PAFlow

Structure-based drug design asks a generative model to place a drug-like molecule inside a known protein pocket, producing continuous 3D coordinates and discrete atom types together. Two problems have limited the generators built for this task. Diffusion models follow a highly stochastic denoising trajectory, so molecular quality varies run to run and sampling needs on the order of a thousand steps. And non-autoregressive methods decide how many atoms to generate by drawing from a distribution fitted to reference ligands, ignoring the geometry of the pocket being targeted.

PAFlow addresses both. It replaces stochastic denoising with flow matching, integrating an ordinary differential equation from noise to molecule with an Euler solver, and derives a new conditional flow matching formulation so that discrete atom types travel in the same framework as coordinates rather than through a bolted-on categorical mechanism. On top of that sit two prior-knowledge components: an interaction predictor that steers the vector field toward higher-affinity poses, and an atom number predictor that sizes the molecule from pocket geometry alone.

The model comes from Shikui Tu and Lei Xu's group at Shanghai Jiao Tong University, posted to arXiv in September 2025 and accepted at NeurIPS 2025. It joins a growing group of flow-based pocket-conditioned generators in the catalog alongside MolPIF, DrugFlow, and FLOWR.root.

#Key Features

• Flow matching for both modalities: Atomic coordinates follow the variance-preserving probability path, while a newly derived conditional flow matching form handles discrete atom types, giving a single ODE-based generation process instead of a hybrid scheme.
• Prior interaction guidance: A protein-ligand interaction predictor trained on normalized CrossDocked2020 binding affinities reads the network's final atom embeddings and nudges the coordinate vector field toward tighter-binding poses at every sampling step.
• Pocket-aware atom count: An atom number predictor takes pocket volume, surface area, protein atom count, and space size and predicts how many atoms the ligand should have, removing the dependence on reference-ligand size priors.
• Noise-injected size prediction: Adding small Gaussian noise to the predicted atom count regularizes against overconfident point estimates and improves generated affinity over both a noise-free predictor and predefined sampling.
• Fast sampling: Fifty ODE steps suffice, against roughly a thousand for diffusion baselines; generating 100 molecules takes 717 seconds versus 3,968 for TargetDiff.

#Technical Details

The vector field is parameterized by an SE(3)-equivariant network of nine equivariant layers, each a transformer with hidden dimension 128 and 16 attention heads, operating on the joint protein-ligand atom graph with the protein center of mass shifted to the origin. The interaction predictor is a two-layer MLP over pooled atom embeddings; the atom number predictor is a separate four-layer MLP trained on 98,000-plus pocket descriptor samples whose volume and surface area come from PyKVFinder. Training uses the standard CrossDocked2020 split — 100,000 complexes filtered at 1 Å pose RMSD and 30% sequence identity, with 100 held-out pockets — on a single NVIDIA A100 40 GB.

Across 100 molecules generated per test pocket, PAFlow reports a mean Vina Score of -8.31 kcal/mol (median -8.92), mean Vina Min of -8.79, and mean Vina Dock of -9.46, with 80.8% of generated molecules binding better than the reference ligand. Against the strongest diffusion baseline, ALiDiff, that is a 17.5% improvement on mean Vina Score, and PAFlow gives the best median affinity on 77% of the 100 targets. Drug-likeness and synthesizability are held rather than improved, at mean QED 0.49 and mean SA 0.57 with diversity 0.71. An ablation stripping out both prior components leaves a model differing from TargetDiff only in using ODE sampling instead of denoising, and it still beats TargetDiff on affinity.

#Applications

The intended use is pocket-directed hit generation: given a protein structure with a defined binding site, a medicinal chemistry team can sample fresh chemotypes without a reference ligand to set molecular size, which makes the model applicable to targets that have structure but no known binders. The tunable step count lets users trade speed for quality — 20 steps still outperforms every published baseline on affinity metrics — suiting virtual screening campaigns where throughput across many pockets matters more than per-molecule polish.

#Impact

PAFlow shows that flow matching, with an appropriate probability path for each variable type, outperforms diffusion on the standard structure-based drug design benchmark while cutting sampling cost by an order of magnitude, and that conditioning molecule size on pocket geometry is a cheap fix for a mismatch the field had accepted. Several caveats bound the result. All evaluation is computational, resting on docking scores and cheminformatics proxies rather than measured binding, and the gains are concentrated in affinity while drug-likeness and synthetic accessibility stay at baseline levels. The generator occasionally produces chemically implausible structures, including improper double bonds and large or fused rings, and it treats the protein as rigid. Training and sampling code are public, but the repository ships no LICENSE file and pretrained checkpoints are distributed through Google Drive rather than a model hub.