Loading…
Pocket-conditioned 3D ligand generation for structure-based drug design, with variance-exploding diffusion that keeps generated geometry low-strain.
No providers recorded yet. Browse all providers
A generative model asked to fill a protein pocket can almost always be pushed toward a better docking score: turn up the conditioning pressure and the sampler finds atom arrangements AutoDock Vina likes. What comes back often carries bond lengths and torsions no chemist would draw, and the strain needed to hold that pose swamps whatever predicted affinity was gained. Image diffusion works over bounded pixels or a latent space, where too much guidance merely overexposes a picture; structure-based drug design denoises in unconstrained Euclidean coordinates, so nothing stops guidance from dragging geometry off the manifold of plausible conformations.
PocketVE, from Peining Zhang and Jinbo Bi at the University of Connecticut, treats that trade-off as a property of the generative backbone rather than something to clean up afterwards. It is a protein-pocket-conditioned variance-exploding (VE) diffusion model for 3D molecule generation: one network, trained once on CrossDocked2020, that samples full-atom ligands inside whatever pocket is handed to it. Three choices separate it from the SE(3)-equivariant diffusion models it shares a lineage with — an EDM-style VE parameterization that keeps coordinate denoising stable across noise scales, classifier-free guidance over several molecular properties at once, and a small Gaussian perturbation of the pocket atoms during training.
The work extends the same group's earlier unconditional VE diffusion model for 3D molecules into the target-aware setting opened by TargetDiff. It reuses the target-aware equivariant network of the TAGMol baseline as its denoiser while replacing the diffusion parameterization, guidance mechanism, sampling procedure and optimizer around it, so the result is a retrained checkpoint rather than a sampling-time wrapper.
Training uses the standard CrossDocked2020 split for target-aware generation — poses filtered at 1 Å RMSD with protein pairs above 30% sequence identity removed — for 400,000 steps at batch size 4, with Muon on the hidden two-dimensional weight matrices and AdamW elsewhere. Evaluation covers 100 test proteins at 100 molecules each, scored under the GenBench3D protocol alongside Vina docking, PoseCheck diagnostics and pocket-permutation controls. Against the TAGMol architectural baseline, PocketVE raises Valid_3D from 58.6 to 80.6 and cuts median strain energy from 457.4 to 127.9, with a median Vina score of −7.89 and a clash-free rate of 94.0%. Relative to TargetDiff, a 10,000-replicate paired bootstrap puts the Valid_3D gain at 2.25 percentage points (95% CI −0.75 to 5.13) and the strain reduction at 178.1 (95% CI 150.4 to 203.4). Sampling the full benchmark takes 234 seconds on a single A100 at guidance scale 5. The guidance sweep shows the trade-off directly: raising the scale improves Vina, QED and SA, but past a moderate setting Valid_3D falls and distributional fidelity against reference ligand geometry drifts.
The model is aimed at hit generation for targets with a solved or modeled binding site and no known chemical series: a medicinal chemist supplies pocket coordinates and gets back 3D poses that go straight into docking, MMFF relaxation and visual triage without a separate conformer-repair step. The property-bin conditioning suits the common case where potency must be balanced against drug-likeness and synthetic accessibility at once, since all three are set by choosing bins rather than by calibrating a classifier per objective.
PocketVE is a preprint awaiting peer review, and no code repository or model weights have been published — the authors state that training, sampling and evaluation code accompanies the paper as supplementary material. Its contribution is less a leaderboard position than a demonstration that geometric quality and conditional steering are coupled, and better managed in the noise schedule and coordinate parameterization than by filtering outputs. The authors are explicit about the limits: classifier-free guidance doubles the network evaluations per guided step, the preferred guidance scale can be target-dependent, Gaussian jitter is a crude stand-in for protein flexibility, and every result rests on docking-based affinity proxies over one filtered CrossDocked2020 split, with no experimental validation.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.