A structural biology community assessment of AlphaFold2 applications

Abstract

Most proteins fold into 3D structures that determine how they function and orchestrate the biological processes of the cell. Recent developments in computational methods for protein structure predictions have reached the accuracy of experimentally determined models. Although this has been independently verified, the implementation of these methods across structural-biology applications remains to be tested. Here, we evaluate the use of AlphaFold2 (AF2) predictions in the study of characteristic structural elements; the impact of missense variants; function and ligand binding site predictions; modeling of interactions; and modeling of experimental structural data. For 11 proteomes, an average of 25% additional residues can be confidently modeled when compared with homology modeling, identifying structural features rarely seen in the Protein Data Bank. AF2-based predictions of protein disorder and complexes surpass dedicated tools, and AF2 models can be used across diverse applications equally well compared with experimentally determined structures, when the confidence metrics are critically considered. In summary, we find that these advances are likely to have a transformative impact in structural biology and broader life-science research.

Accurate protein stability predictions from homology models

“…We conclude by suggesting that these kinds of analyses should be applied more broadly, and to a wider range of data and proteins, and point out a recent assessment of performance of multiple methods on AlphaFold2 [32] models, compared to measurements from MAVE experiments [33] . Also, it is our hope that future developments of stability prediction methods, as well as other methods to analyze protein structures, will be developed with applications on homology models or predicted structures in mind.…”

Buffering of genetic dominance by allele-specific protein complex assembly

“…The “RepairPDB” command was first run to minimise structures followed by the “BuildModel” command on the repaired structures. The final Gibbs free energy change was calculated as the average of ten replicates, and in subsequent analyses residues with pLDDT < 50, which are predicted to be disordered in solution (Akdel et al 2022), were excluded.…”

“…We thought that the low helix content in the LOF group may be indicative of high protein disorder, as disorder is naturally anticorrelated with structured regions in proteins. To investigate this, we took advantage of the structure-averaged pLDDT (predicted local-distance difference test) values from the AlphaFold predicted human protein models, which are highly predictive of intrinsic disorder (Akdel et al 2022;Tunyasuvunakool et al 2021). As expected, LOF homomers and heteromers exhibit the lowest median pLDDT value (median = 75.7, Figure S3D), significantly lower than DN subunits (median = 77, p = 1.6 × 10 -5 , Dunn's test of multiple comparisons with Holm-Bonferroni correction) and AR subunits (median = 83.3, p = 1.4 × 10 -6 ).…”

Cross-protein transfer learning substantially improves disease variant prediction

“…We used AlphaFold2 predicted structures from the AlphaFold2 human proteome database for all proteins in this study (Methods) [26]. There has been considerable interest in using AlphaFold2 structures for missense variant effect prediction [32][33][34][35]. The specific features we tested include predicted amino acid distributions from multiple versions of the deep neural network ProteinMPNN (which takes structure as input) [36], as well as a hand-designed feature that combines a known structure with conservation in the EVE MSA (Methods).…”