Dissecting AlphaFold’s Capabilities with Limited Sequence Information

Abstract Protein structure prediction, a fundamental challenge in computational biology, aims to predict a protein’s 3D structure from its amino acid sequence. This structure is pivotal for elucidating protein functions, interactions, and driving innovations in drug discovery and enzyme engineering. AlphaFold2, a powerful deep learning model, has revolutionized this field by leveraging phylogenetic information from multiple sequence alignments (MSAs) to achieve remarkable accuracy in protein structure prediction. However, a key question remains: how well does AlphaFold2 understand protein structures? This study investigates AlphaFold2’s capabilities when relying primarily on high-quality template structures, without the additional information provided by MSAs. By designing experiments that probe local and global structural understanding, we aimed to dissect its dependence on specific features and its ability to handle missing information. Our findings revealed AlphaFold2’s reliance on sterically valid C- β atoms for correctly interpreting structural templates. Additionally, we observed its remarkable ability to recover 3D structures from certain perturbations and the negligible impact of the previous structure in recycling. Collectively, these results support the hypothesis that AlphaFold2 has learned an accurate local biophysical energy function. However, this function seems most effective for local interactions. Our work significantly advances understanding of how deep learning models predict protein structures and provides valuable guidance for researchers aiming to overcome limitations in these models. protein folding, alphafold, side-chain, interpretability