Open Collections

Abstract

Autoinhibition is a self-regulatory mechanism in cells where proteins can switch themselves “on” or “off” in response to certain stimuli. Autoinhibited proteins exist in an equilibrium between inhibitory and non-inhibitory conformations, shifting conformation in response to certain stimuli. Its regulatory use in multiple cellular pathways makes it an ideal drug target and an important component of multiple biomolecular disciplines, but our understanding is hampered by challenges in obtaining experimental structures. Standard methods are time-consuming, expensive, and inconsistent. AlphaFold2, an AI program designed to predict protein structures, can now generate highly accurate structures from only peptide sequences. Its predictions are informed by learned evolutionary and structural data and it has been shown that manipulating its inputs leads it to predict alternative conformations. I here investigate AlphaFold2's ability to predict autoinhibited protein structures in the hopes of obtaining both their inhibited and non-inhibited conformations. I find that AlphaFold2 is less accurate and less confident in predicting autoinhibited proteins than non-autoinhibited proteins and, though it can be manipulated to produce alternative conformations, these predictions are rarely accurate to known experimental structures. In addition, I investigated the link between its predictions and the evolutionary data provided to it, finding that a greater depth of evolutionary information leads to predictions closer to experimental structures and that a prediction's conformation is dependent upon the evolutionary information provided.