bhumika-tewari-282006/ayurvedic-affinity-training-data

Ayurvedic Affinity Training Data

The real, processed protein-ligand binding-affinity training set used to train the companion model bhumika-tewari-282006/ayurvedic-drug-discovery-affinity-model (RandomForest regressor for pKd prediction).

What this is

181 real protein-ligand complexes from PDBBind v2013-core, each with:

• a real PDB structure identifier (pdb_code)
• the real ligand SMILES
• the real experimentally-measured binding affinity (pkd)
• 39 real, computed features: 10 docking-derived terms from an actual AutoDock Vina --score_only run on the complex's real crystal pose, 18 RDKit QSAR/physicochemical descriptors, and 11 Ayurvedic/phytochemical-class heuristic features (mostly zero/default here, since PDBBind complexes are not Ayurvedic phytochemicals — included so the exact same 39-feature vector shape is used at both training and inference time against real Ayurvedic compound data downstream).

This is not synthetic or fabricated data. Every value in every row comes from a real crystal structure, a real docking run, or a real cheminformatics calculation — nothing here was generated, interpolated, or estimated to fill a gap.

How it was built

Source complexes: real PDBBind v2013-core (real crystal structures, real experimentally-measured affinities). Of 195 candidate complexes considered, 181 (93%) were successfully processed end-to-end:

1. Ligand and receptor prepared to PDBQT via OpenBabel from the real crystal structure.
2. Real AutoDock Vina --score_only run against the ligand's real crystal pose, producing real physics-based energy terms (van der Waals, hydrogen-bonding, hydrophobic, electrostatic) — not a simulated or approximated score.
3. RDKit-computed QSAR descriptors from the real ligand structure.
4. 14 candidate complexes were honestly excluded rather than patched with fabricated values: 10 had a ligand SDF that RDKit genuinely could not parse, 1 had a real Vina --score_only exit failure, and 2 had no usable experimental label.

Splits used for training

The companion model was trained with a scaffold split (Bemis-Murcko), not a random split, with a Tanimoto-similarity leakage check between splits (max cross-split similarity found: 0.915, one flagged pair — logged, not hidden):

Split	Rows
Train	127
Validation	18
Test (held out)	36

This dataset ships as one flat 181-row CSV (training_data_real.csv); the train/validation/test assignment used for the companion model's published metrics is not re-derivable from this file alone without re-running the same scaffold-split procedure — reported here for transparency about how the model's numbers were produced, not as a ready-made column in the file.

Columns

pdb_code, smiles, pkd (the real identifier, structure, and label), followed by the 39 feature columns in the same order the companion model expects:

• Docking-derived (10): vina_affinity, ligand_efficiency, intermolecular_energy, torsional_energy, desolvation_energy, num_hbonds, num_hydrophobic_contacts, num_vdw_contacts, electrostatic_score, pocket_occupancy
• RDKit QSAR (18): molecular_weight, logp, hbd, hba, tpsa, rotatable_bonds, heavy_atom_count, ring_count, aromatic_ring_count, fraction_csp3, num_aliphatic_rings, num_heterocycles, formal_charge, num_radical_electrons, bertz_ct, balaban_j, bcut2d_mwhi, qed_score
• Ayurvedic heuristic (11): phytochemical_class_flavonoid, phytochemical_class_alkaloid, phytochemical_class_terpenoid, phytochemical_class_phenolic, phytochemical_class_saponin, phytochemical_class_tannin, is_privileged_scaffold, traditional_use_score, ayurvedic_dosha_vata, ayurvedic_dosha_pitta, ayurvedic_dosha_kapha

Known honest limitation

At 181 rows this is a genuinely small training set by modern ML standards — the companion model's modest test-set R² (≈0.01) is a direct, disclosed consequence of this size, not a data-quality problem. A materially stronger model would need substantially more labeled real binding data, which is why this dataset (and the model card) report these numbers plainly rather than omitting or smoothing them over.

License

PDBBind's own terms govern the underlying structures and affinity values; this processed/featurized derivative is shared under CC-BY-4.0 for research and educational use. This is a computational research artifact — not a source of clinical, therapeutic, or safety claims about any compound.