Antibody-antigen structure prediction, complex scoring and all-atom binder design in one MIT-licensed stack that runs on GPUs and Ascend NPUs.
No providers recorded yet. Browse all providers
An antibody's multiple sequence alignment tells you about immunoglobulin frameworks, not about which epitope this particular clone binds: the CDR loops that do the recognizing are hypervariable by construction and carry almost no coevolutionary signal. That is why antibody-antigen complexes remain the worst-served interface class in all-atom co-folding, with the FoldBench benchmark reporting failure rates above 50% even for models that handle protein-protein and protein-ligand interfaces well.
TorchX, released by Changping Laboratory with Huawei on 17 September 2026, attacks that gap by leaving the AlphaFold 3 architecture alone and rebuilding the training distribution and the objective around the interface. It ships as three modes over one network: TorchFold predicts antibody-antigen complexes, TorchScore reuses the same weights to evaluate coordinates someone else generated, and TorchCraft inverts a frozen all-atom predictor — either the TorchFold checkpoint or AlphaFold 3's own parameters — to design binders. The design half follows the predictor-inversion pattern established by BoltzDesign1 — gradients flow onto sequence logits, never onto the weights. The name is unrelated to the PyTorch TorchX job launcher and to the TorchXRayVision imaging toolkit.
Inference, training and design code are MIT-licensed in one repository with parallel GPU and Ascend NPU trees. The TorchFold antibody-antigen checkpoint and the released training data are not distributed alongside them: both sit in a single Hugging Face dataset repository whose access is granted by manual review.
Distillation expanded the non-redundant antibody-antigen training set from 5.0k experimental examples to 22.9k after the first round and 24.5k after the second, drawn from computational designs, patent and private discovery collections, ASD and CoV-AbDab. Ranked success on FoldBench-AB (DockQ > 0.23) rose from 34.0% for the from-scratch model to 65.9% and then 70.1%, with matching evaluations on CPL-28, PXMeter-AB, 2026ARK-AB and AbAg2526 against AlphaFold 3, Protenix-v2 and OpenDDE under a shared MSA pipeline. Sampling 100 seeds raises ranked success to 72.4% and oracle success to 93.6%. Confidence is calibrated well enough for triage: at ipTM ≥ 0.8, 292 of 297 pooled predictions clear DockQ 0.23. Accuracy falls with antigen size, from 74% of interfaces recovered below 50 residues to 38% beyond 600. TorchScore's ipTM correlates with DockQ at Pearson 0.869 on AbAg PDB55.
TorchCraft produced nanomolar binders against four minibinder targets and four VHH targets, measured by biolayer interferometry on sequences taken directly from its output with no inverse-folding redesign; CDR-restricted IgLM guidance supplies an antibody sequence prior for the VHH mode. Every design the preprint reports — the experimentally validated binders and the computational benchmarks alike — was generated with pretrained AlphaFold 3 weights, not with the released TorchFold checkpoint. Those computational benchmarks cover 12 minibinder targets against BoltzGen and RFdiffusion, 15 VHH targets adding Germinal, and eight cyclic-peptide targets, with ligand-binding designs assessed by AutoDock Vina and Boltz-2. A 500-residue design takes about 16.6 minutes on an Ascend 910B with fused attention and 7.6 minutes on an A100 with cuEquivariance kernels. Fold, score and design all accept either the released TorchFold checkpoint or official AlphaFold 3 parameters obtained from DeepMind.
The stack targets the antibody discovery loop end to end: modelling a complex to locate an epitope before committing to crystallography, ranking candidate pairings from a campaign by ipTM without regenerating structures, and designing minibinders or framework-conditioned nanobodies against targets such as IFNAR2 and IL-7Rα, where measured affinities reached the nanomolar range. The training code is MIT-licensed, and the checkpoint and the public portion of the training data are offered through a Hugging Face gate that is approved by hand and whose terms bar redistribution, so groups holding antibody-antigen structures of their own can fine-tune rather than treat the model as a fixed service — and the Ascend branch makes that practical on domestic Chinese compute.
TorchX is an open all-atom biomolecular stack built to run natively on Ascend NPUs, and its permissive licensing of the full code path — inference, training and design — contrasts with the restricted parameter access governing the architecture it inherits, though the checkpoint itself is handed out only through a manually approved gate. The evidence sits at different levels of scrutiny: the TorchCraft design work is an arXiv preprint awaiting peer review, and the TorchFold prediction results are documented on the project site rather than in a paper. The training corpus behind that gate includes patent and private discovery collections that cannot be redistributed, so the released data does not fully reproduce the distillation. The gains are also specific — this is an antibody-antigen specialist, and its advantage over general co-folding models is claimed on that interface class alone.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.