Scripts/Buchwald_prediction.py

import matplotlib

matplotlib.use('TkAgg')
import pandas as pd
import numpy as np
from rdkit import Chem
from rdkit.Chem import AllChem
import h2o
from h2o.estimators import H2OGradientBoostingEstimator

# Initialize H2O
h2o.init()


def featurize_smiles(smiles):
    try:
        mol = Chem.MolFromSmiles(smiles)
        fingerprint = AllChem.GetMorganFingerprintAsBitVect(mol, radius=2, nBits=1024)
        features = np.array(fingerprint)
        return features
    except Exception as e:
        print(f"Error featurizing SMILES: {smiles} -> {e}")
        return np.zeros(1024)  # Return a zero vector if there's an issue


# Specify the input and output column indices
input_column_index1 = 5  # Replace with index for the first reactant SMILES
input_column_index2 = 6  # Replace with index for the second reactant SMILES
output_column_index = 3  # Replace with your output (product) column index

# Prepare a H2OFrame for the first train-test split
file_path = '/Users/colestephens/Desktop/CodeFolders/ML_Dataset_Curation_Project/Split_data/buchwald.xlsx'
splits = pd.ExcelFile(file_path)

# Let's use the first split as an example
train_key = 'train_split_0'
test_key = 'test_split_0'

# Read the train and test sets
train_df = pd.read_excel(splits, sheet_name=train_key)
test_df = pd.read_excel(splits, sheet_name=test_key)

# Determine the column names using indices
input_column_name1 = train_df.columns[input_column_index1]
input_column_name2 = train_df.columns[input_column_index2]
output_column_name = train_df.columns[output_column_index]

# Featurize the SMILES columns for both reactants and concatenate the features
train_features1 = np.array([featurize_smiles(smiles) for smiles in train_df[input_column_name1]])
train_features2 = np.array([featurize_smiles(smiles) for smiles in train_df[input_column_name2]])
train_features = np.hstack((train_features1, train_features2))

test_features1 = np.array([featurize_smiles(smiles) for smiles in test_df[input_column_name1]])
test_features2 = np.array([featurize_smiles(smiles) for smiles in test_df[input_column_name2]])
test_features = np.hstack((test_features1, test_features2))

# Print message after successful featurization
print("Successfully converted SMILES to Morgan fingerprints for train and test sets.")

# Convert to H2OFrames
train_X = pd.DataFrame(train_features)
test_X = pd.DataFrame(test_features)

# Combine features with the target column
train_h2o = h2o.H2OFrame(pd.concat([train_X, train_df[[output_column_name]]], axis=1))
test_h2o = h2o.H2OFrame(pd.concat([test_X, test_df[[output_column_name]]], axis=1))

# Set the target and features
y = output_column_name
X = train_h2o.columns[:-1]  # All columns except the last one

# Initialize and train a Gradient Boosting model
gbm_model = H2OGradientBoostingEstimator(
    ntrees=50,
    max_depth=6,
    learn_rate=0.1,
    seed=1
)

gbm_model.train(x=X, y=y, training_frame=train_h2o)


# Function to predict the product from two input SMILES
def predict_product_from_smiles(smiles_str1, smiles_str2):
    # Featurize both input SMILES strings
    features1 = featurize_smiles(smiles_str1)
    features2 = featurize_smiles(smiles_str2)

    # Concatenate the features
    features = np.hstack((features1, features2))

    # Convert to a DataFrame
    feature_df = pd.DataFrame([features])

    # Create H2OFrame from the feature DataFrame
    feature_h2o = h2o.H2OFrame(feature_df)

    # Predict using the trained model
    prediction = gbm_model.predict(feature_h2o)

    # Retrieve the first element from the prediction column
    predicted_value = prediction.as_data_frame()['predict'][0]

    # Print the prediction result
    print(f"Predicted output for input SMILES '{smiles_str1}' + '{smiles_str2}': {predicted_value}")


# Example usage
example_smiles1 = 'BrC1=CC=CC=C1'  # Replace with a first reactant SMILES string
example_smiles2 = 'NC(C1=CC=CC=C1)=O'  # Replace with a second reactant SMILES string
predict_product_from_smiles(example_smiles1, example_smiles2)

# Shutdown H2O
h2o.shutdown(prompt=False)