#!/usr/bin/env python3 """Prepare additional M1 datasets for Exp 1 and Exp 4. Exp 1 (NegBioDB vs. random negatives): - negbiodb_m1_uniform_random.parquet — untested pairs, uniform sampling - negbiodb_m1_degree_matched.parquet — untested pairs, degree-distribution matched Exp 4 (Node degree bias — DDB split): - negbiodb_m1_balanced_ddb.parquet — M1 balanced + split_degree_balanced column Usage: uv run python scripts/prepare_exp_data.py uv run python scripts/prepare_exp_data.py --data-dir exports/ --db data/negbiodb.db Prerequisite: - exports/negbiodb_m1_balanced.parquet (from export_ml_dataset.py) - exports/chembl_positives_pchembl6.parquet - data/negbiodb.db (Exp 1 only) """ from __future__ import annotations import argparse import logging import sys from pathlib import Path import pandas as pd logging.basicConfig( level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s", datefmt="%H:%M:%S", ) logger = logging.getLogger(__name__) # Root of the project ROOT = Path(__file__).parent.parent def prepare_exp1_uniform( db_path: Path, positives: pd.DataFrame, output_dir: Path, seed: int = 42 ) -> Path: """Generate M1 dataset with uniform random negatives (Exp 1 control A).""" from negbiodb.export import generate_uniform_random_negatives logger.info("Generating uniform random negatives (Exp 1 control A)...") result = generate_uniform_random_negatives( negbiodb_path=db_path, positives=positives, n_samples=len(positives), # 1:1 balanced output_dir=output_dir, seed=seed, ) out_path = Path(result["path"]) logger.info( "Saved %d rows to %s (pos=%d, neg=%d)", result["total"], out_path.name, result["n_pos"], result["n_neg"], ) return out_path def prepare_exp1_degree_matched( db_path: Path, positives: pd.DataFrame, output_dir: Path, seed: int = 42 ) -> Path: """Generate M1 dataset with degree-matched random negatives (Exp 1 control B).""" from negbiodb.export import generate_degree_matched_negatives logger.info("Generating degree-matched random negatives (Exp 1 control B)...") result = generate_degree_matched_negatives( negbiodb_path=db_path, positives=positives, n_samples=len(positives), # 1:1 balanced output_dir=output_dir, seed=seed, ) out_path = Path(result["path"]) logger.info( "Saved %d rows to %s (pos=%d, neg=%d)", result["total"], out_path.name, result["n_pos"], result["n_neg"], ) return out_path def prepare_exp4_ddb( m1_path: Path, output_dir: Path, seed: int = 42 ) -> Path: """Add full-task split_degree_balanced column to M1 balanced for Exp 4. The DDB split is computed on the merged M1 benchmark graph itself so both positives and negatives are assigned under the same degree-balanced policy. """ from negbiodb.export import add_degree_balanced_split logger.info("Building M1 balanced + DDB split (Exp 4)...") m1 = pd.read_parquet(m1_path) logger.info("M1 balanced loaded: %d rows", len(m1)) m1_ddb = add_degree_balanced_split(m1, seed=seed) logger.info( "DDB split distribution (all rows): %s", m1_ddb["split_degree_balanced"].value_counts().to_dict(), ) by_label_same = ( m1_ddb.assign(same_as_random=m1_ddb["split_degree_balanced"] == m1_ddb["split_random"]) .groupby("Y", dropna=False)["same_as_random"] .mean() .to_dict() ) logger.info("Fraction matching split_random by label: %s", by_label_same) out_path = output_dir / "negbiodb_m1_balanced_ddb.parquet" m1_ddb.to_parquet(out_path, index=False) logger.info("Saved %d rows to %s", len(m1_ddb), out_path.name) return out_path def verify_schema(path: Path, reference_path: Path) -> None: """Check that output schema matches the reference M1 file.""" import pyarrow.parquet as pq ref_cols = set(pq.read_schema(reference_path).names) out_cols = set(pq.read_schema(path).names) extra = out_cols - ref_cols missing = ref_cols - out_cols if extra or missing: logger.warning( "%s schema differs from reference — extra: %s, missing: %s", path.name, extra, missing, ) else: logger.info("%s schema matches reference M1 ✓", path.name) def main(argv: list[str] | None = None) -> int: parser = argparse.ArgumentParser(description="Prepare Exp 1 and Exp 4 datasets") parser.add_argument( "--data-dir", type=Path, default=ROOT / "exports", help="Directory containing M1 parquets (default: exports/)" ) parser.add_argument( "--db", type=Path, default=ROOT / "data" / "negbiodb.db", help="Path to negbiodb.db SQLite file" ) parser.add_argument( "--seed", type=int, default=42, help="Random seed for negative sampling (default: 42)" ) parser.add_argument( "--skip-exp1", action="store_true", help="Skip Exp 1 random negative generation (slow, ~30 min)" ) parser.add_argument( "--skip-exp4", action="store_true", help="Skip Exp 4 DDB split preparation" ) args = parser.parse_args(argv) data_dir: Path = args.data_dir db_path: Path = args.db # Verify inputs exist m1_balanced = data_dir / "negbiodb_m1_balanced.parquet" positives_path = data_dir / "chembl_positives_pchembl6.parquet" required_inputs = [m1_balanced] if not args.skip_exp1: required_inputs.append(positives_path) for p in required_inputs: if not p.exists(): logger.error("Required file missing: %s", p) return 1 if not args.skip_exp1 and not db_path.exists(): logger.error("Database not found: %s", db_path) return 1 logger.info("Using data_dir=%s, db=%s, seed=%d", data_dir, db_path, args.seed) # --- Exp 1: random negatives ---------------------------------------- if not args.skip_exp1: positives = pd.read_parquet(positives_path) logger.info("Loaded %d ChEMBL positives", len(positives)) out_uniform = prepare_exp1_uniform(db_path, positives, data_dir, args.seed) verify_schema(out_uniform, m1_balanced) out_deg = prepare_exp1_degree_matched(db_path, positives, data_dir, args.seed) verify_schema(out_deg, m1_balanced) else: logger.info("Skipping Exp 1 (--skip-exp1)") # --- Exp 4: DDB split ----------------------------------------------- if not args.skip_exp4: out_ddb = prepare_exp4_ddb(m1_balanced, data_dir, args.seed) else: logger.info("Skipping Exp 4 (--skip-exp4)") logger.info("Done.") return 0 if __name__ == "__main__": sys.exit(main())