TheSelective: Dual Affinity-Guided Diffusion for Selective Molecular Generation

Diffusion-based generative models have shown promise in Structure-Based Drug Design (SBDD) by generating 3D ligands with strong binding affinities. However, existing models primarily focus on maximizing absolute potency toward a single target, neglecting selectivity: the ability to maximize the affinity gap between a target and potential off-targets. To address this, we propose TheSelective, a novel framework that enhances molecular selectivity via an asymmetric dual-guidance mechanism. Our approach captures on-target interactions through complex-based graphs while estimating off-target affinity via cross-attention between independent protein and ligand embeddings, bypassing the need for off-target docked structures. By jointly applying these dual-affinity signals to both atom types and spatial coordinates, TheSelective steers the generative process toward a maximized affinity gap across both chemical and geometric spaces. Experiments on the CrossDocked2020 dataset demonstrate that our model generates ligands with superior selectivity against both structurally similar and dissimilar off-targets. We validate the effectiveness of our framework through comprehensive ablation studies and case analyses, highlighting its potential for real-world drug discovery. The code is available at https://github.com/DannyJPark/TheSelective