Temporal Perturbation Scanning: AI-Driven Deconstruction of Universal Biomolecular Recognition Mechanisms

Understanding the physicochemical principles governing intermolecular recognition remains a fundamental challenge across biochemistry, drug discovery, and synthetic biology. While molecular dynamics simulations capture interaction dynamics and machine learning models predict binding affinities, neither provides a systematic approach to deconstruct recognition mechanisms into their fundamental physical components across time. Here we present Temporal Perturbation Scanning (TPS), an AI framework that universally decomposes intermolecular interactions into electrostatic, hydrophobic, and steric contributions across binding stages. Our method: i) combines graph neural networks with interpretable, multi-faceted in silico mutagenesis to automatically process biomolecular complexes, ii) applies targeted physicochemical perturbations, and iii) quantifies residue-specific importance through statistical effect sizes. Crucially, TPS achieves this mechanistic insight with orders-of-magnitude greater computational efficiency than traditional mutation simulations, extracting equivalent information from single trajectories that would otherwise require dozens of independent simulations. Applied to neurotoxin-receptor and other biomolecular systems, TPS reveals fundamentally different engagement strategies and stage-dependent interaction patterns, demonstrating its capacity to capture mechanistic variation across complexes. Beyond resolving these biological insights, our framework provides a generalizable computational platform for mechanistic analysis across diverse molecular interactions, enabling researchers to move beyond predicting whether molecules interact to understanding how and why they recognize one another through interpretable, physicochemical decomposition across time.