Loading…
Flexible protein-ligand docking that repacks the pocket side chains from the backbone and the ligand graph before a physics sampler places the ligand.
No providers recorded yet. Browse all providers
Most docking programs hold the receptor rigid, which is fine when a holo crystal structure of the very complex you are modelling is already in hand — and useless in the situation docking actually exists for. Given an apo structure, or a predicted one from AlphaFold 2 or ESMFold, the pocket side chains sit in rotamers that were never induced by the ligand, and they block the site. No amount of ligand sampling recovers from a tyrosine pointing the wrong way.
ApoDock, developed at Xiamen University with the Tan Kah Kee Innovation Laboratory and Putian University, treats that as a packing problem to solve before docking rather than a flexibility term to bolt onto the search. Two learned modules bracket a conventional sampler. ApoPack generates plausible holo side-chain conformations for the pocket from the backbone plus a 2D graph of the ligand — the ligand's identity conditions the rotamers, but its pose is not yet known. Gnina or Smina then docks into each repacked pocket with ordinary physics-based scoring, and ApoScore, a mixture density network, re-ranks the resulting poses.
The modular design is the point. End-to-end deep learning dockers such as DiffDock place ligand atoms directly and frequently produce poses that score well on RMSD but violate basic chemistry, with clashes and distorted bond geometry. By keeping a physics-based sampler in the loop and using the networks only for pocket conformation and pose ranking, ApoDock gets induced-fit behaviour without giving up physical validity.
ApoPack encodes the ligand as an atom graph updated by message passing, encodes pocket residues as nodes linked to their 30 nearest neighbours, and lets the two exchange information through cross-attention layers before an MLP head emits the chi-angle distributions. ApoScore pairs the same ligand encoder with a geometric vector perceptron graph network over the pocket. Both were trained on PDBbind v2020 under an EquiBind-style time split — 16,334 training complexes, 964 for validation, and 363 deposited after 2019 held out — and both ship as fixed checkpoints that run on new complexes without refitting.
On that time-split test set ApoDock reaches top-1 success (ligand RMSD below 2 Å) of 58.56% with Gnina and 56.90% with Smina while predicting side chains rather than reading them off the holo structure. On PoseBusters it reaches 52% top-1, 48% once pose-validity checks are applied — the small gap being the signature of physics-based sampling. The harder tests are the realistic ones: on the Apo2Holo set of 32 apo-holo pairs it lifts Group 1 accuracy from 14.8% for Gnina alone to 42.6%, and across a cross-docking benchmark of 94 pockets and 4,399 ligands from 38.8% to 46.6%. ApoScore alone scores 95.1% top-1 docking power on CASF-2016 with an enrichment factor of 29.15 at 1%. On ESMFold-modelled structures it reports 28.5% top-1 success, the best of the methods compared and ahead of DiffDock at 21.7%.
The system targets the ordinary virtual screening situation where no holo structure exists: docking libraries against apo crystal structures, predicted models, or a holo structure solved with a different ligand. ApoPack is also usable on its own as a fast ligand-aware repacking step ahead of any docking or scoring pipeline, and ApoScore as a drop-in rescoring function for poses generated elsewhere.
ApoDock is a reminder that the productive division of labour in structure-based drug design may not be end-to-end. Decades of tuning went into physics-based samplers and their treatment of steric feasibility; what they cannot do is anticipate the receptor conformation the ligand will induce, and that is what a network trained on complexes does well. The main limitation is scope: ApoPack moves side chains only, and on targets whose backbone shifts by more than 1.5 Å between apo and holo forms the approach shows no advantage. Code and all checkpoints are released under the MIT license, though running the pipeline requires downloading a Gnina or Smina binary separately.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.