Inference-Time Guidance in Pocket-Conditioned Molecular Diffusion Models: Limits in Preventing Steric Clashes
P. Bärtschi · 2026
Recent advances in AI-based protein structure prediction have accelerated structure-based drug design by enabling computational models to exploit detailed three-dimensional protein–ligand interactions. In parallel, geometric deep learning has enabled generative models that directly design pocket-conditioned ligands with physically plausible 3D structures. Among these approaches, SE(3)-equivariant diffusion models have emerged as a powerful framework for jointly modeling atomic coordinates and chemical features during molecular generation. However, these models frequently produce unrealistic ligand–pocket interactions, particularly steric clashes that violate fundamental geometric constraints. In this thesis, we investigate the origin of steric clashes in diffusion-based ligand generation and analyze their structural characteristics using continuous distance-based and pocket-component–specific metrics. Building on the score-based diffusion formulation, we introduce a side-chain repulsive guidance mechanism that steers ligand atoms away from protein side chains during sampling. Although the proposed model achieves competitive performance compared to recent state-of-the-art methods, our experiments show that inference-time repulsive guidance does not reduce steric clash frequency, suggesting that mitigating such failures likely requires stronger inductive biases incorporated directly into the training objective.