Development of machine learning models for the precise identification of potential molecules active against pancreatic Cancer

Pancreatic cancer is considered one of the deadliest cancers worldwide due to its insidious onset and late-stage diagnosis, requiring novel therapeutic candidates effective at the advanced stages of the disease. Rapid identification of potential lead molecules is necessary to accelerate the process, which can be achieved using advanced computational techniques. In this context, we report herein the development of robust machine learning models to predict the anti-pancreatic cancer activity of compounds using a comprehensive dataset of 21,056 molecules with unique SMILES and ChEMBL IDs. We systematically trained and evaluated a range of classifiers, including Logistic Regression, eXtreme Gradient Boosting, Random Forest, K-Nearest Neighbors, Support Vector Classifier, Artificial Neural Network, Custom K-Nearest Neighbors, and their ensemble combinations. These models were tested across four molecular fingerprints and 217 descriptors. The optimized ensemble models demonstrated superior performance compared to individual models. To improve the reliable identification of inactive compounds, we developed a custom Tversky-similarity-based model that better excludes inactive molecules. All models' performance was measured using Matthews Correlation Coefficient (MCC), accuracy, and recall, showing stable predictions across cross-validation folds. The five-voting classifier achieved the highest accuracy of 88% and an MCC of 0.66 for fingerprint-based models, with Morgan fingerprints proving most effective for this task. Similarly, the descriptor-based model with four-voting classifiers achieved an accuracy of 0.88 and an MCC of 0.75. To enhance chemical interpretability, Shapley Additive Explanations (SHAP) identified the twenty most influential molecular features driving predictions, which were validated using seven clinically approved drugs. Overall, the present study provides a computationally efficient, robust framework for the rational design and identification of potential molecules to accelerate therapeutic development against pancreatic cancer.