TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1)
arXiv (cs.AI) · 2026 · arXiv:2609.19770v1, submitted 17 September 2026 · Preprint
What this source contributed
Records in this release that cite this source. A record can cite several sources.
Results
Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.
Claims
Statements this source supports, each with where in the source it was checked.
- Apparent dissociation constant of a selected minibinder design by biolayer interferometry: Minibinder candidate pool, filters and selection: Candidate pool 3,000 TorchCraft candidates were generated per minibinder target, filtered by radius of gyration below 17, binder pTM above 0.7 and TorchCraft ipTM above 0.7, then ranked by Rosetta Interface Analyzer using the composite score 100 x TorchCraft ipTM minus the Rosetta interface score, and the highest-scoring candidates were taken to the wet lab. The number expressed and tested is not stated. (Section 2.2; Appendix A.2.1 and A.2.2.1)
- Apparent dissociation constant of a selected VHH design by biolayer interferometry: VHH candidate pool, filters and selection: Candidate pool 12,500 TorchCraft candidates were generated per VHH target, filtered by CDR interface rate above 0.6, CDR3 hotspot coverage rate above 0 and TorchCraft ipTM above 0.7, then ranked by Rosetta Interface Analyzer using the composite score 100 x TorchCraft ipTM minus the Rosetta interface score, and the top-ranked designs were selected for experimental validation. The number expressed and tested is not stated. (Section 2.3; Appendix A.2.1 and A.2.2.2)
- ESM2-150M perplexity as a VHH expression classifier: The expression proxy threshold and its two printed values: Operating threshold Designs with a total expression of at least 0.08 mg were classified as high-expression cases, ESM2-150M perplexity was swept as a one-dimensional classifier and the operating threshold chosen by maximising the Matthews correlation coefficient. The threshold is printed as 6.31 in Appendix A.4.3.2 and as 6.309 in the Supplementary Figure S4b caption. (Appendix A.4.3.2; Supplementary Figure S4 caption, panel b)
- TorchCraft 8-target cyclic peptide design benchmark: Benchmark reported only as a figure: Reported values Figure 4c compares computational pass rates for BoltzGen, RFpeptides and TorchCraft across eight target systems in head-to-tail and disulfide-linked settings, with optional ProteinMPNN polishing shown separately; Supplementary Figure S6 shows the cyclisation outcome fractions and Supplementary Figure S7 the Rosetta interface scores. 100 designs were generated per method per target in the head-to-tail setting. None of these figures prints numbers, so no value was extracted. (Figure 4c and its caption; Supplementary Figures S6 and S7; Appendix A.4.3.3)
- TorchCraft 10-ligand pocket design benchmark: Benchmark reported only as a figure: Reported values Figures 5c and 5d show AutoDock Vina docking scores and Boltz-2 affinity predictions for designed pockets across the 10-ligand panel, comparing BoltzDesign with TorchCraft and a separately analysed LigandMPNN-redesigned TorchCraft variant; Supplementary Figure S9 shows TorchScore pass rates for TorchCraft. 500 designs of 160 residues were generated per method per ligand. None of these figures prints numbers, so no value was extracted. (Figures 5c and 5d and their captions; Supplementary Figures S8 and S9; Appendix A.4.3.4)
- TorchCraft 12-target minibinder design benchmark: Benchmark reported only as a figure: Reported values Figure 2f compares computational pass rates for TorchCraft, BoltzGen and RFdiffusion across 12 protein targets, aggregated over binder lengths of 75, 120, 160 and 200 residues, with an optional ProteinMPNN-polished TorchCraft setting shown separately. 100 designs were generated per method, per length, per target. The figure prints no numbers, so no value was extracted. (Figure 2f and its caption; Appendix A.4.3.1)
- TorchCraft 15-target VHH design benchmark: Benchmark reported only as a figure: Reported values Figure 3i compares computational pass rates for TorchCraft with RFantibody, BoltzGen and Germinal across 15 targets and five VHH frameworks under a joint TorchScore-confidence and ESM2-perplexity criterion, with an optional AbMPNN-polished setting shown separately and a framework-stratified version in Supplementary Figure S5. 100 designs were generated per method, per framework, per target. Neither figure prints numbers, so no value was extracted. (Figure 3i and its caption; Supplementary Figure S5; Appendix A.4.3.2)
- TorchCraft minibinder design, raw output: The amino-acid composition regulariser and its reference: Sequence regulariser An amino-acid composition regulariser matches the average residue distribution of designed minibinders to a reference estimated from ProteinMPNN-redesigned samples; its weight is 5 in Supplementary Table S1. The preprint reports that it changed sequence composition, raised ESMFold confidence across the 12-target benchmark and, in two PDGFRb campaigns, that the regularised round expressed better than the preceding unregularised round. Those comparisons are shown only as figures without printed numbers. (Section 2.2; Supplementary Table S1 row 'Minibinder'; Supplementary Figure S1)
- TorchCraft on Ascend 910B without fused attention: The 500-residue NPU run without fused attention used batch size 1: Measurement exception The 500-residue timing for the Ascend 910B without fused attention was measured at batch size 1 because batch size 2 exceeded memory. Every other cell of the table is a batch-size-2 run divided by 2. (Supplementary Table S9 footnote: '* Batch size 1; batch size 2 exceeded memory (OOM).')
- TorchCraft runtime benchmark across sequence lengths: How the runtime figures were measured: Measurement setup Runtime is reported in minutes per design, obtained by dividing the total wall-clock time of a batch-size-2 run by 2, on an Ascend 910B with 64 GB with and without fused attention and on an A100 with 80 GB with and without cuequivariance kernels. Batch size 2 is the recommended setting and GPU timings are the mean of three independent runs. (Supplementary Table S9 caption)
- TorchCraft: Binding does not establish the predicted interface is correct: Evidence limit The preprint states that binding measurements do not by themselves establish the accuracy of the predicted interfaces. (Discussion, paragraph beginning 'The current study has several limitations')
- TorchCraft: Every reported TorchCraft experiment used pretrained AlphaFold 3 weights: Evaluated weight configuration All designs reported in the preprint, both the experimentally validated minibinders and VHHs and the computational benchmarks, were generated with pretrained AlphaFold 3 weights through the TorchFold implementation. (Abstract ("Using pretrained AlphaFold3 weights"); Section 1 final paragraph ("All designs reported here, including experimentally validated binders and computational benchmarks, were generated using pretrained AF3 weights"); Appendix A.1.1 ("All designs reported in this study used pretrained AF3 weights"); Figure 1 caption ("Experimental campaigns and computational benchmarks reported here used the same AF3-weight setting"))
- TorchCraft: Experimental validation covers minibinders and VHHs only: Experimental validation scope Experimental validation is confined to minibinders and VHHs; the cyclic-peptide and small-molecule results remain computational. (Discussion, paragraph beginning 'The current study has several limitations'; Sections 2.4 and 2.5)
- TorchCraft: Implementation changes that make repeated predictor evaluation affordable: Implementation optimisations Repeated predictor evaluation is supported by multithreaded conformer generation with per-thread random state, reentrant gradient checkpointing for the design graph and non-reentrant checkpointing for the template and MSA modules, caching of trunk conditioning tensors across denoising steps within one prediction, a modified NPU fusion-attention operator, and optional cuequivariance kernels on GPU. (Appendix A.3.1 to A.3.5; Supplementary Table S9)
- TorchCraft: Nothing in this batch was reproduced: Reproduction status No design run, score, docking calculation or binding measurement was reproduced for this batch. Every value is a transcription from the pinned preprint, checked against the sources named on each record. (Whole batch; see notes.md)
- TorchCraft: The shared four-stage optimisation schedule: Optimisation schedule The sequence logits pass through up to four stages: an optional soft warm-up using softmax probabilities, a linearly scheduled interpolation from raw logits to probabilities, temperature annealing from an initial to a minimum temperature, and straight-through hard optimisation on a one-hot representation with gradients taken through the low-temperature softmax. Minibinder and VHH runs used temperature annealing followed by straight-through optimisation; cyclic-peptide runs used all four stages; small-molecule runs began with a ligand-free warm-up before the ligand was added and the interface weight was raised. (Appendix A.1.2 and A.1.3; Algorithm 1; Supplementary Table S1)
- TorchCraft: The tested sequences were raw TorchCraft outputs: Tested sequence provenance The experimentally tested minibinder and VHH sequences were raw TorchCraft outputs, obtained without post hoc MPNN sequence redesign. (Sections 2.2 and 2.3; Appendix A.2.1; Discussion)
- TorchCraft: TorchCraft trains no model weights of its own: Training status TorchCraft introduces no trained generative model: the predictor is frozen, only the editable sequence logits are updated, and the final sequence is obtained by discrete decoding. (Abstract ("optimizes sequence logits through a frozen all-atom predictor"); Appendix A.1.1 ("the weighted design objective is differentiated with respect to L. Only the editable sequence variables are updated"); Appendix A.1.2 Algorithm 1)
- TorchFold on the gated TorchFold antibody-distilled checkpoint: No reported result used the antibody-distilled TorchFold checkpoint: Performance claim scope No design or benchmark result in the preprint was produced with the antibody-specialised TorchFold checkpoint. The Discussion names its use for CDR optimisation as a direction still to be tested with matched comparisons. (Appendix A.1.1; Discussion final paragraph ("Its specialized checkpoint may offer more informative guidance during CDR optimization. Matched comparisons ... will be needed to test this potential"))
- TorchScore: Predictor-derived pass rates are not experimental success rates: Interpretation limit The preprint states that predictor-derived pass rates are sensitive to the evaluator and should not be interpreted as experimental success rates. (Discussion, paragraph beginning 'The current study has several limitations')
- TorchScore: TorchScore uses the design weights and does not refold sequences: Scoring dependency TorchScore is a score-only loading of TorchFold: it bypasses the diffusion structure module, trains no separate scoring network, and in this study used the same pretrained AlphaFold 3 weights used for design. It scores supplied coordinates, so designs rebuilt with FlowPacker were scored from those coordinates and were not refolded from sequence. (Appendix A.4.2 ("No separate scoring network is trained. In this study, TorchScore used pretrained AF3 weights, the same checkpoint used for design ... they were not refolded from sequence by TorchScore"); Figure 2f and Figure 3i captions)
Models, benchmarks and other records
Models (3)
Methods (2)
Configurations (17)
- ESM2-150M perplexity as a VHH expression classifier
- TorchCraft disulfide-linked cyclic-peptide design with ProteinMPNN polishing
- TorchCraft disulfide-linked cyclic-peptide design, raw output
- TorchCraft framework-conditioned VHH design, raw output
- TorchCraft head-to-tail cyclic-peptide design with ProteinMPNN polishing
- TorchCraft head-to-tail cyclic-peptide design, raw output
- TorchCraft ligand-conditioned pocket design with LigandMPNN redesign
- TorchCraft ligand-conditioned pocket design, raw output
- TorchCraft minibinder design with ProteinMPNN polishing
- TorchCraft minibinder design, raw output
- TorchCraft on A100 with cuequivariance kernels
- TorchCraft on A100 without cuequivariance kernels
- TorchCraft on Ascend 910B with fused attention
- TorchCraft on Ascend 910B without fused attention
- TorchCraft VHH design with AbMPNN polishing
- TorchFold on pretrained AlphaFold 3 parameters
- TorchFold on the gated TorchFold antibody-distilled checkpoint
Pipelines (1)
Benchmarks (5)
Tasks (6)
Protocols (16)
- Apparent dissociation constant of a selected minibinder design by biolayer interferometry
- Apparent dissociation constant of a selected VHH design by biolayer interferometry
- Cyclic peptide computational pass rate under cyclisation and TorchScore confidence
- Docking score of designed ligand pockets
- Ligand-pocket computational pass rate under TorchScore confidence
- Minibinder computational pass rate under TorchScore confidence
- Predicted affinity of designed ligand pockets
- TorchCraft runtime per design at 100 residues
- TorchCraft runtime per design at 200 residues
- TorchCraft runtime per design at 300 residues
- TorchCraft runtime per design at 400 residues
- TorchCraft runtime per design at 500 residues
- VHH computational pass rate under TorchScore confidence and ESM2 perplexity
- VHH expression against ESM2-150M perplexity: Pearson correlation
- VHH expression against ESM2-150M perplexity: Spearman correlation
- VHH expression classification by ESM2-150M perplexity: Matthews correlation coefficient
Evaluators (1)
Datasets (8)
- Eight target systems of the cyclic peptide design benchmark
- Fifteen protein targets of the VHH design benchmark
- Four protein targets of the minibinder validation campaigns
- Four protein targets of the VHH validation campaigns
- Retrospective VHH expression measurements
- Ten small-molecule ligands of the pocket design benchmark
- TorchCraft runtime benchmark design runs
- Twelve protein targets of the minibinder design benchmark
Dataset subsets (8)
- BHRF1 measurements of the VHH validation campaign
- EFNA5 measurements of the VHH validation campaign
- IFNAR2 measurements of the minibinder validation campaign
- IL-7Ra measurements of the minibinder validation campaign
- PD-L1 measurements of the minibinder validation campaign
- PDGFRb measurements of the minibinder validation campaign
- PDGFRb measurements of the VHH validation campaign
- S100A4 measurements of the VHH validation campaign
Evaluations (31)
- ESM2-150M perplexity against measured VHH expression: Matthews correlation coefficient
- ESM2-150M perplexity against measured VHH expression: Pearson correlation
- ESM2-150M perplexity against measured VHH expression: Spearman rank correlation
- Measured binding of a selected TorchCraft minibinder to IFNAR2
- Measured binding of a selected TorchCraft minibinder to IL-7Ra
- Measured binding of a selected TorchCraft minibinder to PD-L1
- Measured binding of a selected TorchCraft minibinder to PDGFRb
- Measured binding of a selected TorchCraft VHH to BHRF1
- Measured binding of a selected TorchCraft VHH to EFNA5
- Measured binding of a selected TorchCraft VHH to PDGFRb
- Measured binding of a selected TorchCraft VHH to S100A4
- TorchCraft runtime at 100 residues, GPU w/ cuequivariance
- TorchCraft runtime at 100 residues, GPU w/o cuequivariance
- TorchCraft runtime at 100 residues, NPU w/ FA
- TorchCraft runtime at 100 residues, NPU w/o FA
- TorchCraft runtime at 200 residues, GPU w/ cuequivariance
- TorchCraft runtime at 200 residues, GPU w/o cuequivariance
- TorchCraft runtime at 200 residues, NPU w/ FA
- TorchCraft runtime at 200 residues, NPU w/o FA
- TorchCraft runtime at 300 residues, GPU w/ cuequivariance
- TorchCraft runtime at 300 residues, GPU w/o cuequivariance
- TorchCraft runtime at 300 residues, NPU w/ FA
- TorchCraft runtime at 300 residues, NPU w/o FA
- TorchCraft runtime at 400 residues, GPU w/ cuequivariance
- TorchCraft runtime at 400 residues, GPU w/o cuequivariance
Sources (6)
Evidence
Source checking verifies the cited claim or transcription. It does not establish independent reproduction.
Evidence table
Inspect claims, sources and review details
Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.
One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.
31 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Article licence arXiv.org perpetual, non-exclusive license 1.0 (http://arxiv.org/licenses/nonexclusive-distrib/1.0/), the only licence the abstract page links. No open reuse or redistribution licence is offered. Source metadata | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) No field-specific location recorded Version: arXiv:2609.19770v1, submitted 17 September 2026 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Artifact SHA-256 f8145b5b95552dd30996eda2940507c0f32d09adfc60f44f28458e7872788f21 Source metadata | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) No field-specific location recorded Version: arXiv:2609.19770v1, submitted 17 September 2026 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Artifact url https://arxiv.org/pdf/2609.19770v1 Source metadata | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) No field-specific location recorded Version: arXiv:2609.19770v1, submitted 17 September 2026 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Extraction method pdftotext -layout text layer; Table S9 by extract/parse_runtime_table.py, the other values by hand transcription Source metadata | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) No field-specific location recorded Version: arXiv:2609.19770v1, submitted 17 September 2026 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Hash scope SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. Source metadata | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) No field-specific location recorded Version: arXiv:2609.19770v1, submitted 17 September 2026 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Licence arXiv.org perpetual, non-exclusive license 1.0 (http://arxiv.org/licenses/nonexclusive-distrib/1.0/), the only licence the abstract page links. No open reuse or redistribution licence is offered. Source metadata | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) No field-specific location recorded Version: arXiv:2609.19770v1, submitted 17 September 2026 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
Limitations2 values
| TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) No field-specific location recorded Version: arXiv:2609.19770v1, submitted 17 September 2026 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Media type application/pdf Source metadata | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) No field-specific location recorded Version: arXiv:2609.19770v1, submitted 17 September 2026 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Missing metadata: doi: note The abstract page serves no citation_doi meta tag for this preprint. Source metadata | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) No field-specific location recorded Version: arXiv:2609.19770v1, submitted 17 September 2026 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Missing metadata: doi: reason unavailable Source metadata | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) No field-specific location recorded Version: arXiv:2609.19770v1, submitted 17 September 2026 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
Sources and history
Release 2026-11-21-adc15377fab2 · Record review: source checked
The TorchCraft preprint, which reports the design framework, the experimental minibinder and VHH campaigns, the computational benchmarks and the full methods.
Technical metadata and extraction receipts
Stable ID: torchcraft-20261115-source-arxiv-v1
- areas
- protein-structure; molecular-interactions
- contexts
- research
- article licence
- arXiv.org perpetual, non-exclusive license 1.0 (http://arxiv.org/licenses/nonexclusive-distrib/1.0/), the only licence the abstract page links. No open reuse or redistribution licence is offered.
- artifact sha256
- f8145b5b95552dd30996eda2940507c0f32d09adfc60f44f28458e7872788f21
- artifact url
- https://arxiv.org/pdf/2609.19770v1
- extraction method
- pdftotext -layout text layer; Table S9 by extract/parse_runtime_table.py, the other values by hand transcription
- hash scope
- SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here.
- licence
- arXiv.org perpetual, non-exclusive license 1.0 (http://arxiv.org/licenses/nonexclusive-distrib/1.0/), the only licence the abstract page links. No open reuse or redistribution licence is offered.
- limitations
- Developer-reported: the authors designed, ran and measured everything reported here.; No independent reproduction was attempted for this batch.
- media type
- application/pdf
- missing metadata
- doi: note: The abstract page serves no citation_doi meta tag for this preprint.; reason: unavailable
- publication status
- preprint
- retrieved at
- 2026-11-15T23:36:17Z
- review
- artifact sha256: f8145b5b95552dd30996eda2940507c0f32d09adfc60f44f28458e7872788f21; method: source-hash-verification; automated-source-review; method note: Re-retrieved and hashed by the reviewer. All seven pinned artifacts were re-retrieved by the reviewer on 2026-11-16T00:08Z. Six came back byte-identical to the hash on their source record. The Hugging Face page did not, and cannot: three consecutive fetches gave three different hashes. Normalising its per-request hf-sanitized CSS class hashes leaves one further per-request difference, window.requestId, and the page also prints a download counter; with those normalised the markup is identical. Its licence tag, gate text and dataset card matched the recorded quotations word for word, so the page is a per-retrieval stamp, not a changed page. The pdftotext -layout text layer of the refetched PDF hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b, the value the extractor recorded.; retrieval url: https://arxiv.org/pdf/2609.19770v1; reviewed at: 2026-11-16T01:05:00Z; reviewer: claude; reviewer note: Claude (Opus 5) review agent, independent of the extractor; no human review claimed. Every pinned artifact was re-retrieved and hashed, and every stored value re-read against it. Nothing was reproduced: no design run, score, docking calculation or binding measurement was repeated, and no value was recomputed from source data.
- url
- https://arxiv.org/pdf/2609.19770v1
- venue
- arXiv (cs.AI)
- version
- arXiv:2609.19770v1, submitted 17 September 2026
- year
- 2026
Related records
- companion of: arXiv abstract page for 2609.19770v1
- companion of: Hugging Face dataset page TorchX-CPL/TorchFold
- companion of: TorchX project site (torchx-cpl.github.io)
- companion of: TorchX torchcraft_gpu/README.md at commit ebf9dc6
- companion of: TorchX repository LICENSE file at commit ebf9dc6
- companion of: TorchX repository README at commit ebf9dc6