Current structure predictors are not learning the physics of protein folding

Abstract

Predicting the native state of a protein has long been considered a gateway problem for understanding protein folding. Recent advances in structural modelling driven by deep learning have achieved unprecedented success at predicting a protein’s crystal structure, but it is not clear if these models are learning the physics of how proteins dynamically fold into their equilibrium structure or are just accurate knowledge-based predictors of the final state. In this work, we compare the pathways generated by state-of-the-art protein structure prediction methods to experimental data about protein folding pathways. The methods considered were AlphaFold 2, RoseTTAFold, trRosetta, RaptorX, DMPfold, EVfold, SAINT2 and Rosetta. We find evidence that their simulated dynamics capture some information about the folding pathway, but their predictive ability is worse than a trivial classifier using sequence-agnostic features like chain length. The folding trajectories produced are also uncorrelated with experimental observables such as intermediate structures and the folding rate constant. These results suggest that recent advances in structure prediction do not yet provide an enhanced understanding of protein folding. Supplementary information Supplementary data are available at Bioinformatics online.

Membranome 3.0: Database of single‐pass membrane proteins with AlphaFold models

“…At the 14th Critical Assessment of protein Structure Prediction (CASP), the AlphaFold 2 (AF2) system 12 outperformed all other computational methods, producing models rivaling experimental structures. 12 , 13 , 14 , 15 Although AlphaFold 2 has a limited applicability in modeling protein dynamics, multiple conformational states, or effects of mutations, 16 , 17 the approach is thought to achieve a significant progress in predicting the structure of a single protein chain. 12 Public availability of the AF2 source code and the recent release of the AlphaFold DataBase 18 with over half a million protein models, including the full proteomes of 16 model organisms and 32 pathogens, are starting to have a transformative impact on structural biology.…”

“… 12 Public availability of the AF2 source code and the recent release of the AlphaFold DataBase 18 with over half a million protein models, including the full proteomes of 16 model organisms and 32 pathogens, are starting to have a transformative impact on structural biology. 13 , 15 , 16 Despite the varying quality of AF2‐generated models, they have been added to high‐quality authoritative resources for protein sequences, structures, and functional information, such as UniProt 4 and PDBsum. 19 The availability of high‐accuracy predictions for a significant portion of many organisms' proteomes is a novel source of information into bitopic proteins.…”

AlphaFold2 fails to predict protein fold switching

“…It is notable, however, that AlphaFold2’s predictions miss the ground state of fold switchers 30% of the time. This is further evidence that its predictions are primarily rooted in sophisticated pattern recognition, not protein biophysics 37 .…”

“…and kinetics (what pathways do proteins traverse between unfolded and folded states?) 4; 36 , deep learning approaches reveal apparent properties of experimentally determined protein structure rather than biophysical pathways 37 . Thus, it is not surprising that its predictions often fail for proteins whose properties are not fully apparent from solved protein structures, such as IDPs 5-7 .…”

Assessing Protein Homology Models with Docking Reproducibility

“…14 This approach has not provided much information on the molecular level details of a protein model. While AlphaFold 2 does not elucidate the physical manner in which a protein folds, 15 it has been shown to predict protein backbones with high accuracy when compared to experimental structures, however, less attention has been paid to side-chain structure in AlphaFold 2 models. A recent paper reported that for the three CASP14 proteins that had bound ligands of S-adenosylmethionine or adenosine-5'-diphosphate in their experimental structures, the homology model had the same location of binding but the binding pose was different than the experimental structure.…”