AlphaFold Model Quality Self-Assessment Improvement Via Deep Graph Learning

In the past several years significant advances have been made in the field of deep-learning based computational modeling of proteins, with DeepMind’s AlphaFold2 being among the most prominent. Alongside the atom coordinates, these computationally modeled protein structures typically contain self-confidence metrics that can be used to gauge the relative modeling quality of individual residues, or the protein as a whole. Unfortunately, these scores are not always accurate, and may sometimes annotate poorly modeled regions of the protein as high confidence. Here, we introduce EQA-Fold to address this problem. EQA-Fold overhauls the LDDT prediction head of AlphaFold to provide more accurate self-confidence scores. We show that EQA-Fold is able to provide more accurate self-confidence scores than the standard AlphaFold architectures, as well as recent Model Quality Assessment protocols.