AlphaFold Models Illuminate Half of Dark Human Proteins

Abstract

We investigate the use of confidence scores to predict the accuracy of a given AlphaFold model for drug discovery. Predicted accuracy is improved by eliminating confidence scores below 80, due to effects of disorder. 95% of models corresponding to a set of recent crystal structures are accurate at the fold level. Conformational discordance in the training set has a more significant effect on accuracy than sequence divergence. We propose criteria for models and residues that are possibly useful for virtual screening, by which AlphaFold provides models for half of understudied (dark) human proteins and two-thirds of residues in those models.