TorchX Changping Laboratory

Open-source · GPU & Ascend NPU

TorchX

A Unified Framework for
Predicting, Designing, and Scoring Biomolecular Interactions

TorchFold Paper TorchCraft Paper Code Web Server Hugging Face

1. TorchFold

Improving antibody–antigen structure prediction through large-scale distillation of sequence pairs

Mingchen Chen, Yunlong Cao, and TorchFold Team

Changping Laboratory, Beijing, China

Corresponding authors: mingchenchen@cpl.ac.cn, ylcao@cpl.ac.cn

Changping Laboratory, 2026

Paper Hugging Face

Overview of TorchFold

TorchFold improves antibody–antigen prediction through three training components:

  • An interface-specific loss concentrates the coordinate objective on the binding surface.
  • A revised stage-2 noise schedule moves from stronger corruption to structural refinement.
  • Two rounds of distillation convert antibody–antigen sequence pairs into confidence-filtered structural labels.
TorchFold overview
Figure 1. TorchFold overview: architecture, interface supervision, revised noise schedule, and two-round sequence-pair distillation.

Two-round sequence-pair distillation

Accepted distillates add predicted epitope coverage and interface-contact patterns beyond those represented in SAbDab, broadening the recognition geometries available during training.

The from-scratch model was trained on Huawei Ascend NPUs and reached 34.0% ranked FoldBench success. It then predicted sequence pairs from computational designs, patent and private discovery collections, ASD, and CoV-AbDab; high-confidence complexes became first-round labels. Fine-tuning on these labels and experimental structures reached 65.9%.

The first-round model revisited medium-confidence candidates. Newly accepted structures expanded the final Ab–Ag training collection to about 24.5k examples and raised success to 70.1%.

FoldBench success versus total Ab-Ag training examples
Figure 3. Two-round antibody–antigen distillation. FoldBench Ab–Ag success versus total Ab–Ag training examples (×10³). From-scratch 5.0k, 34.0%; stage 1 22.9k, 65.9%; stage 2 24.5k, 70.1%.

Antibody–antigen structure prediction

TorchFold was compared with AlphaFold3, Protenix-v2, and OpenDDE using a shared AF3 MSA pipeline. Success is defined at DockQ > 0.23; ranked and oracle results report the top-scored and best-sampled interfaces, respectively.

  • Ranked FoldBench-AB success is 70.1%, with a substantial share of medium- and high-quality interfaces.
  • The same evaluation is reported on CPL-28, PXMeter-AB, 2026ARK-AB, and AbAg2526 under identical settings.
  • Oracle results exceed ranked results across all collections, showing that additional successful interfaces are present in the sampled ensembles.
  • As a checkpoint-transfer experiment, AF3-TF loads the TorchFold-AbAg checkpoint into a compatibility-adapted AF3 implementation. Its success rates closely track TorchFold across all five benchmarks, showing that the learned antibody–antigen prediction capability is retained across implementations.
DockQ success on five Ab-Ag benchmarks
Figure 4. Antibody–antigen DockQ success. (a) Ranking-based and (b) oracle-based success on FoldBench-AB, 2026ARK-AB, AbAg2526, CPL-28 and PXMeter-AB. AF3-TF denotes AF3 loaded with the TorchFold-AbAg checkpoint.

DockQ success-rate curve

Cumulative DockQ curves show that a substantial fraction of top-ranked predictions on FoldBench-AB, CPL-28, and PXMeter-AB fall in the medium- and high-quality regimes, not only above the acceptable cutoff. The AF3-TF curves remain close to TorchFold across the five benchmarks.

Cumulative DockQ success-rate curves
Figure 5. Cumulative DockQ success-rate curves. Fraction of interfaces whose top-ranked prediction reaches DockQ ≥ the cutoff.

Test-time scaling

More inference-time samples improve both ranked and oracle FoldBench-AB success. At 100 seeds, TorchFold reaches 72.4% ranked success and 93.6% oracle success, showing substantial sampling headroom.

ipTM is a precise filter of successful interfaces

Across five pooled benchmarks, ipTM provides a precise filter for successful interfaces:

  • At ipTM ≥ 0.8, 292/297 predictions have DockQ > 0.23 (98.3%), and 288/297 exceed DockQ 0.49.
  • Mean and median DockQ are 0.80 and 0.84 in this high-confidence subset.
  • At ipTM ≥ 0.7, precision remains 94.3% (416/441).
ipTM versus DockQ
Figure 7. Ranked ipTM versus DockQ, pooled across benchmarks. Rank-top-1 antibody–antigen pair ipTM against DockQ for AlphaFold3, Protenix-v2, OpenDDE and TorchFold (956 interfaces).

Accuracy by antigen length

TorchFold recovers 74% of interfaces for antigens up to 50 residues and 38% beyond 600 residues. Accuracy decreases with antigen length across the evaluated models.

DockQ success versus antigen length
Figure 8. DockQ success versus antigen length. Interfaces pooled across FoldBench-AB, 2026ARK-AB, AbAg2526, CPL-28 and PXMeter-AB, binned by antigen residue count.

Case studies

Three crystallographic examples illustrate accurate epitope localization and binding orientation. Experimental coordinates are gray, TorchFold cyan, and OpenDDE green.

  • 7TXW: both Pfs25 interfaces are recovered (DockQ 0.90 and 0.85).
  • 8AV2: the crystallographic leptin-receptor site is recovered (DockQ 0.88).
  • 9JBQ: the PcrV binding mode is recovered (DockQ 0.96).
Three antibody-antigen case studies
Figure 9. Recognition of three antibody–antigen epitopes. Top, TorchFold (cyan); bottom, OpenDDE (green), each aligned to the experimental antigen and overlaid on the crystallographic antibody (gray).

TorchScore for antibody–antigen structure evaluation

TorchScore bypasses diffusion generation and evaluates supplied coordinates through the TorchFold confidence pathway. It uses the same antibody–antigen model weights, with no separately trained network.

  • On AbAg PDB55, TorchScore ipTM correlates with DockQ at Pearson 0.869 (RMSE 0.187).
  • The score can be used to screen and rank candidate complexes without regenerating structures.
TorchScore on AbAg PDB55
Figure 10. ipTM versus DockQ on AbAg PDB55. TorchScore loaded with the TorchFold Ab–Ag checkpoint. Dashed lines: Acceptable, Medium and High DockQ thresholds.

2. TorchCraft

TorchCraft: Unified binder design by inverting an all-atom structure predictor

TorchCraft Team

Changping Laboratory, Beijing, China

Corresponding author: mingchenchen@cpl.ac.cn

Changping Laboratory, 2026

Paper

A frozen predictor as the design engine

TorchCraft optimizes editable sequence logits through a frozen predictor, using gradients from structural and sequence objectives.

  • Confidence, contact, geometry, and sequence priors guide optimization.
  • Continuous sequence representations are progressively discretized.
  • The output is a designed sequence with its predicted all-atom structure.
TorchCraft hallucination trajectory
Figure 11. Predictor-inversion design trajectory.
  • One optimization engine covers minibinders, VHHs, cyclic peptides, and ligand-binding proteins.
  • Tasks differ in editable positions, constraints, losses, and schedules.
TorchCraft framework
Figure 12. TorchCraft overview.

Minibinder design and validation

  • Four design campaigns produced binders measured by multi-concentration BLI.
  • Apparent dissociation constants were in the nanomolar range.
  • Tested sequences came directly from TorchCraft without inverse-folding redesign.
Minibinder wet-lab validation
Figure 13. Predicted minibinder complexes and multi-concentration BLI measurements.

Framework-conditioned VHH design

  • A predefined framework is fixed while the CDR regions are optimized.
  • CDR-restricted IgLM guidance introduces an antibody sequence prior.
  • Guidance improves IgLM likelihood and OASis humanness scores.
CDR-restricted IgLM guidance
Figure 15. CDR-restricted language-model guidance and sequence analysis.
  • Four VHH campaigns produced binders measured by multi-concentration BLI.
  • Tested sequences came directly from framework-conditioned TorchCraft designs.
VHH wet-lab validation
Figure 16. Predicted framework-conditioned VHH complexes and BLI measurements.
  • A 15-target benchmark compares TorchCraft with RFantibody, BoltzGen, and Germinal.
  • Optional AbMPNN redesign is evaluated separately.
VHH design benchmark
Figure 17. TorchScore-based VHH benchmark across 15 targets, comparing BoltzGen, RFantibody, Germinal, TorchCraft and TorchCraft with AbMPNN.

Cyclic peptides and ligand-binding proteins

  • Head-to-tail peptides use cyclic relative positional encoding.
  • Disulfide-linked peptides fix the cysteines that define cyclization.
  • An eight-target benchmark compares TorchCraft with RFpeptides and BoltzGen.
Cyclic peptide design
Figure 18. Head-to-tail and disulfide-linked cyclic-peptide design.

Runtime on GPU and Ascend NPU

Table S9. TorchCraft runtime across sequence lengths on NPU (with/without fused attention, FA) and GPU (with/without cuequivariance). Batch size 2 is the recommended setting. Runtime is reported in minutes per design, obtained by dividing the total wall-clock time of a batch-size-2 run by 2. GPU timings are the mean of three independent runs.
Code Operator Machine 100 aa 200 aa 300 aa 400 aa 500 aa
NPU w/ FA 910B (64G) 3.1 3.75 6.23 11.29 16.56
NPU w/o FA 910B (64G) 3.0 4.13 8.63 14.64 33.47∗
GPU w/ cuequivariance A100 (80G) 3.92 4.27 4.54 5.85 7.59
GPU w/o cuequivariance A100 (80G) 4.76 7.96 13.08 21.65 33.3

∗ Batch size 1; batch size 2 exceeded memory (OOM).

Citations

How to cite

TorchFold

Copy BibTeX 
@misc{chen2026torchfold,
  title        = {TorchFold: Distillation of diverse antibody-antigen interfaces improves structure prediction},
  author       = {{TorchFold Team}},
  year         = {2026},
  note         = {Changping Laboratory},
  url          = {https://torchx-cpl.github.io}
}

TorchCraft

Copy BibTeX 
@misc{chen2026torchcraft,
  title        = {TorchCraft: Hallucination-based all-atom protein design with TorchFold},
  author       = {{TorchCraft Team}},
  year         = {2026},
  note         = {Changping Laboratory},
  url          = {https://torchx-cpl.github.io}
}