Beyond Geometry: Functionally Grounded Molecule Generation for Structure-Based Drug Design

Ruifeng Li, Mingqian Li, Zichang Jin, Shiyang Zhang, Zesen Lyu, Yao Yang, Qiang Zhang, Hongyang Chen

Keywords: structure-based ligand generation

TL;DR: we introduce a Functionally Grounded Molecule Generation Network (FGMOL) that operates within a unified structure-function alignment space

Abstract: Structure-based drug design aims to generate 3D ligands that bind stably to specific protein pockets. While recent generative models have improved by incorporating the geometry of protein pockets, they still overlook the biochemical functional interactions between proteins and ligands. Crucial interactions include hydrogen bonds, hydrophobic interactions, and $\pi-\pi$ stacking, which are essential for binding affinity and structural stability. This oversight leads to strained, high-energy ligands that may geometrically fit but functionally misalign with the binding site. To bridge the gap, we introduce a Functionally Grounded Molecule Generation Network (FGMOL) that operates in a unified structure-function alignment framework, enabling molecular generation to align with protein-ligand interactions, extending beyond mere geometric fitting. Our design of \method introduces: (1) Interaction-Aware Embedding, which annotates protein atoms with explicit interaction types and feed them into SE(3)-equivariant neural networks; (2) Interaction-Informed Motif Alignment, which leverages differentiable clustering and Sinkhorn matching to align protein-ligand functional motifs; and (3) Interaction-Guided Generation with Bayesian Flow Network, which jointly models ligand coordinates and atom types via Bayesian updates in continuous space, conditioned on protein-guided cross-attention. Experiments on the CrossDocked2020 benchmark demonstrate that \method surpasses prior state-of-the-art methods in binding affinity, and notably reduces strain energy by over 20%, while maintaining high synthetic accessibility—highlighting its advantage in interaction-aware ligand generation.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 3031