Loading…
Protein complex structure prediction conditioned on sparse restraints from crosslinking, covalent labeling, NMR shifts and mutational scans.
No providers recorded yet. Browse all providers
An in-cell crosslinking experiment reports that two lysines came within 30 Å of each other. A covalent-labeling run reports which residues stopped being chemically modified once a partner bound. A deep mutational scan reports which substitutions abolish binding. Each is a hard constraint on where the interface must be, and none of them is a structure — so in a conventional workflow the evidence sits beside the model rather than inside it.
GRASP (Generalized Restraints Assisted Structure Predictor), developed at Changping Laboratory and Peking University with Huawei, builds that channel into the network. It reduces every experiment to two primitives: residue-pair restraints (RPRs), which bound the distance between two named residues, and interface restraints (IRs), which mark a single residue as sitting at the interface. Crosslinking mass spectrometry and NMR NOESY yield RPRs; covalent labeling, chemical shift perturbation and deep mutational scanning yield IRs. Both are fed straight into an AlphaFold-Multimer backbone.
The design choice that matters is that GRASP is fine-tuned once rather than optimized per target. Earlier restrained-docking work from the same group reached the same goal by inverting a frozen predictor with gradient optimization on each complex, which is expensive and capped at roughly 1,200 residues. GRASP instead ships fixed checkpoints and treats restraints as ordinary inference-time inputs, so batch modeling of an entire crosslinked interactome becomes practical.
Inference runs five checkpoints. Four were fine-tuned from the AlphaFold-Multimer v2.3 model-1 checkpoint for up to 22,000 steps on 64 Ascend 910A processors under four added restraint-related loss terms, snapshotted at 8k, 14k, 20k and 22k steps. The fifth was fine-tuned from MEGAFold-Multimer, a MindSpore reimplementation of AlphaFold-Multimer that the authors trained from scratch for 46,200 steps on 128 Ascend 910A chips. Training used the PSP dataset — 570k experimental structures released before 8 October 2021 plus a 760k distillation set — with a 4:1 multimer-to-monomer ratio. Each checkpoint runs with five random seeds, and predictions are ranked by pLDDT and restraint recall. On a benchmark of 713 interfaces across 313 hard complexes, two inter-chain contact RPRs raised mean DockQ to 0.35 with 52.7% of cases above the acceptable threshold of 0.23; 4, 10 and 20 IRs gave 0.24, 0.34 and 0.41. On simulated crosslinks at 2% sequence coverage, GRASP reached 0.21 against HADDOCK's 0.08 and AlphaLink's 0.11. Across 84 SARS-CoV-2 RBD–antibody complexes restrained by experimental deep mutational scanning, it reached a median DockQ of 0.25 and 53.6% success, against 0.07 and 39.3% for AlphaFold 3. A JAX port handles complexes beyond 1,280 residues on A100 GPUs.
The clearest fit is antibody engineering, where antigen–antibody interfaces remain hard to predict from sequence and where escape-mapping data already exists. The second is integrative structural biology: modeling multi-subunit assemblies such as the APOBEC3G–VCBC complex or the KCNQ1–KCNE1 channel from a mixture of crosslinks, mutagenesis, paramagnetic relaxation enhancement and density maps. The third is in-situ interactomics — 158 inter-protein crosslinks from targeted mitochondrial crosslinking were mapped to 144 interactions and modeled in batch, yielding 119 dimer structures that passed confidence and restraint filters.
GRASP appeared on bioRxiv in September 2024 and in Nature Methods in 2025, with code released under Apache 2.0 and weights and evaluation datasets deposited on OSF. It gives experimentalists a route from sparse, noisy measurements to a testable complex model without hand-built docking protocols. Its limits are worth stating plainly: the backbone is an AlphaFold-Multimer-era network rather than an all-atom generative one, restrained prediction without restraints performs no better than the base model, and AlphaFold 3 still wins on 38 of the 84 deep mutational scanning cases — prompting the authors to propose a joint protocol that accepts an AlphaFold 3 model when it already agrees with the experimental data and falls back to GRASP when it does not. Changping Laboratory and Huawei have filed for a patent covering the method.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.