Do Deep Learning Models for Co-Folding Learn the Physics of Protein-Ligand Interactions?

Co-folding models are the newest innovation in deep-learning-based protein-ligand structure prediction. The recent publications of RoseTTAFold All-Atom and AlphaFold 3 have shown high-quality results on predicting the structures of proteins interacting with small-molecules, other proteins and nucleic-acids. Despite these advanced capabilities and broad potential, the current study presents critical findings that question these models adherence to fundamental physical principles and its susceptibility to overfitting specific protein families. Through adversarial examples based on established physical, chemical, and biological principles, we demonstrate notable discrepancies in protein-ligand structural predictions when subjected to biologically plausible perturbations. These discrepancies reveal a significant divergence from expected physical behaviors, indicating potential overfitting to particular data subsets within its training corpus. Our findings underscore the models limitations in generalizing effectively across diverse biological 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.