Interaction-Aware Machine Learning Interatomic Potential Guidance for Diffusion-Based 3D Ligand Generation in Protein Pockets

Abstract

Protein-conditioned three-dimensional (3D) generative models can directlyconstruct ligands in target binding pockets, but learning the structuraldistribution of known complexes does not ensure that sampled poses occupyfavorable regions of the underlying energy landscape. Recent guidancestrategies have begun to introduce affinity or force-field information duringgeneration, yet a key issue remains unresolved, i.e. the force acting on ligandatoms mixes ligand-intrinsic conformational relaxation withprotein-environment-induced interactions, and these contributions need not beequally informative for target-oriented molecular design. Here, a training-free, inference-time framework integrates the machine learninginteratomic potential (MLIP) into the late-stage reverse diffusion of thepretrained generative model. A chemistry-aware proxy enables MLIP evaluation ofintermediate diffusion states, while an interaction-aware force decompositionseparates ligand-intrinsic and pocket-induced contributions to enhancephysics-guided molecular generation. Across eight protein-ligand systems, interaction-aware guidance consistentlyoutperformed the unguided generation baseline, yielding improvedligand-pocket interaction energies and reduced pocket-constrained ligandstrain. Independent evaluations showed the same trend, with guidedconformations achieving better AutoDock Vina scores and more favorablesingle-frame MM/GBSA energy estimates, while the PoseBusters pass rateincreased relative to unguided structures. Ablation analysis further showedthat direct complex-force guidance mainly favors ligand relaxation, whereasinteraction-aware decomposition more effectively improvestarget-recognition-related properties. This record contains the guided-generation code, the six evaluation pipelines,and the generated ligand structures (4800 conformations; eight targets, threeconditions, 200 molecules each). Directory names correspond to the conditionsin the manuscript: native DiffGui (unguided baseline), Complex force(complex-force guidance), and Interaction-weighted force(interaction-weighted guidance).