TorchCraft: Unified binder design by inverting an all-atom structure predictor
TorchCraft Team, Yu Liu, Zhouhanyu Shen, Zhengyi Li, Xikun Huang, Jiaqi Liu, Shuxian Gao, Qilin Yu, et al.
cs.AI · cs.CE · cs.LG
Quick verdict
Unified AF3-inversion binder design; wet-lab minibinder/VHH validation is credible, but self-scored benchmarks and unreported hit rates temper the claims.
Summary
TorchCraft is a sequence-optimization framework that inverts a frozen all-atom structure predictor — AlphaFold 3 weights executed through TorchFold, the authors' PyTorch/NPU reimplementation — to design binders in four formats: minibinders, framework-conditioned VHHs, head-to-tail and disulfide-linked cyclic peptides, and ligand-conditioned pockets. A four-stage schedule (soft warm-up, logit-to-soft transition, temperature annealing, straight-through) updates sequence logits against confidence, distogram-contact, geometric, and sequence-prior losses. Experimentally, minibinders against IFNAR2, IL-7Rα, PDGFRβ, and PD-L1 showed apparent BLI affinities of 60.2, 8.55, 2.04, and 23.3 nM, and VHHs against BHRF1, EFNA5, PDGFRβ, and S100A4 spanned 24.3–354 nM, all without post hoc sequence redesign. Computational benchmarks report pass rates competitive with RFdiffusion, RFantibody, BoltzGen, Germinal, and RFpeptides. The main limitations are that headline benchmark scores (TorchScore) come from the same AF3 weights used for design — Rosetta distributions in supplements only partially offset this; experimental evidence is BLI-only against largely bivalent Fc-fusion targets with no structural or epitope validation; per-campaign hit rates and numbers tested are unreported; and cyclic-peptide plus ligand results remain computational. Code is stated to be available on GitHub.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single frozen all-atom predictor (AF3 weights via TorchFold) can be inverted with task-specific masks, losses, and schedules to design minibinders, framework-conditioned VHHs, cyclic peptides, and ligand-binding proteins.
- Raw TorchCraft outputs achieved experimentally measured BLI binding without post hoc sequence redesign: minibinders at 2.04–60.2 nM across four targets and VHHs at 24.3–354 nM across four targets.
- In computational benchmarks, TorchCraft's pass rates are competitive with RFdiffusion, RFantibody, BoltzGen, Germinal, and RFpeptides across protein, VHH, and cyclic-peptide settings.
- Task-specific sequence regularization (amino-acid composition, CDR-restricted IgLM) improves sequence-quality proxies such as ESMFold confidence, SAbDab agreement, and OASis humanness.
Strengths
Unredesigned wet-lab validation in two formats
Eight campaigns (4 minibinder §2.2, 4 VHH §2.3) yielded apparent BLI KDs of 60.2/8.55/2.04/23.3 nM and 24.3/122/268/354 nM; tested sequences were raw TorchCraft outputs (A.2.2), a strong functional demonstration of all-atom predictor inversion.
Fully specified, reusable optimization machinery
Algorithm 1's four-stage schedule, all loss formulas (A.1.4, Algorithms 2–3), per-task hyperparameters (Table S1), and candidate filters (ROG<17, pTM/ipTM>0.7, Rosetta composite in A.2.2) make the protocol unusually reproducible for a methods paper.
Antibody-specific sequence priors with measurable effect
Honest scoping of evidence
Engineering practicality
Weaknesses
Benchmark circularity
Thin experimental statistics
Affinity evidence limited to BLI on bivalent targets
Self-referential pass gate for VHHs
Half the claimed scope is computational only
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
TorchCraft (arXiv:2609.19770; Changping Laboratory) is a predictor-inversion framework for de novo binder design built on a frozen all-atom structure predictor: pretrained AlphaFold 3 weights executed through TorchFold (ref. [21]), the authors' PyTorch/NPU reimplementation. Sequence logits over editable positions are updated by SGD against confidence losses (pLDDT, pTM, ipTM, PDE, PAE/iPAE), distogram-contact objectives, geometric regularizers (ROG proxy, isotropy via Gram-matrix eigenvalues, helix/β terms), and sequence priors (ProteinMPNN-derived composition; CDR-restricted IgLM infilling). One engine serves four task families (Supplementary Table S1): minibinders, framework-conditioned VHHs, head-to-tail and disulfide-linked cyclic peptides, and ligand-conditioned pockets. The headline claims are (i) experimentally measured binding of raw, unredesigned outputs — apparent BLI KDs of 60.2, 8.55, 2.04 and 23.3 nM for minibinders against IFNAR2, IL-7Rα, PDGFRβ and PD-L1 (§2.2), and 24.3–354 nM for VHHs against BHRF1, EFNA5, PDGFRβ and S100A4 (§2.3); (ii) computational pass rates competitive with RFdiffusion, RFantibody, BoltzGen, Germinal and RFpeptides (§2.2–2.4); and (iii) sequence regularization improving sequence-realism proxies.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is TorchCraft?
A computational binder-design framework that optimizes a binder's amino-acid sequence by backpropagating through a frozen all-atom structure predictor (AlphaFold 3 weights run via TorchFold). One optimization engine handles minibinders, VHH nanobodies, cyclic peptides, and small-molecule-binding proteins, differing only in editable positions, constraints, and loss terms.
Did TorchCraft designs work experimentally?
Yes, for two formats. Minibinders against IFNAR2, IL-7Rα, PDGFRβ and PD-L1 showed apparent binding affinities of 60.2, 8.55, 2.04 and 23.3 nM in BLI assays, and VHHs against BHRF1, EFNA5, PDGFRβ and S100A4 ranged from 24.3 to 354 nM — all tested as raw outputs without later sequence redesign. Evidence is BLI-only; the cyclic-peptide and ligand-pocket results are computational.
How is TorchCraft different from BindCraft?
BindCraft inverts AlphaFold2, which covers protein components. TorchCraft inverts AlphaFold 3 through the TorchFold reimplementation, adding all-atom molecular context (ligands, nucleic acids), framework-conditioned VHH design with antibody-specific IgLM priors, and cyclic-peptide and ligand-pocket tasks inside the same optimization framework.
Is the TorchCraft code available?
The paper states that the TorchCraft inference implementation and design protocols are available at github.com/Mingchenchen/TorchX, with a project page at torchx-cpl.github.io. Design datasets, seeds, and scored outputs are not explicitly stated as released.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure panels are raster images; every number cited in this review comes from captions, tables, or body text present in the extraction, not from reading plots.
- Supplementary figures S1–S9 are captioned in the text, but their underlying distributions cannot be independently verified from the extraction.
- The extraction runs unbroken through the Figure S9 caption, which is the final entry in the Appendix Contents; no truncation was detected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.