PharmacoBridge

Structure-based generative drug design usually conditions on a protein pocket, handing the model an entire binding site and asking it to work out implicitly which atoms matter. Most of the pocket does not: only a handful of chemical features — a hydrogen-bond donor here, an aromatic ring there, a hydrophobe in between — actually make the interaction. A pharmacophore is exactly that shortlist, a spatial arrangement of the features required for binding, and medicinal chemists have used it as a design constraint for decades.

PharmacoBridge, from Conghao Wang and Jagath C. Rajapakse at Nanyang Technological University, makes that shortlist the generative condition directly in three dimensions. Rather than treating pharmacophores as a latent code and emitting SMILES, it learns a diffusion bridge between two point clouds: the terminal state is the pharmacophore arrangement, and the initial state is the molecule that satisfies it. Doob's h-transform adds a drift term guaranteeing the process reaches a fixed endpoint, so a trained model can be run in reverse to sample molecular structures from any pharmacophore hypothesis supplied at inference time. The whole map is SE(3)-equivariant, which means the generated geometry rotates and translates with the input rather than having to relearn every orientation.

The work was posted to bioRxiv in December 2024 and to arXiv, with code released under an MIT license. It sits alongside PharmaDiff, which conditions a 2D molecular-graph diffusion model on a pharmacophore hypothesis; PharmacoBridge instead keeps the design entirely in 3D space, so the spatial arrangement is never discarded and re-derived.

#Key Features

• Pharmacophore as the generative endpoint: A bridge process fixes both ends of the trajectory, so sampling starts from the requested feature arrangement instead of steering toward it with a soft penalty.
• SE(3)-equivariant denoiser: The score network is an equivariant graph neural network, giving roto-translational equivariance over atom coordinates and invariance over atom features.
• One-shot 3D generation: Molecules emerge as complete point clouds rather than being assembled atom by atom, which avoids compounding errors from an artificial decision order and supports batched sampling.
• Works from ligands or from pockets: Pharmacophores can be extracted from a known active with no target structure, or identified from a solved binding complex, covering both ligand-based and structure-based design.

#Technical Details

PharmacoBridge is trained on CrossDocked2020 v1.3, filtered to complexes with intimate binding poses (RMSD below 0.1 Å), leaving roughly 242,000 ligands. Splitting by pocket rather than by molecule gives about 197,000 training, 19,000 validation, and 15,000 test molecules with no target leakage. Molecules and their pharmacophores are represented as paired point clouds; the score model is parameterized following the EDM preconditioning scheme with an equivariant GNN backbone, trained on a mean-squared-error objective over coordinates and features (feature loss weighted 10x), and sampled with a deterministic Heun second-order solver. Both variance-preserving and variance-exploding bridge designs were tried, with the VP variant plus aromatic features performing best.

On unconditional generation of 10,000 molecules, that configuration reached 99.96% validity, 91.94% uniqueness, and 100% novelty, against 100% validity but 6.89% uniqueness for EDM and 82.60% for GruM. Under pharmacophore guidance, average matching scores across ten held-out targets ranged from 0.71 to 1.00, above Pocket2Mol (0.08–0.49) and TargetDiff (0.11–0.38) on every target. GNINA docking showed that for seven of those ten targets, over 90% of generated molecules scored better than the original CrossDocked ligand.

#Applications

The model suits hit identification where the useful prior knowledge is a set of interaction features rather than a full receptor structure — orphan targets with known actives but no crystal structure, series where a pharmacophore hypothesis has already been distilled from SAR, and scaffold-hopping campaigns where the required contacts are fixed but the connecting framework should change. Because generation is 3D from the start, output conformers can go straight to docking without an intermediate conformer-enumeration and filtering step.

#Impact

PharmacoBridge shows that diffusion bridges, which fix a distribution at both ends of the trajectory, are a natural fit for conditional molecular design: any structured constraint expressible as a point cloud becomes a terminal state rather than a guidance term. Its results are computational — validity, uniqueness, pharmacophore recovery, and docking scores — with no reported synthesis or binding assays, and the docking comparison uses GNINA scores rather than measured affinities. The MIT-licensed repository carries training, sampling, and evaluation code, while the pretrained checkpoint for the best-performing VP-bridge aromatic-feature configuration is distributed separately, through a Google Drive folder linked from the README and without license terms of its own. As a preprint, the work awaits peer review.