RAPiDock

A small molecule entering a binding pocket has a handful of rotatable bonds; a ten-residue peptide has dozens of backbone and side-chain torsions, and it frequently has no stable fold at all until it binds. That difference is why the docking machinery built for ligands does not carry over. Rigid-fragment search explodes combinatorially over peptide conformers, and folding networks approach the problem from the other side — AlphaFold-Multimer will place a peptide on a protein, but it must be coaxed into producing an ensemble of alternative poses.

RAPiDock, developed at Zhejiang University by Huifeng Zhao and colleagues in the groups of Yu Kang and Tingjun Hou, treats the peptide's flexibility as the object of generation rather than an obstacle to search. Rather than denoising atomic coordinates in Cartesian space, it runs reverse diffusion over a product of the degrees of freedom that actually move: the peptide's rigid-body translation and rotation, its backbone torsions, and its side-chain torsions. This is the formulation DiffDock introduced for rigid ligands, extended to a flexible chain, and it constrains the sampler to geometries that are chemically buildable by construction. Sampling is seeded from three physically favoured backbone conformations — α-helical, extended and polyproline II — so trajectories start inside the region peptides actually occupy. The score network reads a bi-scale graph that represents the system at both atomic and residue resolution, using Clebsch–Gordan tensor products to keep predictions equivariant under rotation.

The model docks: it takes a protein structure (or a pocket) plus a peptide sequence, and returns ranked all-atom complexes. It is not a de novo complex predictor. The work was published in Nature Machine Intelligence in August 2025, with code and weights released a year earlier. It is unrelated to RapidDock, a transformer for protein–small-molecule docking from a different group.

#Key Features

• Non-Euclidean, physically constrained diffusion: Noise is applied to translation, rotation and torsion angles rather than to free coordinates, which compresses the sampling space and keeps bond lengths and angles valid throughout the reverse process.
• Bi-scale graph representation: Atoms and residues are modelled as two coupled levels of the same graph, so side-chain detail is available where it matters without paying all-atom message-passing costs across the whole protein.
• 92 residue types: Post-translationally modified residues (phosphoserine, phosphotyrosine, acetyl-lysine, hydroxyproline), D-amino acids and other non-canonical building blocks are first-class inputs, written inline in the peptide sequence as bracketed codes.
• Separate local and global checkpoints: One trained model docks into a known pocket; a second searches an entire protein surface, recovering multiple distinct binding sites rather than one.
• Library screening from a table: A table of protein–peptide pairs runs as one batched job, which is what makes structure-resolved screening of peptide libraries practical.

#Technical Details

The score model is an E(3)-equivariant graph convolutional network built on Clebsch–Gordan tensor products, four convolution layers deep, with a 5 Å intra-molecular interaction radius and a separate diffusion schedule for each degree of freedom. Protein and peptide residues are embedded with ESM-2 650M; when only a protein sequence is supplied, ESMFold folds it first. Inference runs 16–20 denoising steps, after which generated poses are relaxed and ranked with Rosetta's ref2015 scoring function through PyRosetta. Two checkpoints are distributed, one for local and one for global docking, at 56.6 MB each. On RefPepDB-RecentSet, a time-split test set of 523 complexes, the authors report a 93.7% top-25 success rate — 13.4% higher than AlphaFold-Multimer — at roughly 0.35 seconds per complex, about 270 times faster. Performance is also reported on PepSet, a benchmark of unbound structures.

#Applications

Peptide therapeutics are the immediate target: the paper works through four systems that stand in for common discovery problems — phosphopeptide recognition by the Polo-like kinase 1 polo-box domain, specificity differences between the N- and C-terminal SH2 domains of SHP2, multi-site nuclear localization signal binding on importin-α, and peptide–HLA complexes, where binding mode and affinity both matter for T-cell work. Modified-residue support means phosphopeptides and D-amino-acid analogues can be modelled without leaving the tool, and the batch interface turns structure prediction into a screening step rather than a follow-up to one.

#Impact

RAPiDock addresses a gap structure prediction had not closed: peptide binding modes at a throughput that supports library-scale work, with code under MIT and weights and evaluation datasets deposited on Zenodo. Its constraints are worth stating plainly. A protein structure is required, so global docking on a novel target inherits whatever error ESMFold contributes. The authors implemented a learned confidence model during development but dropped it from the final workflow in favour of Rosetta rescoring, so pose ranking depends on PyRosetta and on a classical energy function rather than on the network itself. And the released code covers inference only — the training pipeline is not part of the repository.