Open Collections

Understanding gene regulatory mechanisms of mouse immune cells using a convolutional neural network

University of British Columbia

Cell differentiation is controlled via complex interactions of genomic regulatory sites such as promoters and enhancers that lead to precise cell type-specific patterns of gene expression. Local chromatin accessibility at these sites is a requirement of regulatory activity, and is therefore an important component of the gene regulation machinery. To understand how DNA sequence drives local chromatin accessibility within the context of immune cell differentiation, we examined a dataset of open chromatin regions (OCRs) derived with the ATAC-seq assay from 81 closely related mouse immune cell types. We selected and optimized a convolutional neural network (CNN), which we named AI-TAC, that takes as input a 250bp DNA sequence of a potential OCR and predicts the relative chromatin activity at that OCR across the 81 different immune cell types in our dataset. Test dataset results showed that for many OCRs, AI-TAC is able to predict chromatin state with a high degree of accuracy, even differentiating between closely related cell types. Using CNN interpretability methods we identified sequence motifs used by the model to make predictions, many of which match closely to known transcription factor (TF) binding sites. The cell type - specific influence assigned to each motif by AI-TAC in many instances recapitulates prior biological knowledge about the role of these TFs in immune cell differentiation, lending credibility to our model and interpretation methods. Finally, we attempt to discern if the model detected any combinatorial activity between TFs that is predictive of chromatin accessibility.

Text

2020-05-04

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/74302

Bioinformatics

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/