zifeng-ai/target-based-drug-screening

Target-based Molecular Screening

This dataset is the completed, leakage-filtered materialization for the target-based-drug-screening task in rsi-for-science/ai4sci-tasks-drugclip. It provides compact model-ready LMDBs, canonical metadata, deterministic split manifests, frozen DUD-E/LIT-PCBA test archives, and indexed test labels.

Scale

Component	Samples
ProFSADB pretraining subset	500,000
BioLiP2/Q-BioLiP train	356,420
BioLiP2/Q-BioLiP development	55,215
DUD-E	102 targets; 22,805 actives; 1,411,214 decoys
LIT-PCBA	15 targets; 10,033 actives; 2,798,737 inactives

ProFSADB was sampled with a fixed seed after structural validation and test-protein filtering. BioLiP2 was not affordability-downsampled: all source IDs remain indexed, with invalid, artifact, and leakage exclusions recorded. BioLiP train/dev is split by 30%-identity protein clusters. See DATASET_REPORT.md for exact counts and distributions.

Published layout

processed/
  profsa_pretrain_500k/data.lmdb
  biolip_train/data.lmdb
  biolip_dev/data.lmdb
  manifest.json
canonical/
  biolip2.parquet
  profsadb.parquet
  *_vs_test.tsv and protein-cluster metadata
splits/
test/
  test_ligands.parquet
  test_targets.parquet
  test_targets.fasta
  manifest.json
test_archives/
  DUD-E.zip
  LIT-PCBA.zip
metadata/
smoke/

The three LMDBs are the compact, model-ready view. Canonical per-example NPZ files are intentionally not duplicated here because the same arrays are already embedded in these LMDBs; their authoritative tables, split IDs, and hashes are retained. Large raw BioLiP2 and ProFSADB source/extraction caches are also omitted. They are reproducible from the pinned source manifest and pipeline. The frozen DUD-E/LIT-PCBA release ZIPs are retained so evaluation is self-contained.

Each LMDB is a single file (subdir=False) with ASCII integer keys and pickled Python dictionaries. BioLiP records contain smi, pocket, atoms, coordinates, pocket_atoms, pocket_coordinates, and label. ProFSADB also preserves lig_atoms_real and lig_coord_real.

Fixed leakage policy

• Protein leakage: MMseqs2 identity strictly above 30%, requiring 80% bidirectional coverage, is excluded against all frozen test targets.
• Ligand leakage: exact test-active identity or Morgan/ECFP4 Tanimoto strictly above 0.60 is excluded from BioLiP2.
• DUD-E and LIT-PCBA target and ligand populations are complete and are not subsampled.

Integrity and provenance

VALIDATION.json and .complete.json are emitted only after the full pipeline passes. processed/manifest.json contains SHA-256 hashes of every LMDB and its canonical inputs. test/manifest.json pins the original THU-ATOM/DrugCLIP_data revision and test archive hashes. metadata/protocol.json and metadata/source_manifest.json pin the processing policy and sources.

The source datasets retain their respective upstream terms. The license: other label is deliberate: it does not replace or relicense DUD-E, LIT-PCBA, BioLiP2/Q-BioLiP, ProFSADB, DrugCLIP, or their source structures. Users must review and comply with those upstream terms.

Smoke validation

smoke/result.json records a tiny NumPy MLP engineering smoke test over real LMDB records. It verifies data loading, forward/backward training, checkpoint writing, dev inference, and DUD-E input traversal. It is not an official benchmark baseline; its DUD-E score must not be used for scientific comparison.