MolPIF

Structure-based drug design asks a generative model to build a small molecule inside a known protein pocket, which means emitting two very different kinds of variable at the same time: continuous 3D atomic coordinates and discrete atom types. Diffusion and flow-matching approaches have generally handled this mismatch by running two separate mechanisms in parallel, or by forcing categorical atom types into a Euclidean space where distances between element labels are not meaningful — a modelling seam that surfaces as implausible geometry and unstable atom-type assignment.

MolPIF closes that seam by moving the generative process out of sample space and into parameter space. Rather than transporting a noisy point cloud toward a molecule, Parameter Interpolation Flow (PIF) interpolates the parameters of the distributions those variables are drawn from: a Gaussian mean per coordinate and a Dirichlet concentration vector per atom type. Interpolating the Gaussian parameters recovers Wasserstein-2 optimal transport for the continuous coordinates, while interpolating the Dirichlet parameters traces Fisher-Rao geodesics on the probability simplex — the geometry appropriate to categorical variables. One mechanism, applied consistently to both modalities.

The model comes from Lingang Laboratory in Shanghai with collaborators at the Shanghai Institute of Materia Medica, ShanghaiTech, Fudan University, and Shanghai Jiao Tong University. It was posted as a preprint in July 2025 and published in Bioinformatics in 2026. It sits alongside catalog entries such as FLOWR.root, which applies SE(3)-equivariant flow matching in sample space, and Apo2Mol, which co-generates the ligand and a relaxed pocket.

#Key Features

• Unified parameter-space generation: Coordinates and atom types are produced by a single interpolation mechanism, avoiding the hybrid continuous/discrete schemes used by earlier pocket-conditioned generators.
• Flexible prior distributions: Because the flow acts on distribution parameters, the prior is a free choice. A Laplace prior yields substructure distributions closer to the reference molecules than the default Gaussian.
• Geometry-enhanced learning: Training mixes full-molecule noising with a masked variant in which a random subset of ligand atoms is left clean, borrowing from masked autoencoding and inpainting so the network learns to read local atomic context.
• Zero-shot pocket conditioning: A single pretrained checkpoint generates molecules for arbitrary new pockets given a protein structure and a reference ligand to define the site, with no per-target retraining.
• Lead optimization modes: A second checkpoint supports fragment growth, linker design, scaffold hopping, and side-chain decoration by fixing selected atom indices and attachment points; across three case-study targets, 25.6–63.3% of optimized ligands scored better than the reference compound.

#Technical Details

An SE(3)-equivariant network predicts the distribution parameters at each step of the flow, conditioned on the pocket atoms. Training uses the standard CrossDocked2020 benchmark split — 99,900 protein-ligand training pairs filtered at 1 Å RMSD, with 100 held-out test pockets separated at 30% sequence identity — the same preprocessed data used by TargetDiff, which makes the comparisons directly interpretable.

On the 100 test pockets, MolPIF reaches a mean Vina Score of -6.64 kcal/mol against -6.59 for MolCRAFT, -5.75 for the autoregressive baseline, -5.47 for TargetDiff, and -5.14 for Pocket2Mol; mean Vina Min is -7.41 and mean Vina Dock is -8.09, the best in each column. Drug likeness and synthesizability hold up alongside the affinity gains, at mean QED 0.59 and mean SA 0.72, with diversity of 0.72. Geometric quality metrics include a stable-atom ratio of 0.9643, a stable-molecule ratio of 0.5366, and a bond-length distribution JSD of 0.2332 against the reference set. All evaluation is computational, using docking scores and cheminformatics proxies rather than measured binding.

#Applications

The practical use is pocket-directed hit generation: given a crystal or predicted structure and a reference ligand marking the site, a medicinal chemist can sample fresh chemotypes and then elaborate a chosen hit in place. Because the same weights apply to any pocket, the model fits early-stage campaigns where a target has structure but little chemical matter, and the masking interface covers the narrower jobs that dominate real projects — growing a fragment into an adjacent subpocket, linking two bound fragments, or swapping a scaffold while holding key contacts fixed.

#Impact

MolPIF's contribution is primarily methodological: it shows that the awkward continuous/ discrete split in molecular generative modelling can be dissolved by flowing over distribution parameters instead of samples, and that doing so buys principled transport geometry for both variable types plus the freedom to swap priors. The gains over the strongest prior baseline are real but incremental on binding affinity, and larger on drug-likeness and geometric fidelity. Two limitations bound the claims: no experimental validation accompanies the in-silico benchmarks, and the checkpoints are distributed through Google Drive rather than a model hub, with the repository declaring GPL-3.0 in its README but shipping no LICENSE file. Peer-reviewed publication, released training code, and two public checkpoints nonetheless make it a reproducible reference point for parameter-space generative modelling.