Pharmacophore model guided 3D molecular generation through diffusion model

Generating molecules that rigorously satisfy three-dimensional (3D) pharmacophore constraints has remained a challenge in de novo drug design, despite rapid progress in molecular generative models. We present DiffPharm, a diffusion-based framework that embeds explicit pharmacophore control into the generative process. Unlike previous models, 3D pharmacophore models are encoded as pharmacophore graphs and utilized as constraints injected into an equivariant transformer architecture during the denoising process. This design enables not only maintaining good pharmacophore alignment for generated conformations but also a unique inpainting capability, allowing structure generation under both substructure and pharmacophore constraints. Our benchmark studies show that DiffPharm can efficiently generate structurally diverse and drug-like molecules that achieve superior pharmacophore match scores as well as docking scores compared to existing methods.