rewirebio.iobenchmarks
Source

Tupikina et al. 2026, Can Folding Models Tell Binders from Bluffers? Evidence from POISK (bioRxiv v1 PDF)

v1, posted 26 October 2026 · Preprint

DOI: 10.64898/2026.10.24.761513Original source

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.

Model or configurationMetricValueDatasetEvaluation
AlphaFold-Multimer with full-Fv ipSAE rankingTop-1 accuracy (retrieval; one patent-supported binder ranked against 49 putative decoys for the same antigen by full-Fv ipSAE, mean over 27 antigen constructs, 20 trials per construct per seed, 25 seeds)0.12 fractionPOISK 1-in-50 screening targets (27 antigen constructs)AlphaFold-Multimer on POISK 1-in-50 screening simulation, top-1 retrieval
AlphaFold-Multimer with full-Fv ipSAE rankingTop-1 accuracy (retrieval; one patent-supported binder ranked against 49 putative decoys for the same antigen by full-Fv ipSAE, mean over 27 antigen constructs, 20 trials per construct per seed, 25 seeds)11.7 percentPOISK 1-in-50 screening targets (27 antigen constructs)AlphaFold-Multimer on POISK 1-in-50 screening simulation, top-1 retrieval
AlphaFold-Multimer with full-Fv ipSAE rankingCohen's d (separation of binder and decoy full-Fv ipSAE score distributions over all 22,494 complexes)0.39 unitlessPOISK co-folding binder-decoy benchmark set (22,494 complexes)AlphaFold-Multimer on POISK binder-decoy separation in full-Fv ipSAE, Cohen's d
AlphaFold-Multimer with full-Fv ipSAE rankingCohen's d (separation of binder and decoy ipTM score distributions over all 22,494 complexes)0.31 unitlessPOISK co-folding binder-decoy benchmark set (22,494 complexes)AlphaFold-Multimer on POISK binder-decoy separation in ipTM, Cohen's d
AlphaFold-Multimer with full-Fv ipSAE rankingSuccess rate (at least one annotated epitope residue with a heavy atom within 4 angstrom of an antibody heavy atom, over 1,232 binder records with mappable patent epitope annotations)50.4 percentPOISK binder records with mappable patent epitope annotations (1,232)AlphaFold-Multimer on POISK epitope contact recovery
AlphaFold-Multimer with full-Fv ipSAE rankingProportion (pairs in which binder and decoy receive numerically identical full-Fv ipSAE, over all 11,247 pairs)22.9% percentPOISK co-folding binder-decoy benchmark set (22,494 complexes)AlphaFold-Multimer on POISK pairs left undecidable by identical ipSAE
AlphaFold-Multimer with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 10% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.846 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)AlphaFold-Multimer on POISK pairwise binder-versus-decoy accuracy at 10% coverage
AlphaFold-Multimer with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 25% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.756 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)AlphaFold-Multimer on POISK pairwise binder-versus-decoy accuracy at 25% coverage
AlphaFold-Multimer with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 50% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.68 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)AlphaFold-Multimer on POISK pairwise binder-versus-decoy accuracy at 50% coverage
AlphaFold-Multimer with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 50% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.679 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)AlphaFold-Multimer on POISK pairwise binder-versus-decoy accuracy at 50% coverage
AlphaFold-Multimer with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 75% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.643 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)AlphaFold-Multimer on POISK pairwise binder-versus-decoy accuracy at 75% coverage
AlphaFold-Multimer with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, 100% coverage of the 11,247 pairs, exact ties at half credit)0.607 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)AlphaFold-Multimer on POISK pairwise binder-versus-decoy accuracy at 100% coverage
Boltz-2 with full-Fv ipSAE rankingTop-1 accuracy (retrieval; one patent-supported binder ranked against 49 putative decoys for the same antigen by full-Fv ipSAE, mean over 27 antigen constructs, 20 trials per construct per seed, 25 seeds)0.03 fractionPOISK 1-in-50 screening targets (27 antigen constructs)Boltz-2 on POISK 1-in-50 screening simulation, top-1 retrieval
Boltz-2 with full-Fv ipSAE rankingTop-1 accuracy (retrieval; one patent-supported binder ranked against 49 putative decoys for the same antigen by full-Fv ipSAE, mean over 27 antigen constructs, 20 trials per construct per seed, 25 seeds)3.1 percentPOISK 1-in-50 screening targets (27 antigen constructs)Boltz-2 on POISK 1-in-50 screening simulation, top-1 retrieval
Boltz-2 with full-Fv ipSAE rankingCohen's d (separation of binder and decoy full-Fv ipSAE score distributions over all 22,494 complexes)0.23 unitlessPOISK co-folding binder-decoy benchmark set (22,494 complexes)Boltz-2 on POISK binder-decoy separation in full-Fv ipSAE, Cohen's d
Boltz-2 with full-Fv ipSAE rankingCohen's d (separation of binder and decoy ipTM score distributions over all 22,494 complexes)0.20 unitlessPOISK co-folding binder-decoy benchmark set (22,494 complexes)Boltz-2 on POISK binder-decoy separation in ipTM, Cohen's d
Boltz-2 with full-Fv ipSAE rankingSuccess rate (at least one annotated epitope residue with a heavy atom within 4 angstrom of an antibody heavy atom, over 1,232 binder records with mappable patent epitope annotations)55.4 percentPOISK binder records with mappable patent epitope annotations (1,232)Boltz-2 on POISK epitope contact recovery
Boltz-2 with full-Fv ipSAE rankingProportion (pairs in which binder and decoy receive numerically identical full-Fv ipSAE, over all 11,247 pairs)4.3% percentPOISK co-folding binder-decoy benchmark set (22,494 complexes)Boltz-2 on POISK pairs left undecidable by identical ipSAE
Boltz-2 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 10% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.641 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)Boltz-2 on POISK pairwise binder-versus-decoy accuracy at 10% coverage
Boltz-2 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 25% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.622 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)Boltz-2 on POISK pairwise binder-versus-decoy accuracy at 25% coverage
Boltz-2 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 50% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.60 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)Boltz-2 on POISK pairwise binder-versus-decoy accuracy at 50% coverage
Boltz-2 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 50% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.602 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)Boltz-2 on POISK pairwise binder-versus-decoy accuracy at 50% coverage
Boltz-2 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 75% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.591 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)Boltz-2 on POISK pairwise binder-versus-decoy accuracy at 75% coverage
Boltz-2 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, 100% coverage of the 11,247 pairs, exact ties at half credit)0.581 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)Boltz-2 on POISK pairwise binder-versus-decoy accuracy at 100% coverage
Cross-model structural consensus, at least three models at DockQ 0.23Proportion (share of the 11,247 pairs eligible for classification with at least three models agreeing at DockQ 0.23 or better)54.2% percentPOISK co-folding binder-decoy benchmark set (22,494 complexes)Cross-model structural consensus (at least three models agreeing at DockQ 0.23 or better) on POISK pair coverage
Cross-model structural consensus, at least three models at DockQ 0.23Paired accuracy (binder identified as the pair member with more agreeing folding-model poses, at least three models agreeing at DockQ 0.23 or better, over the pairs eligible at that threshold)73.1% percentPOISK co-folding binder-decoy benchmark set (22,494 complexes)Cross-model structural consensus (at least three models agreeing at DockQ 0.23 or better) on POISK pairwise accuracy
Cross-model structural consensus, at least three models at DockQ 0.80Proportion (share of the 11,247 pairs eligible for classification with at least three models agreeing at DockQ 0.80 or better)3.0% percentPOISK co-folding binder-decoy benchmark set (22,494 complexes)Cross-model structural consensus (at least three models agreeing at DockQ 0.80 or better) on POISK pair coverage
Cross-model structural consensus, at least three models at DockQ 0.80Paired accuracy (binder identified as the pair member with more agreeing folding-model poses, at least three models agreeing at DockQ 0.80 or better, over the pairs eligible at that threshold)96.8% percentPOISK co-folding binder-decoy benchmark set (22,494 complexes)Cross-model structural consensus (at least three models agreeing at DockQ 0.80 or better) on POISK pairwise accuracy
IntelliFold v2 with full-Fv ipSAE rankingTop-1 accuracy (retrieval; one patent-supported binder ranked against 49 putative decoys for the same antigen by full-Fv ipSAE, mean over 27 antigen constructs, 20 trials per construct per seed, 25 seeds)0.25 fractionPOISK 1-in-50 screening targets (27 antigen constructs)IntelliFold v2 on POISK 1-in-50 screening simulation, top-1 retrieval
IntelliFold v2 with full-Fv ipSAE rankingTop-1 accuracy (retrieval; one patent-supported binder ranked against 49 putative decoys for the same antigen by full-Fv ipSAE, mean over 27 antigen constructs, 20 trials per construct per seed, 25 seeds)25.2 percentPOISK 1-in-50 screening targets (27 antigen constructs)IntelliFold v2 on POISK 1-in-50 screening simulation, top-1 retrieval
IntelliFold v2 with full-Fv ipSAE rankingTop-10 accuracy (retrieval; one patent-supported binder ranked against 49 putative decoys for the same antigen by full-Fv ipSAE, mean over 27 antigen constructs, 20 trials per construct per seed, 25 seeds)56.0% percentPOISK 1-in-50 screening targets (27 antigen constructs)IntelliFold v2 on POISK 1-in-50 screening simulation, top-10 retrieval
IntelliFold v2 with full-Fv ipSAE rankingCohen's d (separation of binder and decoy full-Fv ipSAE score distributions over all 22,494 complexes)0.66 unitlessPOISK co-folding binder-decoy benchmark set (22,494 complexes)IntelliFold v2 on POISK binder-decoy separation in full-Fv ipSAE, Cohen's d
IntelliFold v2 with full-Fv ipSAE rankingCohen's d (separation of binder and decoy ipTM score distributions over all 22,494 complexes)0.60 unitlessPOISK co-folding binder-decoy benchmark set (22,494 complexes)IntelliFold v2 on POISK binder-decoy separation in ipTM, Cohen's d
IntelliFold v2 with full-Fv ipSAE rankingSuccess rate (at least one annotated epitope residue with a heavy atom within 4 angstrom of an antibody heavy atom, over 1,232 binder records with mappable patent epitope annotations)56.5 percentPOISK binder records with mappable patent epitope annotations (1,232)IntelliFold v2 on POISK epitope contact recovery
IntelliFold v2 with full-Fv ipSAE rankingProportion (pairs in which binder and decoy receive numerically identical full-Fv ipSAE, over all 11,247 pairs)23.4% percentPOISK co-folding binder-decoy benchmark set (22,494 complexes)IntelliFold v2 on POISK pairs left undecidable by identical ipSAE
IntelliFold v2 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 10% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.960 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)IntelliFold v2 on POISK pairwise binder-versus-decoy accuracy at 10% coverage
IntelliFold v2 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 25% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.901 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)IntelliFold v2 on POISK pairwise binder-versus-decoy accuracy at 25% coverage
IntelliFold v2 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 50% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.80 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)IntelliFold v2 on POISK pairwise binder-versus-decoy accuracy at 50% coverage
IntelliFold v2 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 50% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.798 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)IntelliFold v2 on POISK pairwise binder-versus-decoy accuracy at 50% coverage
IntelliFold v2 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 75% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.723 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)IntelliFold v2 on POISK pairwise binder-versus-decoy accuracy at 75% coverage
IntelliFold v2 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, 100% coverage of the 11,247 pairs, exact ties at half credit)0.668 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)IntelliFold v2 on POISK pairwise binder-versus-decoy accuracy at 100% coverage
OpenFold3 with full-Fv ipSAE rankingTop-1 accuracy (retrieval; one patent-supported binder ranked against 49 putative decoys for the same antigen by full-Fv ipSAE, mean over 27 antigen constructs, 20 trials per construct per seed, 25 seeds)0.12 fractionPOISK 1-in-50 screening targets (27 antigen constructs)OpenFold3 on POISK 1-in-50 screening simulation, top-1 retrieval
OpenFold3 with full-Fv ipSAE rankingTop-1 accuracy (retrieval; one patent-supported binder ranked against 49 putative decoys for the same antigen by full-Fv ipSAE, mean over 27 antigen constructs, 20 trials per construct per seed, 25 seeds)11.6 percentPOISK 1-in-50 screening targets (27 antigen constructs)OpenFold3 on POISK 1-in-50 screening simulation, top-1 retrieval
OpenFold3 with full-Fv ipSAE rankingCohen's d (separation of binder and decoy full-Fv ipSAE score distributions over all 22,494 complexes)0.43 unitlessPOISK co-folding binder-decoy benchmark set (22,494 complexes)OpenFold3 on POISK binder-decoy separation in full-Fv ipSAE, Cohen's d
OpenFold3 with full-Fv ipSAE rankingCohen's d (separation of binder and decoy ipTM score distributions over all 22,494 complexes)0.33 unitlessPOISK co-folding binder-decoy benchmark set (22,494 complexes)OpenFold3 on POISK binder-decoy separation in ipTM, Cohen's d
OpenFold3 with full-Fv ipSAE rankingSuccess rate (at least one annotated epitope residue with a heavy atom within 4 angstrom of an antibody heavy atom, over 1,232 binder records with mappable patent epitope annotations)52.4 percentPOISK binder records with mappable patent epitope annotations (1,232)OpenFold3 on POISK epitope contact recovery
OpenFold3 with full-Fv ipSAE rankingProportion (pairs in which binder and decoy receive numerically identical full-Fv ipSAE, over all 11,247 pairs)37.9% percentPOISK co-folding binder-decoy benchmark set (22,494 complexes)OpenFold3 on POISK pairs left undecidable by identical ipSAE
OpenFold3 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 10% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.901 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)OpenFold3 on POISK pairwise binder-versus-decoy accuracy at 10% coverage
OpenFold3 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 25% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.780 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)OpenFold3 on POISK pairwise binder-versus-decoy accuracy at 25% coverage
OpenFold3 with full-Fv ipSAE rankingPaired accuracy (binder scored above its matched putative decoy by full-Fv ipSAE, top 50% of the 11,247 pairs ranked by within-pair maximum ipSAE, exact ties at half credit)0.67 fractionPOISK co-folding binder-decoy benchmark set (22,494 complexes)OpenFold3 on POISK pairwise binder-versus-decoy accuracy at 50% coverage

Claims

Statements this source supports, each with where in the source it was checked.

  • links:family:poisk-20261115-model-openfold3: OpenFold3 (MSA): Family OpenFold3 (data/entities/configurations.jsonl, record esmfold2-2026-runs-n-poses-method-openfold3-msa, source_locator "Figure 2C Runs N' Poses, bar label OpenFold3 (MSA); Appendix A.2.10"; data/entities/models.jsonl, record poisk-20261115-model-openfold3)
  • OpenFold3: provenance_note: Provenance note The preprint names this engine only as "OpenFold3" among the six folding models it evaluates. No release tag, commit, weights hash or parameter count is printed, so the record stays at family level and the evaluated setup is on the configuration record. Checked against the store during review: no model record for this engine exists under any other ID, but the store does already hold it as an unparented configuration, esmfold2-2026-runs-n-poses-method-openfold3-msa, named "OpenFold3 (MSA)", whose links array is empty. That configuration should gain a family edge to this record; the edge is not made here because it changes a stored record outside this batch. discovery-model-openfold is OpenFold, a PyTorch implementation of AlphaFold 2 and AlphaFold-Multimer, and is not this engine. (data/entities/configurations.jsonl, record esmfold2-2026-runs-n-poses-method-openfold3-msa; data/entities/models.jsonl, record discovery-model-openfold)
  • POISK epitope contact recovery: metric_implementation: Metric implementation Geometric epitope contact: a record counts as recovered when at least one annotated epitope residue in the folded antigen construct has a heavy atom within 4 angstrom of an antibody heavy atom in the predicted complex. The paper states it as "We counted a record as recovered when at least one annotated epitope residue had a heavy atom within 4Å of an antibody heavy atom". No interface confidence score, ipSAE included, enters this endpoint. (Section 3.3.2, sentence beginning "We counted a record as recovered"; Table 2)
  • POISK pair coverage under cross-model structural consensus: metric_implementation: Metric implementation Cross-model pose agreement: DockQ scores are computed for the Fv-antigen interface of each model's predicted complex, and consensus is the stated number of models predicting the same pose above the stated DockQ threshold. The paper states the decision rule as "For any given pair, a prediction was considered successful if the binder demonstrated structural agreement amongst more models than its decoy". The pair is decided by which member has agreement among more models, not by an interface confidence score; ipSAE does not enter this endpoint. (Section 3.3.1, sentences beginning "To test this, we calculated DockQ scores" and "For any given pair, a prediction was considered successful")
  • POISK pairwise binder-versus-decoy accuracy under cross-model structural consensus: metric_implementation: Metric implementation Cross-model pose agreement: DockQ scores are computed for the Fv-antigen interface of each model's predicted complex, and consensus is the stated number of models predicting the same pose above the stated DockQ threshold. The paper states the decision rule as "For any given pair, a prediction was considered successful if the binder demonstrated structural agreement amongst more models than its decoy". The pair is decided by which member has agreement among more models, not by an interface confidence score; ipSAE does not enter this endpoint. (Section 3.3.1, sentences beginning "To test this, we calculated DockQ scores" and "For any given pair, a prediction was considered successful")
  • POISK: affinity correlation: Affinity correlation Heterogeneous affinity and potency measurements (KD, KA, IC50, EC50) were standardised to pAffinity across the 5,207 binders with reported values that parse unambiguously. Binned comparisons show a marginal trend toward higher ipSAE at higher affinity, but the paper says the signal is too weak to be predictive and that the models appear to capture a binary binding or non-binding distinction rather than a graded measure of interaction strength. (Section 3.2.4; Section 4; Appendix I)
  • POISK: AlphaFold-Multimer screening variability across targets: Screening variability across targets In the 1-in-50 screening simulation, AlphaFold-Multimer has a standard deviation of 13.4 percentage points in top-1 accuracy across the 27 antigen constructs, against 1.5 points across the 25 sampling seeds. For Protenix-v2 the per-construct top-1 accuracy ranges from 2.2% to 70.8%. Between-target variation exceeds sampling variation for every engine. (Table 8, row "SD across targets (pp)", column "AF-M"; Table 8, row "SD across seeds (pp)", column "AF-M"; Appendix H)
  • POISK: Boltz-2 screening variability across targets: Screening variability across targets In the 1-in-50 screening simulation, Boltz-2 has a standard deviation of 3.3 percentage points in top-1 accuracy across the 27 antigen constructs, against 0.8 points across the 25 sampling seeds. For Protenix-v2 the per-construct top-1 accuracy ranges from 2.2% to 70.8%. Between-target variation exceeds sampling variation for every engine. (Table 8, row "SD across targets (pp)", column "Boltz-2"; Table 8, row "SD across seeds (pp)", column "Boltz-2"; Appendix H)
  • POISK: composition: Composition The collection spans over 2,000 unique target sequences, with 91.8% of records carrying explicit target metadata. Canonical monoclonal VH/VL antibodies are 85.6% of entries, bispecific 7.4% and single-domain VHH 3.9%. (Section 3.1 'Dataset')
  • POISK: compute budget: Compute budget Folding all 22,494 complexes across the six models required approximately 9,000 GPU-hours on NVIDIA B200 GPUs, for 674,820 structure predictions in total. The label-noise experiment of Appendix D.2 took a further 250 B200 GPU-hours. (Section 2.3; Appendix L; Appendix D.2)
  • POISK: consensus coverage: Consensus coverage Requiring agreement among five or six models restricts coverage sharply: when all six models must agree, about half the dataset is excluded even at a DockQ threshold of 0.01, and at DockQ 0.23 or better the retained pairs reach at least 90% pairwise accuracy. At k of 3 or more and DockQ 0.80 accuracy reaches 96.8% but only 3.0% of pairs remain eligible. (Section 3.3.1; Appendix K; Figures 7 and 16)
  • POISK: contamination analysis: Contamination analysis Benchmark pairs were grouped by the publication date of the binder's source patent, one bucket before 2018 and six-month buckets to H1 2025, with bucket sizes from 305 to 3,229 pairs and bootstrap 95% confidence intervals. Accuracy does not systematically decline toward more recent patents: Protenix-v2 moves from 0.716 to 0.689 and IntelliFold v2 from 0.654 to 0.669. Splitting at each engine's stated PDB training cutoff, the largest pre-minus-post difference across all six engines is 1.8 percentage points (IntelliFold v2) and 0.6 points for both Protenix releases, against a spread between engines of more than ten points. (Appendix J; Figures 13 and 14)
  • POISK: coverage definition: Coverage definition The headline pairwise accuracy is measured after discarding the half of the pairs for which both members scored lowest. Appendix F states this "is not the accuracy over the benchmark as a whole" and reports accuracy at 100%, 75%, 50%, 25% and 10% coverage, where coverage is the fraction of the 11,247 pairs kept after ranking pairs by the higher of their two ipSAE scores. Exact ties receive half credit. Over all 11,247 pairs Protenix-v2 reaches 0.72, against 0.87 at the midpoint. (Appendix F; Section 3.2.2)
  • POISK: epitope localisation: Epitope localisation Patent epitope residue annotations could be mapped to the folded antigen construct for 1,232 binder records. Even the best-scoring model, Protenix-v2, placed the antibody in contact with the annotated epitope in only 62.8% of cases. The paper concludes that binding discrimination does not establish accurate epitope localization and that a confident prediction is not necessarily a correct one. (Section 3.3.2; Section 4 'Discussion')
  • POISK: evidence tiers: Evidence tiers Among the 11,247 benchmark binders, all 11,247 (100.0%) carry explicit binding evidence and a concrete antigen sequence, 5,967 (53.1%) a reported binding-affinity value, 3,888 (34.6%) any epitope characterisation and 1,232 (11.0%) a residue-level epitope. The tiers overlap and are not successive filters. Within the 3,888 epitope-characterised records, 1,003 are discontinuous or conformational, 826 linear, 501 functional or competitive and 1,558 not further specified. (Appendix C; Table 5)
  • POISK: extraction pipeline: Extraction pipeline Antibody chains were collected from GenBank (heavy, lambda and kappa, 70 to 1000 amino acids) and classified with ANARCI, giving 10,110 eligible US patents. Antigen identification, chain pairing and 36 further fields were extracted from the patent text with the Google Generative AI Python library and the Gemini 3 Pro model, with Pydantic schemas. Missing antigen sequences were looked up in UniProt by target identity and organism. Gemini also acted as auditor: fields it flagged as not explicitly mentioned were excluded, and records flagged for incorrect chain pairing were removed. (Section 2.1 'Data extraction')
  • POISK: folding protocol: Folding protocol All six models used MSAs computed with ColabFold and no templates. The diffusion-based models generated one sample at each of five random seeds; AlphaFold-Multimer used its five parameter models. The highest confidence prediction was selected for analysis. Each of the 22,494 complexes was folded by all six engines with five predictions, giving 674,820 structure predictions and about 9,000 NVIDIA B200 GPU-hours; MSAs were computed once per unique sequence and excluded from the reported timings. (Section 2.3; Appendix L)
  • POISK: IntelliFold v2 screening variability across targets: Screening variability across targets In the 1-in-50 screening simulation, IntelliFold v2 has a standard deviation of 19.0 percentage points in top-1 accuracy across the 27 antigen constructs, against 1.7 points across the 25 sampling seeds. For Protenix-v2 the per-construct top-1 accuracy ranges from 2.2% to 70.8%. Between-target variation exceeds sampling variation for every engine. (Table 8, row "SD across targets (pp)", column "IF v2"; Table 8, row "SD across seeds (pp)", column "IF v2"; Appendix H)
  • POISK: label noise sensitivity: Label noise sensitivity To test whether the engines simply prefer patent-derived sequences, 300 benchmark datapoints were sampled under a target-uniqueness constraint, balanced 150 germline-reverted and 150 OAS-matched, and their targets shuffled so that neither candidate binds the antigen it is folded against; the sampling was repeated under five seeds (42 to 46), giving 30,000 structure predictions and about 250 NVIDIA B200 GPU-hours. On the invalid triplets Protenix-v2 scored the patent-derived sequence higher in 50.3% of surviving pairs against 72.4% on the same pairs with their correct targets, and Boltz-2 52.5% against 55.4%. (Appendix D.2; Table 6 'Mean' column)
  • POISK: manual audit: Manual audit A manual audit of 40 records against their source patents found 19 of 20 benchmark records (95%) and 18 of 20 records from the full dataset correct in every populated field, with 181 of 184 audited field instances correct. Two of the three errors were conservative omissions; the third was a human target labelled Humanized. No record in either sample asserted an antibody-antigen binding pair its patent does not support, which the authors say does not completely rule out their existence elsewhere. (Section 2.1; Appendix B, Table 4)
  • POISK: negative composition: Negative composition Of the 11,247 decoys, 7,647 (68.0%) are retrieved from the Observed Antibody Space with KA-Search and 3,600 (32.0%) are generated by germline reversion with CDR-H3 randomisation. Both classes are paired against human antigens. (Appendix D.1)
  • POISK: negative construction: Negative construction Putative non-binders were built by KA-Search against the Observed Antibody Space for the closest VH CDR match to each binder, paired with the original antigen. For the 32% of records with no close OAS match, germline-encoded regions were reverted to their inferred germline sequences and the CDR-H3 loops randomised by sampling from AbLang predictions while keeping total identity at 85% or above. The final benchmark holds 22,494 records with an equal split between binding and non-binding instances. (Section 2.2)
  • POISK: negative label status: Negative label status The paper states of the non-binding subset: "while the biological context strongly supports the assumption of non-reactivity, these pairs remain putative negatives and have not been experimentally validated for lack of binding." Appendix D adds that neither the tolerance argument nor the naive-repertoire argument is a measurement and that the authors claim no numerical false-negative rate. (Section 2.2; Appendix D.1)
  • POISK: OpenFold3 screening variability across targets: Screening variability across targets In the 1-in-50 screening simulation, OpenFold3 has a standard deviation of 11.4 percentage points in top-1 accuracy across the 27 antigen constructs, against 1.6 points across the 25 sampling seeds. For Protenix-v2 the per-construct top-1 accuracy ranges from 2.2% to 70.8%. Between-target variation exceeds sampling variation for every engine. (Table 8, row "SD across targets (pp)", column "OF3"; Table 8, row "SD across seeds (pp)", column "OF3"; Appendix H)
  • POISK: overlap with other resources: Overlap with other resources A comparison with the Antibody and Nanobody Design Dataset found only 716 shared antibody-antigen entries. (Section 3.1)

Models, benchmarks and other records

Models (3)

Methods (1)

Configurations (8)

Benchmarks (1)

Tasks (1)

Protocols (13)

Datasets (1)

Dataset subsets (3)

Evaluations (66)

Sources (4)

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.

33 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-11-21-c8bea25e5abb
Property and statementOriginal source and locationReview and provenance
Archive sha256
39d2ff104527916aa9b33386ca5d2830d2efd9e14b4d98258af54f07b0598950
Source metadata
Tupikina et al. 2026, Can Folding Models Tell Binders from Bluffers? Evidence from POISK (bioRxiv v1 PDF)

Original source ↗

No field-specific location recorded

Version: v1, posted 26 October 2026
Retrieved: 2026-11-15T23:32:24Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.archive_sha256

Source artifact SHA-256: 526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924

Hash scope: One download of 2026-11-15. The PDF container is not byte-stable: bioRxiv regenerates it on request, so artifact_sha256 is that one download's hash and no refetch reproduces it. Three downloads on 2026-11-16 gave 3389c1bf6d3654c5b2151a07fcfcefc67cf6f36e5f0facfc03c33521d8b84491 twice and d699fd680e3b0097ca80130c288e8d872773dc1e523044f4605e06fc4f89d9bc once. All three yield the same pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, which is the article's stable identity, is archived gzipped in artifacts/biorxiv-v1-layout.txt.gz under archive_sha256, is what extract/build_batch.py asserts, and is what every stored value was read from.

Format: application/pdf, 16 pages

Article licence
CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
Source metadata
Tupikina et al. 2026, Can Folding Models Tell Binders from Bluffers? Evidence from POISK (bioRxiv v1 PDF)

Original source ↗

No field-specific location recorded

Version: v1, posted 26 October 2026
Retrieved: 2026-11-15T23:32:24Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.article_licence

Source artifact SHA-256: 526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924

Hash scope: One download of 2026-11-15. The PDF container is not byte-stable: bioRxiv regenerates it on request, so artifact_sha256 is that one download's hash and no refetch reproduces it. Three downloads on 2026-11-16 gave 3389c1bf6d3654c5b2151a07fcfcefc67cf6f36e5f0facfc03c33521d8b84491 twice and d699fd680e3b0097ca80130c288e8d872773dc1e523044f4605e06fc4f89d9bc once. All three yield the same pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, which is the article's stable identity, is archived gzipped in artifacts/biorxiv-v1-layout.txt.gz under archive_sha256, is what extract/build_batch.py asserts, and is what every stored value was read from.

Format: application/pdf, 16 pages

Artifact format
application/pdf, 16 pages
Source metadata
Tupikina et al. 2026, Can Folding Models Tell Binders from Bluffers? Evidence from POISK (bioRxiv v1 PDF)

Original source ↗

No field-specific location recorded

Version: v1, posted 26 October 2026
Retrieved: 2026-11-15T23:32:24Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.artifact_format

Source artifact SHA-256: 526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924

Hash scope: One download of 2026-11-15. The PDF container is not byte-stable: bioRxiv regenerates it on request, so artifact_sha256 is that one download's hash and no refetch reproduces it. Three downloads on 2026-11-16 gave 3389c1bf6d3654c5b2151a07fcfcefc67cf6f36e5f0facfc03c33521d8b84491 twice and d699fd680e3b0097ca80130c288e8d872773dc1e523044f4605e06fc4f89d9bc once. All three yield the same pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, which is the article's stable identity, is archived gzipped in artifacts/biorxiv-v1-layout.txt.gz under archive_sha256, is what extract/build_batch.py asserts, and is what every stored value was read from.

Format: application/pdf, 16 pages

Artifact SHA-256
526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924
Source metadata
Tupikina et al. 2026, Can Folding Models Tell Binders from Bluffers? Evidence from POISK (bioRxiv v1 PDF)

Original source ↗

No field-specific location recorded

Version: v1, posted 26 October 2026
Retrieved: 2026-11-15T23:32:24Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.artifact_sha256

Source artifact SHA-256: 526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924

Hash scope: One download of 2026-11-15. The PDF container is not byte-stable: bioRxiv regenerates it on request, so artifact_sha256 is that one download's hash and no refetch reproduces it. Three downloads on 2026-11-16 gave 3389c1bf6d3654c5b2151a07fcfcefc67cf6f36e5f0facfc03c33521d8b84491 twice and d699fd680e3b0097ca80130c288e8d872773dc1e523044f4605e06fc4f89d9bc once. All three yield the same pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, which is the article's stable identity, is archived gzipped in artifacts/biorxiv-v1-layout.txt.gz under archive_sha256, is what extract/build_batch.py asserts, and is what every stored value was read from.

Format: application/pdf, 16 pages

Doi
10.64898/2026.10.24.761513
Source metadata
Tupikina et al. 2026, Can Folding Models Tell Binders from Bluffers? Evidence from POISK (bioRxiv v1 PDF)

Original source ↗

No field-specific location recorded

Version: v1, posted 26 October 2026
Retrieved: 2026-11-15T23:32:24Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.doi

Source artifact SHA-256: 526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924

Hash scope: One download of 2026-11-15. The PDF container is not byte-stable: bioRxiv regenerates it on request, so artifact_sha256 is that one download's hash and no refetch reproduces it. Three downloads on 2026-11-16 gave 3389c1bf6d3654c5b2151a07fcfcefc67cf6f36e5f0facfc03c33521d8b84491 twice and d699fd680e3b0097ca80130c288e8d872773dc1e523044f4605e06fc4f89d9bc once. All three yield the same pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, which is the article's stable identity, is archived gzipped in artifacts/biorxiv-v1-layout.txt.gz under archive_sha256, is what extract/build_batch.py asserts, and is what every stored value was read from.

Format: application/pdf, 16 pages

Hash scope
One download of 2026-11-15. The PDF container is not byte-stable: bioRxiv regenerates it on request, so artifact_sha256 is that one download's hash and no refetch reproduces it. Three downloads on 2026-11-16 gave 3389c1bf6d3654c5b2151a07fcfcefc67cf6f36e5f0facfc03c33521d8b84491 twice and d699fd680e3b0097ca80130c288e8d872773dc1e523044f4605e06fc4f89d9bc once. All three yield the same pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, which is the article's stable identity, is archived gzipped in artifacts/biorxiv-v1-layout.txt.gz under archive_sha256, is what extract/build_batch.py asserts, and is what every stored value was read from.
Source metadata
Tupikina et al. 2026, Can Folding Models Tell Binders from Bluffers? Evidence from POISK (bioRxiv v1 PDF)

Original source ↗

No field-specific location recorded

Version: v1, posted 26 October 2026
Retrieved: 2026-11-15T23:32:24Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.hash_scope

Source artifact SHA-256: 526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924

Hash scope: One download of 2026-11-15. The PDF container is not byte-stable: bioRxiv regenerates it on request, so artifact_sha256 is that one download's hash and no refetch reproduces it. Three downloads on 2026-11-16 gave 3389c1bf6d3654c5b2151a07fcfcefc67cf6f36e5f0facfc03c33521d8b84491 twice and d699fd680e3b0097ca80130c288e8d872773dc1e523044f4605e06fc4f89d9bc once. All three yield the same pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, which is the article's stable identity, is archived gzipped in artifacts/biorxiv-v1-layout.txt.gz under archive_sha256, is what extract/build_batch.py asserts, and is what every stored value was read from.

Format: application/pdf, 16 pages

Licence
CC BY 4.0, as printed in the per-page banner of the PDF
Source metadata
Tupikina et al. 2026, Can Folding Models Tell Binders from Bluffers? Evidence from POISK (bioRxiv v1 PDF)

Original source ↗

No field-specific location recorded

Version: v1, posted 26 October 2026
Retrieved: 2026-11-15T23:32:24Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.licence

Source artifact SHA-256: 526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924

Hash scope: One download of 2026-11-15. The PDF container is not byte-stable: bioRxiv regenerates it on request, so artifact_sha256 is that one download's hash and no refetch reproduces it. Three downloads on 2026-11-16 gave 3389c1bf6d3654c5b2151a07fcfcefc67cf6f36e5f0facfc03c33521d8b84491 twice and d699fd680e3b0097ca80130c288e8d872773dc1e523044f4605e06fc4f89d9bc once. All three yield the same pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, which is the article's stable identity, is archived gzipped in artifacts/biorxiv-v1-layout.txt.gz under archive_sha256, is what extract/build_batch.py asserts, and is what every stored value was read from.

Format: application/pdf, 16 pages

Limitations
4 values
  • Preprint, not certified by peer review as posted. The project site states the paper was accepted at NeurIPS 2026; no peer-reviewed version was retrieved.
  • A version 2 of this preprint was posted on 14 November 2026. It was retrieved during review: PDF SHA-256 aeaec593d7aa2d42382a49bd6fadbbccabaa6c084843ea1c7f0c91097dedb0c2, pdftotext -layout text layer SHA-256 82a99bc2e061b7247cf65f7bc9450924b7af5be7ab3ef283d2284ba787ea629c. Every numeric token in version 2 is identical to version 1's apart from an added author affiliation and the two Zenodo record identifiers, which change from 19883049 (dataset version 1.0.1) to 21775623 (version 1.1), so no value in this batch differs between the two versions. Every record stays pinned to version 1, which is the artifact the values were read from.
  • Tables 1 to 9 are images in the bioRxiv JATS XML, so values were read from the PDF text layer, not from XML table markup.
  • bioRxiv regenerates this PDF on request, so its artifact_sha256 is one download's hash and no refetch reproduces it. The stable identity is the pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696.
Source metadata
Tupikina et al. 2026, Can Folding Models Tell Binders from Bluffers? Evidence from POISK (bioRxiv v1 PDF)

Original source ↗

No field-specific location recorded

Version: v1, posted 26 October 2026
Retrieved: 2026-11-15T23:32:24Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.limitations

Source artifact SHA-256: 526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924

Hash scope: One download of 2026-11-15. The PDF container is not byte-stable: bioRxiv regenerates it on request, so artifact_sha256 is that one download's hash and no refetch reproduces it. Three downloads on 2026-11-16 gave 3389c1bf6d3654c5b2151a07fcfcefc67cf6f36e5f0facfc03c33521d8b84491 twice and d699fd680e3b0097ca80130c288e8d872773dc1e523044f4605e06fc4f89d9bc once. All three yield the same pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, which is the article's stable identity, is archived gzipped in artifacts/biorxiv-v1-layout.txt.gz under archive_sha256, is what extract/build_batch.py asserts, and is what every stored value was read from.

Format: application/pdf, 16 pages

Publication status
preprint
Source metadata
Tupikina et al. 2026, Can Folding Models Tell Binders from Bluffers? Evidence from POISK (bioRxiv v1 PDF)

Original source ↗

No field-specific location recorded

Version: v1, posted 26 October 2026
Retrieved: 2026-11-15T23:32:24Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.publication_status

Source artifact SHA-256: 526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924

Hash scope: One download of 2026-11-15. The PDF container is not byte-stable: bioRxiv regenerates it on request, so artifact_sha256 is that one download's hash and no refetch reproduces it. Three downloads on 2026-11-16 gave 3389c1bf6d3654c5b2151a07fcfcefc67cf6f36e5f0facfc03c33521d8b84491 twice and d699fd680e3b0097ca80130c288e8d872773dc1e523044f4605e06fc4f89d9bc once. All three yield the same pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, which is the article's stable identity, is archived gzipped in artifacts/biorxiv-v1-layout.txt.gz under archive_sha256, is what extract/build_batch.py asserts, and is what every stored value was read from.

Format: application/pdf, 16 pages

Retrieval method
curl; bioRxiv rate-limited the first attempts with HTTP 429 and served the PDF on retry
Source metadata
Tupikina et al. 2026, Can Folding Models Tell Binders from Bluffers? Evidence from POISK (bioRxiv v1 PDF)

Original source ↗

No field-specific location recorded

Version: v1, posted 26 October 2026
Retrieved: 2026-11-15T23:32:24Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.retrieval_method

Source artifact SHA-256: 526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924

Hash scope: One download of 2026-11-15. The PDF container is not byte-stable: bioRxiv regenerates it on request, so artifact_sha256 is that one download's hash and no refetch reproduces it. Three downloads on 2026-11-16 gave 3389c1bf6d3654c5b2151a07fcfcefc67cf6f36e5f0facfc03c33521d8b84491 twice and d699fd680e3b0097ca80130c288e8d872773dc1e523044f4605e06fc4f89d9bc once. All three yield the same pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, which is the article's stable identity, is archived gzipped in artifacts/biorxiv-v1-layout.txt.gz under archive_sha256, is what extract/build_batch.py asserts, and is what every stored value was read from.

Format: application/pdf, 16 pages

How to read the evidence

Sources and history

Release 2026-11-21-c8bea25e5abb · Record review: source checked

The bioRxiv preprint that introduces the POISK dataset and benchmarks six folding models as binder-decoy classifiers; every stored value comes from this PDF.

1 source record and release history

Read original source

Download this release (gzip)
Technical metadata and extraction receipts

Stable ID: poisk-20261115-source-manuscript-v1

areas
molecular-interactions; proteins-complexes
contexts
research
archive sha256
39d2ff104527916aa9b33386ca5d2830d2efd9e14b4d98258af54f07b0598950
article licence
CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
artifact format
application/pdf, 16 pages
artifact sha256
526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924
doi
10.64898/2026.10.24.761513
hash scope
One download of 2026-11-15. The PDF container is not byte-stable: bioRxiv regenerates it on request, so artifact_sha256 is that one download's hash and no refetch reproduces it. Three downloads on 2026-11-16 gave 3389c1bf6d3654c5b2151a07fcfcefc67cf6f36e5f0facfc03c33521d8b84491 twice and d699fd680e3b0097ca80130c288e8d872773dc1e523044f4605e06fc4f89d9bc once. All three yield the same pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, which is the article's stable identity, is archived gzipped in artifacts/biorxiv-v1-layout.txt.gz under archive_sha256, is what extract/build_batch.py asserts, and is what every stored value was read from.
licence
CC BY 4.0, as printed in the per-page banner of the PDF
limitations
Preprint, not certified by peer review as posted. The project site states the paper was accepted at NeurIPS 2026; no peer-reviewed version was retrieved.; A version 2 of this preprint was posted on 14 November 2026. It was retrieved during review: PDF SHA-256 aeaec593d7aa2d42382a49bd6fadbbccabaa6c084843ea1c7f0c91097dedb0c2, pdftotext -layout text layer SHA-256 82a99bc2e061b7247cf65f7bc9450924b7af5be7ab3ef283d2284ba787ea629c. Every numeric token in version 2 is identical to version 1's apart from an added author affiliation and the two Zenodo record identifiers, which change from 19883049 (dataset version 1.0.1) to 21775623 (version 1.1), so no value in this batch differs between the two versions. Every record stays pinned to version 1, which is the artifact the values were read from.; Tables 1 to 9 are images in the bioRxiv JATS XML, so values were read from the PDF text layer, not from XML table markup.; bioRxiv regenerates this PDF on request, so its artifact_sha256 is one download's hash and no refetch reproduces it. The stable identity is the pdftotext -layout text layer, SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696.
publication status
preprint
retrieval method
curl; bioRxiv rate-limited the first attempts with HTTP 429 and served the PDF on retry
retrieved at
2026-11-15T23:32:24Z
review
artifact sha256: 526def150fb5ae8f9ac4533efcf16c86e547e636cdb9bfba9a574cd1e2a2f924; date: 2026-11-15; method: source-hash-verification; pdf-text-parse; method note: Extraction pass. Every stored table value was parsed from the pdftotext -layout text layer of the pinned bioRxiv v1 PDF by extract/build_batch.py, which asserts the sha256 of the text layer and then every table caption, row label and column header before reading a cell. Values the paper prints only in prose were transcribed by hand and the script asserts the sentence they come from. No value was read from a figure.; note: The PDF was downloaded on 2026-11-15 and its sha256 recorded. pdftotext (poppler 26.08.0) -layout produced the text layer whose sha256 the extraction script asserts.; retrieval url: https://www.biorxiv.org/content/10.64898/2026.10.24.761513v1.full.pdf; reviewer: claude; reviewer note: Claude (Opus 5) extraction agent; no human review claimed. Extraction only: no independent review is claimed and a different agent reviews this batch. Transcription from the pinned source is not reproduction.
review artifact
pdftotext -layout text layer of three refetched v1 PDFs, all SHA-256 e75c18130ca43770d2f0ed05ae96528759666c0e4fa08f7725f5a32b21187696, byte-identical to the contents of artifacts/biorxiv-v1-layout.txt.gz
review artifact format
text/plain, pdftotext -layout output
review artifact matches original artifact sha256
false
source locator
Whole preprint; per-value locators are on each record
url
https://www.biorxiv.org/content/10.64898/2026.10.24.761513v1.full.pdf
version
v1, posted 26 October 2026
Related records

Suggest a correction