Loading…
De novo protein backbone diffusion model reaching 92% designability without structure-prediction pretraining, sampling in 100 steps on one GPU.
No providers recorded yet. Browse all providers
Protein backbone diffusion has looked like a trade. Borrow a large structure-prediction network — RFdiffusion fine-tunes RoseTTAFold2 — and most generated backbones can be realized by some amino acid sequence, but the generator inherits weights built for a different task and an architecture that is awkward to dissect. Train from scratch, as FrameDiff, Genie and Chroma do, and the model is self-contained but far fewer of its backbones are designable.
Proteus closes that gap by importing the operation that gives structure predictors their accuracy rather than their weights. AlphaFold 2's Evoformer enforces the triangle inequality on pairwise residue distances through triangle attention and triangle multiplication, at an O(N³) cost that is impractical inside a denoiser called at every timestep. Proteus computes that update over a k-nearest-neighbour graph — each residue attends across its K edges instead of all N — which reduces the cost to O(NK²), and gates it with a structural bias built from radial basis functions of the third edge's interatomic distances. Because that bias is read off the current noisy backbone, edge representations are informed by the structure under construction rather than only the reverse.
The model was developed in Longxing Cao's laboratory at Westlake University, with Zhejiang University and Duke University, posted to bioRxiv in February 2024 and published at ICML 2024. The name is shared with two unrelated catalog entries: Proteus, a cell-free DNA gene-expression model from Fred Hutchinson Cancer Center, and NVIDIA's Proteina, a separate backbone generator that benchmarks against this one.
Proteus is a 19.8M-parameter SE(3) diffusion model built on FrameDiff's formulation: each residue is a rigid frame, forward diffusion is Brownian motion on SO(3) and R³, and sampling is Euler–Maruyama integration of the reverse SDE. Training used Protein Data Bank entries with an August 2023 cutoff, splitting oligomeric assemblies into constituent chains rather than discarding them, deduplicating by UniProt ID at 80% sequence overlap, and filtering to 60–512 residues and at most 50% DSSP-assigned loop content — 50,773 chains. On unconditional monomer generation averaged over lengths 100–600, with eight ProteinMPNN sequences per backbone refolded by ESMFold, Proteus reaches 0.921 designability against 0.705 for RFdiffusion, 0.405 for FrameDiff, 0.349 for Genie and 0.174 for Chroma, at a structural diversity of 0.235 where RFdiffusion reaches 0.328. An ablation walks the success rate from 36.9% for a FrameDiff-like baseline to 92.1%, with oligomer augmentation, the structure featurizer, graph triangle attention and triangle multiplication each contributing.
The model serves the first stage of de novo design, where a scaffold is sampled before a sequence is chosen for it: generate backbones at a target length or chain composition, design sequences onto them with ProteinMPNN, then filter by refolding. Cost matters as much as accuracy here — Proteus yields a designable 800-residue backbone in roughly 196 seconds of expected sampling time, a length at which RFdiffusion, Genie, FrameDiff and Chroma each returned none within 200 samples. Large scaffolds, multi-domain proteins and designed nanoparticle components therefore become tractable on modest hardware.
Sixteen designs were synthesized and expressed in E. coli — twelve of 300 residues and four of 500. Nine of the twelve and three of the four eluted as monodisperse size-exclusion peaks at the expected molecular weight, and circular dichroism showed secondary structure matching the design models and retained at 95 °C. Wet-lab confirmation at lengths where competing backbone diffusion models produce little that is designable is what makes the architectural claim credible, and later backbone generators including Proteina now use Proteus as a designability comparator. Two limits are worth stating. Structural diversity trails RFdiffusion, so higher designability is bought partly with a narrower sampled distribution. And the released code covers inference only: the repository ships the paper checkpoint under an MIT license but no training script or curated training set, so the model cannot be retrained from what is published.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.