MolSnapper

A medicinal chemist reading a crystal structure usually knows more than a generative model can be told: which hydrogen bond to the hinge has to survive, which lipophilic subpocket should be filled, which half of a fragment hit is worth growing. Most 3D structure-based generators offer no way to say any of it: they accept a pocket, return molecules, and the only real lever is to retrain on a different set of protein-ligand complexes.

MolSnapper, from the Oxford Protein Informatics Group at the University of Oxford, supplies that lever without touching a model's weights. It constrains a pretrained, unconditional 3D molecule diffusion model's reverse trajectory with 3D pharmacophore points — donor and acceptor positions supplied by hand or extracted from a reference ligand or fragment screen — plus a repulsive term computed from the protein surface. The backbone is MolDiff, pretrained on the GEOM-Drug dataset of drug-like conformers, and it is conditioned without retraining or altering its weights.

That choice changes what data a structure-based generator needs. Conditional models such as TargetDiff learn from protein-ligand complexes, a comparatively scarce and biased resource. MolSnapper keeps the chemistry of a model trained on the far larger space of molecules alone and acquires binding-site awareness only at sampling time, so one checkpoint serves every target.

#Key Features

• Pharmacophore conditioning at sampling time: Fixed donor and acceptor positions enter the parameterized reverse-diffusion distribution through a position mask, so atoms drift toward them and are anchored there in the final steps.
• Clash guidance against the protein: A soft-min surface term contributes a gradient that pushes generated atoms away from pocket atoms — necessary precisely because the backbone was never trained on protein data.
• Atom-type constraints: The element type at each reference position can be pinned, so a design brief specifies the interaction to mimic rather than only where to place an atom.
• Pharmacophore satisfaction filtering: A ligand counts only when its atoms fall within 1 Å of the requested points with a matching feature family, with sampling retried up to three times per reference.
• Scaffold growing and hopping: Because the constraint set is an arbitrary collection of typed 3D points, an existing hit can be held fixed while the rest of the molecule is regenerated around it.

#Technical Details

MolSnapper builds on MolDiff's E(3)-equivariant graph neural network, in which atom positions are denoised as Gaussians and atom types as categorical variables, with a separate pretrained bond predictor guiding bond formation. Conditioning enters in two places: the predicted mean for constrained coordinates is interpolated toward the fixed pharmacophore positions along the pre-defined noise schedule, and the clash-loss gradient is added to the reverse-generation mean, scaled by a user-set coefficient. Reference positions are used unnoised, and sampling initializes atoms randomly around the pharmacophore points; an ablation shows that pairing gives the best shape-and-color similarity.

Evaluation covers CrossDocked2020 (73 filtered test complexes) and the experimental Binding MOAD set (43). On CrossDocked, MolSnapper reaches a 58% PoseBusters pass rate against 27% for SILVR, the closest training-free conditioning method, with SC_RDKit similarity to the reference ligand of 0.721 versus 0.586 and a success rate of 45% versus 7%. Against DiffSBDD, which is trained on complexes, it passes 58% versus 47% on CrossDocked and 57% versus 31% on Binding MOAD, in both cases without redocking the generated poses. Two case studies follow published campaigns: metallo-β-lactamase inhibitor design, where 1,139 unique filtered molecules recapitulated the reference compound's interactions in 55.5% of cases, and scaffold growing on a COVID-19 3CLpro hit.

#Applications

MolSnapper suits the stage of a project where structural knowledge already exists: a fragment hit whose binding mode is solved, a series with a known essential contact, or a target whose pharmacophore is inferred from published inhibitors. Constraints can be picked manually or lifted from experimental data such as a fragment screen, and ligands are generated to respect them without commissioning a target-specific model. Because the poses clear physical-plausibility checks without redocking, output goes straight into triage and scoring rather than through a refinement step that would erase the conditioning.

#Impact

MolSnapper is a worked demonstration that a frozen, unconditional generator can be given a capability it was never trained for, with one mechanism serving de novo design, scaffold hopping and scaffold growth. Pharmacophore-conditioned generators such as PharmaDiff and PharmacoBridge build the constraint into training and work from a hypothesis alone; the contribution here is to combine the pharmacophore with the pocket surface while leaving the generator untouched. The limits are equally clear: the method inherits MolDiff's chemical space, conditioning lowers the unconditional model's raw PoseBusters pass rate from 91% to 58%, and every result is in silico, with no synthesis or assay of the generated compounds. The work was published in the Journal of Chemical Information and Modeling.