Pharmacophore constraints go straight into 3D diffusion: PharmaDiff hits a 0.6990 exact-match rate without a target structure
Synopsis
PharmaDiff turns a 3D pharmacophore hypothesis into an atom-level subgraph and injects it as a condition into diffusion-based generation, producing 3D molecules that follow the prescribed spatial arrangement. Across 250 pharmacophore conditions drawn from the GEOM-Drugs test set it reaches a mean match score of 0.8964 and a perfect match rate of 0.6990, above PGMG, TransPharmer, REINVENT4, ShEPhERD and PhoreGen; its average docking scores on four targets also beat DiffSBDD, which needs the protein structure, though its PoseBusters pass rate is below ShEPhERD and PhoreGen.
Interpretation
Molecule generation is no longer constrained only by 2D similarity or a whole protein pocket; it is conditioned directly on a 3D pharmacophore hypothesis to produce a complete 3D molecular graph. Similarity methods such as REINVENT4's Mol2Mol constrain only similarity to a reference, TransPharmer satisfies the full set of pharmacophore features but ignores their arrangement in 3D space, and PGMG uses a 3D pharmacophore hypothesis yet represents molecules as SMILES and replaces Euclidean distance with shortest-path distance. On 250 pharmacophore conditions randomly drawn from the GEOM-Drugs test set, PharmaDiff reaches a mean match score of 0.8964 and a perfect match rate of 0.6990, versus 0.7430 and 0.3182 for PhoreGen and 0.5160 and 0.1206 for ShEPhERD.
The generation process treats the atoms tied to the pharmacophore as fixed anchors, re-injecting their types, charges and 3D coordinates at every denoising step. The approach adds pharmacophore conditioning to MiDi's SE(3)-equivariant graph-transformer diffusion architecture: pharmacophore features are concatenated with atom types and formal charges, inpainting replaces the pharmacophore atoms' features and positions at the start of each denoising time step, their positions are reintroduced at the end of each transformer layer, and cross-attention lets generated atoms query the pharmacophore atoms. The model is trained on GEOM-Drugs, which holds over 450,000 drug-sized molecules averaging 44.4 atoms; the benchmark generates 25,000 molecules across 250 pharmacophore conditions.
Perspective
The demonstrated scope is a ligand-based generation benchmark on GEOM-Drugs, structure-based comparisons on four targets (BRD4, VEGFR2, CDK6, TGFB1), and two case studies on DDR1 and PI3Kα; all results are computational, with no experimental synthesis or activity assay.
A perfect match rate of 0.6990 means roughly three in ten generated molecules do not exactly reproduce the conditioning pharmacophore, owing to features lost when the local environment of pharmacophore atoms changes and to disconnected fragments when spatially distant features are linked by too few generated atoms; the PoseBusters pass rate is 48.68% on the ligand-based benchmark, below ShEPhERD's 95.97% and PhoreGen's 64.03%, which the authors attribute to a trade-off between pharmacophore fidelity and structural plausibility from fixing pharmacophore atoms.