Current structure predictors are not learning the physics of protein folding

Area of Science:

• Computational Biology
• Biophysics
• Structural Biology

Background:

• Predicting protein native states is crucial for understanding protein folding.
• Deep learning models excel at predicting protein crystal structures.
• The physical basis of protein folding dynamics in these models remains unclear.

Purpose of the Study:

• To evaluate if current protein structure prediction methods capture protein folding physics.
• To compare simulated folding pathways from advanced models with experimental data.

Main Methods:

• Comparison of simulated protein folding pathways from AlphaFold 2, RoseTTAFold, and other leading methods.
• Assessment against experimental observables like intermediate structures and folding rates.
• Evaluation of predictive accuracy against sequence-agnostic features.

Main Results:

• Simulated dynamics from structure prediction methods show limited information about folding pathways.
• Predictive ability of these models for folding pathways is inferior to simple sequence-agnostic classifiers.
• Folding trajectories generated by the models are uncorrelated with experimental folding rates and intermediate structures.

Conclusions:

• Current advanced protein structure prediction models do not yet offer enhanced insights into protein folding mechanisms.
• These models may be primarily knowledge-based predictors of final structures rather than dynamic folding processes.
• Further research is needed to bridge the gap between predicted structures and the physical understanding of protein folding.