PharmacoForge: pharmacophore generation with diffusion models

Structure-based drug design (SBDD) is enhanced by machine learning (ML) to improve both virtual screening and de novo design. Despite advances in ML tools for both strategies, screening remains bounded by time and computational cost, while generative models frequently produce invalid and synthetically inaccessible molecules. Screening time can be improved with pharmacophore search, which quickly identifies ligands in a database that match a pharmacophore query. In this work, we introduce PharmacoForge, a diffusion model for generating 3D pharmacophores conditioned on a protein pocket. Generated pharmacophore queries identify ligands that are guaranteed to be valid, commercially available molecules. We evaluate PharmacoForge against automated pharmacophore generation methods using the LIT-PCBA benchmark and ligand generative models through a docking-based evaluation framework. We further assess pharmacophore quality through a retrospective screening of the DUD-E dataset. PharmacoForge surpasses other pharmacophore generation methods in the LIT-PCBA benchmark, and resulting ligands from pharmacophore queries performed similarly to de novo generated ligands when docking to DUD-E targets and had lower strain energies compared to de novo generated ligands.