Loading…
Light-weight rigid protein-ligand docking model adapting the AlphaFold2 architecture, with a binding affinity module for virtual screening.
No providers recorded yet. Browse all providers
A virtual screen weighs millions of candidate molecules against one target, so what decides whether the screen is possible at all is seconds per complex — not the last half-ångström of pose accuracy. Machine-learning docking has mostly moved the other way. Diffusion samplers such as DiffDock and co-folding models that predict protein and ligand together buy pose quality with runtime, taking seconds to minutes per prediction. QuickBind is a deliberate move back across that trade-off: a small, fast model for rigid blind docking.
Its mechanism is a stripped-down AlphaFold 2. Protein sequence and structure features are concatenated with ligand features into a first-order ("single") representation, from which a pair representation is built using pairwise distances, positional encodings and the ligand's adjacency matrix. Both pass through a modified Evoformer stack that drops column-wise self-attention, since no multiple sequence alignments are used, and then through a structure module that iteratively updates ligand heavy-atom coordinates while the protein is held fixed. Small molecules have no canonical backbone from which to build reference frames, so the authors introduce their own: atoms are reordered by RDKit's canonical ranking, and each heavy atom's frame is anchored by its two lowest-indexed neighbours. Updating those ligand frames inside a gated Invariant Point Attention module — rather than reconstructing them passively — is what carries AlphaFold 2's geometry over to a ligand.
The AlQuraishi lab in Columbia University's Department of Systems Biology, with a co-author at Harvard Medical School, presented QuickBind at the Machine Learning in Computational Biology meeting in 2024; it builds directly on the same group's OpenFold codebase.
QuickBind has 3.0 million parameters across 12 Evoformer blocks and 8 structure module blocks, and was trained on two four-GPU nodes at an effective batch size of 16. Training uses the PDBBind temporal split introduced by EquiBind — 16,379 training and 968 validation complexes published before 2019, with 363 test complexes from 2019 onward — under a combined frame-aligned point error loss applied at every structure module block, plus a Kabsch RMSD loss on the ligand. On that test set QuickBind places 28.9% of predictions within 2 Å symmetry-corrected ligand RMSD, and 17.4% on the 144 complexes whose proteins are absent from training; it leads EquiBind, TANKBind and E3Bind while trailing FABind, DiffDock and NeuralPLexer, all substantially slower. On the harder PoseBusters benchmark of 428 post-2021 complexes it beats other rigid docking methods but trails co-folding models. It omits the post-processing step most ML docking methods apply, so its raw poses generally fail the PoseBusters validity checks on bond lengths and angles; force-field energy minimization repairs most of those failures and slightly raises the success rate.
The intended use is early-stage virtual screening, where a fast pose plus an affinity estimate narrows a large library to a shortlist worth expensive treatment. The authors screened PDBBind ligands against five targets — BACE1, galectin-3, HIV-1 protease, KRAS and a tRNA guanine-methyltransferase — and predicted affinities separated known binders from decoys for the first two and K-Ras, with poses under 2 Å for 73% and 86% of BACE1 and galectin-3 ligands. Because the affinity module consumes embeddings rather than crystal structures, it can be trained on affinity databases far larger than the structural record.
QuickBind is as much a research instrument as a screening tool: small enough to serve as a test bed for new docking ideas, and its interpretability analysis is evidence that a pose-prediction network learns physicochemical descriptors rather than memorising complexes. The authors state its limits plainly. Like other ML docking methods it is trained and evaluated on rigid re-docking against the holo structure co-crystallised with the query ligand — not the situation a screening campaign faces. When the input structure deviates substantially, accuracy collapses: KRAS inputs averaging 7.7 Å backbone RMSD from the true conformation yielded only 17% of poses within 5 Å. Model code and weights are released under the MIT license in the lab's repository.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.