gschottlender/LigQ_2_benchmark

Ligand-Based Target Benchmark Splits

This dataset contains target-level ligand benchmark splits generated from the same pipeline used for the enrichment-factor evaluations.

Each target has one folder per random_state. Inside each split:

• known_actives.csv: active ligands available to the search method. These are the query/reference ligands.
• evaluation_actives.csv: held-out active ligands for the same target. These are the positives to recover during evaluation.
• putative_inactives.csv: selected putative inactive ligands used as negatives in the evaluation pool. These compounds should be interpreted as a chemical background set, not as experimentally confirmed inactive ligands.
• split_metadata.json: target identifiers, split parameters, and counts.

All three ligand CSV files use the same schema:

target_id,random_state,chem_comp_id,smiles

To evaluate a method, use known_actives.csv as the known ligand set, then rank the ligands from evaluation_actives.csv plus putative_inactives.csv. Assign label 1 to rows from evaluation_actives.csv and label 0 to rows from putative_inactives.csv.

Putative inactive sets are defined at the target_id/random_state level. They are sampled from the same target-specific inactive candidate universe, but the number sampled follows a fixed negative-to-positive ratio. Therefore, when the number of evaluation actives changes across random states, the corresponding putative_inactives.csv can also differ. This keeps the evaluation prevalence consistent within each split.

The putative inactive universe is used as a chemically broad background for virtual-screening evaluation. It is built by excluding compounds annotated as active against proteins from the same Pfam family as the target, then sampling from the remaining ligand database. This design supports enrichment-style benchmarking against a large chemical background, while avoiding the stronger claim that every background compound is experimentally inactive for the target.

Splits are independent by target_id/random_state and should not be mixed before per-split scoring. Aggregate metrics can be computed across targets and random states after scoring.

The active split was generated by clustering active ligands with Butina/Tanimoto, keeping one representative per cluster, and then separating representatives into known and evaluation actives. The default split used here has known_frac=0.1, test_to_known_ratio=10, and butina_cutoff=0.8.