| { |
| "@context": { |
| "@language": "en", |
| "@vocab": "https://schema.org/", |
| "citeAs": "cr:citeAs", |
| "column": "cr:column", |
| "conformsTo": "dct:conformsTo", |
| "cr": "http://mlcommons.org/croissant/", |
| "rai": "http://mlcommons.org/croissant/RAI/", |
| "data": { |
| "@id": "cr:data", |
| "@type": "@json" |
| }, |
| "dataType": { |
| "@id": "cr:dataType", |
| "@type": "@vocab" |
| }, |
| "dct": "http://purl.org/dc/terms/", |
| "examples": { |
| "@id": "cr:examples", |
| "@type": "@json" |
| }, |
| "extract": "cr:extract", |
| "field": "cr:field", |
| "fileProperty": "cr:fileProperty", |
| "fileObject": "cr:fileObject", |
| "fileSet": "cr:fileSet", |
| "format": "cr:format", |
| "includes": "cr:includes", |
| "isLiveDataset": "cr:isLiveDataset", |
| "jsonPath": "cr:jsonPath", |
| "key": "cr:key", |
| "md5": "cr:md5", |
| "parentField": "cr:parentField", |
| "path": "cr:path", |
| "recordSet": "cr:recordSet", |
| "references": "cr:references", |
| "regex": "cr:regex", |
| "repeated": "cr:repeated", |
| "replace": "cr:replace", |
| "samplingRate": "cr:samplingRate", |
| "sc": "https://schema.org/", |
| "separator": "cr:separator", |
| "source": "cr:source", |
| "subField": "cr:subField", |
| "transform": "cr:transform" |
| }, |
| "@type": "sc:Dataset", |
| "name": "PROTAC-Bench", |
| "description": "Cold-target evaluation benchmark for PROTAC degradation prediction. 10,748 entries across 173 protein targets with 65 Leave-One-Target-Out (LOTO) folds. Merged from PROTAC-DB 3.0, Ribes et al., and DegradeMaster with canonical SMILES standardization.", |
| "conformsTo": "http://mlcommons.org/croissant/1.0", |
| "license": "https://creativecommons.org/licenses/by/4.0/", |
| "url": "https://huggingface.co/datasets/PROTAC-Bench/protac-bench", |
| "version": "1.0.0", |
| "citeAs": "@inproceedings{protacbench2026, title={PROTAC-Bench: A Cold-Target Benchmark for PROTAC Degradation Prediction}, booktitle={NeurIPS Datasets and Benchmarks Track}, year={2026}}", |
| "datePublished": "2026-05-02", |
| "creator": { |
| "@type": "sc:Organization", |
| "name": "PROTAC-Bench Authors (anonymized for double-blind review)" |
| }, |
| "keywords": [ |
| "PROTAC", |
| "protein degradation", |
| "drug discovery", |
| "benchmark", |
| "cold-target evaluation", |
| "binary classification" |
| ], |
| "rai:dataCollection": "PROTAC-Bench aggregates 10,748 PROTAC-target pairs from three publicly released sources: PROTAC-DB 3.0 (Weng et al., 2023; Nucleic Acids Research), the Ribes et al. (2024) curated benchmark, and DegradeMaster (Liu et al., 2024). Records were de-duplicated on canonical SMILES + UniProt accession pairs. SMILES were standardised with RDKit canonicalisation; targets were mapped to UniProt accessions via UniProt REST API queries on HGNC/UniProt-name strings supplied in source databases.", |
| "rai:dataCollectionType": "Aggregation of pre-existing publicly published datasets; no primary experimental data collection.", |
| "rai:dataCollectionTimeframe": "Source databases span PROTAC publications 2001-2023 (PROTAC-DB 3.0 release); merged corpus frozen 2025-Q4. Temporal split: pre-2022 entries used for training, 2022+ held out for the temporal evaluation fold.", |
| "rai:dataCollectionRawData": "Processed: SMILES are canonicalised, targets are normalised to UniProt accessions, activity labels are binarised (DC50<1 uM OR Dmax>50% -> 1). Raw DC50 / Dmax values are preserved in dc50_nm / dmax_pct columns for users who prefer custom thresholds.", |
| "rai:dataCollectionMissingData": "~38% of entries report only Dmax or only DC50, not both. The binary label is computed from whichever potency endpoint is available. Cell line, assay format, and time-point metadata are NOT included; users needing assay-context-aware modelling should consult the original source publications.", |
| "rai:dataAnnotationProtocol": "Activity labels are inherited from the source databases' published binarisation rules. Each source's primary literature was hand-curated by that source's authors; PROTAC-Bench performs no additional re-annotation. The cross-source label-agreement rate on the 1,247 entries appearing in two or more source DBs is 98.4% (kappa=0.96).", |
| "rai:dataAnnotationPlatform": "No platform - labels propagated from upstream curated databases (PROTAC-DB 3.0 web portal exports, Ribes et al. 2024 supplementary tables, DegradeMaster 2024 release).", |
| "rai:dataAnnotationAnalysis": "Inter-source agreement was measured on the 1,247 entries shared by >=2 source databases: raw agreement 98.4%, Cohen's kappa 0.96. Disagreements (12 cases) were retained as separate rows flagged with source_conflict=true rather than resolved by majority vote, to preserve the upstream signal.", |
| "rai:dataAnnotationPerItemTime": "Not applicable - no per-item human annotation was performed by the PROTAC-Bench authors. Upstream curators do not report per-record annotation timing.", |
| "rai:dataAnnotationDemographics": "Not applicable: labels derive from biochemical assay readouts in source publications, not from human-judgement annotation.", |
| "rai:dataAnnotationTools": "No annotation tools were used - upstream labels were ingested verbatim. Standardisation tooling (not annotation): RDKit 2024.03 for SMILES canonicalisation; UniProt REST API (https://rest.uniprot.org/uniprotkb) for target-name to accession resolution.", |
| "rai:dataPreprocessingProtocol": "SMILES canonicalisation: RDKit MolToSmiles(mol, canonical=True) after MolFromSmiles round-trip with sanitisation. Stereochemistry preserved. Target normalisation: UniProt accessions resolved via the UniProt REST API; entries that fail to resolve to a single canonical accession are dropped (1,043 records, 8.8% of pre-merge total).", |
| "rai:dataPreprocessingImputation": "None. Missing potency values are kept as null; the binary label is computed from whatever potency value is available. Entries with neither DC50 nor Dmax are excluded from the benchmark.", |
| "rai:dataPreprocessingManipulation": "De-duplication on (canonical SMILES, UniProt) tuples; cross-source conflict resolution by majority vote (3 sources) or, if 2 sources conflict (12 cases), retained as separate entries flagged with source_conflict=true.", |
| "rai:dataUseCases": "(1) Benchmarking PROTAC degradation prediction models under cold-target evaluation (held-out UniProt accessions). (2) Studying generalisation decay as molecular similarity to training set decreases. (3) Measuring E3-ligase scaffold transferability (VHL <-> CRBN). (4) Few-shot transfer experiments for low-data targets. NOT INTENDED for direct clinical candidate selection - predictions are research-stage and have not been validated against held-out wet-lab assays beyond the source databases.", |
| "rai:dataLimitation": "(1) E3-ligase imbalance: VHL and CRBN account for 87% of records; performance on rare E3 ligases (RNF114, IAP, MDM2, ...) is data-limited. (2) Target-class imbalance: kinases dominate (47% of entries) due to PROTAC literature focus. (3) Activity-label binarisation discards potency gradient - models cannot learn DC50 ranking. (4) Assay heterogeneity is not encoded - the same compound assayed by different labs at different time-points may receive divergent labels. (5) Publication-positivity bias: inactive PROTACs are systematically under-reported in the literature.", |
| "rai:dataBiases": "Documented biases: (a) chemotype bias toward CRBN/VHL warhead families documented in the cheminformatics literature; (b) target bias toward oncology targets (BCR-ABL, BTK, AR, EGFR, BRD4 are over-represented); (c) lab-of-origin confounding - three labs contribute >40% of records, introducing potential lab-specific assay-condition signatures that models can latch onto (see task14_within_target_cross_lab.json and the 'lab-confound' analysis in the paper).", |
| "rai:dataSocialImpact": "Positive: lowers the entry barrier for ML-driven PROTAC design, enables reproducible benchmarking and reduces wasted wet-lab effort on poorly-generalising models. Negative / dual-use: PROTAC technology in principle enables targeted degradation of arbitrary proteins; however, this dataset contains only published research-stage compounds and provides no novel uplift for misuse beyond what is already in the primary literature. No human-subject data; no privacy concerns.", |
| "rai:personalSensitiveInformation": "None. The dataset contains chemical structures (SMILES), protein identifiers (UniProt accessions), and biochemical activity labels. No human-subject data, no PII, no patient-derived material.", |
| "rai:dataReleaseMaintenancePlan": "Distributed under CC-BY-4.0 via HuggingFace Datasets. Maintained by the PROTAC-Bench authors; versioned releases tagged in the HF repo and the source repository's RELEASE_MANIFEST.md. Issues / corrections accepted via GitHub issues; merged updates tagged as semver minor releases. No deprecation date - long-term maintenance is committed for at least the duration of the NeurIPS 2026 reproducibility window (2026-2028).", |
| "distribution": [ |
| { |
| "@type": "cr:FileObject", |
| "@id": "protac_bench.csv", |
| "name": "protac_bench.csv", |
| "description": "Main dataset: 10,748 PROTAC entries with SMILES, target, E3 ligase type, and binary activity label.", |
| "contentUrl": "data/protac_bench.csv", |
| "encodingFormat": "text/csv", |
| "sha256": "6d273d9fbfb1921f5b2da9ba94a74d46fe787ffd3f6a766058c8d1c3c76d30fa" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "loto_folds.json", |
| "name": "loto_folds.json", |
| "description": "65 pre-computed Leave-One-Target-Out fold assignments with test indices, entry counts, and activity rates.", |
| "contentUrl": "data/loto_folds.json", |
| "encodingFormat": "application/json", |
| "sha256": "61564e68683db7c46424a5b6b58fe25d7cab319ee3dc7a300f86b7611bb4de3c" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "lofo_folds.json", |
| "name": "lofo_folds.json", |
| "description": "Leave-One-Family-Out fold assignments. 22 protein families covering 61 LOFO-eligible targets (LOTO-eligible targets that are mapped to a named family in robustness/lofo.py FAMILY_MAP). Family-level holdout: each fold's test_indices are all rows whose target_uniprot belongs to the family's target set.", |
| "contentUrl": "data/lofo_folds.json", |
| "encodingFormat": "application/json", |
| "sha256": "16d4d2fabe2ac84c0f7528128a49f94832eef92db096edd4f3bf8ee2ed32c8ee" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "cross_lab_folds.json", |
| "name": "cross_lab_folds.json", |
| "description": "Within-target cross-lab fold assignments. 36 targets with >=20 entries, >=3 publications (DOIs), and both classes present. For each (target, paper) holdout where the paper has >=5 entries and both classes, test_indices are the rows matching that target and DOI; train is everything else. Used by the lab-confound analysis (results/task14_within_target_cross_lab.json).", |
| "contentUrl": "data/cross_lab_folds.json", |
| "encodingFormat": "application/json", |
| "sha256": "3beb321d4a2e5e2365501b48a38d885cffc3fecf5b82a99c0c5b268bfc8bd964" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "temporal_prospective_folds.json", |
| "name": "temporal_prospective_folds.json", |
| "description": "Temporal prospective split. Train: rows with publication year < 2023 (1,866 entries). Test: rows with publication year == 2024 (132 entries). 2023 entries are excluded as a temporal gap year. pub_year derived from DOI mapping; rows without resolved pub_year are excluded from the split.", |
| "contentUrl": "data/temporal_prospective_folds.json", |
| "encodingFormat": "application/json", |
| "sha256": "a68950af03b1be29df7ddb47b373e234cd72190fb2101433b2f818f8aa6cc67b" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "admet_scores.csv", |
| "name": "admet_scores.csv", |
| "description": "7-property ADMET cascade scores for all 10,748 entries.", |
| "contentUrl": "data/admet_scores.csv", |
| "encodingFormat": "text/csv", |
| "sha256": "db5c33175ec6c1f1f9608cc230f655b2dd823b1e745df36e34d0616ec056d7e0" |
| } |
| ], |
| "recordSet": [ |
| { |
| "@type": "cr:RecordSet", |
| "@id": "protac_entries", |
| "name": "protac_entries", |
| "description": "Individual PROTAC degradation entries.", |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "protac_entries/smiles", |
| "name": "smiles", |
| "description": "Canonical SMILES representation of the PROTAC molecule.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "protac_bench.csv" |
| }, |
| "extract": { |
| "column": "smiles" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "protac_entries/target_uniprot", |
| "name": "target_uniprot", |
| "description": "UniProt accession ID of the target protein.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "protac_bench.csv" |
| }, |
| "extract": { |
| "column": "target_uniprot" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "protac_entries/e3_type", |
| "name": "e3_type", |
| "description": "E3 ubiquitin ligase type (VHL, CRBN, or Other).", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "protac_bench.csv" |
| }, |
| "extract": { |
| "column": "e3_type" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "protac_entries/label", |
| "name": "label", |
| "description": "Binary activity label (1 = active: DC50 < 1 uM OR Dmax > 50%, 0 = inactive).", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "protac_bench.csv" |
| }, |
| "extract": { |
| "column": "label" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "protac_entries/dc50_nm", |
| "name": "dc50_nm", |
| "description": "Half-maximal degradation concentration in nanomolar (when available).", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "protac_bench.csv" |
| }, |
| "extract": { |
| "column": "dc50_nm" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "protac_entries/dmax_pct", |
| "name": "dmax_pct", |
| "description": "Maximum degradation percentage (when available).", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "protac_bench.csv" |
| }, |
| "extract": { |
| "column": "dmax_pct" |
| } |
| } |
| } |
| ] |
| } |
| ] |
| } |
