Predicting side chain conformations in folded proteins by AlphaFold: Perspective and challenges.

Area of Science:

• * Computational Biology
• * Structural Biology
• * Bioinformatics

Background:

• * AlphaFold revolutionized protein structure prediction from amino acid sequences.
• * A key challenge remains predicting individual amino acid side chain conformations within folded proteins.
• * Understanding side chain conformations is crucial for molecular modeling and drug design.

Purpose of the Study:

• * To assess the accuracy of ColabFold (an AlphaFold2 implementation) and AlphaFold3 in predicting protein side chain conformations.
• * To investigate factors influencing prediction accuracy, such as side chain type and use of structural templates.
• * To explore the integration of AlphaFold with mutation prediction models for analyzing cooperative effects.

Main Methods:

• * Evaluation of ColabFold and AlphaFold3 on benchmark protein sets (Set A and recently released structures).
• * Calculation of prediction errors for various dihedral angles (χ1 to χ3).
• * Application of a Potts model for large-scale mutational scans and subsequent structural analysis using ColabFold.

Main Results:

• * ColabFold shows prediction errors of ~14% for χ1 and ~48% for χ3 dihedral angles on benchmark proteins.
• * Prediction accuracy is higher for nonpolar side chains and slightly improved with structural templates.
• * ColabFold exhibits a bias towards common rotamer states, potentially limiting prediction of rare conformations.
• * AlphaFold3 demonstrates slightly better side chain prediction accuracy than ColabFold.
• * ColabFold maintains similar accuracy for recently released protein structures not used in AlphaFold2 training.

Conclusions:

• * ColabFold and AlphaFold3 can predict protein side chain conformations with notable accuracy, particularly for common states.
• * The models may struggle with rare side chain conformations due to biases in training data.
• * Integrating sequence-based models with AlphaFold offers a novel approach to study mutation-induced structural changes and their impact on fitness.