Comprehensive assessment of AlphaFold's predictions of secondary structure and solvent accessibility at the amino acid-level in eukaryotic, bacterial and archaeal proteins.

Abstract

Numerous sequence-based predictors of the amino acid (AA)-level solvent accessibility (SA) and secondary structure (SS) of proteins have been developed. We empirically investigated whether these two key characteristics of AA-level structure can be accurately predicted from putative structures generated by the popular AlphaFold2. We compared AlphaFold2's results against several representative SS and SA predictors on a large test dataset that covers five distinct taxonomic groups (animals, plants, fungi, bacteria, and archaea). We used a broad collection of metrics that evaluate predictions of the numeric and binary (buried vs. solvent exposed) SA and the 3-state SS at both AA- and SS-region levels. We found that AlphaFold2 generated very accurate results, with high average Q₃ accuracy of 0.928 for the SS prediction and high Pearson Correlation Coefficient (PCC) of 0.815 between its putative and native SA values. AlphaFold2 significantly and consistently outperforms the considered predictors of SA and SS across the five taxonomic groups and both AA and region level evaluations. Moreover, we demonstrated that AlphaFold2 nearly perfectly reconstructs distributions of the sizes and numbers of the SS regions. We also showed that AlphaFold2 substantially improves over the SS and SA predictors when tested on a low sequence similarity test dataset, although its results and results of two other predictors suffer a modest drop in the quality of predicting SS regions. Altogether, our results suggest that AlphaFold2 makes very accurate predictions of SS and SA, which can be easily extracted from 200+ million pre-computed AF2's structure predictions in AlphaFoldDB.