Loading…
Two-stage backbone generator that designs protein domains separately, then weaves them into one long chain with an SE(3) diffusion assembly module.
No providers recorded yet. Browse all providers
De novo backbone generators are reliable at 100 to 300 residues and steadily less so beyond that: at 800 residues, RFdiffusion designs refold to a mean self-consistency TM-score of 0.49. The failure is not in the local structure but in how the parts sit together — long designs come out with fewer distinguishable domains and worse interfaces than native proteins of the same length. Nature never faces this, because it does not build long chains from scratch: roughly 500 domain folds in CATH are recombined into hundreds of thousands of structures, and the interfaces between those domains are what makes new functions possible.
ProteinWeaver takes that recombination literally and splits backbone design in two. First the target length is partitioned into domains, each generated independently by an existing backbone model — the regime where those models are strongest. Then a purpose-trained SE(3) diffusion module assembles the pieces, conditioned on a spliced Cα distance map in which each domain's own distance matrix sits on the diagonal and every inter-domain entry is masked to -1. The model has to invent exactly the part that was missing — the placement of the domains and the interfaces between them — and assembly is flexible rather than rigid, so domains deform as they pack.
The work comes from ByteDance Research, released as an arXiv preprint in November 2024 alongside an ICLR submission. It shares its name with an unrelated Gene Ontology network-visualization webtool from the Ritz lab at Reed College.
The assembly module builds on FrameDiff's SE(3) diffusion formulation over per-residue rigid-body frames, with folding blocks combining an IPA module, a backbone update, and an edge update; the spliced distance map initializes the edge representation, at 256 dimensions for single representations and 128 for pairs. Pretraining used 5,835 multi-domain structures filtered from 22,728 single-chain PDB monomers of 60 to 512 residues below 5 Å resolution, domains parsed by Unidoc and refolded through ESMFold — an ablation shows training on refolded rather than crystallographic domains is what lets the module tolerate generated inputs. Alignment used 100 TMalign-deduplicated domains recombined into 10,000 spliced maps, three samples each, with winner and loser chosen by self-consistency TM-score. Training took 10 days on eight 32 GB V100 GPUs. Assembling RFdiffusion-generated domains yields a median scTM of 0.92 and a mean interface scTM of 0.80; on full backbone design, mean scTM reaches 0.86 against RFdiffusion's 0.76 at 500 residues, 0.79 against 0.66 at 600, and 0.68 against 0.49 at 800, with better novelty scores across 100 to 300 residues as well.
The immediate use is long multi-domain backbone design, where a pipeline is otherwise forced down to shorter targets. Because assembly is conditioned on named domains, it also combines functions that already exist: the paper assembles nanobodies into pairs that could bind two antigens at once, and joins substrate-recruitment scaffolds to enzymes in the geometry shown to improve selectivity. Protein engineers working on multi-specific binders, tethered enzyme systems, and other modular constructs are the natural audience, with the caveat that stage one requires a separate backbone generator whose quality carries through to the assembled result.
ProteinWeaver names domain assembly as a design task in its own right and supplies a benchmark for it, including the deliberately hard case of randomly drawn CATH domains with no reason to fit together. Its wider contribution is methodological: preference alignment, borrowed from language model post-training, applied to a structural diffusion model and beating supervised fine-tuning on the same data — a general recipe for objectives easy to score but hard to write as a loss. The evidence is entirely in silico, scored by refolding rather than experiment, and the enzyme and antibody cases are framed as proof-of-concept. Neither code nor weights have been released, so the results stand as a demonstration of the paradigm rather than a tool others can run.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.