Full-atom flow matching model that generates a ligand and the induced-fit holo pocket together, starting from an apo binding site.
No providers recorded yet. Browse all providers
Generative structure-based drug design almost always assumes the pocket holds still. A model is handed a fixed binding site and asked to invent a molecule for it, which ignores induced fit: side chains and backbone rearrange as a ligand binds, and the holo conformation often differs enough from the apo one that molecules designed against the latter do not fit. Molecular dynamics can sample the missing conformations, but crossing the energy barriers between them is too expensive for a design loop.
DynamicFlow, accepted as a poster at ICLR 2025, learns the transition instead. It comes from a collaboration led by Tsinghua University with the Chinese Academy of Sciences, Peking University, and the drug discovery company StoneWise. The model is a full-atom flow taking an apo pocket plus noise-sampled ligand atoms at time zero and integrating them to a holo pocket plus a bound 3D molecule at time one, generating ligand chemistry, ligand geometry, and pocket motion in one pass rather than in stages.
Flow matching had already been applied to protein conformational change on its own — as in AlphaFlow, which samples ensembles for a protein in isolation. DynamicFlow's departure is coupling that motion to the molecule causing it, which yields a second usable product: the predicted pocket conformations can be handed to conventional rigid-pocket generators as better inputs than the apo structure.
Training data derives from MISATO, which supplies 8 ns of recorded molecular dynamics trajectory for 16,972 protein-ligand complexes taken from PDBbind. Removing oligopeptide ligands leaves 12,695 complexes; pockets are residues within 7 Å of the ligand across 100 trajectory frames, clustered with GROMACS at a 1 Å heavy-atom RMSD threshold and cut to the top ten clusters per complex. Apo counterparts come from AlphaFold Protein Structure Database entries mapped through UniProt, yielding 7,528 aligned apo pockets. Filtering on average ligand RMSD below 3 Å gives the final set: 5,692 apo/holo pairs carrying 46,235 holo-ligand conformations. An atom-level SE(3)-equivariant graph network processes ligand and protein atoms together and pools its pocket features to residue level for an invariant point attention Transformer predicting frames and torsions; the model totals 15.9M parameters and trains on seven weighted losses including a protein-ligand interaction term.
Evaluation uses 50 held-out targets, generating 100 molecules each and scoring the best by AutoDock Vina in minimize mode, with every method given the apo pocket. DynamicFlow-SDE reports a Vina score of -7.65 with QED 0.53 and 52.5% of designs beating the reference ligand; the ODE variant, -7.28 and 51.0%. Reference ligands average -7.84. Baselines on the same apo inputs reach -5.50 for Pocket2Mol, -5.09 for TargetDiff, and -7.55 for IPDiff. Substituting DynamicFlow's refined pockets lifts TargetDiff to -9.00 and its high-affinity rate from 44.0% to 61.2%, and IPDiff to -11.04 and 51.0% to 83.7%. Generating ten molecules takes 35 seconds at 100 function evaluations, against 156 seconds for TargetDiff and 980 for Pocket2Mol.
The model is built for targets where the informative structure is the one that does not exist yet: a protein solved only in its apo form, a cryptic pocket that opens on binding, or a site where docking against the unbound conformation has repeatedly disappointed. It offers two entry points — sampling ligands directly, or using it as a preprocessing step that turns an apo structure into a holo-like one before an established design or docking pipeline. The second mode is lower-friction, since it slots into workflows already in place.
DynamicFlow reframes structure-based design as generating a complex rather than a ligand for a fixed receptor, and the pocket-refinement result is the strongest evidence the reframing pays: two unrelated generators improve substantially on inputs the model produces, a transferable gain rather than a self-contained benchmark win. The bounds are clear. Evaluation covers 50 targets scored entirely in silico, with Vina, QED, and synthetic accessibility standing in for properties only assays can settle; training conformations come from 8 ns simulations, short relative to large conformational transitions; and no code or trained weights have been released, so the results cannot be independently reproduced.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.