pubchem_experiment/make_predictions.py

import chunk
import os
import warnings

import pandas as pd
from rdkit import Chem
# from rdkit.Chem import AllChem
from rdkit.Chem import rdFingerprintGenerator
from sklearn.ensemble import RandomForestClassifier
from tqdm.auto import tqdm
import numpy as np
import clamp
import torch

warnings.filterwarnings("ignore")


def generate_morgan_fingerprints(smiles_list, radius=4, n_bits=4048):
    """
    Generate Morgan fingerprints for a list of SMILES.
    """
    mfpgen = rdFingerprintGenerator.GetMorganGenerator(radius=radius,fpSize=n_bits)
    mols = [Chem.MolFromSmiles(smi) for smi in smiles_list]
    fps = []
    for smiles, mol in zip(smiles_list, mols):
        if mol is None:
            print(smiles)
            fps.append(None)
        else:
            fps.append(mfpgen.GetFingerprintAsNumPy(mol))
    # np.array([mfpgen.GetFingerprintAsNumPy(mol) for mol in mols])
    return fps

def rf(df, train_smiles, test_smiles):
    """
    Train and test RF baseline model.

    Parameters:
        df : pd.DataFrame with 'SMILES' and 'Activity_label' columns
        train_smiles : list of training set smiles
        test_smiles : list of test set smiles
    Returns:
        preds : list of predicted labels for the test set
    """
    train_df = df[df['SMILES'].isin(train_smiles)]
    test_df = df[df['SMILES'].isin(test_smiles)]

    # Generate Morgan fingerprints for training and test sets
    X_train = generate_morgan_fingerprints(train_df['SMILES'])
    X_test = generate_morgan_fingerprints(test_df['SMILES'])

    # Extract labels
    y_train = train_df['Activity'].values 

    # Train a Random Forest Classifier
    clf = RandomForestClassifier(n_estimators=200, random_state=82)
    clf.fit(X_train, y_train)

    # Make predictions on the test set
    try:
        preds = clf.predict_proba(X_test)[:,1]
    except Exception as e:
        print(e)
        print(test_df)
        print(X_test)

    return preds

def fh(smiles_list):
    df = pd.read_csv('data/fh_predictions.csv')
    preds = df[df['SMILES'].isin(smiles_list)]['Prediction'].tolist()
    return preds

def drop_assays_with_limited_data(df, na_min=50, ni_min=100):
    print('Drop assays with not enough datapoints...')
    unique_uids = df['AID'].sort_values().unique() # Sorted unique targets
    activity_counts = df.groupby('AID')['Activity'].value_counts().unstack().fillna(0) # matrix: rows=sorted targets, columns=nactive, ninactives
    mask = ((activity_counts[1] >= na_min) & (activity_counts[0] >= ni_min) ) # Both nactives and ninactives above nmin
    df = df[df['AID'].isin(unique_uids[mask])]
    return df

def run(
        n_actives : int,
        n_inactives : int,
        model : str = 'MHNfs',
        task : str = 'UID',
        input_file : str = '', # todo add path
        output_dir : str = '', # todo add path
        n_repeats : int = 3,
        seed : int = 42
        ):

    # Load data
    data = pd.read_csv(input_file)

    if task == 'AID':
        data = drop_assays_with_limited_data(data, 30, 30)

    # Output dir
    output_dir = os.path.join(output_dir, model, task, f'{n_actives}+{n_inactives}x{n_repeats}')
    print(output_dir)
    os.makedirs(output_dir, exist_ok=True)

    # Tasks
    tasks = data[task].value_counts(ascending=True).index.tolist()
    # print(tasks)

    if model == 'MHNfs':           
        predictor = ActivityPredictor()

    # Iterate over tasks
    for t in tqdm(tasks):

        # Output file
        output_file = os.path.join(output_dir, f'{t}.csv')
        if os.path.exists(output_file):
            continue

        # Data for task
        df = data[data[task] == t]

        # Iterate over replicates
        results = []
        for i in range(n_repeats):
            # Select support sets and test molecules
            actives = df.loc[df['Activity'] == 1, 'SMILES'].sample(n=n_actives, random_state=seed+i).tolist()
            inactives = df.loc[df['Activity'] == 0, 'SMILES'].sample(n=n_inactives, random_state=seed+i).tolist()
            test_smiles = df[~df.SMILES.isin(actives+inactives)].SMILES.tolist()

            if model == 'RF':
                preds = rf(df, actives+inactives, test_smiles)
            else:    
                if len(test_smiles) > 10_000:
                    # MHNfs breaks for over 20_000 datapoints -> Use chunks to make predictions 
                    chunk_size = 10_000
                    chunks = [test_smiles[i:i + chunk_size] for i in range(0, len(test_smiles), chunk_size)]
                    preds = []
                    for chunk in chunks:
                        preds.extend( predictor.predict(chunk, actives, inactives))
                else:
                    preds = predictor.predict(test_smiles, actives, inactives)

            d = {
                'SMILES' : test_smiles, 
                'Label' : df[df.SMILES.isin(test_smiles)].Activity,
                'Prediction' : preds,
                'Fold' : [i] * len(test_smiles) 
            }
            results.append(pd.DataFrame(d))
        
        results = pd.concat(results)
        results.to_csv(output_file, index=False)

if __name__ == '__main__':

    mhnfs_path = # mhnfs_path + '/mhnfs'
    benchmark_path = # project_path

    import sys
    sys.path.append(mhnfs_path)
    from src.prediction_pipeline import ActivityPredictor

    support_sets = [(1,7), (2,6), (4,4)]
    models = ['RF', 'MHNfs']
    tasks =  ['AID', 'UID']
    
    input_file = # preprocessed_data path + '/pubchem24_preprocessed_2.csv.gz'

    for support_set in support_sets:
        for model in models:
            for task in tasks:
                run(*support_set, task=task, model=model, input_file=input_file)