Scientific Credibility

Dataset Cards

Where every model's training data comes from, how it was cleaned and split, and what isn't documented (reported explicitly, not inferred).

Hepatotoxicity

· 1,511 compounds
Origin
ChEMBL v34 + DILIrank (Chen et al., 2016 — Drug Discovery Today).
Total Compounds
1,511
Preprocessing
Canonical SMILES deduplication (RDKit), activity-unit standardization, and manual literature curation where needed. Feature selection and scaling (StandardScaler) are fit on the training fold only, never on the test fold — no data leakage. Endpoint-specific: label binarization; SMOTE oversampling applied inside each CV training fold only (class balance ~60% hepatotoxic before SMOTE).
Splitting Strategy
Scaffold GroupKFold (k=5): Murcko generic scaffolds computed via rdkit.Chem.Scaffolds.MurckoScaffold.MakeScaffoldGeneric(); all compounds sharing a scaffold are kept in the same fold (sklearn GroupKFold). Split indices are archived in data/processed/{endpoint}_split_indices.npz (random_state=42) to guarantee identical splits across experiments. Train/test: 1,057 / 303.
Quality Control
Canonical SMILES dedup; scaffold-split leakage check; decision threshold calibrated via Youden index with a specificity ≥ 20% constraint.
Licensing
Not documented in the current repository (DILIrank component); raw ChEMBL portion not redistributed per ChEMBL's licensing terms.

Nephrotoxicity

· 302 compounds
Origin
FDA DIRIL database + curated literature.
Total Compounds
302
Preprocessing
Canonical SMILES deduplication (RDKit), activity-unit standardization, and manual literature curation where needed. Feature selection and scaling (StandardScaler) are fit on the training fold only, never on the test fold — no data leakage.
Splitting Strategy
Scaffold GroupKFold (k=5): Murcko generic scaffolds computed via rdkit.Chem.Scaffolds.MurckoScaffold.MakeScaffoldGeneric(); all compounds sharing a scaffold are kept in the same fold (sklearn GroupKFold). Split indices are archived in data/processed/{endpoint}_split_indices.npz (random_state=42) to guarantee identical splits across experiments. Train/test: 118 / 31 (as documented — see the model card's note on this not summing to the stated total N).
Quality Control
Canonical SMILES dedup; scaffold-split leakage check.
Licensing
Not documented in the current repository.

Ames Mutagenicity

· 876 compounds
Origin
Curated literature + ChEMBL v34.
Total Compounds
876 (full curated set); the deployed model trains on an active subset — see data/metadata/ames_manifest.json in the backend repo for exact inclusion/exclusion criteria.
Preprocessing
Canonical SMILES deduplication (RDKit), activity-unit standardization, and manual literature curation where needed. Feature selection and scaling (StandardScaler) are fit on the training fold only, never on the test fold — no data leakage.
Splitting Strategy
Scaffold GroupKFold (k=5): Murcko generic scaffolds computed via rdkit.Chem.Scaffolds.MurckoScaffold.MakeScaffoldGeneric(); all compounds sharing a scaffold are kept in the same fold (sklearn GroupKFold). Split indices are archived in data/processed/{endpoint}_split_indices.npz (random_state=42) to guarantee identical splits across experiments.
Quality Control
Canonical SMILES dedup; scaffold-split leakage check; manifest-documented compound inclusion/exclusion criteria.
Licensing
Not documented in the current repository (literature portion); raw ChEMBL portion not redistributed per ChEMBL's licensing terms.

Skin Sensitization

· 423 compounds
Origin
LLNA (Local Lymph Node Assay) and human repeat-insult patch test data (literature + ChEMBL).
Total Compounds
423
Preprocessing
Canonical SMILES deduplication (RDKit), activity-unit standardization, and manual literature curation where needed. Feature selection and scaling (StandardScaler) are fit on the training fold only, never on the test fold — no data leakage.
Splitting Strategy
Scaffold GroupKFold (k=5): Murcko generic scaffolds computed via rdkit.Chem.Scaffolds.MurckoScaffold.MakeScaffoldGeneric(); all compounds sharing a scaffold are kept in the same fold (sklearn GroupKFold). Split indices are archived in data/processed/{endpoint}_split_indices.npz (random_state=42) to guarantee identical splits across experiments.
Quality Control
Canonical SMILES dedup; scaffold-split leakage check. Note: only 30.77% applicability-domain coverage — a QC-relevant limitation for downstream users.
Licensing
Not documented in the current repository.

hERG Cardiotoxicity

· 10,661 compounds
Origin
ChEMBL v34, target CHEMBL240 (hERG binding assays).
Total Compounds
10,661
Preprocessing
Canonical SMILES deduplication (RDKit), activity-unit standardization, and manual literature curation where needed. Feature selection and scaling (StandardScaler) are fit on the training fold only, never on the test fold — no data leakage. Endpoint-specific: activity values in nM converted to pIC50 = −log₁₀(IC50/1e9); duplicates resolved by keeping the median pIC50 per compound.
Splitting Strategy
Scaffold GroupKFold (k=5): Murcko generic scaffolds computed via rdkit.Chem.Scaffolds.MurckoScaffold.MakeScaffoldGeneric(); all compounds sharing a scaffold are kept in the same fold (sklearn GroupKFold). Split indices are archived in data/processed/{endpoint}_split_indices.npz (random_state=42) to guarantee identical splits across experiments.
Quality Control
Canonical SMILES dedup; median-pIC50 duplicate resolution; scaffold-split leakage check (no shared scaffolds between train/test).
Licensing
Raw ChEMBL data is not redistributed by this project due to ChEMBL's own licensing terms (per model_card.md).

Cytotoxicity

· 387 compounds
Origin
Literature + ChEMBL cytotoxicity assays.
Total Compounds
387
Preprocessing
Canonical SMILES deduplication (RDKit), activity-unit standardization, and manual literature curation where needed. Feature selection and scaling (StandardScaler) are fit on the training fold only, never on the test fold — no data leakage.
Splitting Strategy
Scaffold GroupKFold (k=5): Murcko generic scaffolds computed via rdkit.Chem.Scaffolds.MurckoScaffold.MakeScaffoldGeneric(); all compounds sharing a scaffold are kept in the same fold (sklearn GroupKFold). Split indices are archived in data/processed/{endpoint}_split_indices.npz (random_state=42) to guarantee identical splits across experiments. Train/test: 118 / 34 (as documented — see the model card's note on this not summing to the stated total N).
Quality Control
Canonical SMILES dedup; scaffold-split leakage check.
Licensing
Not documented in the current repository (literature portion); raw ChEMBL portion not redistributed per ChEMBL's licensing terms.

Reproductive Toxicity

· 127 compounds
Origin
Curated teratogenicity and reproductive toxicity literature.
Total Compounds
127 — explicitly documented as severely underpowered.
Preprocessing
Canonical SMILES deduplication (RDKit), activity-unit standardization, and manual literature curation where needed. Feature selection and scaling (StandardScaler) are fit on the training fold only, never on the test fold — no data leakage.
Splitting Strategy
Scaffold GroupKFold (k=5): Murcko generic scaffolds computed via rdkit.Chem.Scaffolds.MurckoScaffold.MakeScaffoldGeneric(); all compounds sharing a scaffold are kept in the same fold (sklearn GroupKFold). Split indices are archived in data/processed/{endpoint}_split_indices.npz (random_state=42) to guarantee identical splits across experiments. Train/test: approximately 100 / 26.
Quality Control
Canonical SMILES dedup; scaffold-split leakage check. Documented roadmap target: expand to ≥1,000 compounds via CERAPP/CoMPARA integration before this endpoint could be reconsidered for the integrated score.
Licensing
Not documented in the current repository.