scripts/create_disjoint_splits.py

#!/usr/bin/env python3
"""
Create disjoint splits from drug_rna_cds_sampled_11 dataset.

Creates three datasets:
1. drug_rna_cds_disjoint (fully disjoint if possible)
2. drug_rna_cds_disjoint_rna (RNA-disjoint)
3. drug_rna_cds_disjoint_compound (compound-disjoint)
"""

import pandas as pd
import numpy as np
from pathlib import Path
from collections import defaultdict
import argparse


def analyze_disjoint_feasibility(df):
    """Analyze if fully disjoint split is feasible."""
    print("=" * 80)
    print("ANALYZING DISJOINT SPLIT FEASIBILITY")
    print("=" * 80)

    n_compounds = df['SMILES'].nunique()
    n_rnas = df['RNA_seq'].nunique()
    n_samples = len(df)

    print(f"\nDataset statistics:")
    print(f"  Total samples: {n_samples:,}")
    print(f"  Unique compounds: {n_compounds:,}")
    print(f"  Unique RNAs: {n_rnas:,}")

    # Analyze compound-RNA pairs
    pairs = df.groupby(['SMILES', 'RNA_seq']).size().reset_index(name='count')
    print(f"  Unique (compound, RNA) pairs: {len(pairs):,}")

    # Check samples per entity
    samples_per_compound = df.groupby('SMILES').size()
    samples_per_rna = df.groupby('RNA_seq').size()

    print(f"\n  Samples per compound: mean={samples_per_compound.mean():.1f}, median={samples_per_compound.median():.1f}")
    print(f"  Samples per RNA: mean={samples_per_rna.mean():.1f}, median={samples_per_rna.median():.1f}")

    # Estimate fully disjoint feasibility
    # For fully disjoint: need non-overlapping sets of (compound, RNA) pairs
    # This is only possible if we can partition compounds and RNAs

    # Simple heuristic: check if we have enough samples
    min_samples_needed = int(0.15 * n_samples)  # For test set (15%)

    print(f"\n  Minimum samples needed for test (15%): {min_samples_needed:,}")

    # Check if we can allocate compounds/RNAs
    n_compounds_test = int(0.15 * n_compounds)
    n_rnas_test = int(0.15 * n_rnas)

    print(f"  If we allocate 15% of compounds: {n_compounds_test} compounds")
    print(f"  If we allocate 15% of RNAs: {n_rnas_test} RNAs")
    print(f"  Max possible samples (all pairs): {n_compounds_test * n_rnas_test:,}")

    # Check actual connectivity
    compound_to_rnas = df.groupby('SMILES')['RNA_seq'].apply(set).to_dict()
    rna_to_compounds = df.groupby('RNA_seq')['SMILES'].apply(set).to_dict()

    # Calculate how connected the graph is
    avg_rnas_per_compound = df.groupby('SMILES')['RNA_seq'].nunique().mean()
    avg_compounds_per_rna = df.groupby('RNA_seq')['SMILES'].nunique().mean()

    print(f"\n  Connectivity:")
    print(f"    Avg RNAs per compound: {avg_rnas_per_compound:.1f}")
    print(f"    Avg compounds per RNA: {avg_compounds_per_rna:.1f}")

    # Determine feasibility
    fully_disjoint_feasible = (n_compounds_test * n_rnas_test >= min_samples_needed)

    print(f"\n  Fully disjoint split feasible: {'YES' if fully_disjoint_feasible else 'NO'}")

    return fully_disjoint_feasible


def create_fully_disjoint_split(df, train_ratio=0.7, val_ratio=0.15, seed=42):
    """Create fully disjoint split (compounds AND RNAs don't overlap)."""
    print("\n" + "=" * 80)
    print("CREATING FULLY DISJOINT SPLIT")
    print("=" * 80)

    np.random.seed(seed)

    # Get unique entities
    compounds = df['SMILES'].unique()
    rnas = df['RNA_seq'].unique()

    # Shuffle
    np.random.shuffle(compounds)
    np.random.shuffle(rnas)

    # Split compounds
    n_compounds = len(compounds)
    n_train_comp = int(train_ratio * n_compounds)
    n_val_comp = int(val_ratio * n_compounds)

    train_compounds = set(compounds[:n_train_comp])
    val_compounds = set(compounds[n_train_comp:n_train_comp + n_val_comp])
    test_compounds = set(compounds[n_train_comp + n_val_comp:])

    # Split RNAs
    n_rnas = len(rnas)
    n_train_rna = int(train_ratio * n_rnas)
    n_val_rna = int(val_ratio * n_rnas)

    train_rnas = set(rnas[:n_train_rna])
    val_rnas = set(rnas[n_train_rna:n_train_rna + n_val_rna])
    test_rnas = set(rnas[n_train_rna + n_val_rna:])

    # Assign samples based on BOTH compound and RNA membership
    train_df = df[(df['SMILES'].isin(train_compounds)) & (df['RNA_seq'].isin(train_rnas))].copy()
    val_df = df[(df['SMILES'].isin(val_compounds)) & (df['RNA_seq'].isin(val_rnas))].copy()
    test_df = df[(df['SMILES'].isin(test_compounds)) & (df['RNA_seq'].isin(test_rnas))].copy()

    print(f"\n  Allocated:")
    print(f"    Train: {len(train_compounds)} compounds, {len(train_rnas)} RNAs → {len(train_df)} samples")
    print(f"    Val:   {len(val_compounds)} compounds, {len(val_rnas)} RNAs → {len(val_df)} samples")
    print(f"    Test:  {len(test_compounds)} compounds, {len(test_rnas)} RNAs → {len(test_df)} samples")

    total_samples = len(train_df) + len(val_df) + len(test_df)
    print(f"\n  Total samples retained: {total_samples:,} / {len(df):,} ({100*total_samples/len(df):.1f}%)")

    if total_samples < 0.5 * len(df):
        print(f"\n  ⚠️  Warning: Lost {100*(1-total_samples/len(df)):.1f}% of samples")
        print(f"      Fully disjoint split may not be practical for this dataset")
        return None

    return train_df, val_df, test_df


def create_rna_disjoint_split(df, train_ratio=0.7, val_ratio=0.15, seed=42):
    """Create RNA-disjoint split (RNAs don't overlap across splits)."""
    print("\n" + "=" * 80)
    print("CREATING RNA-DISJOINT SPLIT (Target: 70:15:15)")
    print("=" * 80)

    np.random.seed(seed)

    # Get RNAs with their sample counts
    rna_counts = df.groupby('RNA_seq').size().reset_index(name='count')
    rna_counts = rna_counts.sample(frac=1, random_state=seed).reset_index(drop=True)  # Shuffle

    # Target sample counts
    total_samples = len(df)
    target_train = int(train_ratio * total_samples)
    target_val = int(val_ratio * total_samples)
    target_test = total_samples - target_train - target_val

    print(f"\n  Target sample distribution:")
    print(f"    Train: {target_train:,} ({train_ratio*100:.0f}%)")
    print(f"    Val:   {target_val:,} ({val_ratio*100:.0f}%)")
    print(f"    Test:  {target_test:,} ({(1-train_ratio-val_ratio)*100:.0f}%)")

    # Greedy assignment: assign RNAs to splits to match target ratios
    train_rnas = []
    val_rnas = []
    test_rnas = []

    train_count = 0
    val_count = 0
    test_count = 0

    for _, row in rna_counts.iterrows():
        rna = row['RNA_seq']
        count = row['count']

        # Assign to the split that needs samples most
        train_deficit = target_train - train_count
        val_deficit = target_val - val_count
        test_deficit = target_test - test_count

        if train_deficit >= val_deficit and train_deficit >= test_deficit:
            train_rnas.append(rna)
            train_count += count
        elif val_deficit >= test_deficit:
            val_rnas.append(rna)
            val_count += count
        else:
            test_rnas.append(rna)
            test_count += count

    train_rnas = set(train_rnas)
    val_rnas = set(val_rnas)
    test_rnas = set(test_rnas)

    # Assign samples based on RNA
    train_df = df[df['RNA_seq'].isin(train_rnas)].copy()
    val_df = df[df['RNA_seq'].isin(val_rnas)].copy()
    test_df = df[df['RNA_seq'].isin(test_rnas)].copy()

    print(f"\n  Split RNAs:")
    print(f"    Train: {len(train_rnas)} RNAs ({len(train_df):,} samples, {100*len(train_df)/len(df):.1f}%)")
    print(f"    Val:   {len(val_rnas)} RNAs ({len(val_df):,} samples, {100*len(val_df)/len(df):.1f}%)")
    print(f"    Test:  {len(test_rnas)} RNAs ({len(test_df):,} samples, {100*len(test_df)/len(df):.1f}%)")

    # Verify disjointness
    assert len(train_rnas & val_rnas) == 0, "Train and val RNAs overlap!"
    assert len(train_rnas & test_rnas) == 0, "Train and test RNAs overlap!"
    assert len(val_rnas & test_rnas) == 0, "Val and test RNAs overlap!"

    print(f"\n  ✓ RNA-disjoint verified: no RNA overlap across splits")

    # Check compound overlap (expected)
    train_compounds = set(train_df['SMILES'].unique())
    val_compounds = set(val_df['SMILES'].unique())
    test_compounds = set(test_df['SMILES'].unique())

    overlap_train_val = len(train_compounds & val_compounds)
    overlap_train_test = len(train_compounds & test_compounds)

    print(f"\n  Compound overlap (expected):")
    print(f"    Train-Val: {overlap_train_val} compounds")
    print(f"    Train-Test: {overlap_train_test} compounds")

    return train_df, val_df, test_df


def create_compound_disjoint_split(df, train_ratio=0.7, val_ratio=0.15, seed=42):
    """Create compound-disjoint split (compounds don't overlap across splits)."""
    print("\n" + "=" * 80)
    print("CREATING COMPOUND-DISJOINT SPLIT (Target: 70:15:15)")
    print("=" * 80)

    np.random.seed(seed)

    # Get compounds with their sample counts
    compound_counts = df.groupby('SMILES').size().reset_index(name='count')
    compound_counts = compound_counts.sample(frac=1, random_state=seed).reset_index(drop=True)  # Shuffle

    # Target sample counts
    total_samples = len(df)
    target_train = int(train_ratio * total_samples)
    target_val = int(val_ratio * total_samples)
    target_test = total_samples - target_train - target_val

    print(f"\n  Target sample distribution:")
    print(f"    Train: {target_train:,} ({train_ratio*100:.0f}%)")
    print(f"    Val:   {target_val:,} ({val_ratio*100:.0f}%)")
    print(f"    Test:  {target_test:,} ({(1-train_ratio-val_ratio)*100:.0f}%)")

    # Greedy assignment: assign compounds to splits to match target ratios
    train_compounds = []
    val_compounds = []
    test_compounds = []

    train_count = 0
    val_count = 0
    test_count = 0

    for _, row in compound_counts.iterrows():
        compound = row['SMILES']
        count = row['count']

        # Assign to the split that needs samples most
        train_deficit = target_train - train_count
        val_deficit = target_val - val_count
        test_deficit = target_test - test_count

        if train_deficit >= val_deficit and train_deficit >= test_deficit:
            train_compounds.append(compound)
            train_count += count
        elif val_deficit >= test_deficit:
            val_compounds.append(compound)
            val_count += count
        else:
            test_compounds.append(compound)
            test_count += count

    train_compounds = set(train_compounds)
    val_compounds = set(val_compounds)
    test_compounds = set(test_compounds)

    # Assign samples based on compound
    train_df = df[df['SMILES'].isin(train_compounds)].copy()
    val_df = df[df['SMILES'].isin(val_compounds)].copy()
    test_df = df[df['SMILES'].isin(test_compounds)].copy()

    print(f"\n  Split compounds:")
    print(f"    Train: {len(train_compounds)} compounds ({len(train_df):,} samples, {100*len(train_df)/len(df):.1f}%)")
    print(f"    Val:   {len(val_compounds)} compounds ({len(val_df):,} samples, {100*len(val_df)/len(df):.1f}%)")
    print(f"    Test:  {len(test_compounds)} compounds ({len(test_df):,} samples, {100*len(test_df)/len(df):.1f}%)")

    # Verify disjointness
    assert len(train_compounds & val_compounds) == 0, "Train and val compounds overlap!"
    assert len(train_compounds & test_compounds) == 0, "Train and test compounds overlap!"
    assert len(val_compounds & test_compounds) == 0, "Val and test compounds overlap!"

    print(f"\n  ✓ Compound-disjoint verified: no compound overlap across splits")

    # Check RNA overlap (expected)
    train_rnas = set(train_df['RNA_seq'].unique())
    val_rnas = set(val_df['RNA_seq'].unique())
    test_rnas = set(test_df['RNA_seq'].unique())

    overlap_train_val = len(train_rnas & val_rnas)
    overlap_train_test = len(train_rnas & test_rnas)

    print(f"\n  RNA overlap (expected):")
    print(f"    Train-Val: {overlap_train_val} RNAs")
    print(f"    Train-Test: {overlap_train_test} RNAs")

    return train_df, val_df, test_df


def save_splits(train_df, val_df, test_df, output_dir, dataset_name):
    """Save splits to files."""
    output_path = Path(output_dir)
    output_path.mkdir(parents=True, exist_ok=True)

    # Save CSV files
    train_df.to_csv(output_path / 'train_text.csv', index=False)
    val_df.to_csv(output_path / 'val_text.csv', index=False)
    test_df.to_csv(output_path / 'test_text.csv', index=False)

    # Create report
    report = f"""# {dataset_name} Dataset Report

## Overview
Disjoint splits created from drug_rna_cds_sampled_11 dataset.

## Split Statistics

### Sample Counts
- Train: {len(train_df):,} samples ({100*train_df['label'].mean():.1f}% positive)
- Val: {len(val_df):,} samples ({100*val_df['label'].mean():.1f}% positive)
- Test: {len(test_df):,} samples ({100*test_df['label'].mean():.1f}% positive)
- Total: {len(train_df) + len(val_df) + len(test_df):,} samples

### Unique Entities

#### Compounds
- Train: {train_df['SMILES'].nunique():,}
- Val: {val_df['SMILES'].nunique():,}
- Test: {test_df['SMILES'].nunique():,}

#### RNAs
- Train: {train_df['RNA_seq'].nunique():,}
- Val: {val_df['RNA_seq'].nunique():,}
- Test: {test_df['RNA_seq'].nunique():,}

## Disjointness Properties

### Compound Disjointness
- Train-Val compound overlap: {len(set(train_df['SMILES'].unique()) & set(val_df['SMILES'].unique()))}
- Train-Test compound overlap: {len(set(train_df['SMILES'].unique()) & set(test_df['SMILES'].unique()))}
- Val-Test compound overlap: {len(set(val_df['SMILES'].unique()) & set(test_df['SMILES'].unique()))}

### RNA Disjointness
- Train-Val RNA overlap: {len(set(train_df['RNA_seq'].unique()) & set(val_df['RNA_seq'].unique()))}
- Train-Test RNA overlap: {len(set(train_df['RNA_seq'].unique()) & set(test_df['RNA_seq'].unique()))}
- Val-Test RNA overlap: {len(set(val_df['RNA_seq'].unique()) & set(test_df['RNA_seq'].unique()))}

## Output Format
Same as drug_rna_cds_sampled_11:
- Columns: RNA_ID, Compound_ID, RNA_seq, SMILES, text, label
- Files: train_text.csv, val_text.csv, test_text.csv

## Use Case
This split tests the model's ability to generalize to:
- {'New compounds AND new RNAs' if 'disjoint' in dataset_name and 'rna' not in dataset_name and 'compound' not in dataset_name else 'New RNAs' if 'rna' in dataset_name else 'New compounds' if 'compound' in dataset_name else 'New samples'}

Date: {pd.Timestamp.now().strftime('%Y-%m-%d %H:%M:%S')}
"""

    with open(output_path / 'DATASET_REPORT.md', 'w') as f:
        f.write(report)

    print(f"\n  Saved to: {output_path}")
    print(f"    - train_text.csv")
    print(f"    - val_text.csv")
    print(f"    - test_text.csv")
    print(f"    - DATASET_REPORT.md")


def main():
    parser = argparse.ArgumentParser(
        description="Create disjoint splits from drug_rna_cds_sampled_11"
    )
    parser.add_argument(
        '--input',
        default='datasets/drug_rna_cds_sampled_11',
        help='Input dataset directory'
    )
    parser.add_argument(
        '--seed',
        type=int,
        default=42,
        help='Random seed'
    )

    args = parser.parse_args()

    print("=" * 80)
    print("CREATING DISJOINT SPLITS FROM drug_rna_cds_sampled_11")
    print("=" * 80)

    # Load data
    input_path = Path(args.input)

    # Combine all splits from source
    dfs = []
    for split in ['train', 'val', 'test']:
        csv_path = input_path / f'{split}_text.csv'
        if csv_path.exists():
            dfs.append(pd.read_csv(csv_path))

    df = pd.concat(dfs, ignore_index=True)
    print(f"\nLoaded {len(df):,} samples from {input_path}")

    # Analyze feasibility
    fully_disjoint_feasible = analyze_disjoint_feasibility(df)

    # 1. Try fully disjoint
    if fully_disjoint_feasible:
        print("\n" + "#" * 80)
        print("# CREATING: drug_rna_cds_disjoint (FULLY DISJOINT)")
        print("#" * 80)

        result = create_fully_disjoint_split(df, seed=args.seed)
        if result is not None:
            train_df, val_df, test_df = result
            save_splits(
                train_df, val_df, test_df,
                'datasets/drug_rna_cds_disjoint',
                'drug_rna_cds_disjoint'
            )
        else:
            print("\n  ⚠️  Skipping fully disjoint split - not practical for this dataset")
    else:
        print("\n  ⚠️  Skipping fully disjoint split - not enough samples")

    # 2. Create RNA-disjoint
    print("\n" + "#" * 80)
    print("# CREATING: drug_rna_cds_disjoint_rna (RNA-DISJOINT)")
    print("#" * 80)

    train_df, val_df, test_df = create_rna_disjoint_split(df, seed=args.seed)
    save_splits(
        train_df, val_df, test_df,
        'datasets/drug_rna_cds_disjoint_rna',
        'drug_rna_cds_disjoint_rna'
    )

    # 3. Create compound-disjoint
    print("\n" + "#" * 80)
    print("# CREATING: drug_rna_cds_disjoint_compound (COMPOUND-DISJOINT)")
    print("#" * 80)

    train_df, val_df, test_df = create_compound_disjoint_split(df, seed=args.seed)
    save_splits(
        train_df, val_df, test_df,
        'datasets/drug_rna_cds_disjoint_compound',
        'drug_rna_cds_disjoint_compound'
    )

    print("\n" + "=" * 80)
    print("ALL DISJOINT SPLITS CREATED SUCCESSFULLY")
    print("=" * 80)
    print("\nCreated datasets:")
    if fully_disjoint_feasible:
        print("  1. drug_rna_cds_disjoint - Fully disjoint (compounds AND RNAs)")
    print("  2. drug_rna_cds_disjoint_rna - RNA-disjoint")
    print("  3. drug_rna_cds_disjoint_compound - Compound-disjoint")


if __name__ == "__main__":
    main()