Structural Hotspot Conditioning for Diffusion-Based Molecular Design

Diffusion-based generative models have emerged as powerful tools for structure-based molecular design, yet their effectiveness depends critically on how prior information about protein-ligand interactions is incorporated into the generative process. In this work, we propose and systematically investigate structural hotspot conditioning as a mechanism to inject interaction-relevant chemical knowledge into pocket-conditioned diffusion models. We evaluate three complementary strategies developed for defining atom-level reference points: docking score-based atom selection, structure-based pharmacophore localization, and data-driven prediction of hydrogen-bonding sites. Using MolSnapper as a unified diffusion-based generative model, we benchmark these strategies on a standardized CrossDocked2020 test set, evaluating docking-derived affinity estimates, physicochemical quality, structural plausibility, and chemical diversity. Compared to established baselines, our explicit hotspot-based conditioning systematically alters generative behavior, enabling controllable trade-offs between affinity, chemical feasibility, and diversity. An iterative case study on the shallow allosteric pocket of Ube2T further illustrates how alternating hotspot strategies supports effective exploration and prioritization of ligand candidates for challenging targets. Overall, this work establishes structural hotspot conditioning as a flexible and tunable design variable for guiding diffusion-based molecular generation toward chemically meaningful solutions.