Conformational epistasis impairs AlphaFold structural predictions
Preprint
Abstract
Protein structures have been massively predicted using homologous sequence information. AlphaFold2 (AF2) is a recent breakthrough to predict 3D models using machine learning approaches that reached an outstanding accuracy in recent quality evaluations. However, information derived from extant homologous sequences, as those used by AF2, might not contain enough information to accurately predict protein structure. This limitation could be related to the process known as epistasis, which describes the differential effect of a mutation on the evolutionary trajectory. Clear evidence of conformational epistasis, which has a specific impact on protein structure, was characterized in the evolutionary origin of the glucocorticoid receptor (GR) specificity during its functional divergence from the mineralocorticoid (MR) receptor. In this work we explore how AF2 can reproduce conformations derived from epistatic effects. Using structural clustering and principal component analysis to analyze the structural similarities in 16 and 13 extant GR and MR conformers, respectively, we found that AF2 models for human GR failed to reproduce extant GR conformations. Interestingly, AF2 models for human MR, for which no conformational epistasis was reported, were almost indistinguishable from extant MR. Our results showcase the importance of evolutionary trajectories to predict accurate 3D models.Keywords
This publication has 50 references indexed in Scilit:
- lDDT: a local superposition-free score for comparing protein structures and models using distance difference testsBioinformatics, 2013
- Molecular mechanisms of epistasis within and between genesTrends in Genetics, 2011
- Analyzing protein structure and function using ancestral gene reconstructionCurrent Opinion in Structural Biology, 2010
- Robustness of Ancestral Sequence Reconstruction to Phylogenetic UncertaintyMolecular Biology and Evolution, 2010
- Improved prediction of protein side‐chain conformations with SCWRL4Proteins: Structure, Function, and Bioinformatics, 2009
- Enzyme (re)design: lessons from natural evolution and computationCurrent Opinion in Chemical Biology, 2009
- Epistasis — the essential role of gene interactions in the structure and evolution of genetic systemsNature Reviews Genetics, 2008
- SDR: a database of predicted specificity-determining residues in proteinsNucleic Acids Research, 2008
- PAML 4: Phylogenetic Analysis by Maximum LikelihoodMolecular Biology and Evolution, 2007
- SWISS-MODEL: an automated protein homology-modeling serverNucleic Acids Research, 2003