README.md

Scripts

Self-contained tools that operate on the public artifacts in data/. None of these scripts need anything outside this repository.

audit_tanimoto_leak.py — exhaustive cross-fold Tanimoto audit

Verifies the tan70 / tan60 splits in data/splits/ do not contain any cross-fold Morgan radius-2 2048-bit Tanimoto edges at or above the declared cutoff. The splits were built with a graph constraint that forbids such edges, so a clean deposit produces zero violations.

# full tan70 audit (cutoff 0.70, ~22B comparisons, ~15-45 min)
python scripts/audit_tanimoto_leak.py --split tan70

# full tan60 audit (cutoff 0.60, ~8B comparisons, ~6-20 min)
python scripts/audit_tanimoto_leak.py --split tan60

# fast smoke test on 200 train compounds
python scripts/audit_tanimoto_leak.py --split tan70 --sample 200

Output:

• Prints a per-~2%-progress line showing leaks-so-far and ETA.
• Writes violations (if any) to audit_<split>_leaks.csv next to the CWD, with columns row_idx_train, row_idx_other, fold_other, tanimoto.
• Exits with status 0 on PASS (zero violations) or 1 on FAIL.
• Deletes the empty violations CSV on PASS to keep the directory clean.

Algorithm: O(n_train x n_other) exhaustive — for every train compound, compute Tanimoto against every val and every test compound via RDKit's DataStructs.BulkTanimotoSimilarity. No locality-sensitive hashing or nearest-neighbor index — the point is to make the audit obvious to read, not fast at the asymptotic level.

Dependencies (declare them once with pip install rdkit numpy or via the audit optional extra in pyproject.toml):

• rdkit >= 2025.9.6
• numpy >= 2.0