DiffPharma : A Conditional Diffusion Framework for Interaction-Constrained 3D Molecular Design
Masami SAKO *1, Nobuaki YASUO2, Masakazu SEKIJIMA1
1Department of Computer Science, Institute of Science Tokyo
2Department of Chemical Science and Engineering, Institute of Science Tokyo
[Purpose]
Pharmacophore modeling in structure-based drug design (SBDD) is a key approach in drug discovery, as it captures critical interaction features between the target proteins and ligand molecules. Conventional pharmacophore modeling relies on screening existing molecular databases, which limits the exploration of chemical space. In this work, we propose DiffPharma, a structure-based pharmacophore modeling framework based on a conditional diffusion model to generate molecules that satisfy specified interaction constraints., which generate new molecules that satisfy predefined interaction constraints, thereby expanding the accessible chemical space beyond existing libraries.
[Methods]
DiffPharma is based on the Denoising Diffusion Probabilistic Model (DDPM) and introduces “interaction particles” between proteins and ligands. The model processes hydrogen bonds and hydrophobic interactions through separate equivariant graph neural networks (EGNNs), which are integrated via a multi-path adaptive fusion EGNN (MAP-EGNN).
[Results and Discussion]
As a case study, we focused on the SARS-CoV-2 main protease and generated molecules under two conditions: with and without a hydrogen bond constraint.
・Spatial distribution analysis based on static structures
・Binding stability evaluation by molecular dynamics simulation
・Binding free energy calculation by MM/GBSA method
・Drug-likeness profiling via ADMETlab 3.0
These analyses revealed that DiffPharma successfully generated molecules exhibiting higher binding affinity than the reference ligand molecule (PDB ID: 7GBL).
[Conclusion]
DiffPharma demonstrates that incorporating explicit protein–ligand interaction constraints into 3D molecular generation enables the generation of molecules with structurally and interactively appropriate binding poses, and yields molecules with desirable binding affinity and stability toward the target protein.