NLM DIR Seminar Schedule

Abstract:

AlphaFold and other machine learning methodologies have vastly improved protein structure prediction. Despite their success, question remain how they can be best utilized to advance structure prediction for conformationally dynamic proteins, including those capable of fold switching. Approaches utilizing sequence clustering or masking have been shown to increase the prevalence of alternative conformation predictions, however, with these approaches there can be little correlation between prediction accuracy and confidence. We hypothesize that the basis of these alternative-conformation predictions is sequence associations with the alternative-conformation structures seen in the AlphaFold training set. If this is true, a smaller dataset should provide an ideal candidate for finding which sequence elements are driving predictions, so we turn to fold-switching proteins for testing. Indeed, we find that AlphaFold associates specific sequence features with alternative-conformation structures in its training set. We go on to demonstrate that these predictions have mixed success when compared to experimental results, and that a ground state prediction may be more reliable than an alternative state prediction.