Investigating whether deep learning models for co-folding learn the physics of protein-ligand interactions

Co-folding models represent a major innovation in deep-learning-based protein-ligand structure prediction. The recent publications of RoseTTAFold All-Atom, AlphaFold3, and others have shown high-quality results on predicting the structures of proteins interacting with small-molecules, nucleic-acids, and other proteins. Despite these advanced capabilities and broad potential, the current study presents critical findings that question the adherence of these models to fundamental physical principles. Through adversarial examples based on established physical, chemical, and biological principles, we demonstrate notable discrepancies in protein-ligand structural predictions when subjected to biologically and chemically plausible perturbations. These discrepancies reveal a significant divergence from expected physical behaviors, indicating potential overfitting to particular data features within its training corpus. Our findings underscore the models' limitations in generalizing effectively across diverse protein-ligand structures and highlight the necessity of integrating robust physical and chemical priors in the development of such predictive tools. The results advocate a measured reliance on deep-learning-based models for critical applications in drug discovery and protein engineering, where a deep understanding of the underlying physical and chemical properties is crucial.