication periodPercentage of models020406080100162201711043023311984-1993 1994-2003 2004-2014TropicalTemperateHigh latitudesPublication periodPercentage of models0204060801003080535919129 122131984-1993 1994-2003 2004-2014050100150Publication periodNumber of groups1984-1993 1994-2003 2004-2014010203040Publication periodTime period (years)1984-1993 1994-2003 2004-20141e-021e+001e+021e+041e+06Publication periodArea (km²)  43   Figure 2.5. Number of publications presenting EwE-based studies over the last thirty years. For each decade (1984-1993, 1994-2003, and 2004-2014), a pie chart shows the repartition of the publications by type. The two vertical dotted lines highlight the transition years between the three decades. The black solid line represents the number of all types of scientific publications while the gray solid line represents journal articles only. The dotted line represents the cumulated number of first authors associated to the publications from 1984 to 2014.   Book sectionConference paperJournal articleReportThesis1985 1990 1995 2000 2005 2010 2014010203040506070Publication yearNumber of publications0100200300400Cumulated number of different first authors 44  3. Proposal for a restored definition of the concept of ‘keystone species’  3.1. Introduction The ecological term of ‘keystone’ species, introduced in aquatic food web ecology by R. T. Paine (1969a, Paine 1969b), is based on the analogy between a natural ecosystem and a stone arch. The keystone refers to the stone that maintains all the others in the arch. Both the species and the stone “derive their functional importance to the system as a whole from bidirectional interactions with lower energy levels” (Davic 2003). If the keystone is removed, the structure of the arch is removed as well, as the remaining stones can only form simpler structures, such as a stack or a wall (Platten and Henfrey 2009). Therefore, in the community, as well as in the stone arch, the keystone species fulfills an irreplaceable function by maintaining its structure (Paine 1969a), and only a limited number of species can be keystone (Power et al. 1996).  Since its introduction, the keystone species concept had been applied to a large number of aquatic and terrestrial species, playing a variety of critical roles in the ecosystem (Paine 1995, Power and Mills 1995, Power et al. 1996). The relevance of such an over-used ecological concept was much debated in the literature (Hurlbert 1997), either described as a powerful buzzword in conservation (Barua 2011), or criticized as an ambiguous and inclusive concept (Cottee-Jones and Whittaker 2012). The concept first appeared to be attractive to managers, looking for species-level regulations in a context of ecosystem-based approaches to conservation (Simberloff 1998). The notion seemed valuable for communicating on the fragility and importance of balances in natural systems (Paine 1995, Piraino et al. 2002). Because of its metaphorical nature, which evokes concrete images, the keystone species term was very popular to both the general public and the scientific community, and helped bridging the gap between scientific knowledge, conservation action, and public awareness (Barua 2011). For instance, ‘keystone species’ was the most frequently encountered metaphorical ecological concept in English newspapers during an 8-year period starting in 2000 (Barua 2011), and was used in 1,600 scientific articles’ title or topic description published by  2012 (Cottee-Jones and  45  Whittaker 2012). The keystone species concept was seen as a helpful concept for (i) identifying species whose removal would have unsuspected and dramatic consequences on their communities; (ii) understanding the complex underlying mechanisms of community dynamics; and (iii) considering the whole ecosystem functioning and structure to inform biodiversity conservation (Paine 1995, Power et al. 1996, Simberloff 1998). Paine’s articles introducing the concept and published in 1966 and 1969 were cited 2,509 and 465 times in the peer-reviewed literature, respectively (Cottee-Jones and Whittaker 2012).  Yet, the wide-spread use of the terminology led to misusages, and semantic confusion with other overlapping concepts was introduced. For instance, the keystone species term was defined in only 35% of the English news articles cited above, with less than half of the articles providing accurate definitions (Barua 2011). Ecosystem engineer species were notably found to be often mislabeled as keystone species (Barua 2011, Cottee-Jones and Whittaker 2012). In fact, the keystone species term (like several other scientific terms) had been promoted by the scientific community, before any consensus on its exact meaning and implications for biodiversity conservation was reached, leading to high misrepresentation (Hurlbert 1997, Barua 2011, Cottee-Jones and Whittaker 2012). Consequently, the keystone species concept rapidly evolved to become what might be called a ‘non-concept’ (a concept with uncertain meaning), or a ‘panchestron’ (a concept “which can explain anything and thus nothing at all” (Hardin 1957)) (Hurlbert 1997, Cottee-Jones and Whittaker 2012). As pointed out by Paine (1995), “An answer serving all masters probably provides few useful solutions”. Consequently, the keystone species concept was criticized for its lack of clear and operational definition, preventing its implementation for management and policy-making (Mills et al. 1993, Hurlbert 1997). Often, species depicted as keystone gained preference over the ultimate objective of biodiversity conservation (Barua 2011).  In this study, we argued for the restoration of the over-used concept of keystone species. We intended to clearly define what are and are not ‘keystone species’, and proposed a restored definition of the keystone species concept. To do so, we conducted a thorough literature review and analyzed the origin and evolution of the multiple definitions, terms, and examples associated to the keystone species concept in the peer-reviewed literature. First, we built an evolutionary tree to represent the usages of the concept over time in the scientific literature. Then, we proposed a comparative framework describing the different categories of species of potential  46  interest, including keystone species. We used our framework to differentiate examples of misusages from appropriate applications of the keystone species terminology in the literature. We concluded by a conceptual diagram representing a restrictive and criteria-based definition of the keystone species concept, which was confronted to related notions in functional ecology and food web theory. Lastly, challenges related to the context-dependency and applicability of the keystone species concept were discussed and illustrated. 3.2. Methods We conducted a historical review of the peer-reviewed literature on keystone species. Using Google Scholar, we gathered research articles published in scientific journals, from 1966 to 2013, and written in English. We first selected peer-reviewed articles with a clear focus on keystone species (term explicitly mentioned in their title or topic description). We then extended our selection to articles dealing with concepts overlapping with the keystone species concept (e.g., ecosystem engineers). We identified 101 relevant articles on keystone species, with a broad scope and including detailed semantic, practical and/or theoretical considerations. We used the selected articles to analyze the origins and evolution of the keystone species concept. The multiple definitions, terms, and examples of keystone species in marine, freshwater or terrestrial systems, proposed in the 101 articles were collected and synthetized.  First, we used an evolutionary tree to represent the temporal evolution of the semantic usages of the keystone species concept. We summarized on the tree the different terminologies that were introduced over time in the scientific literature. Then, we developed a comparative framework describing the different categories of species of potential interest to conservation and management, including keystone species. For each category, a detailed definition, illustrative examples and references from the literature were compiled, so as to clearly differentiate the concept of keystone species from overlapping concepts.  Keystone species is an ill-defined concept, so that there are many species once labeled as keystone that probably should not be called so (Zacharias and Roff 2001). Thus, we used our framework to differentiate examples of misusages from appropriate applications of the keystone species concept in the literature. Species frequently labeled as ‘keystone’ in the literature, but not matching all criteria to be sensu stricto keystone species, were labeled as ‘pseudo-keystone’  47  species. We determined the categories to which they would rather correspond by comparing the description of their roles in the literature to the definitions compiled in our conceptual framework. We focused on the best-known examples of sensu stricto keystone species and the most commonly cited example of pseudo-keystone species. We acknowledged that many more examples could be found in the literature but the ones listed in our review were the most representative ones. Besides, the majority of the examples that could be encountered in the literature would correspond to pseudo-keystone species, because of the increasing misusage of the concept.  Lastly, we identified the required attributes for a species to be defined as keystone species. We looked back at the original definition proposed by Paine (1969b): “The patterns of species occurrence, distribution and density are disproportionately affected by the activities of a single species of high trophic status”. Based on the latter definition, we determined the exclusive attributes of keystone species. We used a conceptual diagram to represent the restored definition of the keystone species concept. Our definition includes the four primary web features that were highlighted by Paine (1980): (1) the number of species involved, (2) the nature of their connections, (3) the number of connections per species, and (4) the intensity of connections between species.  In our study, we claimed that the concept of keystone species cannot be dissociated from the notion of species. Therefore, we avoided to use some terms commonly found in the literature, such as ‘keystone role’ or ‘keystone effect’, and restricted ourselves to the ‘keystone species’ or ‘keystoneness’ (i.e. the potential of a species to be keystone) terminology.  3.3. Results 3.3.1. Evolution of the keystone species concept We reviewed the evolution of the keystone species term over time in the peer-reviewed literature and summarized it in an evolutionary tree (Figure 3.1).  Paine (1966) first described the underlying mechanisms of the keystoneness of a predator species in rocky intertidal marine communities. The keystone species played the critical role of  48  maintaining high species diversity and complex trophic structure in the community (Paine 1966). Thus, variations in the keystone species abundance or activity would have greater impacts on the biodiversity and trophic structure, compared to other coexisting species with similar or higher abundance (Paine 1969a, Paine 1969b). Paine’s definition emphasized three main criteria for a species to be a keystone species: a single, native, and predator species. Moreover, Paine’s definition highlighted two main traits of keystone predator: (i) feeding on a highly competitive prey; and (ii) predating on this prey at all stages of the prey life cycle (Paine 1980).  Since Paine’s analogy, the keystone predator concept rapidly expanded to include almost all species playing just about any role in the community, provided that their effects were considered to be significant at the ecosystem scale. After reviewing the usages of the keystone species term in the literature, Mills et al. (1993) proposed a typology comprising several terms describing keystone species, deviating from the original term: ‘keystone predator’, but also ‘keystone prey’, ‘keystone host’ (also called ‘keystone resource’, such as plants or corals), ‘keystone mutualist’ or ‘keystone link’ (e.g., pollinators, seed dispersers), ‘keystone modifier’ (e.g., beavers building dams, termites building mounds), and ‘keystone herbivore’ species (e.g., sea urchins). Bond (1994) also proposed a classification of the different types of keystone species discussed in the literature. Some of the types corresponded to the ones previously described (‘keystone predators’, ‘keystone herbivores’, ‘keystone mutualists’ (including ‘keystone plant resources’), and ‘earth-movers’), while others were newly formulated (‘keystone pathogens’, ‘keystone competitors’, and ‘abiotic processes’) (Bond 1994). The author acknowledged that there was little supporting evidence for some of the terms, which were not explicitly associated to the terminology keystone (e.g., ‘dispersers’, ‘pollinators’, or ‘system processes’), while the best evidence was for ‘keystone predator’ (Bond 1994). Both of the latter typologies illustrates that the concept was applied to a variety of species, occupying different trophic levels and having diverse types of interactions within their communities (not only trophic, but also physical interactions through the habitat) (Mills et al. 1993).  The expansion of the keystone species concept was discussed by the “keystone cops”, a group of experts who attempted to refine the notion during a workshop in 1994 (Paine 1995, Power and Mills 1995). The experts reached a consensus on the following definition: “a species whose impacts on its community or ecosystem are large, and much larger than would be expected from its abundance” (Power and Mills 1995, Power et al. 1996). Thus, any species  49  causing disproportionate effects on its community could be considered as a keystone species, including: predators, competitors, mutualists, seeds or spores dispersers, pollinators, producers, parasites, vectors of disease, and modifiers of habitats or abiotic factors (Paine 1995, Power et al. 1996). The workshop participants concluded that a keystone species may occur in all ecosystems, at all trophic levels, and use all kinds of ecological processes (Paine 1995, Power and Mills 1995, Power et al. 1996). Nevertheless, they acknowledged that examples of keystone species with a low trophic level did not meet Paine’s original criteria, but were essentially based on the large impact of the species, unexpected from their relatively low abundance (Power et al. 1996).  A keystone species had simply become a species with high functional importance (Hurlbert 1997), and was often equally defined and labeled as a ‘key’ species (e.g., Khanina 1998, Higdon 2002). Despite the call of some authors against the misinterpretation of Paine’s original definition (Piraino and Fanelli 1999, Davic 2000), the keystone species concept was applied to an increasing number of species, based on inclusive (re)definitions. For instance, Menge and Freidenburg (2001) listed all kinds of consumers as potential keystone species (comprising predators, parasites, pathogens, herbivores, pollinators, and mutualists, but not plants), while Davic (2002) argued that the concept could be expanded to any species regulating functional diversity, such as primary producers, fungi and bacteria.  The keystone term was progressively detached from the notion of species. In some cases, ‘keystone’ was extended to other notions with a broader scope, such as ‘keystone guilds’ (Brown and Heske 1990), ‘keystone habitats’ (Davidar et al. 2001), ‘keystone structures’ (Tews et al. 2004), ‘keystone populations’ (Perry 2010), ‘keystone communities’ (Mouquet et al. 2012), or ‘keystone species complexes’ (Daily et al. 1993, Ortiz et al. 2013b). In other cases, the keystone term was applied at smaller scales, to the extended concepts of ‘keystone molecules’ (Zimmer and Ferrer 2007) or ‘keystone individuals’ (Modlmeier et al. 2014). Recently, Cottee-Jones and Whittaker (2012) proposed a generic definition: “a keystone species is of demonstrable importance for ecosystem function”, where the term ‘species’ could be replaced with any other ecologically relevant term. Specific traits associated to the historical definition (e.g., single species, disproportional impact and food web structure) were abandoned (Cottee-Jones and Whittaker 2012).   50  3.3.2. Different categories of critical species, including keystone species Seeing the ambiguity around the keystone species terminology, a clear discrimination between the concept of keystone species and all overlapping concepts was required. We extricated the keystone species concept of all the other concepts used in the literature to describe species which are of potential interest, yet not necessarily keystone (Table 3.1, Figure 3.2 and Figure 3.3).  First, the concept of keystone species is sometimes assimilated to ‘invasive’ and ‘endangered’ species. However, these two terms should be differentiated from the other ecological concept listed below. Often because of human activities, invasive species are introduced to a new ecosystem, where they may cause transient disturbance, until they get established or perish. The ecological role of invaders remains uncertain until they manage to establish themselves in the new ecosystem (Bond 2001). So, the potential effects of invasive species occur beyond the evolutionary time scale of the community (Paine 1980). Moreover, once invaders are established in the new ecosystem, they may become abundant and dominate the species assemblage. Therefore, invaders should not be considered as keystone species in the introduced range. We note that invasive species acting as ecosystem engineers were given a different name in the literature, and called ‘transformer species’ (Richardson et al. 2000). Thus, invaders acting as keystone species should also be given another name to differentiate them from native species. Endangered species, such as the ones listed on the IUCN Red List (www.iucnredlist.org), are threatened by extinction in their native range, so that their ecological roles may be impaired (Powles et al. 2000). Such species may have played a keystone or another important role in the ecosystem in the past, when they were abundant enough. Yet, due to anthropogenic or environmental threats, they sometimes become ‘ecologically extinct’ and are not able to fulfill any ecological role anymore (e.g., overhunted sea otters) (Jackson 2008).  3.3.2.1  Focal species versus surrogate species We clearly discriminated two broad categories of species: ‘focal species’ and ‘surrogate species’. Focal species correspond to species which are selected to preferentially focus on, among all species in the community (Armstrong and Caro 2002). Surrogate species designates  51  species that are substitute for something else (Armstrong and Caro 2002), used as tools to achieve another purpose (Caro and O'Doherty 1999), and considered as “shortcuts” for biodiversity management and conservation (Roberge and Angelstam 2004). Thus, a focal species is assumed to be of importance to its ecosystem, whereas a surrogate species may not be of intrinsic interest (Armstrong and Caro 2002). Both terms were introduced at the same period: focal species was first coined by Lambeck (1997), while surrogate species was first used by Caro and O’Doherty (1999). Although the two terminologies imply different assumptions and objectives, they have often been used very loosely in the literature, sometimes as synonyms (Caro 2010). That is why some authors recommended to “define precisely what we are talking about instead of using insider jargon”, such as surrogate or focal species (Armstrong and Caro 2002).  3.3.2.2  Sub-categories of focal species In this study, we defined critical species as focal species of ecological importance that play an important role for the structure, composition or functioning of its ecosystem, so that the absence of these species would lead the loss of many others. Critical species may also be called ‘ecologically important species’ (Perry 2010). Critical species had been compared, and sometimes assimilated, to critical processes (e.g., Bond 1994), since some critical processes may be driven by critical species (Power et al. 1996). Yet, the biotic and abiotic factors influencing ecosystems are qualitatively distinct, so that critical processes should be differentiated from critical species (Menge and Freidenburg 2001). Here, we defined critical processes as abiotic factors that can structure communities, such as fire or frost (Menge and Freidenburg 2001). The adjective ‘key’ has sometimes been used to describe critical species or processes (e.g., Bond 1994, Piraino and Fanelli 1999), but we preferred to not to use it, due to the frequent confusion between ‘key’ and ‘keystone’ in the literature (e.g., Khanina 1998). Critical species may have various ecological functions in the ecosystem, and we therefore differentiated several sub-categories of critical species. The first three sub-categories of critical species (‘keystone’, ‘foundation’ and ‘connector’ species) were defined based on their abundance and position in the food web. In contrast, the last two sub-categories (‘engineer’ and ‘modifier’ species) were defined by their relationships to the physical environment.  52  A species may exert a high influence on the community due to its high abundance, and may be described as a ‘dominant’ species (Power et al. 1996). Alternatively, a species may have strong effects on communities and ecosystems despite its low abundance, and thus be described as a ‘keystone’ species (Power et al. 1996, Menge and Freidenburg 2001). Keystone species are predators which strongly influence the food web structure of the community (Paine 1966, 1969a). Although keystone species may be defined by opposition to dominant species (Power et al. 1996), the disproportionate influence of a species is determined relatively, by comparison with the influence and abundance of the other species (Cottee-Jones and Whittaker 2012). Moreover, the dominance of a species is a relative notion as well, defined by comparing its biomass to the biomass of other species. Thus, the species identified as dominant may vary with the community considered within the ecosystem. For instance, Davic (2003) identified as keystone some species which were dominant within their functional group. Thus, we chose not to use the ‘dominant’ terminology to describe critical species, and restricted its meaning to species of high abundance in its community. Instead, we discriminated two other sub-categories of critical species, defined based on their influence on the community: ‘foundation’ species and ‘connector’ species.  Foundation species are abundant species which define the structure of the community (Dayton 1972). They also create stable conditions for other species and control ecosystem dynamics, by modulating fundamental ecosystem processes (Ellison et al. 2005, Caro 2010). Ecosystems are often named after foundation species, which are commonly low trophic level species involved in bottom-up controls, such as primary producers or habitat-forming species (e.g., redwood forests, seaweed beds or coral reefs) (Ellison et al. 2005, Sala and Knowlton 2006, Caro 2010). Note that foundation species were originally labeled as ‘dominant’ species (Clements 1936).  Connector species maintain the food web structure of the community, through wasp-waist controls (Cury et al. 2000). They generally correspond to abundant consumers of intermediate trophic level, such as small pelagic fish (Menge and Freidenburg 2001). They were originally described as prey supporting many predators and labeled ‘key-industry’ species (Elton 1927), but we called them ‘connector’ since they act as links between species at lower and higher trophic levels.   53  ‘Ecosystem engineers’ were originally described as species that influence the availability of resources to other species by modifying living or non-living materials (Jones et al. 1994). More precisely, two types of engineer species were defined: (i) ‘allogenic engineers’, which modify their habitat by causing the transformation of the physical environment from one state to another, such as beavers; and (ii) ‘autogenic engineers’, which modify their own physical structure, such as corals (Jones et al. 1994, 1997). However, Ellison et al. (2005) established that such discrimination was unnecessary, since autogenic engineers correspond to foundation species, as defined above. Besides, some species defined as engineer species have also been described as keystone species (Barua 2011, Cottee-Jones and Whittaker 2012), notably with the introduction of the ‘keystone modifier’ (Mills et al. 1993) or ‘earth-movers’ (Bond 1994) terms. However, engineer species differ from keystone species since they modify habitat through their non-foraging activities (Jones et al. 1997, Menge and Freidenburg 2001). Thus, to avoid confusion, the alteration of living resources through herbivory should be differentiated from engineering, even when causing important modification of the habitat. Herbivore species have also been associated to keystone species in the literature (Mills et al. 1993, Bond 1994, Menge and Freidenburg 2001). Yet, grazing animals influence the community structure and ecosystem dynamics by both their feeding strategies and their physical alterations of the habitat (e.g., moose or elephant) (Naiman 1988). So, we suggested to differentiate them from both engineer and keystone species and to label them as ‘modifier species’.  A supplementary sub-category of critical species could be considered: ‘mutualist’ species, such as seed-dispersers or pollinators. They have been described as keystone species (Mills et al. 1993), although evidence of their potential influence on the whole community was limited (Bond 1994). Indeed, mutualism consists in a reciprocally beneficial interaction between two species (Stachowicz 2001). However, in aquatic ecosystems, mutualism commonly support foundation species forming habitat (e.g., corals and zooxanthellae), so that the removal of some mutualist species could actually have community-wide consequences (Stachowicz 2001, Hay et al. 2004).   54  3.3.2.3  Sub-categories of surrogate species Surrogate species are commonly described using metaphorical terms, such as ‘flagship’ or ‘umbrella’ species, which are appealing to the general public (Barua 2011). Due to the symbolic nature of the keystone term, keystone species have sometimes been associated to the metaphorical concepts used to designate surrogate species, all defined either as focal (e.g., Zacharias and Roff 2001), or as surrogate species (e.g., Favreau et al. 2006, Sergio et al. 2008). However, keystone species play a critical role in the food web, and thus correspond to a sub-category of focal species. In contrast, surrogate species may not ensure any important ecological function in their ecosystem (Armstrong and Caro 2002). Although some keystone species may be suitable tools for management (Simberloff 1998), effective surrogate species may not be keystone species (Caro and O'Doherty 1999). Moreover, there has been some semantic confusion between the different sub-categories of surrogate species: ‘indicator’, ‘umbrella’ and ‘flagship’ species (Caro and O'Doherty 1999, Caro 2010). However, a species useful as one type of surrogate, to achieve one particular conservation objective, will generally be unusable as another type, since species traits differ from one surrogate species to another (Caro and O'Doherty 1999).  The concept of indicator species was introduced in the literature as an index of other species attributes or environmental conditions, otherwise challenging to estimate (Landres et al. 1988). A dichotomy was later established between ‘composition’ indicator species (also called ‘ecological’ or ‘environmental’ indicator species) used to select areas of high biodiversity to be protected, and ‘condition’ indicator species (also called ‘bio-indicator’ or ‘sentinel species’) used for monitoring changes due to anthropogenic or natural disturbances (Caro and O'Doherty 1999, Zacharias and Roff 2001). For instance, species at the upper end of the food chain, such as marine mammals, are often good indicators of the health of the ecosystem (Bossart 2006).  The umbrella species concept was first coined in the 1980s, then more precisely defined in the 1990s (e.g., Noss 1990) as species with a large distribution area, so that protecting their habitat area would also protect many other naturally co-occurring species (Caro 2003, Roberge and Angelstam 2004). Due to the allometric relationship between body size and home range size, effective umbrella species are large species (Caro and O'Doherty 1999). Umbrella species are used to determine the size of areas to be protected (Caro and O'Doherty 1999, Roberge and Angelstam 2004). In practice, the concept of umbrella species appears to be more suitable for  55  large terrestrial species (such as tiger, jaguar, or bears) (Barua 2011). Large marine species, which often are migratory species, may be more suitable as flagship species (Zacharias and Roff 2001).  Flagship species are charismatic and popular species that serve as ambassadors of broader conservations goals, often the preservation of a particular habitat or ecosystem (Heywood 1995, Caro and O'Doherty 1999, Zacharias and Roff 2001, Caro et al. 2004). The most charismatic species in ecosystems are not always the most critical to the structure and function of communities (Vanclay 1999). Thus, flagship species may be defined as mere strategic tools selected based on their symbolic or marketing value to raise public concern and encourage action and funding (Verissimo et al. 2011). So, the concept of flagship species differ from the other concepts since it is not based on any ecological foundation (i.e. flagship species do not need to meet any ecological criteria) (Caro and O'Doherty 1999, Zacharias and Roff 2001). Flagship species may encompass a wide range of examples, from the giant panda, symbol of species conservation, to the polar bear, emblem of climate change mitigation (Barua 2011).  3.3.3. Examples of sensu stricto keystone versus pseudo-keystone species We listed examples of species commonly cited in the literature, and discriminated the ones corresponding to sensu stricto keystone species (Table 3.2) from pseudo-keystone species (Table 3.3).  The first classical and historical example of sensu stricto keystone species is the carnivorous sea star Pisaster ochraceus (Brandt, 1835). By feeding on the most competitive herbivorous prey in the rocky intertidal community, the mussel Mytilus californianus (Conrad, 1837), the sea star prevents this prey from dominating the available substrate, thereby opening it up to other less competitive prey species (Paine 1966, 1969a, 1980). The second most popular and historical example of keystone species is the sea otter Enhydra lutris (Linnaeus, 1758). Sea otters, by feeding on the sea urchin Strongylocentrotus purpuratus (Stimpson, 1857), prevent the kelp forest community from shifting (through trophic cascades) to an urchin barren dominated state (Estes and Palmisano 1974, Estes et al. 1978). Sea snails, respectively lobsters, were also shown to play the same role of keystone species as sea stars in rocky intertidal ecosystems (Duran and Castilla 1989), respectively sea otters in kelp forests (Mann and Breen 1972,  56  Babcock et al. 1999, Shears and Babcock 2002). Other aquatic examples of keystone species correspond to predatory fish species feeding on herbivore prey, such as basses in freshwater systems (Power et al. 1985) or sea urchin-eater fishes in coral reefs (McClanahan 1995, McClanahan 2000) and rocky reefs (Sala and Zabala 1996, Sala 1997, Clemente et al. 2010). There have been fewer terrestrial examples of sensu stricto keystone species proposed in the literature, with the most common one being the wolf feeding on large herbivores in temperate and boreal forests (Messier and Crête 1985, McLaren and Peterson 1994, Ripple and Larsen 2000, Beschta 2003).  The most cited example of pseudo-keystone species is the North American beaver (Barua 2011). Beavers were often labeled as keystone (habitat) modifier, although they correspond to engineer species (Jones et al. 1994). In contrast with sensu stricto keystone species, there have been many terrestrial examples of pseudo-keystone species. Prairie dogs are the second most cited example of pseudo-keystone species (Barua 2011). The keystoneness of several species of small rodents, such as prairie dogs, has been much debated in the literature. Small rodents, such as prairie dogs, seem to play multiple ecological roles, while mainly impacting the ecosystem as both engineer species and modifier species (Naiman 1988, Davidson et al. 2012, Cosentino et al. 2014). Elephants in African grasslands are a typical example of species cited as keystone herbivores, which would be better described as modifier species (Naiman 1988). Finally, trees in American forests have sometimes been labeled keystone plants, but would rather correspond to foundation species (Ellison et al. 2005).  3.3.4. A restrictive definition for the keystone species concept Based on our comparative framework, we concluded that a keystone species is a high trophic level species, which has a disproportionate impact on the food web structure of its community. However, the confusion between the different concepts used in the literature to describe species of potential interest highlights the necessity of defining more precise criteria for keystone species. The three exclusive attributes of keystone species are: (1) a low abundance, so that it is likely to be a predator located at a high position in the food web; (2) a high trophic impact on its community, due to one (or few) strong and direct trophic interaction(s) established with one (or few) prey species (predation); and (3) a wide trophic impact, due to many other  57  indirect interactions established with many other species (predation release). The other species are themselves (directly or indirectly) interacting with the keystone species’ prey and are located lower in the food web.  The restrictive definition of keystone species may encompass two different types (Figure 3.4). The first type of keystone species (KS type 1) corresponds to a situation where the keystone predator feeds on a highly competitive prey, which is competing for food or space with many other species. After removal of the keystone predator, the highly competitive prey may become a dominant species in its community. The second type (KS type 2) describes a situation where the keystone predator feeds on a prey that feeds on another prey, which provides shelter or resources to many other species. Thus, if the predator is removed, the first prey will increase in abundance (due to predation release) and the second prey and associated species will consequently decline. KS type 1 may be categorized as ‘level 1’, since only one intermediate prey species is involved, and describes the ecological role of sea stars or sea snails in the rocky intertidal zone. In contrast, KS type 2, involving two intermediate prey species, may be categorized as ‘level 2’, and depicts the role of sea-urchin feeders in kelp forests (e.g., sea otters) or coral reefs (e.g., triggerfishes).  3.4. Discussion 3.4.1. A restored concept of keystone species In this study, we demonstrated that the concept of keystone species was ecologically relevant but commonly misused, so that the technical terminology needed to be refined. We proposed a restored definition of the keystone species concept, based on an in-depth review of the scientific literature and developed in four steps. First, we summarized the usages of the keystone species terminology in the scientific literature. We showed the expansion of its meaning over time with an evolutionary tree. Second, we extricated the keystone species concept of all overlapping concepts designating other categories of species. We proposed a comparative and descriptive framework for a clear discrimination between different species of potential interest, including keystone species. Third, we refined the widely-used concept by differentiating appropriate applications from misusages of the term in the literature. A list of the most  58  commonly cited pseudo-keystone species was established. Lastly, we formulated a clear and exclusive definition for keystone species, represented with a conceptual diagram.  We first described keystone species as a sub-category of ecologically important species, or critical species, defined as species with important ecological functions performed by few other species (Perry 2010). A keystone species is a unique species in terms of functional diversity, and “no other species can take over its role” (Perry 2010). Thus, the uniqueness of the keystone species appears to be one of the major factors of its importance. Yet, importance is a relative concept since species uniqueness depends on the characteristics of the other species in the ecosystem (Hurlbert 1997, Perry 2010). Prey species traits, such as competitiveness or vulnerability to predation, may be critical to the keystoneness potential of their predator (Power and Mills 1995). For instance, we identified sea urchins as dominant prey in six of the eight marine examples of keystone species. Thus, sea urchins might determine the keystone role of the few specialist predators which are able to overcome their spines (Pinnegar et al. 2000). Then, we narrowed down the concept of keystone species by defining its critical role in the food web. We defined a keystone species as a species of low abundance and high trophic level, which exerts a high and wide impact on the food web structure of its community. The restored definition of keystone species we proposed in this study is based on restrictive and specific ecological criteria, and corresponds to the original notion of ‘keystone predator’ (Paine 1966, 1969a, Paine 1969b).   Keystone species have been described as strong interactors in the literature (Paine 1980, Menge et al. 1994, Menge and Freidenburg 2001, Sala and Knowlton 2006). A strong interactor species has a large effect on the species with which it interacts, while a weak interactor species has a small effect on the other species (Menge and Freidenburg 2001). Other species categories were also defined as strong interactors, such as foundation species or mutualist species (Menge and Freidenburg 2001, Soulé et al. 2003). However, being a strong interactor is not a characteristic of species but depends on local conditions in the community (Caro 2010), so that weak interactors may play important ecological roles due to changing conditions (Berlow 1999). Moreover, keystone species are not characterized only by one single strong predator-prey interaction, but rather by a particular configuration of strong and weak interactions in the food web (Berlow et al. 2004).   59  Direct effects are defined as the influence of a species on another species resulting from their direct interaction, while indirect effects correspond to the influence of a species on another species with which it does not interact directly (Menge 1995). Direct effects may be caused by trophic (e.g., predation) or non-trophic (e.g., interference competition) interactions (or ‘linkages’) (Paine 1980, Menge 1995). Menge (1995) described several types of indirect effects corresponding to trophic interactions, including (among others): (i) ‘keystone predation’, when a predator indirectly increases the abundance of competitors of its prey via consumption of the highly competitive prey; and (ii) ‘trophic cascade’ (or ‘tri-trophic interaction’), when an increase in plant abundance is caused by the control of herbivores by a predator. Based on the latter definitions, the KS type 1 proposed in this study would correspond to keystone predation, while the KS type 2 would be better described as a trophic cascade. However, Menge (1995) recognized that trophic cascades and keystone predation could also be seen as “different versions of the same type of indirect interactions”, which may actually be comparable. Also, Pinnegar et al. (2000) defined trophic cascades as “predatory interactions involving three trophic levels”, and keystone predation as a “three-species interaction involving competition”. Thus, depending on whether we define trophic cascades based on the number of species interactions involved (including trophic and non-trophic interactions) or the number of trophic interactions only, keystone species may be, either assimilated to, or differentiated from, trophic cascades.  In our approach, we chose to assimilate the KS type 2 to a keystone species, rather than a trophic cascade. Indeed, the keystoneness of the KS type 2 predator species is not only due to predatory interactions cascading through the food web, but may also be due to the indirect effect of habitat facilitation. Habitat facilitation was defined as the situation where a species  indirectly improves the habitat of another species by altering the abundance of a third species that has a negative effect on the habitat (Menge 1995). Thus, here, the KS type 2 predator feeds on the prey that feeds on the host species on which many other species depend, and thus provides habitat facilitation to these species. Alternatively, the KS type 2 situation could be seen as a trophic cascade that involves the removal of a foundation species.   60  3.4.2. Context-dependency issues with the keystone species concept When dealing with keystone species (or species functional roles in general), ecologists face the issue of context-dependency: a species may or may not be keystone depending on the ecosystem where it is encountered (Power et al. 1996, Arponen 2012). Several authors acknowledged that a-priori identifying keystone species with confidence was a difficult task (Paine 1995, Power and Mills 1995, Power et al. 1996). Some authors demonstrated that species taxonomy could determine keystoneness, due to inherent and intrinsic evolutionary attributes (Stouffer et al. 2012). In contrast, other authors argued that species are keystone under certain conditions only, so that there would be no absolute species-specific property to be keystone (Mills et al. 1993, Paine 1995, Power and Mills 1995). Thus, identifying keystone species would require documenting species abundances, diets and interactions over various environmental conditions (Power et al. 1996). For instance, a comparative analysis of two Eastern Pacific rocky intertidal food webs demonstrated regional changes in the pattern of trophic interactions between the same species (Paine 1980). Menge et al. (1994) concluded that there are no systematic keystone species and that the keystoneness status varies across space. However, the variability in space and time of the species ecological roles is not easy to capture, as illustrated by the two following quotes by R.T. Paine: “All species play some role, albeit generally unknown, at some place or time” (Paine 1995); and “Food webs are idealized pictures of complex trophic patterns that change seasonally and geographically” (Paine 1980).  The keystone status of a species depends on a suite of biotic and abiotic factors determining the distribution and abundance of species in space and time (Harley 2011). Species may or may not be keystone, depending on changes in their own population over time (Piraino et al. 2002). Thus, attempts to discriminate between ‘keystone’ species and ‘dominant’ or ‘rare’ species were claimed to be irrelevant, seeing that keystone species could be more or less abundant over time (Christianou and Ebenman 2005). Environmental variability may also influence the ecological role of a particular species in its ecosystem by modifying species interactions strength (Paine 1966, 1980, Power et al. 1996). Abiotic variables may include physical factors, such as temperature or topology (Paine 1966), but also productivity or disturbance (Power et al. 1996). Changes in temperature may induce direct effects on species distribution and abundance, as well as indirect effects on species interactions (Poloczanska et al.  61  2008). As a result, small changes in climate could generate large modifications in community composition and structure (Harley 2011) and influence species keystoneness (Sanford 1999, Christianou and Ebenman 2005). This is particularly the case in rocky intertidal communities where organisms already live close to their thermal tolerance limits (Harley 2011), and where experiments showed that slight fluctuations in water temperature regulate the impact of a keystone predator on its principal prey (Sanford 1999).  3.4.3.  ‘Keystone species’: more than an ecological concept? In our approach, we defined keystone species exclusively based on ecological criteria, but other types of criteria could potentially be considered. Anthropologists suggested including the cultural value of species as a criterion to select keystone species. They introduced the extended concept of ‘cultural keystone species’, defined as a species essential to the sustainability of socio-ecological systems (Cristancho and Vining 2004, Garibaldi and Turner 2004, Platten and Henfrey 2009). Cultural keystone were defined as species whose removal could dramatically change human (and especially indigenous) communities, which depend on them to maintain their cultures, and for which they may have nutritional, medicinal or spiritual values (Cristancho and Vining 2004, Garibaldi and Turner 2004).  However, the cultural keystone species concept has also been criticized in the literature. Davic (2004) argued that the definition of cultural keystone species was in contradiction with the original meaning of the ecological concept, and that culturally important species should not be associated to keystone species, but rather to flagship species. Nunez and Simberloff (2005) discussed the potential risks and benefits of applying the cultural keystone species concept. On one hand, conservation approaches based on cultural keystone species may be more successful due to the increased participation of local human communities (Nuñez and Simberloff 2005). For instance, cultural keystone species may stimulate co-management between indigenous and government stakeholders for a better integration of traditional knowledge (Butler et al. 2012). On the other hand, culturally important species may correspond to invasive species, so that the application of the cultural keystone species concept could turn out to be counter-productive for biological conservation (Nuñez and Simberloff 2005).  62  Biodiversity conservation is facing the Noah’s Ark problem of “efficiently allocating limited funds to conserve biodiversity” (Perry 2010). As a consequence, trade-offs between benefits and costs should be considered before the implementation of a particular approach to conservation (Wiens et al. 2008). Due to their critical role in the ecosystem, keystone species (and other species of potential interest) could be used to prioritize conservation actions (Arponen 2012). Approaches based on focal or surrogate species may reduce costs by reducing the number of species to monitor and simplifying communication (Wiens et al. 2008). However, these approaches may take years to be beneficial (Caro 2010).Thus, conservation policies based on the concept of keystone species (and other species of potential interest) should be prioritized based on relevant economic criteria, such as cost-effectiveness (Arponen 2012).  Despite the complexity and subjectivity of assigning economic values to species, four classic criteria may be used: intrinsic value, aesthetic value, direct value as resources on the marketplace, and indirect value through maintenance of ecosystem services (De Leo and Levin 1997). The latter four criteria were merged into the notion of total economic value of species, and used for identifying ecologically important species, while ensuring the cost-effectiveness of the resultant species protection measures (Perry 2010).  In theory, species prioritization is science-based and objective, but, in practice, the choice of species to be prioritized for conservation is often subjective and based on real-world constraints (Arponen 2012). Indeed, managers need to include practical criteria when using focal or surrogate species for conservation planning and management (Sergio et al. 2008). For instance, avoiding species generating local conflict, focusing on well-known species, and preferentially selecting species with intrinsic value and easy to monitor, may help for a better efficacy of focal and surrogate species as conservation tools (Sergio et al. 2008). Besides, the main issue with management approaches based on species of potential ecological or managerial interest relates to context-dependency (e.g., Zacharias and Roff 2001, Sergio et al. 2008, Caro 2010). Thus, the temporal and spatial scales on which such approaches could be efficiently applied should be investigated and determined (Favreau et al. 2006). Finally, in response to the arguments against the use of the keystone species concept for management (Zacharias and Roff 2001), we would recommend that the concept should be applied on local scale, based on ecosystem-specific thresholds, and only when appropriate and useful regarding the conservation objectives.   63  Table 3.1. Comparative framework describing 12 categories of species of potential interest, based on a literature review.  Concept Original term Reference(s) Definition Interest Reference(s) Example(s) Reference(s) Focal species Focal species (Lambeck 1997) Species which are selected to preferentially focus on, among all species in the community Species of importance to its ecosystem (Armstrong and Caro 2002) See sub-categories below Critical species Ecologically important species (Perry 2010) Focal species of ecological importance, so that the absence of these species would lead the loss of many others Species that play an important role for the structure, composition or functioning of its ecosystem (Perry 2010) See sub-categories below Keystone species Keystone predator (Paine 1966, 1969a) Consumers of high trophic level whose effects are disproportionately large relative to their abundance Species that maintain the food web structure of the community  (Paine 1966, 1969a) See Table 3.2 Connector species Key-industry species (Elton 1927) Abundant consumers of intermediate trophic level, that act as links between species at lower and higher trophic levels Species that maintain the food web structure of the community (Cury et al. 2000, Menge and Freidenburg 2001) Small pelagic fish (Menge and Freidenburg 2001) Foundation species Dominant species; Foundation species; Autogenic engineers (Clements 1936); (Dayton 1972); (Jones et al. 1994) Abundant, low trophic level, primary producers or habitat-forming species Species that define community structure, create stable environmental conditions, and control ecosystem dynamics (Ellison et al. 2005, Caro 2010) Redwood forests; seaweed beds; coral reefs (Ellison et al. 2005, Caro 2010)  64  Concept Original term Reference(s) Definition Interest Reference(s) Example(s) Reference(s) Engineer species Keystone modifiers; Earth-movers; Allogenic engineers (Mills et al. 1993); (Bond 1994); (Jones et al. 1994) Species causing the transformation of the physical environment from one state to another, through their non-foraging activities  Species that cause important modification to the habitat and thus influence the availability of resources to other species (Jones et al. 1994, 1997) Beavers; woodpeckers (Jones et al. 1994) Modifier species Keystone herbivores; Allogenic engineers (Mills et al. 1993); (Jones et al. 1994) Herbivore species that modify the physical environment by both their feeding strategies and their physical alterations of the habitat  Species that cause important modification to the habitat and thus influence the availability of resources to other species (Naiman 1988) Moose; elephant (Naiman 1988) Mutualist species Mobile links; Mutualists (Gilbert 1980); (Boucher et al. 1982) Species that establish a reciprocally beneficial interaction with another species Species that may support habitat-forming species (Stachowicz 2001, Hay et al. 2004) Corals and zooxanthellae; mangroves and sponges (Stachowicz 2001, Hay et al. 2004) Surrogate species Surrogate species (Caro and O'Doherty 1999) Species that are substitute for something else Species which may not ensure any important ecological function, but are used as tools to achieve another purpose (Caro and O'Doherty 1999, Armstrong and Caro 2002) See sub-categories below  65  Concept Original term Reference(s) Definition Interest Reference(s) Example(s) Reference(s) Indicator species Indicator species (Landres et al. 1988) Species that are used as an index of other species attributes or environmental conditions, otherwise challenging to estimate Species that may be used for selecting areas to be protected or monitoring effects of anthropogenic or natural disturbances (Caro and O'Doherty 1999, Zacharias and Roff 2001) Marine mammals (Bossart 2006) Umbrella species Umbrella species (Noss 1990) Species with a large distribution area, so that protecting their habitat area would also protect many other naturally co-occurring species  Species that may be used to determine the size of areas to be protected  (Caro and O'Doherty 1999, Caro 2003, Roberge and Angelstam 2004) Tiger; jaguar; bears (Barua 2011) Flagship species Flagship species (Heywood 1995) Charismatic and popular species Species with a symbolic or marketing value, that may serve as ambassadors of broader conservation goals (Caro and O'Doherty 1999, Caro et al. 2004, Verissimo et al. 2011) Giant panda; polar bear (Barua 2011)    66  Table 3.2. List of examples of species labeled as ‘keystone predator’ or ‘keystone species’ (Sensu Paine 1966, 1969). * indicates that the term ‘keystone’ was not explicitly mentioned by the authors in the references, although the described predator species corresponded to a ‘keystone species’ (Sensu Paine 1966, 1969).  Keystone species Dominant prey Ecosystem type Study area References Bass (Micropterus spp.) Minnow (Campostoma anomalum) Freshwater (stream) North America (Power et al. 1985) * Lobster (Homarus americanus) Sea urchin (Strongylocentrotus spp.) Marine (kelp forest) Eastern Canada (Mann and Breen 1972) Lobster (Jasus edwarsii) Sea urchin (Evechinus chloroticus) Marine (kelp forest) New Zealand (Babcock et al. 1999, Shears and Babcock 2002) Sea otter (Enhydra lutris) Sea urchin (Strongylocentrotus spp.) Marine (kelp forest) California; Alaska (Estes and Palmisano 1974, Estes et al. 1978) Sea snail (Concholepas concholepas) Mussel (Perumytilus purpuratus) Marine (rocky intertidal) Chile (Duran and Castilla 1989) Sea star (Pisaster ochraceus) Mussel (Mytilus californianus) Marine (rocky intertidal) Washington; Oregon (Navarrete and Menge 1996) Triggerfish (Balistapus undulatus) Sea urchin (Echinometra mathaei) Marine (coral reef) Kenya (McClanahan 1995, McClanahan 2000) Triggerfish (Balistes capriscus) Sea urchin (Diadema antillarum) Marine (rocky reef) Canary Islands (Clemente et al. 2010) Seabream (Diplodus spp.) Sea urchin (Paracentrotus lividus) Marine (rocky reef) Mediterranean (Sala and Zabala 1996, Sala 1997) * Wolf (Canis lupus) Elk (Cervus elaphus) Terrestrial (forest) North America (Ripple and Larsen 2000, Beschta 2003) Wolf (Canis lupus) Moose (Alces alces) Terrestrial (forest) North America (Messier and Crête 1985, McLaren and Peterson 1994) *    67  Table 3.3. List of examples of ‘pseudo-keystone’ species, i.e. species labeled as ‘keystone’ in the literature but whose description would rather correspond to other categories of species of potential interest. * indicates that the term ‘keystone’ was used in the references, but the other species categories to which they are associated in this table was also used or described.  Pseudo-keystone species Labels in the literature Ecosystem type Study area Species category References Beavers Keystone modifier (Mills et al. 1993, Cottee-Jones and Whittaker 2012); Earth-mover (Bond 1994); Habitat modifier (Menge and Freidenburg 2001) Freshwater (wetland) North America Engineer species (Jones et al. 1994) Elephants Keystone herbivore (Mills et al. 1993, Bond 1994); Ecosystem engineers (Soulé et al. 2003, Cottee-Jones and Whittaker 2012) Terrestrial (grassland) Africa Modifier species (Naiman 1988) Kangaroo rats Keystone herbivore (Mills et al. 1993); Keystone predator (Bond 1994); Habitat modifier (Menge and Freidenburg 2001) Terrestrial (desert) North America Engineer species; Modifier species (Cosentino et al. 2014)* Prairie dogs Habitat modifier (Menge and Freidenburg 2001) Terrestrial (grassland) North America Engineer species; Modifier species (Naiman 1988); (Davidson et al. 2012) * Trees Keystone plant/host (Mills et al. 1993); Keystone plant resource (Bond 1994, Cottee-Jones and Whittaker 2012); Keystone plant (Menge and Freidenburg 2001); Keystone mutualist (Cottee-Jones and Whittaker 2012)  Terrestrial (forest) America Foundation species (Ellison et al. 2005)   68  Figure 3.1. Evolutionary tree representing the evolution of the keystone species concept over time, based on a literature review. Since the introduction of the original definition (1), the concept has evolved toward an inclusive (2) and then extended (3) definition, and has been applied for non-ecological considerations (4). All terms are organized on the tree based on the year of publication of their corresponding references.   	  !"#$#%&'() Menge 1995Paine 1966; 1969; 1980Mills et al. 1993Bond 1994Menge and Freidenburg 2001Power et al. 1996Power and Mills 1995Paine 1995Davic 2002Davidar et al. 2001Daily et al. 1993Ortiz et al. 2013Tews et al. 2004Zimmer and Ferrer 2007 Perry 2010Mouquetet al. 2012 Cottee-Jones and Whittaker 2012Modlmeier et al. 2014Brown and Heske 1990*+,-Platten and Henffrey 2009Cristancho and Vining 2004Garibaldi and Turner 2004 69 RELATIONSHIPS  CATEGORIES  Figure 3.2. Comparative framework of the 12 categories of species of potential interest, classified by category and represented based on their potential interests (i.e. relevance as a management tool or importance due to ecological role, at the species, community, or habitat scales).    70  Figure 3.3. Comparative framework of the 12 categories of species of potential interest, distributed on trophic level versus biomass scales. The keystone species category is highlighted in bold.    !	 !!"!# $%&'!##!# 71 BEFORE REMOVAL      AFTER REMOVAL        Figure 3.4. Conceptual diagram representing the restored definition of the keystone species concept. Both types of keystone species (KS type 1 and KS type 2) are represented before and after the removal of the keystone species. The different types of arrow represent different types of species interactions and the size of the boxes represents species abundance.   ()		!*	!!	!*()	    ()		!*	!!  	!*+	!*()		  (*!,	!&$#! 	!!#'"#',!!	!-./0, 72 4. Keystone species: toward an operational concept for marine biodiversity conservation 4.1. Introduction The metaphorical terminology of ‘keystone species’ was introduced in aquatic food web ecology by R.T. Paine (1969a, 1969b). A keystone species was first described as a predator maintaining high species diversity and complex interspecific feeding relationships in the community (Paine 1966). Variations in the keystone species abundance or activity would have greater impacts on biodiversity and trophic structure, compared to other coexisting species with similar or higher abundance in the ecosystem (Paine 1969a, Paine 1969b). Since Paine’s analogy, the concept of keystone species rapidly expanded, as it was applied to an ever-growing number of aquatic and terrestrial species, playing a wide variety of critical roles in the ecosystem (Paine 1995, Power and Mills 1995, Power et al. 1996). Many authors discussed the relevance of the keystone species term turned into an inclusive concept (Hurlbert 1997), which has been either described as a powerful buzzword in conservation (Barua 2011), or criticized as an ambiguous and overused concept (Cottee-Jones and Whittaker 2012). As pointed out by Paine (1995): “an answer serving all masters probably provides few useful solutions”. Thus, in this study, a clear, exclusive, and thus operational, definition of the keystone species was applied, based on the original concept of ‘keystone predator’ (Paine 1966, 1969a). More precisely, we defined a keystone species as a predator species which disproportionately influence the food web structure of its community. In other words, we considered that a keystone species correspond to a predator species with a high and wide impact on its food web, despite a low biomass.  Experts reviewed the methods used to identify keystone species and concluded that the most powerful approach was the combination of comparative studies with experimental methods (Power et al. 1996), as conducted by several authors (e.g., Paine 1966, Menge et al. 1994). However, field experiments have practical limitations. Field-based experimental removals of single species are expensive and time-consuming, restricted in scope (both spatially and temporally), and may be challenging with some types of species (either not easily accessible, or under some restrictive protection statuses) (Power et al. 1996, Bond 2001, Libralato et al. 2006).  73 Thus, experimental approaches often require focusing on a few species only, a-priori assumed to be the potential keystone ones, which may introduces bias in the analysis of keystoneness (Libralato et al. 2006). To overcome these difficulties, recent approaches to identify potential keystone species were based on a variety of ecological modeling methods: ecological network analysis (ENA) (Jordán et al. 1999, Estrada 2007, Jordán et al. 2007, 2009, Torres-Alruiz and Rodríguez 2013); loop analysis (e.g., Ortiz et al. 2013a, Ortiz et al. 2013b); community viability analysis (e.g., Christianou and Ebenman 2005, Ebenman and Jonsson 2005, Berg et al. 2011, Stouffer et al. 2012); or food web models (e.g., Okey et al. 2004, Libralato et al. 2006, Eddy et al. 2014).  Different indices measuring the potential of species to be critical to the community were proposed and reviewed in the literature (Jordán et al. 2008, Jordán 2009, Perry 2010, Cottee-Jones and Whittaker 2012). Initially, the proposed indices were based, either on observations from the field (e.g., interaction strength (Paine 1992) or community importance (Mills et al. 1993, Power et al. 1996)), or on theoretical concepts (e.g., general functional importance (Hurlbert 1997) or functional group dominance (Davic 2003)). A suite of structural indices, derived from binary or weighted networks, was then developed: degree (Jordán et al. 2003, Scotti et al. 2007), centrality indices (Estrada 2007), overlap indices (Jordán 2009), topological importance of species (Jordán et al. 2008), trophic function and trophic field overlap (Jordán et al. 2009). Recently, a topo-dynamical criterion, which estimates the consequences of species removal on community dynamics, was proposed in the ENA literature (Torres-Alruiz and Rodríguez 2013). Besides, qualitative indices, based on loop analysis models, were introduced and compared to existing structural and functional indices (Ortiz et al. 2013a, Ortiz et al. 2013b). Three indices explicitly quantifying species keystoneness were proposed: a structural index (K), based on network analysis (Jordan et al. 1999); a dynamic index (KI), based on time-dynamic simulations (Okey et al. 2004); and a functional index (KS), based on mass-balanced modeling (Libralato et al. 2006).   In this study, we used the Ecopath with Ecosim (EwE) modeling approach to estimate species keystoneness and identify potential keystone species in marine communities. EwE is a well-known and widely-applied modeling approach in aquatic ecology (Coll et al. 2008, Christensen et al. 2011). The EwE software is freely accessible (www.ecopath.org) and  74 published EwE models become more and more accessible, so that the approach has been applied to hundreds of aquatic ecosystems worldwide and several meta-analysis, based on EwE models, have already been published (Colléter et al. 2013b). EwE-based studies help understanding the general structure and functioning of aquatic ecosystems, by summarizing the available knowledge on a given ecosystem, deriving its various system properties, and comparing them to other ecosystems (Walters et al. 1997). The EwE modeling approach was primarily developed to answer ‘what if’ questions about policy that could not be addressed with single-species assessment models (Christensen and Walters 2004). Thus, EwE is more suitable to study aquatic food webs than terrestrial ones and has been primarily applied to marine ecosystems. In this study, we only considered EwE models representing marine ecosystems.  EwE may be described as an ecosystem model since it represents feeding relationships between all species occurring in the ecosystem, although it only captures trophic interactions, not nutrient cycling. EwE is based on ECOPATH, a food web model first proposed and applied to estimate the main predator-prey pathways in the marine ecosystem of Northwestern Hawaiian Islands (Polovina 1984). The original ECOPATH model was then developed further (Christensen and Pauly 1992), and turned into the EwE modeling complex, which can be seen as a tool-box offering a large collection of methods to analyze various ecological phenomena. The EwE modeling complex consists of a suite of three main sub-models: (i) Ecopath, static and descriptive; (ii) Ecosim, dynamic and predictive; (iii) Ecospace, spatially explicit, dynamic and predictive (Christensen and Walters 2004). Ecopath is a descriptive model, representing a static snapshot of the ecosystem trophic structure, and used as a parameter estimation methodology in the EwE modeling complex (Walters et al. 1997). Ecosim uses the outputs of Ecopath to produce time-dynamic simulations of changing trophic interactions with changes in the ecosystem due to anthropogenic or environmental disturbances (Walters et al. 1997). Ecosim may be used to fit model predictions to time-series data, and to explore different fisheries management or climate change scenarios (Christensen and Walters 2004). Ecospace replicates outputs of Ecosim over a spatial grid to explicitly account for the spatial aspects of trophic structure, and was primarily designed for exploring the possible consequences of alternative marine protected areas policies (Walters et al. 1999).    75 In this study, we focused on three main questions related to the identification of keystone species in marine ecosystems: (1) What are the limitations of the existing functional indices measuring species keystoneness? (2) Could these limitations explain the inconsistencies and discrepancies observed when applying these indices to modeled food webs? (3) How could we formulate a new functional index overcoming these limitations? We intended to answer these questions based on a meta-analysis of published Ecopath models representing marine food webs. First, models were selected with a scoring method. Second, several indices of keystoneness were formulated, comprising new and existing ones. Third, the indices were applied to the models, and the obtained keystone species were recorded. A preliminary comparison of the results was made to identify potential biases in the indices. Then, two statistical methods were used to select the most suitable index: Spearman rank correlation tests and a classification tree. The new functional index of keystoneness was selected according to the indicated definition of keystone species and applied to a well-known case study.  4.2. Methods 4.2.1. Derivation of indices of keystoneness from Ecopath food web models 4.2.1.1  Selection of the Ecopath models with a scoring method  The data used in this study was extracted from 101 Ecopath models. First, the EcoBase database was used to establish the list of models to select. EcoBase is an online repository, referencing and integrating information from more than 400 published EwE models (Colléter et al. 2013b). We extracted critical metadata on the models stored in EcoBase, and reused some of the metadata as selection criteria in our analysis. Then, the data extracted from the selected models were accessed through another collection of EwE models (Christensen et al. 2011). We collected from EcoBase a suite of metadata including the ecosystem type, the climatic zone, the first and last year, the spatial extent and the reference of the models. We excluded from our analysis all referenced models representing non-marine ecosystem types. Among the marine types, models were selected by applying a scoring method, based on six  76 criteria: (1) an analysis of keystone species (KS) was previously performed with this model; (2) the model represented an established Marine Protected Area (MPA); (3) another (or more) historical version(s) (called ‘sibling’) of the same modeled ecosystem was available; (4) the model was focusing on a species (or group of species) of particular interest; (5) detailed information on species aggregation into functional groups was available for this model; (6) another (or more) geographical version(s) (called ‘Overlapping’) of the same modeled ecosystem was available. We note that criteria 1, 2 and 3 were particularly relevant to our study, whereas criteria 4, 5 and 6 were less significant. Indeed Outcomes from past analyses of keystone species on the same ecosystems could be confronted to our results. Models representing MPAs, or successive historical versions of the same area, could be used to understand fishing impact on marine food webs and keystone species. In contrast, species of particular interest might or might not be keystone species, and information on species aggregation might or might not help identifying keystone species. Overlapping models were actually very few in the EcoBase database. All criteria correspond to metadata directly extracted from EcoBase, apart from criteria 3 and 6, which were refined for the purpose of this study. Sibling models (criteria 3) were more precisely defined here as models with the same geographic location and surface area, but representing different years or time periods. Overlapping models (criteria 6) were defined as models with the same geographic location and time period, but not necessarily the same surface area (i.e. representing the ecosystem at smaller or larger spatial scales). Each criterion was given a score of 1 if true, and 0 if false. For each model, the final score was obtained by summing the scores of the six criteria, with a coefficient of 2 applied to the first three criteria, since they were more relevant to this study. The final scores ranged from 0 to 6. Models were selected only if their final score was equal to or higher than 3. Nonetheless, models with an overall score lower than 3, but a positive score for one of the first three criteria (of higher importance), were also considered. 4.2.1.2  Description of the Ecopath modeling approach  Ecopath describes the feeding relationships between all species occurring in the modeled food web (see Appendix B (Figure B1) for an example). Species are aggregated into functional  77 (or trophic) groups, corresponding to an age/size class of a single species, a single species, or a group of species sharing similar ecological traits (such as size, diet, predator and life cycle) (Christensen et al. 2008). A functional group can also be defined as a group of species with similar effects on ecosystem processes, without redundancy among the other groups (Perry 2010). Thus, a keystone species may be represented by a functional group (De Leo and Levin 1997, Perry 2010), or by the most abundant species within its functional group (Davic 2003). As a result, the terminologies ‘keystone groups’ and ‘keystone species’ are used likewise in this study.  The main assumption in Ecopath is of mass balance over a given time period (typically one year) (Christensen and Walters 2004). For each functional group, Ecopath assumes that the energy input and output are balanced in the ecosystem, and that the system has the same biomass state at the end of the period as it had at the beginning (Walters et al. 1997). The mass-balance constraint serves as a filter for mutually incompatible estimates in the model. All available information collected about the components of the ecosystem pass through the ‘mass balance filter’, which determines what parameters must be to support the current trophic structure and be consistent with observations (Walters et al. 1997, Christensen and Walters 2004).  For each selected Ecopath model, we gathered, from the collection of EwE models (Christensen et al. 2011), the name and biomass (, t·km-2) of each living group (dead groups, such as detritus, were excluded), and deduced the total number of living groups in the model and their total biomass in the ecosystem (>?>, t·km-2). Biomass is usually entered as input data in Ecopath, but may be estimated by the model if missing. As a first data quality check, we recorded the number of biomass estimates computed by Ecopath for each model (@A>!B), and deduced the proportion of estimated biomass. Although we did not use Ecosim in our analysis, we recorded for each selected model, if the Ecopath model was fitted to time series in Ecosim or not (metadata extracted from EcoBase), and used it as a second data quality check. A supplementary parameter, the trophic level (, dimensionless), was collected for each living group in each selected model.  is a fractional number giving the position of each functional group in its food web (generally ranging from 1 to 5), and estimated by Ecopath based on the diet composition of the group and the  of its prey items (starting with a  of 1 assigned to producers and detritus) (Christensen and Pauly 1992). Some of the selected models described ecosystems exploited by one or more fishing fleets. For these models, we collected the total  78 number of fishing fleets in the model and the total catch rate (>?>, t·km-2·year-1) of all fishing fleets in the ecosystem. The mean trophic level of the catch (C, dimensionless), estimated by Ecopath based on the composition of the total catch and the  of the targeted groups (Christensen et al. 2008), was collected as well.  4.2.1.3  Description of the Mixed Trophic Impact (MTI) analysis and derived indices of keystoneness (KS) Several modules were developed and added to the EwE package, such as the ‘Network Analysis’ module, a tool available with the Ecopath model (Christensen et al. 2008). The Network Analysis notably comprises the ‘Mixed Trophic Impact’ (see Appendix B (Figure B2) for an example) and ‘Keystoneness’ (KS) analyses. The Mixed Trophic Impact (MTI) is an economic input-output analysis adapted to ecological networks by Ulanowicz and Puccia (1990), and implemented in Ecopath (Christensen and Walters 2004). The MTI is calculated by constructing an  $  matrix, whose elements *!3. represent, for each pair of groups *D 0. in the modeled food web, the relative impact of the impacting group  on the impacted group 0, by considering direct (predation) and indirect (competition) feeding interactions (Ulanowicz and Puccia 1990, Christensen and Walters 2004).  The elements *!3. of the MTI matrix represent the relative mixed trophic impact of the impacting group  on the impacted group 0, for each pair of groups *D 0. in the modeled food web. The EFG are derived from the net impact *H!3. of group  on group 0, for all the possible pathways that link both groups in the food web (Ulanowicz and Puccia 1990, Libralato et al. 2006). The H!3 are calculated by the difference between positive effects, quantified by a diet composition term *43!., and negative effects, expressed by a host composition term *&4!3. : H!3 	43! < &4!3         Equation 4.1 Where 43! is the proportion of group  in the diet of group 0, and &4!3 is the proportion of group 0 in the consumption by group  (Ulanowicz and Puccia 1990, Christensen and Walters 2004). For dead groups, 43! is set to zero, while for fishing fleets, 43! represents the proportion of group  in the catch of fleet 0 (Christensen et al. 2008). The !3 are estimated for each fishing  79 fleet defined in the model, by considering fishing fleets as predators (Christensen and Walters 2004).  Two alternative indices measuring the potential for being a keystone species, or ‘keystoneness’ (KS), were implemented in the Keystoneness analysis of the Network Analysis module in EwE (Christensen et al. 2008). The first index (called KS#1 in EwE) was proposed by Libralato et al. (2006), and the second one (called KS#2 in EwE) was adapted from a methodology proposed by Power et al. (1996). Both indices are estimated based on the same parameters: a measure of trophic impact *!. derived from the MTI analysis, and a measure of biomass *I!.. ! represents the overall effect of group  on all the other groups in the food web (without including the effect of the group on itself):  !  J∑ !3%3K!          Equation 4.2 Where ! is expressed as the sum of the squared values of !3	of group , paired with each of the other living group 0 in the food web. The mixed trophic impact of group  on itself (!!) are excluded, as well as the mixed trophic impact on dead groups such as detritus (Libralato et al. 2006).  I! corresponds to the contribution of group  to the total biomass in the food web: !  /)∑ /L9LMN           Equation 4.3 Where ! is the biomass of group	, and O the biomass of each of the  living groups in the food web (Power et al. 1996, Libralato et al. 2006).  The KS#1 and KS#2 indices are obtained by combining ! and I! for each group  (Power et al. 1996, Libralato et al. 2006), such as: P!  QRS	T! $ * < !.U        Equation 4.4 P!  QRST! $ * !⁄ .U        Equation 4.5 4.2.1.4  Formulation and application of several KS indices In both KS#1 and KS#2 indices, the keystoneness is estimated based on two components: (i) a component estimating the trophic impact of the group (here called Impact Component (IC));  80 and (ii) a component estimating the biomass of the group (here called Biomass Component (BC)), so that a general equation for an index of keystoneness (KS) may be expressed:  !  QRS	TV4! $ 4!U         Equation 4.6 Based on the general equation of KS (Equation 4.5), we explored several formulations and combinations of IC and BC. We retained three different ICs and four different BCs, and combined the seven components in all possible ways to obtain 12 alternative KS indices (Table 4.1). ICs were calculated using a measure of trophic impact *!. (Equation 4.2), and BCs using a measure of biomass *!. (Equation 4.3). To avoid imbalance between both components, BCs and ICs were defined so that they had value ranges of comparable order of magnitude. One of the ICs and two of the BCs were directly adapted from the literature (Power et al. 1996, Libralato et al. 2006) and called using a subscript capital letters (V4W, 4W, 4(), while the other components were called using subscript numbers, with a zero for the components using ranks (V4# and 4#). Note that two of the 12 KS indices (+ and %) were directly adapted from the literature and correspond to the KS#1 and KS#2 indices described above (Equation 4.4 and Equation 4.5).  The 12 retained KS indices were applied to the 101 selected modeled ecosystems. The names of the groups with the highest value of keystoneness were recorded for each KS index and each model. The !3	values were extracted from Ecopath for each living functional group (dead groups such as detritus were not considered) of each selected model, and ! was calculated as described above (Equation 4.2). I! was derived from the ! values extracted from Ecopath, as described above (Equation 4.3). Contrary to previous approaches (e.g., Libralato et al. 2006), the !3	corresponding to fishing fleets were included in our analysis. We assumed that fishing fleets could be added as a supplementary functional group, if considered as a predator feeding on targeted groups. Therefore, a supplementary group, called ‘All Fleets’, was added as a    functional group, in all selected models representing exploited ecosystems. All defined fishing fleets were combined together into the ‘All Fleets’ group, whose ! and ! were approximated with the >?> and C, respectively. For models with multiple fleets, !3 of All Fleets was obtained by summing up the !3 values over all the defined fishing fleets.   81 4.2.2. Analysis of the contributions of the IC and BC components to the keystoneness estimation 4.2.2.1  Comparison of groups ordered by keystoneness, trophic impact and biomass  To test the 12 KS indices defined above (Table 4.1), we analyzed the balance between IC and BC components for each index. First, we compared, for each KS index and each selected model, the groups ordered by keystoneness, with the groups ordered by trophic impact and biomass, separately. Precisely, we compared the group of the highest keystoneness with: (i) the three groups with the highest values of trophic impact (!); and (ii) the three groups with the lowest values of biomass (!). If the group with the highest keystoneness value also had one of the three highest trophic impact values, then the model was categorized as ‘Match Impact’ for the particular index. In contrast, if the group with the highest keystoneness value also had one of the three lowest biomass values, then the model was categorized as ‘Match Biomass’ for the particular index. If the group with the highest keystoneness did not correspond to any of the three groups, neither with the highest trophic impact, nor with the lowest biomass, then the model was categorized as ‘No Match’ for the particular index. The methodology was applied to each of the 101 selected models, and the number of models in each matching category (Match Biomass, Match Impact or No Match) was recorded, for each KS index. The ‘Overall Match’, obtained for each KS index over all models, corresponded to the matching category of the majority of models (i.e. 51 models or more, for a total of 101 models).  4.2.2.2  Rank correlation tests between KS indices and trophic impact or biomass estimates To analyze further the balance between IC and BC components for each KS index, we applied the Spearman’s rank correlation test (Coleman 2010) using the R software (www.r-project.org/). We tested if there was a monotonic correlation between the rank (in ascending order, because keystone species should have high keystoneness) of the KS index of group , called X*!., and: (1) the rank (in ascending order, because keystone species should have  82 high trophic impact) of the trophic impact of group , called X*!%.Y and (2) the rank (in descending order, since keystone species should have low biomass) of the biomass of group , called X*!.. The Spearman’s rank correlation test was applied twice (once with the trophic impact, and once with the biomass) to each KS index. The rank correlation tests were validated for each model, if and only if, for the particular KS index: (i) the hypothesis Z (i.e. there is a monotonic correlation) of the Spearman’s rank correlation test was validated (i.e. p-value < 0.05) for both tests 1 and 2; (ii) the coefficient of correlation  was positive for both tests 1 and 2; and (iii) the correlation coefficients + and % of tests 1 and 2 were of the same order of magnitude (a constraint was applied to the value of %	 so that %  + [ \]). The methodology was applied to each of the 101 selected models, and we recorded, for each KS index, the number of models with validated rank correlation tests (i.e. satisfying the three conditions specified above). We also ran, for comparison, the same tests without applying any constraint on the order of magnitude of the correlation coefficients. 4.2.3. Selection of the most promising KS index 4.2.3.1  Implementation of a classification tree We implemented a classification tree (Figure 4.1a), as an alternative to the KS indices for the identification of keystone species among the functional groups of each modeled ecosystem. The classification was based on two (log-transformed) parameters: the trophic impact (in squared values) (!%) and biomass (!) of each group . Four main categories were discriminated: ‘Keystone’ (corresponding to groups with high impact and low biomass), ‘Low Impact – Low Biomass’, ‘Low Impact – High Biomass’, ‘High Impact – High Biomass’, and ‘Intermediate’ (corresponding to groups belonging to none of the previous categories). Quartiles values of both !% and ! parameters, over all living groups in each model, were set as thresholds to identify groups belonging to each category. The third quartiles (1^) were set as lower thresholds for delimiting groups with high !% and !, while the first quartiles (1+) as upper thresholds for groups with low !% and !. The other thresholds were defined by the minimum () or maximum (E_`) values of !% and aF recorded for each modeled ecosystem. Then, among the  83 identified keystone species, we separated the ones with a high trophic level from the others. Thus, a supplementary parameter, the trophic level (!) of each group, was used to discriminate the Keystone groups in two sub-categories: ‘Low TL Keystone’ (whose  was lower than the third quartile values of ), and ‘High TL Keystone’ (whose  was equal to or higher than 1^bW). The classification tree was applied to each of the 101 selected models, and a scatter plot representing the classification of each group was produced for each model (Figure 4.1b). The categories identified with the classification tree were recorded for all groups and all models.  4.2.3.2  Comparison between keystone groups identified with the KS indices and with the classification tree  The keystone groups identified with the classification tree were compared to the ones identified with the KS indices. Precisely, for each of the 101 selected models and each of the KS index, we compared the group with the highest value of keystoneness with the KS index to the groups categorized as (Low TL or High TL) Keystone with the classification tree. If the group with the highest KS index value was classified in the (Low TL or High TL) Keystone categories of the classification tree, then the identified keystone group was considered as validated, for the selected model. The methodology was applied to each of the 101 selected models, and we recorded, for each KS index, the number of models with validated keystone group. In addition, once the most promising KS index was selected (see Section 4.2.3.3), the groups identified as keystone (i.e. with the highest value of keystoneness) with the selected index were compared across the 101 selected models. Then, we compared the most frequently identified groups with the selected index and with the classification tree. Since the 101 models followed different rules for naming their respective functional groups, the names of the keystone groups identified with the selected index were standardized before comparison. Groups were renamed (and sometimes aggregated) using generic group names, based on their original designations. The same standardization method was applied to the groups classified as Low TL or High TL Keystone with the classification tree.   84 4.2.3.3  Allocation of overall statuses to the KS indices The conclusions from the rank correlation tests and the classification tree were summarized and confronted by allocating different statuses to each pair of ‘Ecopath model – KS index’ (Figure 4.2). The results obtained from the status allocation were used for the selection of the most promising of the 12 KS indices applied in the analysis. Statuses were allocated according pre-defined rules and following three steps: (1) if the KS index was validated (correlated positively with IC and BC, and with correlation coefficients of the same order of magnitude), then the result from the rank correlation tests was positive, negative if not; (2) if the identified keystone group was validated (group of highest keystoneness with the KS index classified in the Keystone categories of the classification tree), then the result from the classification tree was positive, negative if not; (3) if results from both steps 1 and 2 were positive (respectively negative), then the status was labeled as ‘True Positive’, (respectively ‘True Negative’), but if step 1 was positive and step 2 negative (respectively step 1 negative and step 2 positive), then the status was labeled as ‘False Positive’ (respectively ‘False Negative’). Thus, there were four alternative statuses possible for each ‘Ecopath model – KS index’ pair: True Positive, False Positive, False Negative and True Negative. In other words, when the result from the rank correlation tests (step 1) was positive, the status was labeled as ‘positive’, ‘negative’ if not, and when the result from the classification tree (step 2) was in agreement with step 1, the status was also labeled as ‘True’, ‘False’ if not. For some models, none of the functional groups was classified in the Keystone categories of the classification tree, so that the identified keystone group could not be confirmed or unfirmed (step 2). Thus, a supplementary status was defined for these models, and labeled as ‘No Keystone’. The status allocation was performed for each of the 101 selected models, and the number of models associated with each status (True Positive, False Positive, False Negative, True Negative and No Keystone) was recorded, for each KS index. The ‘Overall Status’, obtained for each KS index over all models, corresponded to the status allocated to the majority of models.   85 4.2.4. Application to a case study A case study was chosen to which both the classification tree described above and the selected KS index were applied. We chose the Ecopath model representing the Prince William Sound, a nearly enclosed embayment located in the northern Gulf of Alaska, for the time period between 1994 and 1996, and covering an area of about 9,000 km2, with a mean depth of 300 m (Okey and Pauly 1999). This Ecopath model was originally built to better understand the structure and functioning of the food web, based on the information provided by a panel of 35 experts (Okey and Pauly 1999). See Appendix B for graphical representations of the Prince William Sound (1994-1996) food web. 4.3. Results 4.3.1. Pool of selected Ecopath models After applying the scoring method, we obtained a selection of 101 Ecopath models, whose final scores ranged from 2 to 6. Analyses of keystone species were previously performed on 23 of the 101 models. In fact, most models were included in the meta-analysis introducing the KS#2 index (Libralato et al. 2006), or were applied the latter index by their authors, while four models were applied alternative original methods proposed by the modelers (Aydin et al. 2002, Kitchell et al. 2002, Okey et al. 2004). The selection comprises 13 models representing MPAs, and only 5 of them are fully unexploited ecosystems (no fishing fleet defined). However, the No Keystone status was allocated to the majority of the MPA models. About a third (39) of the selected models focuses on a (group of) species of particular interest. The selection includes 81 sibling models, but only 2 overlapping models. In fact, only 48 of the sibling models are suitable for comparative analyses, since some models did not have their corresponding siblings selected, others had different species aggregations between siblings, and some siblings were associated to the No Keystone status. See Appendix C (Table C1) for the detailed results of the scoring method for each selected model.   86 Most of the 101 selected models seem to be good quality models in terms of data. A large proportion of the models (78) provide detailed information on species aggregation into functional groups. Approximately one third of the selected models do not have any biomass input data computed by Ecopath, and the proportions of estimated biomass is lower than 25% for the second third, and lower than 75% for the last third (up to 95% for only one model). Besides, the selection comprises 43 Ecopath model fitted to time-series. See Appendix C (Tables C2 and C3) for the detailed metadata and references of each of the 101 selected models. The 101 selected models represent a wide variety of marine ecosystems. Continental shelf (less than 200 m deep) is the most represented type of ecosystem with 41 models. Six other types are represented in smaller proportions: ocean (more than 200 m deep, 19 models), upwelling (16), channel or strait (10), bay or fjord (7), coral reef (7), and coastal lagoon (1). Most modeled ecosystems are located in temperate or tropical zones (58 and 36 models, respectively), and few in polar zones (7). About two thirds of the selected models represent ecosystems over a time period of one year, and the last third covers time periods lasting from 2 to 11 years (up to 40 years for one model). About half of the models describe ecosystems in the recent past (49 models for the 1990s-2000s), while the other half describe ecosystems in more ancient times (26 for the 1970s-1980s, 18 for the 1950s-1960s, and 8 for periods before 1950). The spatial extent of the selected models varies from 2.48 to more than 18,000,000 km2, but most models cover an area smaller than 500,000 km2 (Figure 4.3a). Among the selected models, ecosystems of the ocean and continental shelf types have the wider ranges of spatial extent, whereas the ocean and upwelling types cover the largest areas, followed by continental shelf and channel or strait (Figure 4.3a). Coral reef ecosystems have the wider range and highest values of total biomass (Figure 4.3b), whereas upwelling ecosystems have the wider range of total catch, followed by bay or fjord and channel or strait (Figure 4.3c). The number of living groups in the 101 selected models ranges from 6 to 96, with most models comprising 20 to 40 groups (Figure 4.4a). The number of fishing fleets varies from 0 to 19, with 14 models without fleet, and most models including 1 to 5 fleets (Figure 4.4b). Among models with defined fishing fleets, the mean trophic level of the catch ranges from 1.02 to 4.52, and is comprised between 3 and 3.5 for most models (Figure 4.4c).   87 4.3.2. Balance in the contributions of IC and BC to the KS indices The comparison between the groups ordered by keystoneness and by trophic impact and biomass, applied to the 101 Ecopath models and for each KS index, shows that three of the KS indices out of the 12 seem to be balanced: ^, c, and +% (Table 4.2). For these three indices, the spreading of the models is comparable between the three matching categories, so that none of the categories is associated to the majority of models (51 or more) and no Overall Match is determined. Thus, for these three indices, the group with the highest keystoneness value may also have one of the three highest trophic impact values, or one of the three lowest biomass values, or none of the previous, with similar probabilities. Nonetheless, we note that the number of models is higher (and close to the majority) in the Match Impact category for the ^index. In addition, the comparison shows that both + and % indices (directly adapted from the literature) are unbalanced. + has a significantly higher number of models (91) in the Match Impact category, meaning that this index tends to identify keystone groups mainly based on their (high) trophic impact. In contrast, % has a significantly higher number (81) of models in the Match Biomass category, so that it would identify keystone groups primarily based on their (low) biomass. The Overall Match of the other KS indices corresponds to, either Match Impact, or Match Biomass, determined with a more or less substantial majority of models.  The rank correlation tests are validated for a majority of the 101 models for only one of the 12 KS indices: ^, with 75 models with validated tests (Table 4.3). Both + and % indices show significantly lower numbers of models with validated rank correlation tests, with respectively 40 and 12 models. The other KS indices have less than 10 models with validated tests (apart from dwith 44 models). Thus, the results from the tests demonstrate a better balance between IC and BC components for the ^ index, for which high keystoneness values seem to correlate, with the same order of magnitude, to high trophic impact and low biomass values. When the constraint on the magnitude order of the correlation coefficients is released, the conclusions from the tests are not changed: ^ appears as the most balanced index. The number of models with validated (unconstrained) tests increases for most of the KS indices (apart from e and f, for which it remains equal to zero), yet the ^ index still has the highest number, with 99 models. When unconstrained, the rank correlation tests are validated for a majority of models with four other indices, including + and %, with respectively 95 and 52 models.   88 4.3.3. Selection of the most promising KS index Based on the classification tree (Figure 4.1), we could confirm or not the keystone groups identified with each of the 12 KS indices, for each of the 101 models (Table 4.3). The number of models with a validated keystone group (i.e. belonging to the Keystone categories of the classification tree) ranges from 15 to 59, over all KS indices. The ^ index shows one of the highest scores (with 48 models), whereas the + and % indices show lower scores (with 15 and 34 models, respectively).  Then, by confronting the results from the rank correlation tests and the classification tree, we could allocate alternative statuses to each of the 101 models, for each of the 12 KS indices (Figure 4.2). Since 33 models were allocated a No Keystone status (no functional groups in these models was categorized as Keystone with the classification tree), the majority was reduced to 35 models to determine the Overall Status of each KS index. The outcomes from the status allocation corroborate the selection of the ^ index as the most promising one. Indeed, ^ is the only one index associated with a True Positive Overall Status (with 40 True Positive models), meaning that results from the tests (step 1) and the classification tree (step 2) are both positive for a majority of models (Table 4.3). Two indices, g and f, are allocated a True Negative Overall Status (with 46 and 53 True Negative models, respectively), traducing negative conclusions from both steps 1 and 2. For three of the KS indices (+, % and d), none of the statuses is associated with the majority of models, so that no Overall Status is determined for these indices. Nonetheless, we note that the number of True Negative, respectively False Negative, models is close to the majority for +, respectively % (with 32 models in each case). The remaining KS indices are allocated a False Negative Overall Status, indicating a negative result from the tests while the result from the classification tree is positive.  4.3.4. Identification of potential keystone groups with the selected KS index and the classification tree The selected ^ index was applied to the 101 selected Ecopath models, so that the results from the whole analysis may be analyzed in terms of species (or groups). We compared, across the 101 selected models, the most frequently identified keystone groups with the ^  89 index (i.e. with the highest value of keystoneness) and with the classification tree (i.e. belonging to the High TL or Low TL Keystone categories).  Groups are labeled using 12 generic group names, to standardize their heterogeneous original designations (Figure 4.5a and Figure 4.5b). Fish groups are distinguished into 4 groups: ‘Elasmobranchii’ (including groups of sharks and rays), ‘Gadiformes’ (cod and hake groups), ‘Scombroidei’ (tunas, billfishes, and barracudas groups), and ‘Fishes’ (all the other groups of fishes). We note that the latter group of fish corresponds to a relatively wide range of species, aggregating groups of both large pelagic and small demersal, because they all have low occurrences as identified keystone groups. Mammals groups are discriminated between: ‘Odontoceti’ (comprising groups of dolphins and orcas, as well as unspecified toothed whales), ‘Pinnipedia’ (seals, sea lions, and walruses groups), and ‘Mammals’ (baleen whales groups, and unspecified groups of cetaceans and mammals). The other groups differentiate ‘Seabirds’ (all groups of birds), from ‘Invertebrates’ (comprising crustaceans, cephalopods and unspecified invertebrates). The ‘Zooplankton’, ‘Producers’, and ‘All Fleets’ groups are unchanged from their original designations.  Overall, the groups identified as keystone species with the highest occurrences across models correspond to cartilaginous fishes and toothed whales (Figure 4.5a and Figure 4.5b).The most frequently identified groups with the ^ index, across the 101 models, are: Elasmobranchii, All Fleets, Fishes, and Odontoceti (Figure 4.6a). In comparison, the most frequently identified groups with the classification tree, across the 101 models, are: Elasmobranchii, Pinnipedia, Odontoceti, and Fishes (Figure 4.6b). With the classification tree, the identified keystone groups may be discriminated with their trophic level (Figure 4.6b). Thus, Odontoceti scores higher than Pinnipedia when considering High TL Keystone only. Besides, All Fleets are Low TL Keystone, as well as a substantial proportion of the Seabirds and Fishes. The keystone groups identified with the ^ index may be discriminated based on the status of the models (Figure 4.6a). Thus, the most frequently identified groups associated to a True Positive status are: Elasmobranchii and Odontoceti. In contrast, the most frequently identified groups associated to a No Keystone status are All Fleets and Fishes. Finally, the High TL Keystone groups, such as Scombroidei or Gadiformes, correspond to a True Positive status. See Appendix C (Table C4) for the detailed results obtained with the selected ^ index, for each model.   90 4.3.5. Case study of the Prince William Sound (1994-1996) food web With the selected ^ index, the Ecopath model of the Prince William Sound (1994-1996) show positive results from the rank correlation tests. Also, the same groups are identified as potential keystone species, both with the classification tree (Figure 4.6) and the selected ^ index (Figure 4.7a), in the Prince William Sound (1994-1996) food web. Transient orcas (group number 1) are classified as High TL Keystone and have the highest value of keystoneness with the ^ index. Avian predators (group number 15, mainly representing bald eagle) are categorized as Low TL Keystone and have the second highest keystoneness value. Thus, the status of the selected ^ index applied to the case study model is True Positive. In contrast, with the + and % indices directly adapted from the literature, results from the classification tree are in disagreement with results from the rank correlation tests. More precisely, the groups having the highest values of keystoneness with the + index belong to the High Impact – High Biomass category (Figure 4.7b). With the % index, although transient orcas and avian predators have the two highest keystoneness values, the correlation tests were not validated, seeing that the groups classified as Low Impact – Low Biomass have high keystoneness (Figure 4.7c). 4.4. Discussion 4.4.1. A balanced index of keystoneness for marine food webs In this study, we derived a functional index estimating species keystoneness from a meta-analysis on a selection of 101 Ecopath models. The selection of models was representative of the variety of marine ecosystems worldwide. Also, most models were of good quality in terms of data. The Ecopath-based approach enabled us to use trophic impact and biomass as measurable species traits, and to propose ecosystem-specific thresholds of minimum trophic impact and maximum biomass for species to be keystone. In each modeled food web, species were ranked according to their keystoneness estimates, so that the potential keystone species could be quantitatively identified. The new index of keystoneness proposed in our study (the ^ index )  91 was obtained by multiplying the V4W impact component (adapted from Libralato et al. 2006) by the 4# biomass component (based on descending ranking, so that high biomass corresponds to low BC). The usage of ranking in the biomass component enabled us to prevent “inflating uncertainty and errors in the index because measures for rare species are divided by small numbers” (Cottee-Jones and Whittaker 2012). The ^ index was selected by confronting the results from two statistical methods: (i) Spearman’s rank correlation tests (step 1); and (ii) a classification tree (step 2). The selected index corresponded to the KS index with the highest number of models for which results from both steps 1 and 2 were positive. The standardized Ecopath approach allowed for a comparison of the keystone groups identified across models. The generic groups of species with the highest occurrences across models were cartilaginous fishes and toothed whales, over all the 101 selected models. These species corresponded to the definition of keystone predator applied in this study. As for the case study model, the identification of a toothed whales species (transient orcas) as a potential keystone group was consistent with empirical knowledge on the modeled ecosystem. Although sea otters are known as an historical example of keystone species in kelp forest ecosystems (Estes and Palmisano 1974), transient orcas were identified as keystone species in the Ecopath model of the Prince William Sound (1994-1996). Alaskan transient orcas were observed to be partially feeding on sea otters, themselves mainly feeding on clams but rarely on sea urchins (Estes et al. 1998, Okey and Pauly 1999). Thus, the keystoneness of sea otters was reduced in the Prince William Sound food web, due to both the predation pressure from transient orcas (Estes et al. 1998), and the low abundance of grazers such as sea urchins (Paine 1980).  In this study, we established that the selected index (^) seemed to be more balanced than the ones previously proposed in the literature and implemented in EwE (+ and %). Both + and % indices were applied to several modeled food webs, representing various types of marine ecosystems (e.g., Libralato et al. 2006, Coll and Libralato 2012, Coll et al. 2013). As for the six models selected in this study to which the latter indices were applied, the results from the + index were not convincing for four of the models (Coll et al. 2007, Tsagarakis et al. 2010, Valls et al. 2012, Tecchio et al. 2013), and inconsistent results were obtained when both + and %indices were used. In this study, we confirmed that both + and % indices fail attributing high keystoneness to groups having both low biomass and high trophic impact. The + index frequently attributed high keystoneness to functional groups with high biomass, provided their  92 trophic impact was high. Such groups would rather correspond to abundant species (Coll and Libralato 2012). In contrast, the % index often gave high values for functional groups with low biomass, even though these groups had low trophic impact. Such groups would rather be categorized as rare species (Libralato et al. 2006, Coll et al. 2013). The ^ index proposed in this study was demonstrated to have more balanced contributions between its trophic impact and biomass components in its estimation of species keystoneness. Therefore, the new index may be used to identify keystone species in marine food webs, without over-representing abundant or rare species.  4.4.2. Methodological considerations on the derived functional KS index The MTI analysis, like the Ecopath model, only gives a snapshot of the trophic web at one point in time, and should not be used for drawing conclusions on the consequences of changes in biomass over time. For instance, a decrease in the biomass of a predator might lead to an increase in the biomass of its prey in the short term. In the longer term, it might also result in higher biomasses of other predators, or changes in the predator’s diet facing reduced prey abundance, but these cumulative effects cannot be predicted from the MTI analysis (Christensen and Walters 2004). However, the (positive or negative) !3 corresponds, for each pair of groups *D 0. in the modeled food web, to the relative change (increase or decrease) in the biomass of the impacted group 	0, caused by a slight increase in biomass of the impacting groups  (Libralato et al. 2006).  The main criticism about Ecopath-based approaches, especially when looking at keystone species, deals with the species aggregation into functional groups. Aggregating species is required to obtain smaller and simpler food webs, easier to analyze. A few methods of aggregation are commonly used, yet there are no systematic aggregation rules, which potentially introduce taxonomic bias (Jordán 2009). Other than mass-balance food web models, alternative methods to identify ecologically important species, such as keystone species, are based on Ecological Network Analysis (ENA). ENA-based approaches are used to derive quantitative structural indices describing species position in complex networks of interspecific interactions (e.g., Estrada 2007, Jordán 2009, Jordán et al. 2009). ENA methods consist in topological studies applied in ecology to characterize the positional importance of species in communities, by  93 considering the number of trophic links between species, but not necessarily the strength of the links (Jordán 2009). Such methods were first applied to binary networks only (e.g., Jordán et al. 2003), and then adapted to weighted networks (e.g., Scotti et al. 2007, Jordán et al. 2008), since weighted structural indices were demonstrated to best correlate with functional indices from EwE-based methods (Jordán et al. 2008). Alike food web models, ecological networks are aggregated to some extent, seeing that all species are directly or indirectly interconnected in the food web (Jordán 2009). Besides, ENA studies are based on the Lotka-Volterra assumption of uniform and random distribution of interactions between species (e.g., Christianou and Ebenman 2005). Yet, the latter assumption was proven to be unrealistic, compared to the foraging arena theory implemented in EwE models (Walters and Martell 2004, Walters and Christensen 2007, Ahrens et al. 2012). Finally, community structure and dynamics were shown to be correlated in EwE studies (Libralato et al. 2006). Likewise, another challenge for ENA-based approaches is to understand species importance, not only in terms of network position, but also regarding network dynamics (Jordán 2009).  4.4.3. Perspectives for future directions in keystoneness analyses The major challenge in the quest for keystone species is the estimation of the time and space required to assess potential keystone effects (Paine 1995, Power and Mills 1995). Thus, the temporal and spatial scales considered are determinant to the identification of keystone species. Although static patterns differ from dynamic processes (Paine 1980, Arponen 2012),  it was verified that the keystoneness of species may be inferred without performing time-dynamic analyses (Libralato et al. 2006). Thus, in this study, we chose to compare Ecopath models of past and present ecosystem states, rather than using Ecosim simulations. Such approach was previously applied to a selection of models representing upwelling ecosystems at different time periods (Libralato et al. 2006), or used to analyze the effects of the establishment of an MPA (Coll et al. 2009, Eddy et al. 2014). Here, we preferentially selected models representing the same ecosystem at different temporal periods (sibling models) or spatial scales (overlapping models), in order to integrate temporal and spatial variability. We ended up with very few (only 2) overlapping models, so that we could not analyze the influence of spatial scale. In contrast, our selection of Ecopath models comprised a large number of sibling models, which were used to  94 explore context-dependency issues a little further. We used the selected sibling models to compare the changes in identified keystone groups over time in the same ecosystem. However, no consistent pattern could be identified across our selection of sibling models. A more in-depth analysis would be required to fully understand the interacting factors affecting species keystoneness over time.  Human-induced variability may be a factor of potential influence on species keystoneness. Anthropogenic impacts may lead to the loss of keystone species (first becoming rare, and then disappearing) and the community architecture they maintained, leaving an altered ecosystem where species are less likely to play keystone roles (Paine 1995). On the contrary, Power et al. (1996) argued that loss of species diversity may lead to the remaining species taking on keystone roles. Using the + and % indices, several authors demonstrated that ecosystems highly affected by fishing impact had less prevalence of keystone species, whose role was modified with increased fishing pressure. As an example in the Mediterranean Sea, the proportion of keystone species in fully exploited ecosystems was shown to be less prevalent than in protected (or slightly exploited) ecosystems, while the proportion of abundant species were comparable across ecosystems (Coll and Libralato 2012). A similar pattern was observed globally, in a study considering coastal ecosystems from all over the world ocean (Heymans et al. 2012). Keystone species in non-exploited marine ecosystems may become rare when over-exploitation occurs, such as cetaceans in the Mediterranean Sea (Coll et al. 2009), or groupers in the Galapagos (Okey et al. 2004). Similarly, the keystone species status of lobsters in New Zealand was shown to be impacted by fisheries (Eddy et al. 2014). Consequently, human activities could result in large-scale removals of potential keystone species, without any record of the ecosystem state before the uncontrolled removals (Power et al. 1996). Human-induced climate change may be determinant to the keystone role of species as well (Harley 2011). Through the same mechanisms as in natural variability, human-induced variability may lead to competitive balance shifts and predator removals, and thus alter species interactions (Poloczanska et al. 2008). The fishing fleets defined in the selected Ecopath models were included as human predators in our analysis, so as to consider potentially significant anthropogenic (fishing) impacts on keystone species. The All Fleets group appeared as a frequently identified keystone group with the selected index (Figure 4.5a), which could suggest a significant effect of fishing on  95 keystone species in the corresponding modeled food webs. However, the keystoneness of the All Fleets group was not confirmed with the classification tree, since most corresponding models were allocated a No Keystone status (Figure 4.5a). We observed that the Overall Match and Overall Status of the different KS indices were not modified by the inclusion of fishing fleets in the calculation of the mixed trophic impacts. Nonetheless, the sensitivity of the identified keystone group to the addition of the All Fleets group in the analysis varied among the different KS indices. Ideally, an index of keystoneness would consider context-dependency aspects and identify keystone species differently depending on the status of the ecosystem. Alternatively, a suite of complementary indices, each best adapted to a specific level of exploitation, could be developed and applied according to the situation. So, further analyses could explore the sensitivity of the proposed KS index to context-dependency, by comparing the changes in identified keystone species with the changes in fishing pressure in the modeled food web. In the literature, sensitivity analyses on matrix models representing ecological succession were notably used to quantify the relative importance of species interactions to community composition and proposed as a method to identify keystone species (Tanner et al. 1994). In addition, we could investigate the robustness of the proposed KS index to uncertainty and possible model construction errors. Structural indices for binary food webs, such as the mixed trophic impact, were demonstrated to be robust to uncertainty (Fedor and Vasas 2009), so that the equivalent functional indices for weighted food webs, used in our approach, may be assumed to be robust as well. Yet, the development of a measure of robustness for functional indices of keystoneness, similar to the one applied to structural indices, would constitute interesting questions for future studies.  4.4.4. Toward an index of keystoneness applicable to biodiversity conservation In this study, we developed an operational methodology, directly applicable to marine ecosystems, and possibly adaptable to other types of systems (freshwater or terrestrial). Food web dynamics and keystone species may respond differently, depending on the type of ecosystem (Bond 2001, Link 2002a). The response to species introduction or removal is known to be generally slow in natural ecosystems, but the time scale is thought to be shorter in aquatic ecosystems, especially lakes, than in terrestrial ones (Bond 1994, Power et al. 1996, Bond 2001).  96 Marine food webs differ from other types of food web by their higher connectivity, due to large ontogenetic changes in size and diet of marine species, which are opportunistic feeders in an open foraging range (Link 2002a). Therefore, the detection of keystone mechanisms might be eased in freshwater or terrestrial food webs.  Conserving biodiversity is often a compromise between protecting species, areas, or processes (Simberloff 1998). Among the critical ecological processes maintaining whole communities, some are driven by critical species, which thus have to be identified and used in the “conservation biologist’s toolbox” (Power et al. 1996). Critical species play an important ecological function, performed by few other species in the ecosystem, such as keystone species which maintain the food web structure of their community (Perry 2010). Thus, identifying and protecting critical species may be the only long-term solution to preserve a “working ecosystem rather than a collection of charismatic species” (Jordán 2009, Perry 2010). In this study, we focused on the keystone species category, yet other categories of critical species may be identified with the classification tree. Indeed, species belonging to the High Impact – High Biomass category may correspond to critical food resources in the food web. On the contrary, the Low Impact – Low Biomass category may include some rare species, potentially endangered or ecologically extinct. Lastly, the fourth category of Low Impact – High Biomass species may represent critical habitat species in the ecosystem. In practice, analyzing functional roles in species assemblages is determinant to setting conservation priorities and defining restoration programs (McClanahan 2000, Clemente et al. 2010). The identification of functionally important species, such as keystone species, not only helps developing effective conservation strategies for species-level prioritization, but also better understanding of ecosystem functioning and processes (Jordán 2009, Clemente et al. 2010). Moving toward a fully operational and widely applicable index of species keystoneness would thus constitute one step further in biodiversity conservation.     97 Table 4.1. Equations of the 12 indices of keystoneness (KS), obtained by combining each impact components (IC) with each biomass components (BC). ICs were calculated using a measure of trophic impact *!. (Equation 4.2), and BCs using a measure of biomass *!. (Equation 4.3). The components adapted from the literature were called with a subscript capital letters: V4W and 4W (Libralato et al. 2006), and 4( (Power et al. 1996). The other components were called using subscript numbers, with a zero for the components using ranks: V4# and 4#. The abbreviations ‘X’ and ‘X’ designate the rank of the variable, with a ranking in ascending and descending order, respectively.  KS Indices Impact Components Biomass Components +  	V4W $ 4W %  	V4W $ 4( ^  	V4W $ 4# h  V4W $ 4+ V4W  ! 4W   < ! 4(   !⁄  4#  X*!. 4+   i!⁄  g  V4# $ 4W d  V4# $ 4( c  V4# $ 4# e  V4# $ 4+ V4#  X*!. 4W   < ! 4(   !⁄  4#  X*!. 4+   i!⁄  f  V4+ $ 4W +#  V4+ $ 4( ++  V4+ $ 4# +%  V4+ $ 4+ V4+  !% 4W   < ! 4(   !⁄  4#  X*!. 4+   i!⁄  Where !  jk!3%3K!  !  !∑ O=Ol+    98 Table 4.2. Results from the comparison between groups ordered by keystoneness, and by trophic impact or biomass, for each KS index, over all selected models. Both indices adapted from the literature (+ and %) are highlighted in gray, and the selected index (^) is in bold. The number of models whose group with the highest keystoneness matched one of the three groups with highest trophic impact (!), respectively with the lowest biomass (!), is shown in the first column (Match Impact), respectively in the second column (Match Biomass). The number of models for which the group with the highest keystoneness value matched none of the three groups with, either the highest trophic impact, or the lowest biomass, is shown in the third column (No Match). In the last column (Overall Match), results are expressed as an overall match for each KS index over all models, corresponding to the matching category of the majority of models (i.e. 51 models or more, for a total of 101 models) for the particular index. Note: “_” indicates no majority. See Table 4.1 for detailed equations of each KS index.  KS Index Match Impact Match Biomass No Match Overall Match KS1 91 10 0 Match Impact KS2 5 81 15 Match Biomass KS3 50 28 23 _ KS4 25 54 22 Match Biomass KS5 86 12 3 Match Impact KS6 0 94 7 Match Biomass KS7 32 35 34 _ KS8 11 70 20 Match Biomass KS9 91 10 0 Match Impact KS10 25 54 22 Match Biomass KS11 71 20 10 Match Impact KS12 46 39 16 _  99 Table 4.3. Results from the rank correlation tests (column 2), the classification tree (column 3 – see Figure 4.1) and the status allocation (columns 4 to 8 – see Figure 4.2), for each KS index. Both indices adapted from the literature (+ and %) are highlighted in gray, and the selected index (^) is  in bold. Values correspond to the number of models with: validated rank correlation tests (Correlation); whose keystone group identified with KS index is categorized as Keystone with classification tree (KS Category); and whose associated status corresponds to each of the possible statuses (True Positive, False Positive, False Negative and True Negative). In column 2 (Correlation), the first sub-column gives the results with a constraint applied on the correlation coefficients, and the second sub-column (values shown in parentheses) gives the results with no constraint applied. In column 8 (Overall Status), results are expressed as an overall status for each KS index, corresponding to the status of the majority of models (i.e. 35 models or more, since 33 models over the 101 models in total were allocated a No Keystone status), for the particular index. Note: “_” indicates no majority. See Table 4.1 for detailed equations of each KS index.  KS Index Correlation   KS Category True Positive False Positive False Negative True Negative Overall Status KS1 40 ( 95 ) 15 7 21 8 32 _ KS2 12 ( 52 ) 34 6 5 32 25 _ KS3 75 ( 99 ) 48 40 12 14 2 True Positive KS4 0 ( 0 ) 50 0 0 55 13 False Negative KS5 0 ( 0 ) 22 0 0 22 46 True Negative KS6 44 ( 50 ) 22 17 24 9 18 _ KS7 9 ( 14 ) 59 7 1 59 1 False Negative KS8 0 ( 51 ) 43 0 0 48 20 False Negative KS9 0 ( 8 ) 15 0 0 15 53 True Negative KS10 1 ( 1 ) 50 0 1 55 12 False Negative KS11 2 ( 8 ) 35 1 0 37 30 False Negative KS12 3 ( 7 ) 47 2 1 51 14 False Negative  100   Figure 4.1. Identification of keystone species with a classification tree. (a) Classification tree for the identification of keystone species among the functional groups of each modeled ecosystem. The classification is based on two (log-transformed) parameters: the biomass (!) and trophic impact (in squared values) (!%) of each group. Groups are classified using the first quartile (1+) and third quartile (1^) values of the parameters over all living groups in each model, as lower and upper thresholds. Four main categories are discriminated: Keystone (i.e. groups with high impact and low biomass), Low Impact – Low Biomass, Low Impact – High Biomass, High Impact – High Biomass, and Intermediate. The Keystone category is subdivided into two sub-categories: Low TL Keystone and High TL Keystone, by comparing the trophic level (!) of the group to the third quartile values of . The diamond-shaped boxes represent yes/no questions, and the rectangle-shaped boxes the categories identified after answering the questions. (b) Simplified graphical representation of the classification tree on a scatter plot, with the biomass (!) on the x-axis, and the trophic impact (in squared values) (!%) on the y-axis, both in log values. The upper-left box represents the area where the (Low TL and High TL) Keystone groups are located, the upper-right box where the High Impact – High Biomass groups are, the lower-left box where the Low Impact – Low Biomass groups are, and the lower-right one where Low Impact – High Biomass groups are. Each category box is defined based on the classification tree, using the minimum (), maximum (m), first quartiles (1+) and third quartiles (1^) values (see Figure 4.1a). Intermediate groups are spread on the rest of the plotting area (shaded in gray).  HIGH TL KEYSTONELOW IMPACT -HIGH BIOMASSHIGH IMPACT -HIGH BIOMASSyesyesyes yesyesyesnononoLOW TL KEYSTONEyesnono*no*no** INTERMEDIATELOW IMPACT -LOW BIOMASSKEYSTONEHIGH IMPACT HIGH BIOMASSLOW IMPACT LOW BIOMASSLOW IMPACT HIGH BIOMASSINTERMEDIATE(a) (b)  101  Figure 4.2. Rules for status allocation to each pair of ‘Ecopath model – KS index’. The statuses were allocated in three steps: (1) if the KS index was validated, then the result from the rank correlation tests was positive, negative if not; (2) if the identified keystone group was validated, then the result from the classification tree was positive, negative if not; (3) if results from step 1 and 2 were both positive (respectively both negative), then the status was labeled as ‘True Positive’, (respectively ‘True Negative’), but if step 1 was positive and step 2 negative (respectively step 1 negative and step 2 positive), then the status was labeled as ‘False Positive’ (respectively ‘False Negative’). Four alternative statuses are possible for each ‘Ecopath model – KS index’ pair: True Positive, False Positive, False Negative and True Negative. A positive result for step 1 means that the KS index was validated for the selected model, because it was correlated positively with IC and BC, with correlation coefficients of the same order of magnitude. A positive result for step 2 means that the identified keystone group was validated for the selected model, since the group of highest keystoneness with the KS index was categorized as Keystone with the classification tree.   1. Result from the rank correlation tests is:2. Result from the classification tree is: 3. Status of the ‘Ecopath model – KS index’ pair is: negativeTRUE POSITIVEFALSE POSITIVEFALSE NEGATIVETRUE NEGATIVE→ Step 1 & 2 both positive→ Step 1 positive & step 2 negative→ Step 1 negative & step 2 positive→ Step 1 & 2 both negativepositive negativenegativepositivepositive 102  Figure 4.3. Boxplots of (a) spatial extent (Area, thousands of km2), (b) total biomass (>?>, t·km-2), and (c) total catch (>?>, t·km-2·year-1), for all selected Ecopath models, categorized by type of modeled ecosystem: bay / fjord (B/F), channel / strait (C/S), coastal lagoon (CL), continental shelf (CS), coral reef (CR), ocean (Oc), or upwelling (Up). The horizontal line in each box represents the median number of models.   B/F C/S CL CS CR Oc Up0100200300400500Spatial ExtentB/F C/S CL CS CR Oc Up01000200030004000Total BiomassB/F C/S CL CS CR Oc Up010203040Ecosystem TypeTotal Catch 103  Figure 4.4. Boxplots of (a) number of living groups, (b) number of fishing fleets, and (c) mean trophic level of the catch (C), for all selected Ecopath models. The horizontal line in each box represents the median number of models.  020406080100Number of Groups05101520Number of Fleets1.01.52.02.53.03.54.04.5Mean TL of Catch   104  Figure 4.5. Bar plots of the groups identified as keystone with (a) the selected ^ index, and (b) the classification tree. Values are expressed in number of models, over the 101 selected Ecopath models. Based on their original names in each model, the identified keystone groups are aggregated into 12 broader groups: All Fleets, Elasmobranchii, Fishes, Gadiformes, Invertebrates, Mammals, Odontoceti, Pinnipedia, Producers, Scombroidei, Birds, and Zooplankton. With the ^ index, the identified keystone groups may be discriminated based on the status of the model: True Positive (TP), False Positive or False Negative or True Negative (FP / FN / TN), and No Keystone (NKS). With the classification tree, the identified keystone groups may be classified as, either High TL Keystone (High TL KS), or Low TL Keystone (Low TL KS).  All Fleets Elasmobranchii Fishes Gadiformes Invertebrates Mammals Odontoceti Pinnipedia Producers Scombroidei Birds ZooplanktonNumber of Models05101520253051525TPFP / FN / TNNKSAll Fleets Elasmobranchii Fishes Gadiformes Invertebrates Mammals Odontoceti Pinnipedia Producers Scombroidei Birds ZooplanktonGroup namesNumber of Models051015202551525High TL KSLow TL KS 105  Figure 4.6. Application of the classification tree (see Figure 4.1) to the Prince William Sound (1994-1996) food web (Okey and Pauly 1999). The scatter plot shows the biomass (!) on the x-axis, and the trophic impact (in squared values) (!%) on the y-axis, both in log values. Each dot is a functional group in the model, identified with a group number and a group name (indicated in the legend). Groups belonging to each category are displayed in the same color as the corresponding category box: Keystone groups in pink (upper-left box), High Impact – High Biomass groups in orange (upper-right box), Low Impact – Low Biomass groups in green (lower-left box), and Low Impact – High Biomass groups in blue (lower-right box). Keystone groups are subdivided in two categories: Low TL Keystone in dark pink and High TL Keystone in pink. The All Fleets group, for which ! was approximated with the total catch (>?>), corresponds to the aggregation of all the fishing fleets defined in the model.  -6 -4 -2 0 2 4-8-6-4-20Biomass (in log)Trophic Impact (in log)1 2345678910111213141549161718 19202122 232425262728293031323334353637383940414243444546 4748Group Names1: TransientOrca2: SalmonSharks3: ResidentOrca4: SleeperSharks5: Halibut6: Pinnipeds7: Porpoise8: Lingcod9: AdultAtooth10: AdultSalmon11: Pacif icCod12: Sablefish13: JuvAtooth14: Spinydogfish15: AvianPredators16: Octopods17: Seabirds18: DeepDemersals19: Pollock1plus20: Rockfish21: BaleenWhales22: SalmonFry0to1223: NearshoreDemersal24: Squid25: Eulachon26: Seaotters27: DeepEpibent28: Capelin29: AdultHerring30: Pollock031: InverteaterBird32: Sandlance33: ShalLgEpibent34: JuvHerring35: Jellies36: DeepSmInfauna37: NearOmniZoo38: OmniZooplankton39: ShalSmInfauna40: Meiofauna41: DeepLgInfauna42: ShalSmEpibent43: ShalLgInfauna44: NearHerbiZoo45: HerbiZooplankton46: NearPhytoplankton47: OffshorePhyto48: MacroalgaeGras49: AllFleets 106  Figure 4.7. Application of the (a) ^, (b) +, and (c) % indices (see Table 4.1) to the Prince William Sound (1994-1996) food web (Okey and Pauly 1999). Each scatter plot shows the rank of the functional groups, based on its index of keystoneness in ascending order (keystone species have high ranks), on the x-axis, and the trophic level (!) of the functional groups on the y-axis. Each dot is a functional group in the model, identified with a group number and a group name (indicated in the legend). The All Fleets group corresponds to the aggregation of all the fishing fleets defined in the model. Groups are displayed in the color of the category they were associated to on the classification tree (see Figure 4.6).  1 2 3 4 501020304050Trophic LevelRank (ICLBCP)123456789101112131415491617181920212223242526272829303132333435363738394041424344454647481 2 3 4 501020304050Trophic LevelRank (ICLBC0)123456789101112131415491617181920212223242526272829303132333435363738394041424344454647481 2 3 4 501020304050Rank (ICLBCL)12345678910111213141549161718192021222324252627282930313233343536373839404142434445464748Group Names1: TransientOrca2: SalmonSharks3: ResidentOrca4: SleeperSharks5: Halibut6: Pinnipeds7: Porpoise8: Lingcod9: AdultAtooth10: AdultSalmon11: PacificCod12: Sablefish13: JuvAtooth14: Spinydogfish15: AvianPredators16: Octopods17: Seabirds18: DeepDemersals19: Pollock1plus20: Rockfish21: BaleenWhales22: SalmonFry0to1223: NearshoreDemersal24: Squid25: Eulachon26: Seaotters27: DeepEpibent28: Capelin29: AdultHerring30: Pollock031: InverteaterBird32: Sandlance33: ShalLgEpibent34: JuvHerring35: Jellies36: DeepSmInfauna37: NearOmniZoo38: OmniZooplankton39: ShalSmInfauna40: Meiofauna41: DeepLgInfauna42: ShalSmEpibent43: ShalLgInfauna44: NearHerbiZoo45: HerbiZooplankton46: NearPhytoplankton47: OffshorePhyto48: MacroalgaeGras49: AllFleets 107 5. Selectivity of fish feeding: potentials, challenges and recommendations 5.1. Introduction Diet compositions for fish species are usually derived from stomach contents data, which are complex, time-consuming, and expensive to collect (Link 2004, Albouy et al. 2011). The main difficulty in stomach contents analysis is the taxonomic identification of all prey items. Thus, diet composition data are often scarce in the literature, either available for well-known fish species only, or lacking taxonomic details. When no data is available, fish diet composition is either obtained from literature data on similar species, or arbitrarily defined based on expert knowledge. As a consequence, fish diet composition is one of the most poorly estimated parameters in quantitative food web analyses (Link 2004). In the absence of data, predicting fish diet composition may be challenging, since several factors may influence the diet composition of fish species. Indeed, most fish are generalist feeders, i.e. feed on a broad range of prey items, and also are opportunists when they feed, i.e. they take advantage of available sources of food in their environment (Gerking 1994, Link 2004). Nevertheless, when several prey items are available, fish select certain prey items among all available based on some constraints. Feeding selectivity may be defined as the active or instinctive choice of eating some of all the possible prey items (Link 2004). The challenge is to identify the main constraints that influence fish feeding selectivity.  Size was identified as an important parameter in predator-prey interactions (Cohen et al. 1993, Thiebaux and Dickie 1993, Gill 2003). However, characteristics other than size may influence prey preference and use by fish species (Gerking 1994), such as swimming capabilities (Webb 1984, Palomares and Pauly 1989). Prey preference may be defined as the favored choice of the predator for the most suitable prey items, whereas prey utilization (or diet composition) corresponds to the actual preference, coupled with prey availability (Link 2004). In a few cases, fish diet composition was predicted based on a suite of characteristics describing predator and  108 prey species (Sibbing and Nagelkerke 2001, Link 2004). However, there is no strong consensus in the literature on good predictors of fish diet composition.   In this study, contrary to most studies in the literature presenting field-based approaches, we used one of the most important open-access, digital and cross-disciplinary data source on fish biodiversity: FishBase (www.fishbase.org). FishBase is a biodiversity information system (BIS), consisting of three components: a database, a website, and analytical tools (Palomares and Bailly 2011). The prototype of FishBase was created in 1988 and the first version of FishBase was released in 1993 (Froese and Pauly 2000). Then, FishBase released CD-ROMs annually from 1994 to 2000 and went online in 1998 (Froese and Pauly 2000). FishBase is an information repository compiling expert-reviewed data describing more than 32,000 fish species of the world, extracted from over 40,000 scientific publications (Froese 2011). The taxonomic authority list and standard nomenclature reference for species scientific names used in FishBase is the “Catalog of Fishes” released by Eschmeyer in 1998 (Palomares and Bailly 2011).  Initially, FishBase was created to fulfill the needs of classical fisheries management, but it has rapidly evolved to include other aspects related to fish biology and ecology, notably with the development of ecosystem-based fisheries management approaches (Palomares and Bailly 2011). Thus, FishBase has progressively become one of the largest databases documenting global fish biodiversity, and is now a well-known and widely-used source of information in the scientific community (Palomares and Bailly 2011). The primary intention of FishBase was to make “a large amount of empirical data […] available […] at one’s fingertips” (Palomares and Bailly 2011).   In our approach, we intended to identify predictors of fish feeding selectivity by using information readily available in FishBase. More precisely, we intended to test a wide range of parameters (morphological but also anatomical, behavioral or ecological) describing fish species, which have been provided in the literature and encoded in FishBase. First, we selected in FishBase available information on diets and species characteristics of marine and estuarine fish. We focused on trophic interactions between fish species, thus we selected information for piscivorous fish (fish preying on fish) and their corresponding fish prey only. The scope of the study was defined by the availability of the diet data in FishBase, which was mostly  109 representative of the North Sea. We considered a wide range of parameters describing the defined pool of predator and prey fish species. We selected, compiled and formatted the diet data and species parameters and performed statistical analyses so as to: (i) discriminate predator species based on their diets; and (ii) correlate diets to predator and prey species traits described by significant parameters.  5.2. Methods 5.2.1. Definition of the species pool We started by defining the pool of fish species to be considered for the analyses. In FishBase, we selected marine and/or estuarine species, whose status was native or introduced or endemic, and whose life stage included adults. We restricted our selection to species associated with the North Sea and Baltic Seas ecosystems. Indeed, the International Council for the Exploration of the Sea (ICES) ‘Year of the Stomach’ dataset was the largest dataset on fish diets recorded in FishBase (version accessed on January 31 2013) and comprises datasets of fish stomach contents for the North Sea (ICES 2010). The data was collected in 1981, 1985-1986, and 1991, then standardized, quality-controlled and made downloadable online in 2010, and it was integrated into the FishBase database in 2012. Besides, the ‘Year of the Stomach’ dataset has previously been used to develop Ecopath with Ecosim models of the North Sea (Christensen 1995b, Mackinson and Daskalov 2007). The North Sea ecosystems accounted for 73% of all the recorded diet analyses in FishBase, 65% of all the diet records, and 88% of all the recorded diet items identified as fish species in FishBase. In contrast, the Baltic Sea represented about 0.1% of all the recorded diet analyses or records in FishBase. We obtained a pool of 238 fish species, corresponding to 85 families, 36 orders and five classes. In comparison, the total number of species associated with the North and Baltic Seas in FishBase is of 203 and 173, respectively.   110 5.2.2. Compilation of the diet composition from FishBase 5.2.2.1  Selection of the diet data In FishBase, we selected the diet data associated with the previously defined pool of 238 fish species. We selected the diets collected in the North and Baltic Seas, and for which the sample size and proportion of empty stomachs were both available, and the sample stage corresponded to adult predator species. We obtained a collection of 44,217 diets (identified by a unique diet code), representing 87% of all the diets stored in the DIET table in FishBase and associated with 163,328 diet items. Then, we screened out the diet items to select only the ones corresponding to adult fish prey species. Of the 163,328 diet items, only 23% corresponded to finfish and only 14% were identified as finfish species. Thus, we obtained a collection of 22,517 diet percents, corresponding to the proportion of each of the adult fish prey identified as species in the diet of the selected adult fish predator species, and representing 11% of all the diet percents stored in the DIET_ITEMS table in FishBase. The selected diet data was referenced from three different sources (Arntz 1978, Malyshev 1980, ICES 2010), though 99.98% of the data came from ICES (2010). Also, the selected diet data was from the North Sea, except for one record which was from the Baltic Sea. Most of the records from the Baltic Sea were not associated with an identified fish prey species, and thus were not retained.  5.2.2.2  Calculation of weighted averaged diet percentage for each pair of predator-prey species The selected diet data was first analyzed in terms of predator and prey species: 31 unique predator species (in 14 families, eight orders and two classes) and 87 unique prey species (in 38 families, 17 orders, and three classes) included in the 22,517 diet percents. Several estimates of diet percents, obtained from different sampling events, were associated with the same pair of predator-prey species. Thus, a weighted averaged diet percentage was calculated for each unique predator-prey pair. Sample sizes were first corrected to take into account the proportion of empty sampled stomachs. Then, the total sample size (accounting for all the different sampling events) was calculated for each predator. The diet percents were weighted by the proportion of total  111 sample size, and the average value of weighted diet percentage was obtained for each predator-prey pair. The standard error of the mean '" was also calculated, based on the equation:  '"	n!3o  	J N*p8qN.$∑ *r:()8L,rs:()8.tL uv8      Equation 5.1 Where !3 is the weighted average diet percentage for the predator 0 – prey  pair; !3O is the weighted diet percentage for the predator 0 – prey  pair estimated from the sampling event X; and 3 is the total number of sampling events. Based on the 22,517 diet percents from FishBase, we obtained 501 weighted averaged diet percentages and corresponding '".  5.2.2.3  Conversion of weighted averaged diet percentages into diet composition expressed in percent and rank Of the 501 calculated weighted average diet percentages, the 132 values obtained for the pairs of selected predator-prey species (see Section 5.2.3) were converted into a diet composition matrix, with the predator species in columns and the prey species in rows. Some predator-prey pairs, with no weighted average diet percentage associated with them, were assumed not to be interacting and attributed zero values. Then, the weighted average diet percentages were summed to one for each predator, so as to obtain a total of 100% for each matrix column. The diet compositions were also expressed in ranks, using ties for equal weighted average diet percentages. Two parameters were then derived from the diet composition matrix: the number of prey for each predator species (Prey No), and the predator with the highest proportion of the prey for each prey species (Predator Max).  5.2.3. Compilation of the species parameters from FishBase 5.2.3.1  Selection of FishBase fields describing the predator and prey species Based on a literature review, we selected 159 fields in FishBase corresponding to biological or ecological species parameters. For each field, we first extracted from FishBase the records associated with each of the 31 predator species comprised in the diet data. When more  112 than one record was associated with the same species in FishBase, the average value was calculated. Records were missing for some species. Thus, among the 31 predator, we selected the species for which more than two thirds of the selected fields were available, and obtained a smaller pool of 14 predator species. Then, among the 159 selected fields, only the ones that were available for all species in the pool were kept. Also, among the available selected fields, some proved to be not discriminative between the 14 predator species and were excluded. Thus, we intended to obtain the optimal selection, comprising the maximum number of fields for the maximum number of species. We obtained a final selection comprising 33 fields describing 14 predator species. We proceeded the same way for prey, and obtained 33 prey species and 30 associated fields. The selection did not comprise exactly the same fields for the predator and the prey species. The lists of predator and prey species associated with the selection of FishBase fields were compared to the ones associated with the diet percents selection. Only the species listed on both selections were kept. Thus, we ended up with 12 predator species and 31 prey species, associated with 132 weighted averaged diet percentages and described by 31 fields for the predator, and 30 for the prey. Indeed, two fields were not discriminative anymore with the reduced selection of predator species and were excluded.  5.2.3.2  Formatting of the selected predator and prey species parameters The selected FishBase fields were then formatted so as to perform statistical analyses. Both selections for the prey and the predator species comprised quantitative, qualitative, and Boolean variable types. We chose to convert all fields to qualitative variables. Recorded numbers for quantitative variables were converted to classes, using quartiles for a homogeneous spread of the species in each class. Yes/no records for Boolean variables were also converted to classes, by aggregating some fields so as to have at least one species in each class. After formatting, we ended up with 21 qualitative parameters for the 12 predator species and 18 qualitative parameters for the 31 prey species.   113 5.2.4. Statistical analyses We performed multivariate analyses so as to describe and formalize the relationships between the diets, considered as measured variables, and the biological or ecological parameters, considered as explicative variables. We chose to use two exploratory methods, based on linear algebra and using matrix decomposition and orthogonal projection: Principal Component Analysis (PCA) and Correspondence Analysis (CA). The PCA is suitable for quantitative variables, while the CA is the equivalent of a PCA for categorical data and commonly applied to contingency tables. Both analyses were performed using the ‘ade4’ library of the R software (www.r-project.org/).  First, we performed a centered PCA using the diet composition matrix, with the weighted average diet percentages expressed in rank, the predator species as individuals and the prey species as variables. Moreover, we projected each of the 21 parameters selected for predators, as well as the number of prey, as supplementary variables of the PCA. We intended to identify distinct groups of predators, corresponding to different predator types, and correlated them to biological or ecological traits. Then, we performed a CA using the selected biological and ecological parameters formatted as contingency tables, for the predator and prey species separately. We also used the parameters derived from the diet composition matrix, i.e. the number of prey for each predator (Prey No) and the predator with the maximum percentage for each prey (Predator Max). Thus, we intended to identify based on species traits, distinct groups of predator and prey, respectively, that we could relate to the diets.  5.3. Results 5.3.1. Pool of predator and prey species Based on both the diet composition and species parameters compiled from FishBase, we obtained a pool of 12 predator species and 31 prey species (Table 5.1). All species selected as predators are also included as prey, except for one species, the tusk (Brosme brosme), so that the selection comprises 32 different species in total. All 32 species belong to the Actinopterygii  114 class. The 12 predator species correspond to members of three different orders: Perciformes (one species), Pleuronectiformes (one species) and Gadiformes (ten species). Among the members of the Gadiformes, there are seven species belonging to the Gadidae family, two species belonging to the Lotidae family, and one species belonging to the Merlucciidae family. The prey species comprise members of seven different orders (1 Anguilliformes, 1 Atheriniformes, 1 Beloniformes, 3 Clupeiformes, 12 Gadiformes, 6 Perciformes, and 7 Pleuronectiformes) and 16 different families.  5.3.2. Diet composition matrix Of the 501 weighted average diet percentages and associated standard errors of the mean calculated, only 132 are selected for the analysis (see Appendix D). The diet composition matrix for the pairs of selected predator and prey species is expressed in percentage and in rank (Table 5.2a and Table 5.2b). The number of prey species varies substantially between predator species (Table 5.2a). Two predator species show a large number of prey species: 30 for Atlantic cod (J69) and 24 for whiting (J29). Three predator species show an intermediate number of prey species: 15 for Atlantic mackerel (J118), 14 for haddock (J1381), and 10 for saithe (J1343). The other predator species have a small number of prey species, and tusk (J51) has the lowest number, with only two species. The predator species with the highest proportion of the prey varies between prey species (Table 5.2a). Only one predator does not appear as the species with the maximum value for any prey, saithe (J1343), while whiting (J29) and Atlantic cod (J69) appear the most frequently (for seven prey species each).  5.3.3. Biological and ecological parameters After the final selection and formatting of the FishBase fields, we obtained 21 parameters for predator and 18 parameters for prey (Table 5.3 and see Appendix E for details). All parameters have been converted to categorical variables, with defined classes or categories, for the purpose of this study. The specific parameters comprise morphological traits of three types: (i) descriptive, including the shape of the body (BS), shape of the caudal fin (CS), attributes of the dorsal fin (DA), shape of the forehead (FS), position of the mouth (MP), attributes of the  115 pelvic fins (PA), specific attributes (such as barbels) (SA), striking features (features may be body shape, fins or mouth) (SF), ventral position (VP and VP2); (ii) morphometric, including the maximum length (ML); and (iii) meristic, including the number of anal fins (AFN), number of barbels (BN), number of dorsal fins (DFN), maximum number of dorsal spines (DSMa), minimum number of dorsal spines (DSMi), and number of keels (KN). The selection also comprises parameters describing the swimming capabilities of the species, including the aspect ratio (AAR, averaged for each species) and swimming Mode (SM), as well as their defensive abilities, with the Dangerous (D) parameter. Several ecological parameters are also selected, describing the preferred type of ecosystem (EsT), environment (ET), and habitat (HT and FWHT) of the species, as well as its maximum depth (DRD). FishBase records corresponding to each selected (and converted when necessary) specific parameter, plus the supplementary parameters derived from the diets (PreyNo and PredatorGroup), are presented separately for the selected predator and prey species (Table 5.4 and Table 5.5). Note that shorthand were used for some of the records, which consisted of long strings of characters, sometimes including symbols, and were shorthanded for programming purposes (original formulations are provided in Table 5.3).  5.3.4. Identification of predator groups The PCA on the diet composition matrix expressed in rank, with the predator species as individuals, shows that predator groups may be defined based on their diet, and may be related to some specific traits. As observed on the scree plot of Eigen values (Figure 5.1a), the first component contributes to 75% of the total inertia, and the second and third components represent 12% and 8%, respectively. The scattering of the predator species on the factor map of the individuals shows that several groups of predators may be discriminated (Figure 5.1b). When specific parameters are projected as supplementary variables, the distinction between the predator groups appears more clearly. The predator species are spread along the first axis according to their number of prey (PreyNo) (Figure 5.2a and Figure 5.2b). The ordination of the predators on the second axis appears to be related to their maximum length (ML) (Figure 5.3a). The separations between individuals are clearer when they are projected on the plan formed by  116 the axes 2 and 3 (Figure 5.3b). Along the third axis, individuals seem to be discriminated based on their average aspect ratio (AAR) (Figure 5.4a and Figure 5.4b).  Based on the results from the PCA (Figure 5.2, Figure 5.3 and Figure 5.4), five groups of predators may be defined. Group A consists of whiting (J29) with high number of prey, medium size and medium aspect ratio. Group B corresponds to tusk (J51), which has a very low number of prey. Group C consists of Atlantic cod (J69), with high number of prey, very large size and high aspect ratio. Group D comprises Atlantic mackerel (J118), saithe (J1343), and haddock (J1381), all having a medium number of prey and a very high aspect ratio. The other species compose group E and all have a low number of prey. Thus, the ‘Predator Max’ parameter, derived from the diet data for each prey species, may be translated into a predator group (PredatorGroup) parameter, by replacing the identified predator species by the predator group to which it belongs (Table 5.5).  5.3.5. Discrimination of prey and predator based on their species traits The CA on the prey parameters shows that some prey species may be discriminated based on a few parameters. The different levels of each parameter (considered as variables) are projected on the factor map of the individuals (prey species), in the plan formed by the axes 1 and 2 (Figure 5.5). On the upper right side of the plot, a group of prey species may be discriminated based on its number of anal fins (AFN equals two), number of barbels (BN equals one), and number of dorsal fins (DFN equals three). On the lower right side, another group of species may be distinguished according to their minimum number of dorsal spines (DSMi equals six to ten), their environment type (ET is pelagic and neritic), and their freshwater habitat type (FWHT is estuaries and lakes). On the left side, there is one distinct species with a pointed caudal fin, which may be associated with several other species, depending on the parameters considered: (i) either the dorsal fins attributes (DA is continuous with caudal fin), the pelvic fins attributes (PA is suppressed/absent) and the ventral fins position (VP is absent); or (ii) the body shape (BS is eel-like); or (iii) the ecosystem type (EsT is saltwater and brackish water and freshwater). When the ‘Predator Group’ parameter is projected, there seems to be a dichotomy between prey species whose predator belongs to the groups A and D on the left side, and those whose predator belongs to the groups B, C, and E on the right side, although the distinction is not  117 clear (Figure 5.6). The CA provides a poor representation of the data, since the first and second axes contribute to only 13% and 12% of the total inertia, respectively.  The CA on the predator parameters does not help to understand the structure of the data. The distribution of the predator species is mainly driven by two species with very distinct traits, while all the other species are aggregated (Figure 5.7). On the upper side of axis 1, one species (Atlantic mackerel, J118) is discriminated from all the other species, based on its dorsal spines maximum and minimum numbers, environment type, keels number, and swimming mode. In contrast, on the right side of the axis 2, another species (megrim, J28) is discriminated due to its body shape, dorsal fin attributes, environment type, striking features, swimming mode, and ventral position. Thus, the swimming mode could explain the spreading of the predator species, with one ‘carangiform’ species, one ‘anguilliform’ species and ten other corresponding to ‘subcarangiform’ species. However, the CA provides a rather poor representation of the data, since the first and second axes contribute to 26% and 21%of the total inertia, respectively. 5.4. Discussion In this study, we could discriminate several groups of predators (piscivorous fish species) based on the number of prey species they were feeding on, as well as their size and swimming capabilities (defined by the average aspect ratio of their caudal fin). Size and caudal ratio have previously been identified as potential predictors of diet in marine fish assemblages in the literature (e.g., Labropoulou and Markakis 1998, Albouy et al. 2011). However, we could not clearly correlate the predator groups to the biological and ecological traits describing their prey species. Difficulties in the identification of good predictors of fish feeding selectivity may be due to different reasons.  5.4.1. Methodological considerations First, even if the general approach to analyze fish feeding selectivity is the same between studies, the results may be influenced by the chosen methods and variables of each study (Motta et al. 1995, Norton 1995). For instance, inconsistent results between ecomorphological studies  118 were assumed to be due to inconsistent methods (Motta et al. 1995). The first issue consists of a-priori choosing the appropriate morphological traits (Clifton and Motta 1998), and the second of applying the appropriate statistical analysis. A multitude of multivariate analyses are possible, such as the rather simple PCA and CA performed in this study. More complex methods could have revealed clearer patterns in the data structure in our case. However, in the literature, the most complicated statistical methods that were applied often did not provide conclusive or significant results, apart from the identification of broadly defined trophic guilds (e.g., Boyle and Horn 2006, Albouy et al. 2011).  Ideally, a standard method could be proposed and applied so as to obtain results that would be comparable between studies (Motta et al. 1995). Such method could be based on species traits that would be universally applicable, easily estimated or readily available. In fact, ecomorphological studies for fish commonly focus on one particular taxa or ecosystem, and universality has been very little explored. Thus, in our study, we chose to use FishBase, with the aim of ensuring a wide applicability and eased repeatability of the approach. Originally, we intended to consider as many parameters as possible for as many species as possible. However, data availability was a limiting factor, even for such a large database like FishBase, and we had to use a much reduced selection of species and associated information on diets and traits. Nevertheless, it might actually be better to focus on a few parameters only when analyzing factors of fish feeding selectivity (Norton 1995).  Moreover, the selected data had to be reformatted after extraction from FishBase. Predator diets were averaged by predator species and converted to ranks, while species parameters were converted to qualitative variables using pre-defined classes. The transformation of the data was done so as to make the data readily usable to perform statistical analyses, but could have had an influence on the results of our analyses. For instance, the morphological parameters used in this study consisted mostly of choice fields extracted from FishBase, whereas morphometric data are often used in the literature to describe fish morphology (e.g., Maldonado et al. 2009, Oliveira et al. 2010, Albouy et al. 2011). However, the influence of body size on these morphological measurements must be taken into account when testing them as potential predictors of fish diet (Ibañez et al. 2007).   119 5.4.2. Potentials and limitations of the use of FishBase One of the objectives of this study was to use FishBase as a data source to provide new insights on fish feeding selectivity, and it was partially achieved. Using FishBase exclusively, we were able to compile a list of predator and prey species and their associated diets and parameters to be used for performing statistical analyses. The main advantage of using FishBase was that the information was centralized, evaluated, and standardized, notably in terms of species. Thus, the species names associated with both the diet percents and biological or ecological parameters were verified and consistent between the different original sources. Moreover, most information is encoded in FishBase as numeric or choice fields, which facilitates reuse and analyses, contrary to free text (Palomares and Bailly 2011). Lastly, FishBase is structured so as to provide a very detailed description of the recorded diet information, which allowed us to screen out records for which sample size, percentage of empty stomach or prey species names were missing. In that sense, FishBase reached its objective of making empirical data, assembled from already published literature, readily available to facilitate their reuse (Palomares and Bailly 2011).  Nevertheless, some aspects of the reuse of the information extracted from FishBase were more challenging, especially regarding the biological and ecological parameters. Indeed, after the selection of the diet data, we ended up with a pool of 31 predator and 87 prey species. However, after the selection of the parameters, the pool was reduced to 12 predator and 31 prey species. Also, of the 159 FishBase fields of potential interest initially considered for the analysis, only 21 parameters were retained for the predator, and 18 for the prey. Many parameters were missing for many species: only 72% of the parameters were available for the best documented half of the predator species selection based on diets (i.e. 15 predator over 31); and only 64% for the best documented third of the prey species selection based on diets (i.e. 33 prey over 87). Missing parameters in FishBase simply may not be available in the literature, and thus could not be documented in FishBase. Thus, our initial pre-selection of 159 FishBase fields could be an optimistic estimation of the current scientific knowledge. Nevertheless, some fields may also be empty in FishBase because they are not considered as ‘active fields’ and thus are not filled or updated on a regular basis, such as for some morphological parameters (Froese and Pauly 2000). FishBase fields of potential interest to our approach, which could not be used due to low  120 availability for the selected species, included the cross section of the body, the caudal attributes, the body weight, morphometric measurements and descriptors of the teeth, gill rakers and scales.  Moreover, some fields were actually available in FishBase, but they could not be reused in our analyses since they were not discriminative, i.e. they had the exact same value for all the selected species. Thus, only 27% of the available fields for the selected predator species, respectively 25% for the prey, happened to be useful to differentiate species. The question of defining parameters to discriminate species is critical when using choice fields, i.e., if choice fields are defined to include every possible case, it may result in slowing down the encoding process (Palomares and Bailly 2011). In FishBase, the minimum number of choices is proposed, including the most frequent cases only (Froese and Pauly 2000). For specific cases, the choice field indicates ‘other’ and details are provided as free text in related remark fields (Froese and Pauly 2000), but such fields are more difficult to reuse. For instance, the FishBase fields that were not reused in our approach because they were not discriminative included the type of mouth, the type of eyes, or the attributes of the fins.  5.4.3. Challenges in predicting fish feeding selectivity First, it is important to emphasize that correlation is different from causality (Motta et al. 1995), and approaches such as the one presented here may only describe correlations and assume causality between species traits and diet composition (Norton 1991, 1995). Based on the ecomorphological hypothesis of a correlation between species morphology and ecology, the morphology of the feeding apparatus could be a good predictor of prey utilization for fish (Motta et al. 1995, Norton 1995). However, there is contrasting evidence from the literature to verify the hypothesis of dietary-morphological relationships in fish. Thus, several studies concluded that fish morphology may predict the way fish are feeding (prey preference or feeding mode), but not the prey they feed on (prey utilization or diet composition) (Motta 1988, Motta et al. 1995, Clifton and Motta 1998, Barnett et al. 2006). Some morphological traits that can be directly related to diet may be good predictors of diet, especially if they are biomechanically relevant (e.g., crushing ability related to hard-shelled prey) (Clifton and Motta 1998). Yet, in many cases, the ecomorphological traits could be used to predict some general aspects of fish feeding only (i.e. trophic guilds, e.g, Boyle and Horn 2006,  121 Ibañez et al. 2007, Oliveira et al. 2010, Albouy et al. 2011). Possibly, predicting broad trophic guilds of fish species could be relatively easier than, for instance, discriminating piscivorous fish species based on their prey species, as attempted in this study. There may be a greater variability between species belonging to different trophic guilds, rather than between species within the piscivorous trophic guild.  Several factors may influence species morphology, which may as a result be very distantly related to diet (Motta et al. 1995, Norton 1995, Labropoulou and Markakis 1998). Indeed, evolutionary solutions are multiple, and so are the morphological strategies to feeding (Motta 1988). Besides, fish morphology determines swimming capabilities used in foraging activities, but also predation escape (Motta 1988), with a trade-off between prey profitability and predation risk in fish feeding (Labropoulou and Markakis 1998). Thus, in our pool of selected species, all predator species (except for one) were prey items in the diet composition of other predators. Also, both fish morphology and diet may be influenced by habitat utilization (Motta et al. 1995, Barnett et al. 2006). Lastly, the necessity to consider phylogenetic relationships between species in ecomorphological studies is a matter of controversy (Clifton and Motta 1998). Evolutionary history was demonstrated to be more important than ecological history as a factor of influence on fish morphology (Labropoulou and Markakis 1998), especially when many parameters are considered (Norton 1995). It was also argued that species sharing a common ancestor should not be considered statistically independent (Ibañez et al. 2007). Nevertheless, predator-prey interactions may be highly variable between individuals of the same species due to learning capabilities of fish (Kieffer and Colgan 1992, Warburton 2003).  Behavior may be more important than morphology to predict fish feeding selectivity and determine diet composition (Grossman 1986). Despite constraining morphology, fish are able to adapt their diet based on local conditions thanks to their behavioral flexibility (Labropoulou and Markakis 1998, Ibañez et al. 2007). There is a trade-off between morphological and dietary specialization for more efficient feeding, and flexibility allowed by a generalist diet (Norton 1995). Prey utilization is a combination of prey preference and prey availability (Link 2004). Thus, when we predict prey preference based on diet, we ignore the influence of prey availability (Barnett et al. 2006) (likewise, we ignore prey selectivity when we predict diet from prey abundance only). For instance, size is commonly used to predict species interactions in size spectra models (Jennings et al. 2008). Indeed, the relationship between prey size and predator  122 size has been demonstrated, notably for some fish species in the North Sea (Floeter and Temming 2003, 2005). Our study confirms the potential correlation between the length of the predator fish and their diet composition. However, previous studies also highlighted that the influence of prey availability is stronger than prey size preference on fish diet composition (e.g., North Sea cod; Floeter and Temming 2003).  5.4.4. Recommendations for future studies Data restriction was the main challenge in our study, as we intended to predict diet composition at the species level, for piscivorous fish species only. The challenges of defining the desirable completeness and level of details in provided information are inherent to any BIS, such as FishBase (Palomares and Bailly 2011). Thus, our approach may highlight some of the strengths and the weaknesses in the current structure and content of FishBase. For instance, some of the fields considered as potentially relevant to our study, but which were not available or discriminative and thus could not be reused, could receive more attention. Overall, we would recommend that future studies relying on FishBase perform an in-depth assessment of the availability and level of details of the FishBase data to be reused, and possibly contribute to the refining of some data if necessary, prior to perform analyses. As FishBase invites, accepts, and acts quickly on critical comments (Palomares and Bailly 2011), we believe that the potentials of the database to be used for future studies are promising.    123 Table 5.1. Nomenclature and taxonomy classification for the 12 predator and 31 prey species selected from FishBase. Code corresponds to the code attributed to each species in FishBase. Code Predator Prey Common Name Genus Species Family Order Class 22 no yes Reticulated dragonet Callionymus reticulatus Callionymidae Perciformes Actinopterygii 23 no yes Dragonet Callionymus lyra Callionymidae Perciformes Actinopterygii 24 no yes Atlantic herring Clupea harengus Clupeidae Clupeiformes Actinopterygii 26 no yes Witch flounder Glyptocephalus cynoglossus Pleuronectidae Pleuronectiformes Actinopterygii 28 yes yes Megrim Lepidorhombus whiffiagonis Scophthalmidae Pleuronectiformes Actinopterygii 29 yes yes Whiting Merlangius merlangus Gadidae Gadiformes Actinopterygii 30 yes yes European hake Merluccius merluccius Merlucciidae Gadiformes Actinopterygii 31 no yes Blue whiting Micromesistius poutassou Gadidae Gadiformes Actinopterygii 33 yes yes Ling Molva molva Lotidae Gadiformes Actinopterygii 34 yes yes Pollack Pollachius pollachius Gadidae Gadiformes Actinopterygii 35 no yes European eel Anguilla anguilla Anguillidae Anguilliformes Actinopterygii 38 no yes Small sandeel Ammodytes tobianus Ammodytidae Perciformes Actinopterygii 39 no yes Transparent goby Aphia minuta Gobiidae Perciformes Actinopterygii 41 no yes Mediterranean scaldfish Arnoglossus laterna Bothidae Pleuronectiformes Actinopterygii 45 no yes Sand smelt Atherina presbyter Atherinidae Atheriniformes Actinopterygii 47 no yes Garfish Belone belone Belonidae Beloniformes Actinopterygii 51 yes no Tusk Brosme brosme Lotidae Gadiformes Actinopterygii 53 no yes Solenette Buglossidium luteum Soleidae Pleuronectiformes Actinopterygii 60 no yes Crystal goby Crystallogobius linearis Gobiidae Perciformes Actinopterygii 66 no yes European anchovy Engraulis encrasicolus Engraulidae Clupeiformes Actinopterygii 69 yes yes Atlantic cod Gadus morhua Gadidae Gadiformes Actinopterygii 118 yes yes Atlantic mackerel Scomber scombrus Scombridae Perciformes Actinopterygii 481 no yes Poor cod Trisopterus minutus Gadidae Gadiformes Actinopterygii 525 no yes Common sole Solea solea Soleidae Pleuronectiformes Actinopterygii 1023 yes yes Norway pout Trisopterus esmarkii Gadidae Gadiformes Actinopterygii 1341 no yes European flounder Platichthys flesus Pleuronectidae Pleuronectiformes Actinopterygii 1342 no yes European plaice Pleuronectes platessa Pleuronectidae Pleuronectiformes Actinopterygii  124 Code Predator Prey Common Name Genus Species Family Order Class 1343 yes yes Saithe Pollachius virens Gadidae Gadiformes Actinopterygii 1357 no yes European sprat Sprattus sprattus Clupeidae Clupeiformes Actinopterygii 1367 yes yes Pouting Trisopterus luscus Gadidae Gadiformes Actinopterygii 1381 yes yes Haddock Melanogrammus aeglefinus Gadidae Gadiformes Actinopterygii 1876 no yes Three-bearded rockling Gaidropsarus vulgaris Lotidae Gadiformes Actinopterygii  125 Table 5.2. Diet composition matrix for selected predator-prey pairs. Species are indicated with their species code in FishBase, with a capital  for prey in rows, and with a capital 0 for predator in columns. (a)  values summed to one and expressed in percentage. “_” indicates null values. The last row (Prey No) indicates, for each predator species, the number of prey species for which the diet percentages are not null. The last column (Predator Max) indicates, for each prey species, the predator species which has the highest the diet percentage. (b)  values expressed in rank. Ties were attributed the same rank. Null values were ranked last.  J28 J29 J30 J33 J34 J51 J69 J118 J1023 J1343 J1367 J1381 Predator Max I22 _ 5.9 _ _ _ _ 1.1 _ _ _ _ _ J29 I23 11.6 1.3 _ _ _ _ 1.2 2.6 _ _ 16.2 4.6 J1367 I24 16.6 5.4 15.9 8.1 11.5 50.0 2.6 11.1 14.2 16.4 1.8 10.4 J51 I26 _ 0.7 _ _ _ _ 1.4 0.1 _ 0.3 _ 3.2 J1381 I28 _ 3.7 _ _ _ _ 1.1 _ _ _ _ _ J29 I29 10.9 4.1 9.2 14.1 9.4 _ 2.7 19.7 _ 3.2 39.6 5.5 J1367 I30 _ 2.2 _ _ _ _ _ _ _ _ _ _ J29 I31 _ 5.9 _ _ _ _ 15.4 6.2 _ 6.3 _ _ J69 I33 _ _ _ _ _ _ 1.5 _ _ _ _ _ J69 I34 _ _ _ _ _ _ 12.1 _ _ _ _ 2.3 J69 I35 _ 5.8 _ _ _ _ 1.8 _ _ _ _ _ J29 I38 _ 4.0 _ _ _ _ 9.9 20.5 _ _ _ 18.7 J118 I39 _ _ _ _ _ _ 0.3 0.1 41.0 _ _ 7.6 J1023 I41 _ 0.5 _ _ _ _ 0.3 0.1 _ _ _ _ J29 I45 _ _ _ _ _ _ 6.2 _ _ _ _ _ J69 I47 _ 8.7 _ _ _ _ 4.4 _ _ _ _ _ J29 I53 _ 3.1 0.2 _ _ _ 0.6 _ _ _ 10.7 _ J1367 I60 _ 1.4 _ _ _ _ 0.4 _ 18.1 _ _ _ J1023 I66 _ _ _ _ _ _ 0.8 _ _ _ _ _ J69 I69 14.3 2.4 _ 21.0 34.0 50.0 2.2 7.4 _ 10.8 _ 3.4 J51 I118 _ 3.3 15.7 19.8 _ _ 2.7 1.2 _ 19.3 _ _ J33 I481 _ 2.6 18.5 _ _ _ 1.4 3.2 _ 10.7 9.7 _ J30 I525 _ 3.1 _ _ _ _ 1.0 _ _ _ _ 3.2 J1381 I1023 13.4 7.5 15.4 16.0 32.8 _ 5.0 11.5 22.8 19.9 _ 14.1 J34 I1341 _ 14.6 _ _ _ _ 2.2 _ _ _ _ _ J29 I1342 _ _ _ _ _ _ 0.9 4.3 _ _ _ 0.3 J118 I1343 _ 1.2 _ _ _ _ 12.5 _ _ _ _ 4.2 J69 I1357 _ 5.3 18.5 _ _ _ 1.8 8.2 3.9 3.6 22.0 9.6 J1367 I1367 16.6 3.2 _ _ _ _ 1.0 _ _ _ _ _ J28 I1381 16.6 3.9 6.5 21.0 12.2 _ 3.6 3.7 _ 9.4 _ 13.0 J33 I1876 _ _ _ _ _ _ 1.8 _ _ _ _ _ J69 Prey No 7 24 8 6 5 2 30 15 5 10 6 14    (a)  126  J28 J29 J30 J33 J34 J51 J69 J118 J1023 J1343 J1367 J1381 I22 6 5 8 6 6 2 22 16 6 11 7 15 I23 4 21 8 6 6 2 20 11 6 11 3 8 I24 1 7 2 5 4 1 11 4 4 3 6 4 I26 6 23 8 6 6 2 18 13 6 10 7 11 I28 6 12 8 6 6 2 21 16 6 11 7 15 I29 5 9 5 4 5 2 9 2 6 9 1 7 I30 6 19 8 6 6 2 31 16 6 11 7 15 I31 6 4 8 6 6 2 1 7 6 7 7 15 I33 6 25 8 6 6 2 17 16 6 11 7 15 I34 6 25 8 6 6 2 3 16 6 11 7 13 I35 6 6 8 6 6 2 14 16 6 11 7 15 I38 6 10 8 6 6 2 4 1 6 11 7 1 I39 6 25 8 6 6 2 29 15 1 11 7 6 I41 6 24 8 6 6 2 30 14 6 11 7 15 I45 6 25 8 6 6 2 5 16 6 11 7 15 I47 6 2 8 6 6 2 7 16 6 11 7 15 I53 6 16 7 6 6 2 27 16 6 11 4 15 I60 6 20 8 6 6 2 28 16 3 11 7 15 I66 6 25 8 6 6 2 26 16 6 11 7 15 I69 2 18 8 1 1 1 13 6 6 4 7 10 I118 6 13 3 2 6 2 10 12 6 2 7 15 I481 6 17 1 6 6 2 19 10 6 5 5 15 I525 6 15 8 6 6 2 24 16 6 11 7 12 I1023 3 3 4 3 2 2 6 3 2 1 7 2 I1341 6 1 8 6 6 2 12 16 6 11 7 15 I1342 6 25 8 6 6 2 25 8 6 11 7 14 I1343 6 22 8 6 6 2 2 16 6 11 7 9 I1357 6 8 1 6 6 2 15 5 5 8 2 5 I1367 1 14 8 6 6 2 23 16 6 11 7 15 I1381 1 11 6 1 3 2 8 9 6 6 7 3 I1876 6 25 8 6 6 2 16 16 6 11 7 15   (b)  127 Table 5.3. List of biological and ecological parameters selected for predator and prey species, with associated acronyms and classes or categories. Adapted from FishBase (see Appendix E for details).  Acronym Parameter Classes/Categories Predator Prey AAR Average Aspect Ratio low (0-0.93); medium (0.94-1.19); high (1.20-1.35); very high (1.36-3.05) yes no AFN Anal Fins Number one; two yes yes BN Barbels Number zero; one; four no yes BS Body Shape eel-like; elongated; fusiform/normal ; short and/or deep yes yes CS Caudal Fin Shape forked; truncate; pointed yes yes D Dangerous reports of ciguatera poisoning; harmless no yes DA Dorsal Fin Attributes continuous with caudal fin; extending over most of the back length; first fin ray very elongated; none; origin on head nearly above eye; other yes yes DFN Dorsal Fins Number one; two; three yes yes DRD Depth Range Deep shallow (0-274); medium (275-524); deep (525-999); very deep (1000-1075) yes no DSMa Dorsal Spines Maximum none (0); one to five (1-5); six to ten (6-10); eleven to fifteen (11-15) yes yes DSMi Dorsal Spines Minimum none (0); one to five (1-5); six to ten (6-10); eleven to fifteen (11-15) yes yes EsT Ecosystem Type saltwater; saltwater and brackish water; saltwater and freshwater yes yes ET Environment Type bathydemersal; bathypelagic; benthopelagic; demersal; pelagic-neritic; pelagic-oceanic yes yes FS Forehead Shape convex; straight yes no FWHT Freshwater Habitat Type none; estuaries; estuaries and lakes; estuaries and stream no yes HT Habitat Type oceanic; oceanic and neritic; oceanic and neritic and intertidal yes yes KN Keels Number zero; two yes no ML Maximum Length very small (0-24); small (25-49); medium (50-99); large (100-149); very large (150-200)  yes yes MP Mouth Position sub-terminal/inferior; superior; terminal yes yes PA Pelvic Fins Attributes asymmetric in size/position; normal; other; suppressed/absent no yes SA Specific Attributes barbels; none yes no SF Striking Features flatfish; fins; mouth; none yes yes SM Swimming Mode anguilliform; carangiform; subcarangiform yes no VP Ventral Fins Position abdominal; absent; jugular; thoracic yes yes VP2 Ventral Fins Position 2 before origin of D1; behind origin of D1 yes no  128 Table 5.4. Parameters selected for the 12 selected predator species. FishBase records are presented using shorthand (see Table 5.3 for the original classes/categories).  Code AAR AFN BS CS DA DFN DRD DSMa DSMi EsT ET J28 medium one shortdeep truncate extending one deep none none salt bathydem J29 medium two fusiform truncate none three shallow none none salt benthopel J30 low one elongated truncate none two verydeep none none salt dem J33 low one elongated truncate none two verydeep none none salt dem J34 high two fusiform forked none three shallow none none salt benthopel J51 high one elongated truncate continuous one verydeep none none salt dem J69 high two fusiform truncate none three deep none none saltbrack benthopel J118 veryhigh one fusiform forked none two verydeep elevtofift sixtoten saltbrack pelner J1023 medium two fusiform truncate none three medium none none salt benthopel J1343 veryhigh two fusiform forked none three medium none none salt dem J1367 low two fusiform truncate none three shallow none none saltbrack benthopel J1381 veryhigh one fusiform truncate none three medium none none salt dem  Code FS HT KN ML MP SA SF SM VP VP2 PreyNo J28 straight ocener zero medium superior none flatfish anguilliform jugular behind low J29 straight ocener zero medium terminal none none subcarangiform thoracic before high J30 straight ocener zero large terminal none none subcarangiform thoracic before low J33 straight ocener zero verylarge terminal none none subcarangiform thoracic before low J34 straight ocener zero large superior none none subcarangiform thoracic before low J51 straight oce zero large terminal barbels none subcarangiform thoracic before verylow J69 straight ocenerint zero verylarge terminal barbels none subcarangiform thoracic before high J118 straight ocener two medium terminal none none carangiform thoracic before medium J1023 straight oce zero small superior barbels none subcarangiform thoracic before low J1343 straight oce zero large terminal barbels none subcarangiform thoracic before medium J1367 convex oce zero small terminal barbels none subcarangiform thoracic before low J1381 convex oce zero large inferior barbels none subcarangiform thoracic before medium  129 Table 5.5. Parameters selected for the 31 selected prey species. FishBase records are presented using shorthand (see Table 5.3 for the original classes/categories).  Code AFN BN BS CS D DA DFN I22 one zero fusiform truncate harmless other two I23 one zero fusiform truncate harmless other two I24 one zero fusiform forked harmless none one I26 one zero shortdeep truncate harmless extending one I28 one zero shortdeep truncate harmless extending one I29 two zero fusiform truncate harmless none three I30 one zero elongated truncate harmless none two I31 two zero elongated truncate harmless none three I33 one zero elongated truncate harmless none two I34 two zero fusiform forked harmless none three I35 one zero eellike pointed harmless continuous one I38 one zero elongated forked harmless continuous one I39 one zero fusiform truncate harmless none two I41 one zero shortdeep truncate harmless extending one I45 one zero elongated forked harmless none two I47 one zero eellike truncate harmless none one I53 one zero elongated truncate harmless extending one I60 one zero fusiform truncate harmless other two I66 one zero elongated forked harmless none one I69 two one fusiform truncate harmless none three I118 one zero fusiform forked harmless none two I481 two one fusiform truncate harmless none three I525 one zero shortdeep truncate harmless origin one I1023 two one fusiform truncate harmless none three I1341 one zero shortdeep truncate harmless origin one I1342 one zero shortdeep truncate harmless origin one I1343 two one fusiform forked harmless none three I1357 one zero fusiform forked ciguatera none one I1367 two one fusiform truncate harmless none three I1381 one one fusiform truncate harmless none three I1876 one four elongated truncate harmless elongated two    130 Code DSMa DSMi EsT ET FWHT HT I22 onetofive onetofive saltbrack dem est ocenerint I23 onetofive onetofive salt dem none ocener I24 none none saltbrack benthopel eststr ocenerint I26 none none salt dem none oce I28 none none salt bathydem none ocener I29 none none salt benthopel none ocener I30 none none salt dem none ocener I31 none none salt bathypel none ocener I33 none none salt dem none ocener I34 none none salt benthopel none ocener I35 none none saltbrackfresh dem eststr oce I38 none none saltbrack dem est ocenerint I39 sixtoten onetofive saltbrack pelner est ocener I41 none none salt dem none ocener I45 sixtoten sixtoten saltbrack pelner estlak ocenerint I47 none none saltbrack peloceanic none ocener I53 none none salt dem none oce I60 onetofive onetofive salt dem none ocener I66 none none saltbrack pelner estlak ocener I69 none none saltbrack benthopel est ocenerint I118 elevtofift sixtoten saltbrack pelner none ocener I481 none none salt benthopel none oce I525 none none saltbrack dem est ocenerint I1023 none none salt benthopel none oce I1341 none none saltbrackfresh dem eststr ocenerint I1342 none none saltbrack dem est ocenerint I1343 none none salt dem none oce I1357 none none saltbrack pelner est ocener I1367 none none saltbrack benthopel none oce I1381 none none salt dem none oce I1876 none none salt dem none oce    131 Code ML MP PA SF VP PredatorGroup I22 verysmall terminal normal fins jugular A I23 small terminal other fins jugular E I24 small superior normal none abdominal B I26 medium superior normal flatfish jugular D I28 medium superior normal flatfish jugular A I29 medium terminal normal none thoracic E I30 large terminal normal none thoracic A I31 medium terminal normal none thoracic C I33 verylarge terminal normal none thoracic C I34 large superior normal none thoracic C I35 medium terminal absent fins absent A I38 verysmall superior absent none absent D I39 verysmall terminal normal none thoracic E I41 small terminal asymmetric flatfish thoracic A I45 verysmall terminal normal none abdominal C I47 medium terminal normal mouth abdominal A I53 verysmall inferior normal flatfish jugular E I60 verysmall superior normal mouth thoracic E I66 verysmall inferior normal none abdominal C I69 verylarge terminal normal none thoracic B I118 medium terminal normal none thoracic E I481 small terminal normal none thoracic E I525 medium inferior normal flatfish jugular D I1023 small superior normal none thoracic E I1341 medium superior normal flatfish thoracic A I1342 large superior normal flatfish thoracic D I1343 large terminal normal none thoracic C I1357 verysmall terminal normal none abdominal E I1367 small terminal normal none thoracic E I1381 large inferior normal none thoracic E I1876 medium terminal normal fins thoracic C  132  Figure 5.1. Results from the PCA on the diet matrix expressed in rank, with the predator species as individuals: (a) scree plot of Eigen values and (b) scatter diagram of the coordinates (or factor map) of the individuals.  1 2 3 4 5 6 7 8 9 10 11ComponentsEigen values in % of total inertia020406080100 d = 10  axis 1 - axis 2  J28  J29  J30  J33  J34  J51  J69  J118  J1023  J1343  J1367  J1381 (a) (b)  133   Figure 5.2. Results from the PCA: factor map of the individuals with the number of prey (PreyNo) projected as a supplementary variable on the plan formed by (a) axes 1 and 2 and (b) axes 1 and 3.    d = 10  J28  J29  J30  J33  J34  J51  J69  J118  J1023  J1343  J1367  J1381  axis 1 - axis 2 PreyNo  high  low  medium  verylow  d = 10  J28  J29  J30  J33  J34  J51  J69  J118  J1023  J1343  J1367  J1381  axis 1 - axis 3 PreyNo  high  low  medium  verylow (a) (b)  134   Figure 5.3. Results from the PCA: factor map of the individuals with the maximum length (ML) projected as a supplementary variable on the plan formed by (a) axes 1 and 2 and (b) axes 2 and 3.    d = 10  J28  J29  J30  J33  J34  J51  J69  J118  J1023  J1343  J1367  J1381  axis 1 - axis 2 ML  large  medium  small  verylarge  d = 10  J28  J29  J30  J33  J34  J51  J69  J118  J1023  J1343  J1367  J1381  axis 2 - axis 3 ML  large  medium  small  verylarge (a) (b)  135   Figure 5.4. Results from the PCA: factor map of the individuals with the average aspect ratio (AAR) projected as a supplementary variable on the plan formed by (a) axes 1 and 3 and (b) axes 2 and 3.   d = 10  J28  J29  J30  J33  J34  J51  J69  J118  J1023  J1343  J1367  J1381  axis 1 - axis 3 AAR  high  low  medium  veryhigh  d = 10  J28  J29  J30  J33  J34  J51  J69  J118  J1023  J1343  J1367  J1381  axis 2 - axis 3 AAR  high  low  medium  veryhigh (a) (b)  136  Figure 5.5. Results from the CA on the prey parameters, with the prey species as individuals: factor map of the individuals with the different levels of each parameter projected on the plan formed by the axes 1 and 2. Arrows of different colors indicate species that may be discriminated.   AFN  one  tw o  BN  four  one  zero  BS  eellike  elongated  fusiform  shortdeep  CS  forked  pointed  truncate  D  ciguatera  harmless  DA  continuous  elongated  extending  none  origin  other  DFN  one  three  tw o  DSMa  elevtofif t  none  onetofive  sixtoten  DSMi  none  onetofive  sixtoten  EsT  salt  saltbrack  saltbrackfresh  ET  bathydem  bathypel  benthopel  dem  pelner  peloceanic  FWHT  est  estlak  eststr  none  HT  oce  ocener  ocenerint  ML  large  medium  small  verylarge  verysmall  MP  inferior  superior  terminal  PA  absent  asymmetric  normal  other  SF  f ins  f latf ish  mouth  none  VP  abdominal  absent  jugular  thoracic  137  Figure 5.6. Results from the CA on the prey parameters: factor map of the individuals with the group of the predator (PredatorGroup) with the highest diet composition value projected as a supplementary variable on the plan formed by the axes 1 and 2.   d = 0.5  axis 1 - axis 2 PredatorGroup  I22  I23  I24  I26  I28  I29  I30  I31  I33  I34  I35  I38  I39  I41  I45  I47  I53  I60  I66  I69  I118  I481  I525  I1023  I1341  I1342  I1343  I1357  I1367  I1381  I1876  A  B  C  D  E  138  Figure 5.7. Results from the CA on the predator parameters, with the predator species as individuals: factor map of the individuals with the different levels of each parameter projected on the plan formed by the axes 1 and 2. Arrows of different colors indicate species that may be discriminated.   AAR  high  low   medium  veryhigh  AFN  one  tw o  BS  elongated  fusiform  shortdeep  CS  forked  truncate  DA  continuous  extending  none  DFN  one  three  tw o  DRD  deep  medium  shallow   verydeep  DSMa  elevtof ift  none  DSMi  none  sixtoten  EsT  salt  saltbrack  ET  bathydem  benthopel  dem  pelner  FS  convex  straight  HT  oce  ocener  ocen rint  KN  tw o  zero  ML  large  medium  small  verylarge  MP  inferior  superior  terminal  SA  barbels  none  SF  flatf ish  none  SM  anguilliform  carangiform  subcarangiform  VP  jugular  thoracic  VP2  before  behind  139 6. Conclusion In this thesis, I performed comprehensive literature reviews and global meta-analyses to build on existing knowledge for providing new insights on trophic interaction between marine species (MSTI). In the Introduction (Chapter 1), I briefly summarized current knowledge on the importance of MSTI and the approaches used to model and predict them. Trophic interactions correspond to feeding relationships between predators and prey, and may have significant effects on marine communities and ecosystems through changes in species abundance (Chapter 1). Anthropogenic drivers, such as fishing activities or human-induced climate change, were identified as important factors of change in MSTI (Chapter 1). The removal of top predators by fishing results in the loss of trophic interactions, and may have consequences for ecosystem functions and services (‘trophic downgrading’, Estes et al. 2011). Ecosystem-based management approaches were introduced to maintain desirable state and processes at the ecosystem level, and ecosystem models were used as tools for simultaneously considering all components of the ecosystem (Chapter 1). The Ecopath with Ecosim (EwE) trophodynamic model was notably developed to inform ecosystem-based fisheries management (Christensen and Walters 2005, Christensen and Walters 2011).  In the subsequent chapters of my thesis, the analysis of MSTI was structured around three main concepts: feeding selectivity, food web and keystone species. Each concept may be expressed and measured by applying some specific parameters: diet composition (DC), mixed trophic impact (MTI), and index of keystoneness (KS), respectively. These parameters describe the importance of trophic interactions at the species level, and are directly related to one another: the KS index is based on MTI values, which are derived from DC estimates. All parameters were obtained from global information repositories, used as sources of a large amount of digital and standardized data and metadata. DC estimates were calculated from referenced diet records encoded in FishBase, while estimates of MTI and KS index were derived from published EwE models compiled in EcoBase.   Chapter 2 provided an overview of the EwE applications worldwide over the last 30 years, and gathered in the EcoBase repository. Based on the compilation of 433 models associated with 397 publications and 23 sets of metadata, I described how the EwE modeling  140 approach has been used to study MSTI. Such a comprehensive review has not previously been done in the literature. Overall, the majority of EwE applications corresponded to Ecopath models of (temperate or tropical) marine ecosystems located in the (Northern or Central) Atlantic Ocean, which were developed to analyze ecosystem functioning and fisheries. Through time, the number of published EwE models has been growing, so that all FAO areas comprise at least one model in 2014. Over the last three decades, the complexity and spatial scale of the models have been increasing (larger number of groups and wider range of model area), and more and more Ecopath models were built to represent a given ecosystem at different periods of time. While a significant number of dynamic (Ecosim) versions of the models have been developed, spatially-explicit (Ecospace) versions have remained sparse. The research questions addressed using EwE models and the types of modeled ecosystems have progressively become more diverse (e.g., high latitudes, terrestrial systems). Thus, Chapter 2 demonstrated the growing contribution of the EwE modeling approach to the scientific literature.  The second outcome of Chapter 2 is the standardization and digitization of the metadata describing the EwE models and the corresponding modeled ecosystems. Open-access to these metadata will enable new meta-analyses to be performed, by using them as criteria and applying simple scoring methods to select the models of potential interest (such as what I presented in Chapter 4). Meta-analyses are a growing type of approach in EwE modeling, since they allow getting new insights from current knowledge, by reusing and comparing existing EwE models or EwE-based studies (e.g., Coll et al. 2012, Pikitch et al. 2012, Heymans et al. 2014). In theory, a wide range of metadata could be defined to select published ecosystem models of interest for conducting meta-analyses: general descriptors of the models (such as the ones compiled in Chapter 2), but also input and output parameters (e.g., system indices). However, input and output of EwE models have not yet been made available in EcoBase. The encoding of the data derived from EwE models (with open-access agreement) is the next step in the development of the EcoBase repository (Colléter et al. 2013b).  The description of some metadata (of potentially high relevance) compiled in Chapter 2 was sometimes lacking accuracy or clarity for many models so that the metadata could not be used. Thus, one possible limiting factor to the development of EwE-based meta-analyses could be the varying degree to which models live up to best modeling practices (FAO 2008). Unless standards in modeling practices are defined and enforced, it will remain the modelers’  141 responsibility to provide detailed information on the objectives, quality and robustness of the models. Nevertheless, the development of EcoBase may encourage modelers to ensure that they give sufficient information to make their models reusable (Colléter et al. 2013b). For instance, in Chapter 2, I identified several metadata that should be systematically provided in the publications presenting the models: location, objectives, area, time period, units, software version, Pedigree index, trophic structure, and environmental conditions. Providing these metadata would facilitate the reuse of EwE models, and thus increase the scientific value of EwE-based publications. Moreover, methodologies allowing diagnostics of EwE models, such as PREBAL (Link 2010), which consists of a set of standard criteria to evaluate network models before the balancing step, could be applied as standard tools when building a model. The Ecopath Research and Development Consortium is working on promoting such best modeling practices within the EwE community (Steenbeek et al. 2014).  Chapter 3 and 4 focused on the concept of keystone species, a metaphorical and ecological term which illustrates the importance of MSTI for ecosystem structure and functioning. In Chapter 3, I analyzed the origin and evolution of the multiple definitions, terms, and examples associated to keystone species in the peer-reviewed literature. I demonstrated the expansion of the concept over time, which had become an over-used concept with a catch-all definition. Originally applied to predator species in aquatic food webs, the concept was later progressively applied to wide variety of species playing various roles in the food web (e.g., keystone prey, keystone competitor, keystone plant), and was subsequently applied at smaller or larger scales and detached from the notion of species (e.g., keystone habitats, keystone communities, keystone individuals) (Chapter 3). Several typologies were proposed in the literature, introducing ambiguity and confusion around the definition of keystone species (e.g., Mills et al. 1993, Bond 1994). In Chapter 3, I argued for a restored definition for keystone species, based on a historical review of the usages and misusages of the term. First, I placed the keystone species concept in a larger conceptual framework, describing different categories of species of potential ecological or managerial interest. Keystone species were notably compared to and differentiated from flagship species, ecosystem engineer species, and foundation species. Then, I proposed an operational definition for keystone species, based on restrictive and specific ecological criteria. I defined a keystone species as a species of high trophic level and low  142 biomass, which exerts a high and wide impact on the food web, relative to the other species in the community (Chapter 3). The restored definition is consistent with the original concept of ‘keystone predator’ (Paine 1966, 1969a, Paine 1969b). Moreover, while the keystone species concept has been described as a catch-all and so useless concept (e.g., Cottee-Jones and Whittaker 2012), the restored definition, based on precise and operational criteria, may help prevent from the possible demise of the concept.  Chapter 3 is based on a literature review, which was not exhaustive given that not every single publication mentioning the term ‘keystone’ was considered, but which was rather focusing on and taking as starting point previous detailed review articles (e.g., Bond 1994, Power et al. 1996, Menge and Freidenburg 2001). I selected the publications that proposed accurate (although sometimes contradictory) definitions and specific examples of the keystone species concept. Also, since the concept has been applied in all types of system (marine, freshwater and terrestrial), the proposed review was not restricted to marine applications. Nonetheless, the overall focus of my thesis being on marine systems, detailed considerations of the concept were slightly more focused on marine species.  Empirical validation of the definition proposed in Chapter 3 may still be debatable, since it is built on a conceptual framework and theoretical considerations (Cottee-Jones and Whittaker 2012). Nonetheless, the restored definition intended to solve the terminological confusion around the keystone species concept and other (more or less) related ecological concepts (Caro 2010). Thus, the proposed conceptual framework may serve to differentiate examples of misusages from appropriate applications of the keystone species terminology in the literature. Among the numerous examples labeled as keystone species in the literature, I selected a few seminal (the most frequently cited) examples of sensu stricto keystone species and pseudo keystone species (Chapter 3). Other case studies, not listed in Chapter 3, may be compared to the definitions established in the suggested framework and identified either as true keystone species or as another species category. Besides, the historical review I conducted in Chapter 3 illustrates the importance of consensus definition and operational criteria for the applicability and usefulness of ecological concepts such as keystone species. Some concepts describing other species categories of potential interest, considered in Chapter 3, were also recently reviewed and refined in the literature: wasp-waist species (Atkinson et al. 2014), ecosystem engineer species (Romero et al. 2014), and umbrella species (Branton and Richardson 2011).   143  In Chapter 4, I applied the restored definition proposed in Chapter 3 and proposed a new functional index of keystoneness (KS) to identify potential keystone species in marine food webs. The novel KS index is calculated by combining a component estimating trophic impact and a component measuring biomass proportion of the species, both derived from the Ecopath modeling approach (Chapter 4). I selected the proposed index from a suite of 12 alternative KS indices (including two indices directly adapted from the literature). The KS indices were tested by applying them to a collection of 101 published Ecopath models selected from EcoBase, and by confronting two statistical methods (rank correlation tests and a classification tree). I selected the ^	index, which obtained positive results with both methods for the majority of the models used in the meta-analysis (Chapter 4). Moreover, I demonstrated that the selected index performed better than existing KS indices in the literature (Power et al. 1996, Libralato et al. 2006) and gave results more consistent with the theoretical definition of keystone species (Chapter 3). Thus, the outcomes of Chapter 4 help improve our understanding of the importance of MSTI, as well as our modeling methods to study them.  Although keystone species is a much debated concept (e.g, Cottee-Jones and Whittaker 2012), it has been widely applied in the literature (e.g., Lewinsohn and Cagnolo 2012), and frequently implemented in recently published EwE applications (e.g., Eddy et al. 2014). The Keystoneness Analysis plug-in, consisting of the two existing KS indices proposed in the literature (Power et al. 1996, Libralato et al. 2006), was used in 11% (47 models) of the models compiled in EcoBase (Chapter 2). Indeed, the concept of keystone species is useful to managers since it potentially allows for the management of a single focal species with the aim of maintaining the whole ecosystem (Perry 2010). The new KS index proposed in Chapter 4 was expressed by following the same approach as in existing indices, so that it could be easily implemented in the EwE plug-in. Thus, the proposed KS index may be reused to inform marine biodiversity conservation, and may also be applied to freshwater and terrestrial systems. Chapter 4 showed that the contributions of the trophic impact and biomass components were better balanced in ^ than in the previous indices. Thus, the proposed KS index helps addressing some of the inconsistencies and weaknesses of existing indices that have been observed in the literature (e.g., Coll and Libralato 2012, Valls et al. 2012). In addition, the classification tree proposed in Chapter 4 may also be implemented in EwE and serve as a novel representation of  144 the modeled food web, through a continuum of species defined by their trophic interactions and abundance. Thus, potential keystone species could be quickly identified and compared to other species in the community, as well as abundant or rare species (along the biomass axis) and strong or weak interactors (along the trophic impact axis) in the food web. The benefits of using classification trees in ecological studies, compared to more traditional statistical techniques, notably include the ease of construction, exploration, and interpretation of complex ecological patterns and processes (De'ath and Fabricius 2000). Moreover, classification trees, as well as Spearman rank correlation tests, can deal with non-linear relationships (De'ath and Fabricius 2000).  Chapter 4 constitutes the first meta-analysis of Ecopath models based on metadata from EcoBase reused as selection criteria. Moreover, the application of the selected KS index to a large number of models, representing a wide variety of marine ecosystems, allows drawing some general conclusions. Notably, results from the ^	index suggested that toothed whales and sharks are the most common keystone species in marine food webs, which may be confirmed by empirical observations, such as for the Prince William Sound food web (Chapter 4). However, uncertainties in the structure (e.g., the aggregation of the species in functional groups) and outcomes of the Ecopath models may have an influence on the estimation of the KS index (Link et al. 2012). More precisely, the performance of highly aggregated models may be reduced, especially if the species playing critical ecological roles, such as keystone species, are over-aggregated (Fulton et al. 2003). Thus, further analyses should focus on the sensitivity of the KS index to the degree of trophic aggregation in the modeled food webs. A minimum number of functional groups as well as a maximum level of aggregation of high trophic level groups might be required for the KS index to be robust. In comparison, structural indices derived from Ecological Network Analyses and used to estimate keystoneness were demonstrated to be robust to uncertainty related to model construction errors (Fedor and Vasas 2009). Model uncertainty may also be analyzed in terms of sensitivity of the KS index to the estimates of biomass and diet composition of the species, since the index is calculated from these two input parameters. Ecopath models were previously demonstrated to be more sensitive to biomass than diet composition (Essington 2007). Besides, semi-quantitative approaches could be used to test the sensitivity of the KS index to MTI values (Rochette et al. 2009).  145 As briefly reviewed in Chapter 1, human activities represent strong drivers of change in the oceans and have significant and synergistic effects on MSTI. As a consequence, fishing and climate change may impact keystone species (e.g., Harley 2011, Eddy et al. 2014). As discussed in Chapters 3 and 4, context-dependency is one of the major challenges for the identification and management of critical species, such as keystone species. Thus, further analyses could focus on the sensitivity of the KS index to different levels of fishing pressure or changes in environmental conditions. More precisely, uncertainty in the outcomes of the analyses performed in Chapter 4 could be assessed. Using statistical methods, the statuses obtained (including the No Keystone status) and the keystone groups identified for each model could be correlated with the characteristics of the modeled ecosystems. Besides, contrary to previous studies, fishing fleets defined in exploited modeled ecosystems were included (roughly aggregated as one predator group) in the keystoneness analyses presented in Chapter 4. Thus, the category associated to the All Fleet group in the classification tree could potentially be correlated with the type, size, or level of exploitation of the modeled ecosystems. The relevance of integrating economic (e.g., fisheries) and ecological systems has been discussed in the literature (e.g., O'Neill and Kahn 2000).   Chapter 5 presented exploratory analyses of feeding selectivity in predatory fish species, which could allow for more formal predictions of MSTI. My intention with Chapter 5 was to build on existing large datasets, provided in the FishBase biodiversity information system, to identify predictors of diet composition (DC) in piscivorous fish species. Correlations between standardized data from stomach contents analyses and a suite of morphological, behavioral and ecological traits (both extracted from FishBase) were tested using multivariate analyses (Principal Components Analysis and Correspondence Analysis). Size and swimming abilities were identified as influencing factors on fish feeding selectivity in the literature (e.g., Labropoulou and Markakis 1998, Albouy et al. 2011). The outcomes of Chapter 5 confirmed the potential significance of these two parameters for determining prey utilization in predatory fish species. However, no clear or consistent pattern was observed, so that no correlation could be established, between the other species characteristics and DC with the selected datasets.  Diet composition is as a critical parameter when identifying critical species in the food web, such as keystone species, since it is used to define trophic interactions. Yet, as briefly  146 introduced in Chapter 1, there are multiple challenges in estimating accurate diet compositions, notably for fish species. The approach presented in this thesis illustrates the challenges related to diet composition data restrictions. Indeed, I chose to use FishBase as data source, with the aim of maximizing available data while ensuring a wide applicability and eased repeatability of the approach. Yet, despite using a large dataset available through FishBase, the taxonomic aggregation of diet information limited the use of available data (to the North Sea ecosystem in my case). To overcome data limitations, it has recently been proposed that frequency of occurrence (expressed in number) could be preferred to diet composition (expressed in volume or biomass), when detailed information is not required, so as to avoid the inherent problems with the collection and estimation of diet composition (Baker et al. 2014). As an alternative, the continuous integration of a growing number of stomach contents analyses in FishBase could help improving data availability. For instance, recent initiatives, such as the Gulf of Mexico Species Interaction (GoMexSI) database and webpage (Simons et al. 2013) and the Integrated Database and Portal for Fish Stomach Records (Dapstom, Pinnegar 2014), will certainly help making progress in the storage and dissemination of diet composition data, and could eventually be integrated into the global FishBase database. In addition, in terms of potential predictors considered in Chapter 5, the required completeness of the data reduced the dataset to a limited number of well-studied species only. Species traits of potential interest to my approach, which could not be used due to low availability or formatting restrictions in FishBase, included descriptors of fish body shape, eyes, mouth, teeth, gill rakers and scales, fin attributes, and morphometric measurements (Chapter 5). Some more flexible statistical methods could have been applied to allow for less restrictive requirements on these parameters. However, the validation of model predictions could have been challenged by natural variability in fish feeding (Link et al. 2012).   In conclusion, the EwE approach has been widely applied to develop marine ecosystem models worldwide. Each of these models provides insights for a better understanding of MSTI, and EwE-based meta-analysis help improve our understanding even further. However, EwE models are data-driven and require detailed data on diet composition. Thus, data availability may be a critical issue when considering some aspects of trophic interactions, especially for modeling and predictions at the species level. That is why open-access sources and data sharing initiatives  147 within the scientific community, notably through the use of online digital information repositories, are critical. Standardized and discoverable data and models, collected and stored in open-access repositories, are needed for the development of global meta-analyses in marine ecology.    148 Bibliography Ahrens, R. N., C. J. Walters, and V. Christensen. 2012. Foraging arena theory. Fish and fisheries 13:41-59. Albouy, C., F. Guilhaumon, S. Villeger, M. Mouchet, L. Mercier, J. M. Culioli, J. A. Tomasini, F. Le Loc'h, and D. Mouillot. 2011. Predicting trophic guild and diet overlap from functional traits: statistics, opportunities and limitations for marine ecology. Marine Ecology Progress Series 436:17-28. Appeltans, W., S. T. Ahyong, G. Anderson, M. V. Angel, T. Artois, N. Bailly, R. Bamber, A. Barber, I. Bartsch, and A. Berta. 2012. The magnitude of global marine species diversity. Current Biology 22:2189-2202. Armstrong, D., and T. Caro. 2002. Focal and surrogate species: getting the language right. Conservation Biology 16:285-287. Arntz, W. E. 1978. The food of adult cod (Gadus morhua L.) in the Western Baltic. Meeresforsch 26:60-69. Arponen, A. 2012. Prioritizing species for conservation planning. Biodiversity and Conservation 21:875-893. Arreguín-Sánchez, F. 2011. Ecosystem dynamics under “top-down” and “bottom-up” control situations generated by intensive harvesting rates Hidrobiológica 21:323-332. Atkinson, A., S. L. Hill, M. Barange, E. A. Pakhomov, D. Raubenheimer, K. Schmidt, S. J. Simpson, and C. Reiss. 2014. Sardine cycles, krill declines, and locust plagues: revisiting ‘wasp-waist’ food webs. Trends in Ecology and Evolution 29:309-316. Aydin, K. Y., V. Lapko, V. Radchenko, and P. Livingston. 2002. A comparison of the eastern Bering and western Bering Sea shelf and slope ecosystems through the use of mass-balance food web models. US Department of Commerce, National Oceanic and Atmospheric Administration, National Marine Fisheries Service, Alaska Fisheries Science Center. Babcock, R. C., S. Kelly, N. T. Shears, J. W. Walker, and T. J. Willis. 1999. Changes in community structure in temperate marine reserves. Marine Ecology Progress Series 189:125-134. Baker, R., A. Buckland, and M. Sheaves. 2014. Fish gut content analysis: robust measures of diet composition. Fish and fisheries 15:170-177. Barnett, A., D. R. Bellwood, and A. S. Hoey. 2006. Trophic ecomorphology of cardinalfish. Marine Ecology Progress Series 322:249-257. Barua, M. 2011. Mobilizing metaphors: the popular use of keystone, flagship and umbrella species concepts. Biodiversity and Conservation 20:1427-1440. Berg, S., M. Christianou, T. Jonsson, and B. Ebenman. 2011. Using sensitivity analysis to identify keystone species and keystone links in size-based food webs. Oikos 120:510-519. Berlow, E. L. 1999. Strong effects of weak interactions in ecological communities. Nature 398:330-334. Berlow, E. L., A. M. Neutel, J. E. Cohen, P. C. De Ruiter, B. Ebenman, M. Emmerson, J. W. Fox, V. A. Jansen, J. Iwan Jones, and G. D. Kokkoris. 2004. Interaction strengths in food webs: issues and opportunities. Journal of Animal Ecology 73:585-598.  149 Beschta, R. L. 2003. Cottonwoods, elk, and wolves in the Lamar Valley of Yellowstone National Park. Ecological Applications 13:1295-1309. Bond, W. 1994. Keystone species. Pages 237-253 in E.-D. Schulze and H. A. Mooney, editors. Biodiversity and ecosystem function. Springer. Bond, W. 2001. Keystone Species--Hunting the Snark? Science 292:63-64. Booth, S., and D. Zeller. 2005. Mercury, food webs, and marine mammals: implications of diet and climate change for human health. Environmental Health Perspectives 113:521-526. Bossart, G. D. 2006. Marine mammals as sentinel species for oceans and human health. Oceanography 19:134-137. Boucher, D. H., S. James, and K. H. Keeler. 1982. The ecology of mutualism. Annual Review of Ecology and Systematics 13:315-347. Boyle, K. S., and M. H. Horn. 2006. Comparison of feeding guild structure and ecomorphology of intertidal fish assemblages from central California and central Chile. Marine Ecology Progress Series 319:65-84. Bracken, M. E., B. E. Bracken, and L. D. Rogers-Bennett. 2007. Species diversity and foundation species: Potential indicators of fisheries yields and marine ecosystem functioning. CalCOFI Report, Vol. 48, California Cooperative Oceanic Fisheries Investigations. Branton, M., and J. S. Richardson. 2011. Assessing the Value of the Umbrella-Species Concept for Conservation Planning with Meta-Analysis. Conservation Biology 25:9-20. Brown, J. H., and E. J. Heske. 1990. Control of a desert-grassland transition by a keystone rodent guild. Science 250:1705-1707. Butchart, S. H., M. Walpole, B. Collen, A. van Strien, J. P. Scharlemann, R. E. Almond, J. E. Baillie, B. Bomhard, C. Brown, and J. Bruno. 2010. Global biodiversity: indicators of recent declines. Science 328:1164-1168. Butler, J. R., A. Tawake, T. Skewes, L. Tawake, and V. McGrath. 2012. Integrating traditional ecological knowledge and fisheries management in the Torres Strait, Australia: the catalytic role of turtles and dugong as cultural keystone species. Ecology and Society 17:1-19. Cardinale, B. J., J. E. Duffy, A. Gonzalez, D. U. Hooper, C. Perrings, P. Venail, A. Narwani, G. M. Mace, D. Tilman, and D. A. Wardle. 2012. Biodiversity loss and its impact on humanity. Nature 486:59-67. Carlton, J. T., J. B. Geller, M. L. Reaka-Kudla, and E. A. Norse. 1999. Historical extinctions in the sea. Annual Review of Ecology and Systematics 30:515-538. Caro, T. 2003. Umbrella species: critique and lessons from East Africa. Animal Conservation 6:171-181. Caro, T. 2010. Conservation by Proxy: Indicator, Umbrella, Keystone, Flagship, and Other Surrogate Species. Island Press. Caro, T., A. Engilis, E. Fitzherbert, and T. Gardner. 2004. Preliminary assessment of the flagship species concept at a small scale. Animal Conservation 7:63-70. Caro, T. M., and G. O'Doherty. 1999. On the use of surrogate species in conservation biology. Conservation Biology 13:805-814. Casey, J. M., and R. A. Myers. 1998. Near extinction of a large, widely distributed fish. Science 281:690-692.  150 Cheung, W. W., J. L. Sarmiento, J. Dunne, T. L. Frölicher, V. W. Lam, M. D. Palomares, R. Watson, and D. Pauly. 2012. Shrinking of fishes exacerbates impacts of global ocean changes on marine ecosystems. Nature Climate Change 3:254-258. Cheung, W. W. L., C. Close, V. W. Y. Lam, R. Watson, and D. Pauly. 2008. Application of macroecological theory to predict effects of climate change on global fisheries potential. Marine Ecology Progress Series 365:187-197. Cheung, W. W. L., J. Dunne, J. L. Sarmiento, and D. Pauly. 2011. Integrating ecophysiology and plankton dynamics into projected changes in maximum fisheries catch potential under climate change in the Northeast Atlantic. ICES Journal of Marine Science 68:1008-1018. Cheung, W. W. L., V. W. Y. Lam, J. L. Sarmiento, K. Kearney, R. Watson, and D. Pauly. 2009. Projecting global marine biodiversity impacts under climate change scenarios. Fish and fisheries 10:235-251. Cheung, W. W. L., V. W. Y. Lam, J. L. Sarmiento, K. Kearney, R. Watson, D. Zeller, and D. Pauly. 2010. Large-scale redistribution of maximum fisheries catch potential in the global ocean under climate change. Global Change Biology 16:24-35. Christensen, V. 1995a. Ecosystem maturity - towards quantification. Ecological Modelling 77:3-32. Christensen, V. 1995b. A Model of Throphic Interactions in the North Sea in 1981, the Year of the Stomach. Dana 11:1-28. Christensen, V. 1996. Managing fisheries involving predator and prey species. Reviews in Fish Biology and Fisheries 6:417-442. Christensen, V., M. Coll, C. Piroddi, J. Steenbeek, J. Buszowski, and D. Pauly. 2014a. A century of fish biomass decline in the ocean. Marine Ecology Progress Series 512:155-166. Christensen, V., M. Coll, J. Steenbeek, J. Buszowski, D. Chagaris, and C. J. Walters. 2014b. Representing Variable Habitat Quality in a Spatial Food Web Model. Ecosystems 17:1397-1412. Christensen, V., L. Garces, G. Silvestre, and D. Pauly. 2003a. Fisheries Impact on the South China Sea Large Marine Ecosystem: A Preliminary Analysis using Spatially-Explicit Methodology. Pages 51-62 in Assessment, management and future directions for coastal fisheries in Asian countries. WorldFish Center Conference Proceedings. Christensen, V., S. Guenette, J. J. Heymans, C. J. Walters, R. Watson, D. Zeller, and D. Pauly. 2003b. Hundred year decline of North Atlantic predatory fishes. Fish and fisheries 4:1-24. Christensen, V., and D. Pauly. 1992. The ECOPATH II - a software for balancing steady-state ecosystem models and calculating network characteristics. Ecological Modelling 61:169-185. Christensen, V., and D. Pauly. 1993a. Flow characteristics of aquatic ecosystems. Pages 338-352 in V. Christensen and D. Pauly, editors. Trophic models of aquatic ecosystems. ICLARM Conference Proceedings, Manila, Philippines. Christensen, V., and D. Pauly. 1993b. Trophic models of aquatic ecosystems. ICLARM Conference Proceedings, Manila, Philippines. Christensen, V., C. Piroddi, M. Coll, J. Steenbeek, J. Buszowski, and P. Daniel. 2011. Fish biomass in the world ocean: a century of decline. Fisheries Centre Working Paper Series, Fisheries Center, The University of British Columbia, Vancouver. Christensen, V., and C. Walters. 2004. Ecopath with Ecosim: methods, capabilities and limitations. Ecological Modelling 172:109-139.  151 Christensen, V., and C. Walters. 2005. Using ecosystem modeling for fisheries management: where are we? Pages 1-16 in ICES Annual Science Conference, Aberdeen, Scotland. Christensen, V., C. Walters, R. Ahrens, J. Alder, J. Buszowski, L. B. Christensen, W. W. L. Cheung, J. Dunne, F. Froese, V. Karpouzi, K. Kaschner, K. Kearney, S. Lai, V. W. Y. Lam, M. L. D. Palomares, R. Peters-Mason, C. Piroddi, J. L. Sarmiento, J. Steenbeek, U. R. Sumaila, R. Watson, D. Zeller, and D. Pauly. 2009. Database-driven models of the world's Large Marine Ecosystems. Ecological Modelling 220:1984-1996. Christensen, V., C. Walters, and D. Pauly. 2005. Ecopath with Ecosim: a user's guide. November 2005 Edition. Fisheries Centre, University of British Columbia, Vancouver, BC, Canada. Christensen, V., C. Walters, D. Pauly, and R. Forrest. 2008. Ecopath with Ecosim version 6 User Guide. Lenfest Ocean Futures Project. Christensen, V., and C. J. Walters. 2011. Progress in the use of ecosystem modeling for fisheries management. Pages 189-205 in V. Christensen and J. L. Maclean, editors. Ecosystem Approaches to Fisheries: a Global Perspective. Cambridge University Press, Cambridge. Christianou, M., and B. Ebenman. 2005. Keystone species and vulnerable species in ecological communities: strong or weak interactors? Journal of Theoretical Biology 235:95-103. Clemente, S., J. C. Hernández, A. Rodríguez, and A. Brito. 2010. Identifying keystone predators and the importance of preserving functional diversity in sublittoral rocky-bottom areas. Marine Ecology Progress Series 413:55-67. Clements, F. E. 1936. Nature and structure of the climax. Journal of ecology 24:252-284. Clifton, K. B., and P. J. Motta. 1998. Feeding morphology, diet, and ecomorphological relationships among five Caribbean labrids (Teleostei, Labridae). Copeia 1998:953-966. Cohen, J. E., S. L. Pimm, P. Yodzis, and J. Saldaña. 1993. Body sizes of animal predators and animal prey in food webs. Journal of Animal Ecology 62:67-78. Coleman, J. S. M. 2010. Spearman Rank Order Correlation. Pages 1405-1408 in N. J. Salkind, editor. Encyclopedia of Research Design. SAGE Publications, Inc., Thousand Oaks. Coll, M., A. Bundy, and L. J. Shannon. 2008. Ecosystem modelling using the Ecopath with Ecosim approach. Pages 225-291 in B. Megrey and E. Moksness, editors. Computers in Fisheries Research, Second edition. Springer. Coll, M., and S. Libralato. 2012. Contributions of food web modelling to the ecosystem approach to marine resource management in the Mediterranean Sea. Fish and fisheries 13:60-88. Coll, M., J. Navarro, R. J. Olson, and V. Christensen. 2012. Assessing the trophic position and ecological role of squids in marine ecosystems by means of food-web models. Deep Sea Research Part II: Topical Studies in Oceanography 95:21-36. Coll, M., J. Navarro, and I. Palomera. 2013. Ecological role, fishing impact, and management options for the recovery of a Mediterranean endemic skate by means of food web models. Biological Conservation 157:108-120. Coll, M., A. Santojanni, I. Palomera, and E. Arneri. 2009. Food-web changes in the Adriatic Sea over the last three decades. Marine Ecology Progress Series 381:17-37. Coll, M., A. Santojanni, I. Palomera, S. Tudela, and E. Arneri. 2007. An ecological model of the Northern and Central Adriatic Sea: Analysis of ecosystem structure and fishing impacts. Journal of Marine Systems 67:119-154. Coll, M., and J. Steenbeek. 2014. New software plug-in to calculate biodiversity and conservation-based indicators from EwE food web models. Pages 68-69 in J. Steenbeek, C. Piroddi, M. Coll, J. J. Heymans, S. Villasante, and V. Christensen, editors. Ecopath 30  152 Years Conference Proceedings: Extended Abstracts. Fisheries Centre Research Reports. Fisheries Centre, University of British Columbia, Vancouver, B.C., Canada. Colléter, M., J. Guitton, and D. Gascuel. 2013a. An Introduction to the EcoTroph R Package: Analyzing Aquatic Ecosystem Trophic Networks. R JOURNAL 5:98-107. Colléter, M., A. Valls, J. Guitton, L. Morissette, F. Arreguín-Sánchez, V. Christensen, D. Gascuel, and D. Pauly. 2013b. EcoBase: a repository solution to gather and communicate information from EwE models. University of British Columbia, Vancouver, B.C., Canada. Conover, D. O., and S. B. Munch. 2002. Sustaining fisheries yields over evolutionary time scales. Science 297:94-96. Coombs, A. P. 2004. Marine mammals and human health in the Eastern Bering Sea: Using an ecosystem-based food web model to track PCBs. MSc thesis. University of British Columbia, Vancouver. Cosentino, B. J., R. L. Schooley, B. T. Bestelmeyer, J. F. Kelly, and J. M. Coffman. 2014. Constraints and time lags for recovery of a keystone species (Dipodomys spectabilis) after landscape restoration. Landscape Ecology 29:665-675. Costello, C., D. Ovando, R. Hilborn, S. D. Gaines, O. Deschenes, and S. E. Lester. 2012. Status and solutions for the world’s unassessed fisheries. Science 338:517-520. Costello, M. J., M. Coll, R. Danovaro, P. Halpin, H. Ojaveer, and P. Miloslavich. 2010. A census of marine biodiversity knowledge, resources, and future challenges. PLoS ONE 5:e12110. Costello, M. J., R. M. May, and N. E. Stork. 2013. Can we name Earth's species before they go extinct? Science 339:413-416. Cottee-Jones, H. E. W., and R. J. Whittaker. 2012. The keystone species concept: a critical appraisal. Frontiers of Biogeography 4:117-127. Cristancho, S., and J. Vining. 2004. Culturally defined keystone species. Human Ecology Review 11:153-164. Crutzen, P. J., and W. Steffen. 2003. How long have we been in the Anthropocene era? Climatic Change 61:251-257. Cury, P., A. Bakun, R. J. Crawford, A. Jarre, R. A. Quiñones, L. J. Shannon, and H. M. Verheye. 2000. Small pelagics in upwelling systems: patterns of interaction and structural changes in “wasp-waist” ecosystems. ICES Journal of Marine Science 57:603-618. Cushing, D. 1990. Plankton production and year-class strength in fish populations: an update of the match/mismatch hypothesis. Advances in marine biology 26:249-293. Daily, G. C., P. R. Ehrlich, and N. M. Haddad. 1993. Double keystone bird in a keystone species complex. Proceedings of the National Academy of Sciences 90:592-594. Dalgleish, R., E. Molero, R. Kidd, M. Jansen, D. Past, A. Robl, B. Mons, C. Diaz, A. Mons, and A. J. Brookes. 2012. Solving bottlenecks in data sharing in the life sciences. Human Mutation 33:1494-1496. Daskalov, G. M., A. N. Grishin, S. Rodionov, and V. Mihneva. 2007. Trophic cascades triggered by overfishing reveal possible mechanisms of ecosystem regime shifts. Proceedings of the National Academy of Sciences 104:10518-10523. Davic, R. D. 2000. Ecological dominants vs. keystone species: a call for reason. Conservation Ecology 4:r2. Davic, R. D. 2002. Herbivores as keystone predators. Conservation Ecology 6:r8.  153 Davic, R. D. 2003. Linking keystone species and functional groups: a new operational definition of the keystone species concept. Conservation Ecology 7:r11. Davic, R. D. 2004. Epistemology, culture, and keystone species. Ecology and Society 9:r1. Davidar, P., K. Yoganand, and T. Ganesh. 2001. Distribution of forest birds in the Andaman islands: importance of key habitats. Journal of Biogeography 28:663-671. Davidson, A. D., J. K. Detling, and J. H. Brown. 2012. Ecological roles and conservation challenges of social, burrowing, herbivorous mammals in the world's grasslands. Frontiers in Ecology and the Environment 10:477-486. Dawson, T. P., S. T. Jackson, J. I. House, I. C. Prentice, and G. M. Mace. 2011. Beyond predictions: biodiversity conservation in a changing climate. Science 332:53-58. Dayton, P. K. 1972. Toward an understanding of community resilience and the potential effects of enrichments to the benthos at McMurdo Sound, Antarctica. Pages 81-96 in Proceedings of the colloquium on conservation problems in Antarctica. Allen Press Lawrence, Kansas, USA. De'ath, G., and K. E. Fabricius. 2000. Classification and regression trees: A powerful yet simple technique for ecological data analysis. Ecology 81:3178-3192. De Leo, G. A., and S. Levin. 1997. The multifaceted aspects of ecosystem integrity. Conservation Ecology 1:3. Doney, S. C. 2010. The growing human footprint on coastal and open-ocean biogeochemistry. Science 328:1512-1516. Doney, S. C., M. Ruckelshaus, J. E. Duffy, J. P. Barry, F. Chan, C. A. English, H. M. Galindo, J. M. Grebmeier, A. B. Hollowed, and N. Knowlton. 2012. Climate change impacts on marine ecosystems. Annual Review of Marine Science 4:11-37. Duffy, J. E., B. J. Cardinale, K. E. France, P. B. McIntyre, E. Thebault, and M. Loreau. 2007. The functional role of biodiversity in ecosystems: incorporating trophic complexity. Ecology Letters 10:522-538. Dulvy, N. K., Y. Sadovy, and J. D. Reynolds. 2003. Extinction vulnerability in marine populations. Fish and fisheries 4:25-64. Duran, L., and J. C. Castilla. 1989. Variation and persistence of the middle rocky intertidal community of central Chile, with and without human harvesting. Marine Biology 103:555-562. Durant, J. M., D. Ø. Hjermann, G. Ottersen, and N. C. Stenseth. 2007. Climate and the match or mismatch between predator requirements and resource availability. Climate Research 33:271-283. Ebenman, B., and T. Jonsson. 2005. Using community viability analysis to identify fragile systems and keystone species. Trends in Ecology and Evolution 20:568-575. Eddy, T. D., T. J. Pitcher, A. B. MacDiarmid, T. T. Byfield, J. C. Tam, T. T. Jones, J. J. Bell, and J. Gardner. 2014. Lobsters as keystone: Only in unfished ecosystems? Ecological Modelling 275:48-72. Edwards, M., and A. J. Richardson. 2004. Impact of climate change on marine pelagic phenology and trophic mismatch. Nature 430:881-884. Edwards, P. N. 2010. A vast machine: Computer models, climate data, and the politics of global warming. Mit Press. Ellison, A. M., M. S. Bank, B. D. Clinton, E. A. Colburn, K. Elliott, C. R. Ford, D. R. Foster, B. D. Kloeppel, and J. D. Knoepp. 2005. Loss of foundation species: consequences for the structure and dynamics of forested ecosystems. Frontiers in Ecology 3:479-486.  154 Elton, C. S. 1927. The animal community. Animal ecology. Sidgwick & Jackson, London. Espinoza-Tenorio, A., M. Wolff, M. H. Taylor, and I. Espejel. 2012. What model suits ecosystem-based fisheries management? A plea for a structured modeling process. Reviews in Fish Biology and Fisheries 22:81-94. Essington, T. E. 2007. Evaluating the sensitivity of a trophic mass-balance model (Ecopath) to imprecise data inputs. Canadian Journal of Fisheries and Aquatic Sciences 64:628-637. Estes, J. A., and J. F. Palmisano. 1974. Sea otters: their role in structuring nearshore communities. Science 185:1058-1060. Estes, J. A., J. Terborgh, J. S. Brashares, M. E. Power, J. Berger, W. J. Bond, S. R. Carpenter, T. E. Essington, R. D. Holt, and J. B. Jackson. 2011. Trophic downgrading of planet Earth. Science 333:301-306. Estes, J. A., M. T. Tinker, T. M. Williams, and D. F. Doak. 1998. Killer whale predation on sea otters linking oceanic and nearshore ecosystems. Science 282:473-476. Estes, J. E., N. S. Smith, and J. F. Palmisano. 1978. Sea otter predation and community organization in the western Aleutian Islands, Alaska. Ecology 59:822-833. Estrada, E. 2007. Characterization of topological keystone species: local, global and “meso-scale” centralities in food webs. Ecological Complexity 4:48-57. Fabry, V. J., B. A. Seibel, R. A. Feely, and J. C. Orr. 2008. Impacts of ocean acidification on marine fauna and ecosystem processes. ICES Journal of Marine Science 65:414-432. FAO. 2008. Fisheries management. 2. The ecosystem approach to fisheries. 2.1 Best practices in ecosystem modelling for informing an ecosystem approach to fisheries. FAO, Rome. FAO. 2014. The State of World Fisheries and Aquaculture. FAO, Rome. Favreau, J. M., C. A. Drew, G. R. Hess, M. J. Rubino, F. H. Koch, and K. A. Eschelbach. 2006. Recommendations for assessing the effectiveness of surrogate species approaches. Biodiversity and Conservation 15:3949-3969. Fedor, A., and V. Vasas. 2009. The robustness of keystone indices in food webs. Journal of Theoretical Biology 260:372-378. Floeter, J., and A. Temming. 2003. Explaining diet composition of North Sea cod (Gadus morhua): prey size preference vs. prey availability. Canadian Journal of Fisheries and Aquatic Sciences 60:140-150. Floeter, J., and A. Temming. 2005. Analysis of prey size preference of North Sea whiting, saithe, and grey gurnard. ICES Journal of Marine Science 62:897-907. Folke, C., S. Carpenter, B. Walker, M. Scheffer, T. Elmqvist, L. Gunderson, and C. Holling. 2004. Regime shifts, resilience, and biodiversity in ecosystem management. Annual Review of Ecology, Evolution, and Systematics 35:557-581. Francis, R. C., M. A. Hixon, M. E. Clarke, S. A. Murawski, and S. Ralston. 2007. Ten commandments for ecosystem-based fisheries scientists. Fisheries 32:217-233. Fretzer, S. 2014. Implementing the habitats directive in Germany: Case conventions versus Ecopath, Ecosim and Ecospace. Pages 64-65 in J. Steenbeek, C. Piroddi, M. Coll, J. J. Heymans, S. Villasante, and V. Christensen, editors. Ecopath 30 Years Conference Proceedings: Extended Abstracts. Fisheries Centre Research Reports. Fisheries Centre, University of British Columbia, Vancouver, B.C., Canada. Froese, R. 2011. The science in FishBase. Pages 47-52 in V. Christensen and J. L. Maclean, editors. Ecosystem Approaches to Fisheries: a Global Perspective. Cambridge University Press, Cambridge.  155 Froese, R., and D. Pauly. 2000. FishBase 2000: Concepts, designs and data sources. ICLARM, Los Banos, Laguna, Philippines. Fulton, E. A. 2010. Approaches to end-to-end ecosystem models. Journal of Marine Systems 81:171-183. Fulton, E. A., J. S. Link, I. C. Kaplan, M. SavinawRolland, P. Johnson, C. Ainsworth, P. Horne, R. Gorton, R. J. Gamble, and A. D. Smith. 2011. Lessons in modelling and management of marine ecosystems: the Atlantis experience. Fish and fisheries 12:171-188. Fulton, E. A., A. D. Smith, and C. R. Johnson. 2003. Effect of complexity on marine ecosystem models. Marine Ecology Progress Series 253:1-16. Garibaldi, A., and N. Turner. 2004. Cultural keystone species: implications for ecological conservation and restoration. Ecology and Society 9:1. Gasche, L., and D. Gascuel. 2013. EcoTroph: a simple model to assess fishery interactions and their impacts on ecosystems. ICES Journal of Marine Science 70:498-510. Gascuel, D. 2005. The trophic-level based model: A theoretical approach of fishing effects on marine ecosystems. Ecological Modelling 189:315-332. Gascuel, D., S. Guénette, and D. Pauly. 2011. The trophic-level based ecosystem modelling approach: Theoretical overview and practical uses. ICES Journal of Marine Science 68:1403-1416. Gascuel, D., L. Morissette, M. L. D. Palomares, and V. Christensen. 2008. Trophic flow kinetics in marine ecosystems: Toward a theorical approach to ecosystem functioning. Ecological Modelling 217:33-47. Gascuel, D., and D. Pauly. 2009. EcoTroph: Modelling marine ecosystem functioning and impact of fishing. Ecological Modelling 220:2885-2898. Gerking, S. D. 1994. Feeding ecology of fish. Academic Press, New York. Gilbert, L. E. 1980. Food web organization and conservation of neotropical diversity. Pages 11-33 in M. Soule and B. Wilcox, editors. Conservation biology. An evolutionary-ecological perspective. Sinauer, Sunderland, MA. Gill, A. B. 2003. The dynamics of prey choice in fish: the importance of prey size and satiation. Journal of Fish Biology 63 (Supplement A):105-116. Gotelli, N. J., and R. K. Colwell. 2001. Quantifying biodiversity: procedures and pitfalls in the measurement and comparison of species richness. Ecology Letters 4:379-391. Gray, J. S. 1997. Marine biodiversity: patterns, threats and conservation needs. Biodiversity and Conservation 6:153-175. Grimm, V., and S. F. Railsback. 2005. Individual-based modeling and ecology. Princeton university press. Grossman, G. D. 1986. Food resource partitioning in a rocky intertidal fish assemblage. Journal of Zoology 1:317-355. Gruber, N. 2011. Warming up, turning sour, losing breath: ocean biogeochemistry under global change. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 369:1980-1996. Hall, M. A., D. L. Alverson, and K. I. Metuzals. 2000. By-catch: problems and solutions. Marine Pollution Bulletin 41:204-219. Halpern, B. S., S. Walbridge, K. A. Selkoe, C. V. Kappel, F. Micheli, C. D'Agrosa, J. F. Bruno, K. S. Casey, C. Ebert, H. E. Fox, R. Fujita, D. Heinemann, H. S. Lenihan, E. M. P. Madin, M. T. Perry, E. R. Selig, M. Spalding, R. Steneck, and R. Watson. 2008. A global map of human impact on marine ecosystems. Science 319:948-952.  156 Hampton, S. E., C. A. Strasser, J. J. Tewksbury, W. K. Gram, A. E. Budden, A. L. Batcheller, C. S. Duke, and J. H. Porter. 2013. Big data and the future of ecology. Frontiers in Ecology and the Environment 11:156-162. Hardin, G. 1957. The threat of clarity. American Journal of Psychiatry 114:392-396. Harley, C. D. 2011. Climate change, keystone predation, and biodiversity loss. Science 334:1124-1127. Harley, C. D. G., A. R. Hughes, K. M. Hultgren, B. G. Miner, C. J. B. Sorte, C. S. Thornber, L. F. Rodriguez, L. Tomanek, and S. L. Williams. 2006. The impacts of climate change in coastal marine systems. Ecology Letters 9:228-241. Hay, M. E., J. D. Parker, D. E. Burkepile, C. C. Caudill, A. E. Wilson, Z. P. Hallinan, and A. D. Chequer. 2004. Mutualisms and aquatic community structure: the enemy of my enemy is my friend. Annual Review of Ecology, Evolution, and Systematics 35:175-197. Heymans, J. J., M. Coll, S. Libralato, and V. Christensen. 2012. Ecopath Theory, Modeling, and Application to Coastal Ecosystems. Pages 93-113 in E. Wolanski and D. McLusky, editors. Treatise on Estuarine and Coastal Science. Academic Press. Heymans, J. J., M. Coll, S. Libralato, L. Morissette, and V. Christensen. 2014. Global Patterns in Ecological Indicators of Marine Food Webs: A Modelling Approach. PLoS ONE 9:e95845. Heywood, V. H. 1995. Global biodiversity assessment. Cambridge University Press. Higdon, J. W. 2002. Functionally dominant herbivores as keystone species. Conservation Ecology 6:r4. Hoegh-Guldberg, O., and J. F. Bruno. 2010. The impact of climate change on the world’s marine ecosystems. Science 328:1523-1528. Hoegh-Guldberg, O., P. Mumby, A. Hooten, R. Steneck, P. Greenfield, E. Gomez, C. Harvell, P. Sale, A. Edwards, and K. Caldeira. 2007. Coral reefs under rapid climate change and ocean acidification. Science 318:1737-1742. Hollowed, A. B., N. Bax, R. Beamish, J. Collie, M. Fogarty, P. Livingston, J. Pope, and J. C. Rice. 2000. Are multispecies models an improvement on single-species models for measuring fishing impacts on marine ecosystems? ICES Journal of Marine Science 57:707-719. Holt, R. D. 1977. Predation, apparent competition, and the structure of prey communities. Theoretical population biology 12:197-229. Hooper, D. U., E. C. Adair, B. J. Cardinale, J. E. Byrnes, B. A. Hungate, K. L. Matulich, A. Gonzalez, J. E. Duffy, L. Gamfeldt, and M. I. O’Connor. 2012. A global synthesis reveals biodiversity loss as a major driver of ecosystem change. Nature 486:105-108. Hunsicker, M. E., L. Ciannelli, K. M. Bailey, J. A. Buckel, J. Wilson White, J. S. Link, T. E. Essington, S. Gaichas, T. W. Anderson, and R. D. Brodeur. 2011. Functional responses and scaling in predator–prey interactions of marine fishes: contemporary issues and emerging concepts. Ecology Letters 14:1288-1299. Hurlbert, S. 1997. Functional importance vs keystoneness: reformulating some questions in theoretical biocenology. Australian Journal of Ecology 22:369-382. Ibañez, C., P. A. Tedesco, R. Bigorne, B. Hugueny, M. Pouilly, C. Zepita, J. Zubieta, and T. Oberdorff. 2007. Dietary-morphological relationships in fish assemblages of small forested streams in the Bolivian Amazon. Aquatic Living Resources 20:131-142. ICES. 2010. Stomach Dataset. ICES-CIEM Stomach Data (1981-2010). Copenhagen. Available online at http://ecosystemdata.ices.dk/stomachdata/index.aspx. Consulted on 2012-10-22.  157 Jackson, J. B. 2008. Ecological extinction and evolution in the brave new ocean. Proceedings of the National Academy of Sciences 105:11458-11465. Jackson, J. B. C., M. X. Kirby, W. H. Berger, K. A. Bjorndal, L. W. Bostford, B. J. Bourque, R. H. Bradbury, R. Cooke, J. Erlandson, J. A. Estes, T. P. Hughes, S. Kidwell, C. B. Lange, H. S. Lenihan, J. M. Pandolfi, C. H. Peterson, R. S. Steneck, M. J. Tegner, and R. R. Warner. 2001. Historical overfishing and the recent collapse of coastal ecosystems. Science 293:629-637. Jarre-Teichmann, A. 1998. The potential role of mass balance models for the management of upwelling ecosystems. Ecological Applications 8:S93-S103. Jennings, S., F. Mélin, J. L. Blanchard, R. M. Forster, N. K. Dulvy, and R. W. Wilson. 2008. Global-scale predictions of community and ecosystem properties from simple ecological theory. Proceedings of the Royal Society B: Biological Sciences 275:1375-1383. Jones, C. G., J. H. Lawton, and M. Shachak. 1994. Organisms as ecosystem engineers. Oikos 69:373-386. Jones, C. G., J. H. Lawton, and M. Shachak. 1997. Positive and negative effects of organisms as physical ecosystem engineers. Ecology 78:1946-1957. Jordán, F. 2009. Keystone species and food webs. Philosophical Transactions of the Royal Society B: Biological Sciences 364:1733-1741. Jordán, F., Z. Benedek, and J. Podani. 2007. Quantifying positional importance in food webs: a comparison of centrality indices. Ecological Modelling 205:270-275. Jordán, F., W.-c. Liu, and Á. Mike. 2009. Trophic field overlap: a new approach to quantify keystone species. Ecological Modelling 220:2899-2907. Jordán, F., W.-C. Liu, and J. van Veen. 2003. Quantifying the importance of species and their interactions in a host-parasitoid community. Community Ecology 4:79-88. Jordán, F., T. A. Okey, B. Bauer, and S. Libralato. 2008. Identifying important species: linking structure and function in ecological networks. Ecological Modelling 216:75-80. Jordán, F., A. Takács-Sánta, and I. Molnár. 1999. A reliability theoretical quest for keystones. Oikos 86:453-462. Kaiser, M. J., J. S. Collie, S. J. Hall, S. Jennings, and I. R. Poiner. 2002. Modification of marine habitats by trawling activities: prognosis and solutions. Fish and fisheries 3:114-136. Keeling, R. F., A. Körtzinger, and N. Gruber. 2010. Ocean deoxygenation in a warming world. Annual Review of Marine Science 2:199-229. Kelling, S., W. M. Hochachka, D. Fink, M. Riedewald, R. Caruana, G. Ballard, and G. Hooker. 2009. Data Intensive Science: A New Paradigm for Biodiversity Studies. BioScience 59:613-620. Khanina, L. 1998. Determining keystone species. Conservation Ecology 2:R2. Kieffer, J. D., and P. W. Colgan. 1992. The role of learning in fish behaviour. Reviews in Fish Biology and Fisheries 2:125-143. Kitchell, J. F., T. E. Essington, C. H. Boggs, D. E. Schindler, and C. J. Walters. 2002. The role of sharks and longline fisheries in a pelagic ecosystem of the central Pacific. Ecosystems 5:202-216. Knowlton, N. 2004. Multiple "stable" states and the conservation of marine ecosystems. Progress in Oceanography 60:387-396. Labropoulou, M., and G. Markakis. 1998. Morphological-dietary relationships within two assemblages of marine demersal fishes. Environmental Biology of Fishes 51:309-319.  158 Laevastu, T., and H. A. Larkins. 1981. Marine fisheries ecosystem: its quantitative evaluation and management. Fishing News Books, Farnham, Great Britain. Lambeck, R. J. 1997. Focal species: a multi-species umbrella for nature conservation. Conservation Biology 11:849-856. Landres, P. B., J. Verner, and J. W. Thomas. 1988. Ecological uses of vertebrate indicator species: a critique. Conservation Biology 2:316-328. Law, R. 2007. Fisheries-induced evolution: present status and future directions. Marine Ecology Progress Series 335:271-277. Lehodey, P., I. Senina, and R. Murtugudde. 2008. A spatial ecosystem and populations dynamics model (SEAPODYM)–Modeling of tuna and tuna-like populations. Progress in Oceanography 78:304-318. Lewinsohn, T. M., and L. Cagnolo. 2012. Keystones in a tangled bank. Science 335:1449-1451. Libralato, S., V. Christensen, and D. Pauly. 2006. A method for identifying keystone species in food web models. Ecological Modelling 195:153-171. Lindeman, R. L. 1942. The trophic-dynamic aspect of ecology. Ecology 23:399-417. Link, J. 2002a. Does food web theory work for marine ecosystems? Marine Ecology Progress Series 230. Link, J. S. 2002b. What does ecosystem-based fisheries management mean? Fisheries 27:18-21. Link, J. S. 2004. A general model of selectivity for fish feeding: a rank proportion algorithm. Transactions of the American Fisheries Society 133:655-673. Link, J. S. 2010. Adding rigor to ecological network models by evaluating a set of pre-balance diagnostics: A plea for PREBAL. Ecological Modelling 221:1580-1591. Link, J. S., and H. I. Browman. 2014. Integrating what? Levels of marine ecosystem-based assessment and management. ICES Journal of Marine Science:1-4. Link, J. S., T. Ihde, C. Harvey, S. K. Gaichas, J. Field, J. Brodziak, H. Townsend, and R. Peterman. 2012. Dealing with uncertainty in ecosystem models: The paradox of use for living marine resource management. Progress in Oceanography 102:102-114. Mace, G. M., K. Norris, and A. H. Fitter. 2012. Biodiversity and ecosystem services: a multilayered relationship. Trends in Ecology and Evolution 27:19-26. Mackinson, S., and G. Daskalov. 2007. An ecosystem model of the North Sea to support an ecosystem approach to fisheries management: description and parameterisation. Science Series Technical Report 142, Cefas Lowestoft. Magnusson, K. G. 1995. An overview of the multispecies VPA - theory and applications. Reviews in Fish Biology and Fisheries 5:195-212. Maldonado, E., N. Hubert, P. Sagnes, and B. De Mérona. 2009. Morphology–diet relationships in four killifishes (Teleostei, Cyprinodontidae, Orestias) from Lake Titicaca. Journal of Fish Biology 74:502-520. Malyshev, V. I. 1980. Evaluation of the food requirements of the North Sea whiting, Odontogadus merlangus. Journal of Ichthyology 20:68-74. Mann, K., and P. Breen. 1972. The relation between lobster abundance, sea urchins, and kelp beds. Journal of the Fisheries Board of Canada 29:603-605. Martínez, M., A. Intralawan, G. Vázquez, O. Pérez-Maqueo, P. Sutton, and R. Landgrave. 2007. The coasts of our world: Ecological, economic and social importance. Ecological Economics 63:254-272. May, R. M., J. R. Beddington, C. W. Clark, S. J. Holt, and R. M. Laws. 1979. Management of multispecies fisheries. Science 205:267-277.  159 McClanahan, T. 2000. Recovery of a coral reef keystone predator, Balistapus undulatus, in East African marine parks. Biological Conservation 94:191-198. McClanahan, T. R. 1995. Fish predators and scavengers of the sea urchinEchinometra mathaei in Kenyan coral-reef marine parks. Environmental Biology of Fishes 43:187-193. McLaren, B., and R. Peterson. 1994. Wolves, moose, and tree rings on Isle Royale. Science 266:1555-1558. Menge, B. A. 1995. Indirect effects in marine rocky intertidal interaction webs: patterns and importance. Ecological Monographs 65:21-74. Menge, B. A., E. L. Berlow, C. A. Blanchette, S. A. Navarrete, and S. B. Yamada. 1994. The keystone species concept: variation in interaction strength in a rocky intertidal habitat. Ecological Monographs 64:249-286. Menge, B. A., and T. L. Freidenburg. 2001. Keystone species. Pages 613-631  Encyclopedia of Biodiversity. Volume 3. Messier, F., and M. Crête. 1985. Moose-wolf dynamics and the natural regulation of moose populations. Oecologia 65:503-512. Michener, W. H., and M. B. Jones. 2012. Ecoinformatics: supporting ecology as a data-intensive science. Trends in Ecology and Evolution 27:85-93. Mills, L. S., M. E. Soule, and D. F. Doak. 1993. The keystone-species concept in ecology and conservation. BioScience 43:219-224. Modlmeier, A. P., C. N. Keiser, J. V. Watters, A. Sih, and J. N. Pruitt. 2014. The keystone individual concept: an ecological and evolutionary overview. Animal Behaviour 89:53-62. Motta, P. J. 1988. Functional morphology of the feeding apparatus of ten species of Pacific butterflyfishes (Perciformes, Chaetodontidae): an ecomorphological approach. Environmental Biology of Fishes 22:39-67. Motta, P. J., K. B. Clifton, P. Hernandez, and B. T. Eggold. 1995. Ecomorphological correlates in ten species of subtropical seagrass fishes: diet and microhabitat utilization. Environmental Biology of Fishes 44:37-60. Mouquet, N., D. Gravel, F. Massol, and V. Calcagno. 2012. Extending the concept of keystone species to communities and ecosystems. Ecology Letters 16:1-8. Naiman, R. J. 1988. Animal influences on ecosystem dynamics. BioScience 38:750-752. Navarrete, S. A., and B. A. Menge. 1996. Keystone predation and interaction strength: interactive effects of predators on their main prey. Ecological Monographs 66:409-429. Norton, S. F. 1991. Capture success and diet of cottid fishes: the role of predator morphology and attack kinematics. Ecology 72:1807-1819. Norton, S. F. 1995. A functional approach to ecomorphological patterns of feeding in cottid fishes. Environmental Biology of Fishes 44:61-78. Noss, R. F. 1990. Indicators for monitoring biodiversity: a hierarchical approach. Conservation Biology 4:355-364. Nuñez, M. A., and D. Simberloff. 2005. Invasive species and the cultural keystone species concept. Ecology and Society 10:r4. Nye, J. A., J. S. Link, J. A. Hare, and W. J. Overholtz. 2009. Changing spatial distribution of fish stocks in relation to climate and population size on the Northeast United States continental shelf. Marine Ecology Progress Series 393:111-129. O'Neill, R. V., and J. R. Kahn. 2000. Homo economus as a keystone species. BioScience 50:333-337.  160 Odum, W. E., and E. J. Heald. 1975. The detritus-based food web of an estuarine mangrove community. Pages 265-286 in L. E. Cronin, editor. Estuarine Research: Chemistry, biology, and the estuarine system. Academic Press, New York. Okey, T. A., S. Banks, A. F. Born, R. H. Bustamante, M. Calvopiña, G. J. Edgar, E. Espinoza, J. M. Fariña, L. E. Garske, and G. K. Reck. 2004. A trophic model of a Galápagos subtidal rocky reef for evaluating fisheries and conservation strategies. Ecological Modelling 172:383-401. Okey, T. A., and D. Pauly. 1999. Trophic mass-balance model of Alaska's Prince William Sound ecosystem, for the post-spill period 1994-1996. Fisheries Centre Research Reports, University of British Columbia, Vancouver, Canada. Oliveira, E. F., E. Goulart, L. Breda, C. V. Minte-Vera, L. R. d. S. Paiva, and M. R. Vismara. 2010. Ecomorphological patterns of the fish assemblage in a tropical floodplain: effects of trophic, spatial and phylogenetic structures. Neotropical Ichthyology 8:569-586. Orr, J. C., V. J. Fabry, O. Aumont, L. Bopp, S. C. Doney, R. A. Feely, A. Gnanadesikan, N. Gruber, A. Ishida, and F. Joos. 2005. Anthropogenic ocean acidification over the twenty-first century and its impact on calcifying organisms. Nature 437:681-686. Ortiz, M., L. Campos, F. Berrios, F. Rodriguez, B. Hermosillo, and J. González. 2013a. Network properties and keystoneness assessment in different intertidal communities dominated by two ecosystem engineer species (SE Pacific coast): A comparative analysis. Ecological Modelling 250:307-318. Ortiz, M., R. Levins, L. Campos, F. Berrios, F. Campos, F. Jordán, B. Hermosillo, J. Gonzalez, and F. Rodriguez. 2013b. Identify