A curated multi-endpoint drug-protein bioactivity dataset with construct-aware protein targets

• IC50, Ki, Kd, and EC50 quantify different aspects of activity, so their reported values cannot generally be converted using the numerical value alone.The resource therefore retains them as separate prediction targets.Source: §1 · pp. 1–2
• Individual measurements retain their reported relations and available assay and source information, while repeated observations are summarized separately within each endpoint.The interaction tables also identify conflicting classification evidence and widely dispersed exact values.Source: §1 · pp. 2–3
• 4,204 drug–accession groups had classification outcomes that depended on experimental sequence across the four endpoint-specific interaction views.Protein entities combine a reviewed accession with its normalized experimental sequence, preserving constructs that accession-only grouping would merge.Source: §1 · pp. 2–3
• The release preserves identifiers and provenance for constructing cold-drug, cold-protein-entity, cold-UniProt, temporal, and source-document-grouped splits.These splits address different generalization questions; canonical SMILES also support user-defined molecular-scaffold grouping.Source: §1 · pp. 2–3
• Curate-DTI provides separate IC50, Ki, Kd, and EC50 measurement and interaction tables, preserving exact and inequality-qualified evidence for endpoint-specific classification and regression.The endpoints are processed independently rather than combined into a common regression target.Source: §1 · pp. 2–3
• Standardized ligand structures are shared across endpoints, and the release preserves available assay and source-document provenance.These connected records support endpoint-specific benchmarking without prescribing molecular descriptors, protein encoders, or prediction architectures.Source: §1 · p. 3

2 Methods

• Measurements for IC50, Ki, Kd, and EC50 are processed independently through validation, ligand standardization, sequence-aware protein definition, and interaction construction.Multi-chain measurements are retained separately with available chain-level information.Source: §2 · p. 3, Figure 1 · p. 4

2.1 Input data and release control

• The 1 July 2026 BindingDB export and a dated Swiss-Prot reference were treated as fixed inputs, with release-specific row identifiers and experimental target sequences used in processing.The release manifest records filenames, download dates, file sizes, and SHA-256 checksums.Source: §2.1 · pp. 4–5

2.2 Computational environment

• Version-pinned dependencies, cached deterministic processing, reconstruction instructions, and validation-report scripts accompany the deposited preprocessing code.The pipeline did not use random sampling, model training, or stochastic data transformations.Source: §2.2 · p. 5

2.3 Measurement selection

• Endpoint fields were processed independently because a source row may report more than one endpoint, and the endpoints were not combined into a common regression target.Table 1 summarizes the number of measurements retained after each sequential processing stage.Source: §2.3 · p. 5, Table 1 · p. 6
• 2,061,581 IC50, 597,553 Ki, 126,972 Kd, and 259,974 EC50 observations were retained for the model-ready single-chain tables.Multi-chain observations were excluded from those tables because sequences were not available for all constituent chains, but retained in archival tables.Source: §2.3 · p. 5
• A retained single-chain observation required a ligand SMILES, reviewed Swiss-Prot accession, positive finite endpoint value or inequality boundary, and valid reported experimental sequence.Measurements were also excluded if their ligand structure could not be parsed or converted into a valid standardized parent structure.Source: §2.3 · p. 5

2.4 Relation-aware endpoint parsing and task labels

• IC50, Ki, Kd and EC50 observations are parsed independently into endpoint type, relation and positive numerical boundary, preserving the original value and source-row identifier.Exact observations are eligible for regression; inequality-qualified observations remain evidence but are not treated as exact regression observations.Source: §2.4 · p. 6
• Table 2 reports retained measurement relations; right-censored observations give lower boundaries and left-censored observations give upper boundaries.The relation categories distinguish exact values from inequality-qualified evidence.Source: Table 2 · p. 6, §2.4 · p. 6
• At 1,000 nM, an observation receives a class only when its exact value or inequality boundary places every possible value on one side of the threshold.For example, E < 500 nM supports class 1, whereas E < 5,000 nM is ambiguous.Source: §2.4 · pp. 6–7, Table 3 · p. 7
• Exact measurements within pE ∈[3, 12] are retained for regression, while inequality-qualified measurements are not subject to this range restriction.The range limits are operational quality-control thresholds, not universal biological or physical boundaries.Source: §2.4 · pp. 6–7
• The same numerical threshold defines tasks across endpoints but does not make their biological interpretations equivalent; class 0 is not proof of universal inactivity.Source: §2.4 · p. 7

2.5 Ligand-structure standardization

• Canonical isomeric SMILES define ligand identity after RDKit cleanup, parent-fragment selection, charge neutralization and sanitization; defined stereochemistry is preserved.Unspecified stereochemistry is not inferred.Source: §2.5 · p. 7
• Structures are excluded only when they cannot be parsed or standardized into a valid parent structure; molecular size and carbon content are not inclusion criteria.Source: §2.5 · p. 7

2.6 Construct-aware protein entities

• Protein identity combines the reviewed UniProt accession with the normalized experimental sequence, so distinct sequences under one accession remain separate entities.Sequences are normalized by removing whitespace and converting letters to uppercase.Source: §2.6 · pp. 7–8
• Experimental sequences are annotated against their canonical Swiss-Prot sequence as fragments, variants or extensions, but these descriptive annotations do not determine protein identity.The comparison uses direct matching where applicable and pairwise alignment otherwise.Source: §2.6 · p. 8
• 9,668 sequence-defined protein entities represent 8,336 reviewed UniProt accessions, and all retained measurements have a BindingDB-reported experimental target sequence.Swiss-Prot sequences serve as canonical references for validation, annotation and sequence comparison.Source: §2.6 · p. 8

2.7 Interaction aggregation and task views

• IC50, Ki, Kd and EC50 are aggregated independently, with one interaction per endpoint-specific Drug ID and Protein Entity ID combination.Each interaction reports counts of exact, left-censored and right-censored measurements.Source: §2.7 · p. 8
• Classification interactions receive class 1 or class 0 only when evidence supports that class without evidence for the other; conflicting evidence is flagged, and ambiguous-only interactions lack a valid label.Source: §2.7 · p. 8
• Regression targets use the median exact measurement in nM, transformed to the endpoint-specific p-value; interactions without exact measurements have no valid regression target.Source: §2.7 · p. 8
• Interactions spanning at least 100-fold across exact measurements are flagged as dispersed but retained.A 100-fold span is equivalent to a difference of at least two pE units.Source: §2.7 · p. 8

2.8 Measurement provenance and source grouping

• Source documents are grouped using normalized DOI, PMID, patent number or BindingDB entry DOI, with a release-specific fallback when none is available.The fallback prevents unrelated records with missing provenance from sharing one group.Source: §2.8 · p. 9
• Keeping source-document identifiers enables measurements from the same article or patent to remain in one split, reducing information leakage across evaluation sets.Source: §2.8 · p. 9

2.9 Release organization and validation

• The release includes shared tables, endpoint-specific views and separate archival tables for multi-chain measurements; validation checks passed for identifiers, references, labels, targets and reported row counts.Parquet tables use Zstandard compression and are accompanied by a column dictionary, manifest, exclusion reports and statistical reports.Source: §2.9 · p. 9

3.1 Dataset composition

• 2,919,717 measurements yielded 2,275,683 endpoint-specific interactions, while construct-aware processing retained 63,341 interactions that accession-only aggregation would merge.The release includes 1,291,872 standardized ligands and 9,668 sequence-defined protein entities across 8,336 reviewed UniProt accessions.Source: §3.1 · pp. 9–10

3.2 Organism coverage

• 431 organism labels are represented; human proteins account for 6,050 entities and 1,950,735 interaction records (85.72%).No organism-based restriction was applied, and the protein table retains organism names and taxonomy identifiers for selecting organism-specific datasets.Source: §3.2 · p. 10

3.3 Source provenance

• 53,046 source groups cover the 2,919,717 retained measurements, with only 705 measurements (0.02%) requiring a fallback identifier.Article DOI groups 1,420,613 measurements (48.66%), while patents account for 1,228,477 (42.08%).Source: §3.3 · p. 12

4.1 Released data

• Four connected record types link individual measurements to shared ligand and protein-entity tables, with repeated evidence aggregated separately by endpoint.The combined tables contain 2,919,717 measurements, 1,291,872 ligands, 9,668 protein entities, and 2,275,683 interactions.Source: §4.1 · p. 12, Figure 3 · p. 13

4.2 Measurement records

• Measurement records retain endpoint values, normalized relations, threshold-classification evidence, and available assay and source provenance.Inequality-qualified measurements do not have exact regression targets; endpoint-specific measurement files are also provided.Source: §4.2 · p. 14

4.3 Ligand records

• Ligand records provide standardized structures, identifiers, and calculated molecular properties, which were descriptive and did not restrict dataset inclusion.Chemical identity within the release is determined by standardized canonical isomeric SMILES rather than names or external identifiers.Source: §4.3 · p. 14

4.4 Protein-entity records

• Protein-entity records link each reviewed UniProt accession to its BindingDB-reported experimental sequence and corresponding canonical Swiss-Prot sequence.Construct-characterization fields describe sequence relationships; protein identity is determined by accession and normalized experimental sequence.Source: §4.4 · pp. 14–15

4.5 Construct-aware interaction records

• Interaction records aggregate measurements by endpoint, ligand, and protein entity; classification and regression validity are recorded separately.Classification uses valid binary targets, while regression targets are calculated only from exact measurements; relation and boundary fields remain available for interval-aware modeling.Source: §4.5 · p. 15

4.6 Archival records and supporting files

• Supplementary archives retain endpoint-explicit observations and available chain-level information for multi-chain targets, which are excluded from model-ready interaction tables.A column dictionary, release manifest, exclusion and attrition reports, construct audits, endpoint summaries, and validation reports accompany the release.Source: §4.6 · pp. 15–16

5.1 Integrity of the released dataset

• All final identifier, connection, schema, and task-field checks passed, and the preprocessing workflow deterministically regenerated the reported counts.Checks also confirmed unique identifiers, valid linked reference records, and agreement between combined and endpoint-specific table counts.Source: §5.1 · p. 16

5.2 Validation of task targets and repeated evidence

• 70,000 interactions lacked valid binary labels: 38,687 had conflicting class evidence and 31,313 had only ambiguous measurements.Conflicting classification records were retained but excluded from the valid classification view, and regression targets required at least one exact measurement.Source: §5.2 · p. 16

5.3 Validation of construct-aware aggregation

• Construct-aware aggregation produced 2,275,683 interactions versus 2,212,342 under accession-only aggregation, preserving 63,341 endpoint-specific interactions.Accession-only aggregation also raised classification conflicts from 38,687 to 42,312, and 4,204 drug–accession groups had outcomes dependent on experimental sequence.Source: §5.3 · p. 17

6.1 Selecting endpoints and prediction tasks

• Classification uses y class only for classification-valid records, while regression uses endpoint-matched p-values only for regression-valid records.Conflicting or exclusively ambiguous evidence is not a standard binary target; dispersed exact measurements may be handled according to study objectives.Source: §6.1 · pp. 17–18
• 78,449 drug–protein pairs occur under multiple endpoints, so cross-endpoint experiments should group identical Drug ID–Protein Entity ID pairs to prevent train–evaluation overlap.The interaction tables are recommended for modeling; measurement tables retain observations, assay information, inequalities, repeated evidence, and provenance.Source: §6.1 · p. 18

6.2 Filtering organisms and protein targets

• Organism subsets should use Taxon ID rather than free-text names; for example, human interactions use Taxon ID = 9606.The Organism field is provided for interpretation and reporting.Source: §6.2 · p. 18
• Protein Entity ID separates experimental sequences, whereas UniProt ID groups their constructs; cold-UniProt splits therefore test generalization to unseen accessions more strictly.A cold-protein-entity split can place another construct from the same UniProt accession in a different partition.Source: §6.2 · p. 18

6.3 Provenance and leakage-aware evaluation

• Source-grouped partitions keep measurements from a source together before aggregation, while temporal splits use publication year and random splits may separate related records.Report the endpoint, task-validity filter, organism subset, identity level, and split procedure for reproducible interpretation and comparison.Source: §6.3 · pp. 18–19

6.4 Limitations

• The 1,000 nM cutoff is operational, class 0 is threshold-negative rather than confirmed absence, and EC50 alone does not identify agonist versus antagonist responses.Source: §6.4 · p. 19
• Reported sequences cannot recover undocumented construct differences; canonical-sequence matches do not prove full-length use, and affinity-tag-only variants remain distinct under strict identity.For more conservative cold-protein evaluation, group all entities sharing a UniProt ID.Source: §6.4 · p. 19
• Source and assay metadata are incomplete for some records, particularly patent-derived measurements.Inspect measurement-level provenance and quality fields when creating application-specific subsets.Source: §6.4 · p. 19

Data Availability

• The deposit contains endpoint-specific measurement and interaction tables, shared ligand and protein tables, archival multi-chain records, documentation, validation summaries, and integrity checksums.Source: Data Availability · p. 19
• The deposit is restricted during peer review and is to become publicly available upon publication under the Creative Commons Attribution–ShareAlike 4.0 International licence.Source: Data Availability · p. 19
• The manifest identifies original BindingDB and UniProtKB/Swiss-Prot source files by release, filename, and checksum, but does not redistribute them as unmodified files.Source: Data Availability · p. 19