tim-shu/drug-affinity-multisource

pKi: complete molecular structures and baseline comparison

The corpus contains 249,613 pairs. The common baseline cohort contains 204,847 pairs for 162 human proteins. A pair is identified by its complete canonical isomeric SMILES and protein, retaining every recorded disconnected component. Different salt forms are separate pairs with separate measurement aggregates. No record is removed because several complete forms share a selected component.

from datasets import load_dataset
data = load_dataset("tim-shu/drug-affinity-multisource", "baseline_benchmark")

affinity_pki contains all curated pairs; benchmark_target marks supported proteins. molecules and source_observations are reference tables; their single train split is a storage convention. Join their structural IDs to the pair table's actual scaffold_split.

Sources and counts in every candidate

Source	Total	Train	Validation	Test
ChEMBL	175,255	139,761 (79.7%)	17,102 (9.8%)	18,392 (10.5%)
BindingDB	26,165	20,803 (79.5%)	2,908 (11.1%)	2,454 (9.4%)
DrugCentral	3,427	2,766 (80.7%)	417 (12.2%)	244 (7.1%)

Train rows train target-specific models. All 20,427 validation pairs select the setup; test is evaluated after selection is recorded. Proteins need at least 300 pairs overall, 100 train, 10 validation and 10 test. All candidates receive identical rows, labels and groups.

Label preparation

The response is pKi = −log10(Ki in mol/L) = 9 − log10(Ki in nM). Larger pKi means stronger binding.

• ChEMBL 37: exact positive Ki in nM, binding assay, single human protein, confidence 9, no annotated mutation, standard_flag=1, potential_duplicate=0, data_validity_comment absent or Manually validated. Confidence describes target assignment, not Ki accuracy. Labels are rebuilt from individual activities and complete canonical_smiles, without previously component-aggregated labels.
• BindingDB 202610: native literature/patent records, one human protein, exact positive Ki; ChEMBL imports excluded. Experimental sequence ≥50 residues must match a contiguous segment of the canonical human UniProt sequence. Convert Ki in nM to pKi.
• DrugCentral: join structures and activities; human targets, ACT_TYPE=Ki, RELATION="=". ACT_VALUE is already pKi, with no second logarithm.

Only invalid or empty SMILES fail structural checks. Canonicalization preserves all components, charge and specified stereochemistry. No size, element-list, component-count or carbon-presence exclusions apply. Within-source repeats of the same complete structure/protein use median pKi; ranges above 2 pKi are quarantined. Cross-source copies use ChEMBL > BindingDB > DrugCentral priority. Auxiliary disagreements above 2 pKi exclude pairs without a ChEMBL label; ChEMBL labels retain their flags. Provenance is preserved in source and measurement identifiers.

Sources were chosen for quantitative Ki, traceable structures/proteins, useful coverage and open reusable data. PubChem was excluded for insufficient independent coverage; DAVIS (Kd), KIBA (mixed score), Open Targets (disease associations) and ClinicalTrials.gov (clinical studies) do not supply this task's Ki labels. No authorized redistributable PDBbind or DrugBank extraction was obtained. UniProt supports identity checks, not additional affinity labels or RF features.

Common split and feature-only ablations

Murcko scaffolds are computed from complete records. Co-occurring nonempty component scaffolds, and all records sharing any of them, are joined into a common group. Fully ringless structures share ACYCLIC. No representative component is selected for curation or splitting. SHA256(seed=42, group) assigns groups to nominal 80/10/10 train/validation/test; actual pair proportions differ. Exact selected-component identities are additionally checked to be disjoint across partitions before training, without dropping rows.

Only feature generation derives parent_smiles: largest carbon-containing component by heavy-atom count, falling back to the largest component of any composition, with canonical SMILES resolving ties. This experimental heuristic cannot change identity, pKi, retained rows or split. Main curated pairs and source observations contain only complete identities. Different forms remain separate even if their experimental fingerprints match.

RF uses chiral Morgan fingerprints, one forest per protein. Seventeen RF setups test full/selected-component fingerprints, eight extra properties, properties alone, fragment counts, larger minimum leaves, 5% train-bit dropout, fingerprint radius and vector length. The radius ablation compares 1, 2 and 3 at 2048 bits on complete structures, both without and with the same eight properties. The length ablation compares 1024, 2048 and 4096 bits at radius 2 in the same two feature families. For the length ablation, the configured number of features sampled at each tree split is held at 45: sqrt gives 45 for 2048 and 2056 features, while the new lengths explicitly use max_features=45. These are separate comparisons, not a full radius-by-length grid. Other ablations and neighbor heuristics use radius 2 and 2048 bits. Four heuristics use per-target training mean, median, nearest-neighbor and similarity-weighted 5-neighbor pKi. Identifiers, sources and measured affinities are not features.

Frozen result

Selected by target-macro validation MAE averaged over seeds 42/43/44: RF_full_descriptors. Effective Morgan radius: 2; vector length: 2048 bits. Fixed final seed: 42. Validation MAE 0.5941 pKi; test MAE across three seeds 0.5762 ± 0.0008 pKi. The next validation candidate is RF_parent_descriptors, gap 0.0001 pKi. SD measures RF randomness, not uncertainty across splits. Selection establishes the best measured result among these candidates on this split, not universally optimal fingerprint settings.

Labels and splitting were rebuilt; scores are not directly comparable with older component-grouped snapshots. Some underlying data were examined in earlier experiments, so this is not an independent blind evaluation. Affinity predictions do not establish therapeutic efficacy.

baseline_study/final_baseline.json fixes the recipe and input checksums. Per-target metrics, predictions, checks and executed EDA.ipynb accompany the snapshot. In prepared/benchmark/fingerprints.npz, radius-2, 2048-bit matrices full and parent align to feature_index in structures.parquet. baseline_study/prepared/radius_fingerprints.npz adds full_radius_1 and full_radius_3 with the same rows and 2048-bit width. baseline_study/prepared/size_fingerprints.npz adds full_bits_1024 and full_bits_4096, both at radius 2 with the same row alignment. baseline_study/prepared/extra_features.npz contains full_descriptors, parent_descriptors, full_counts. The eight properties are molecular weight, logP, TPSA, HBD, HBA, rotatable bonds, rings and heavy atoms.

Reproduction and attribution

Use baseline_study/reproduction/ as project root; install requirements.txt. Copy prepared/benchmark/ to data/multisource_v5/benchmark/ there. Run python scripts/run_baseline_study.py. It recalculates features, compares candidates and saves forests in output/final_baseline_models/v4/. Raw downloads are unnecessary for this rerun. Curation input files and SHA-256 appear in metadata/curation/preparation.json.

The transformed collection is CC BY-SA 4.0; original source notices and attribution remain applicable. See LICENSES.md for source grants and citations. This revision retains complete structures and recomputes labels, groups, features and evaluations. manifest.json records every file checksum. Data licenses do not establish a software license.