Integrative in Silico modeling for mTOR inhibition: From ridge classifiers to descriptor-free deep neural networks

• mTOR inhibition is a key strategy for cancer treatment. • QSAR models successfully predict mTOR inhibitor activity using ML techniques. • Development of QSAR models utilizing both classical ML and DL methodologies. • Establishing robust baselines for future model development and comparisons. Inhibiting the mammalian Target of Rapamycin (mTOR) represents a promising strategy in cancer therapy due to its crucial role in cell growth, survival, and metabolism. Using a variety of quantitative structure-activity relationship (QSAR) models, we present a comprehensive comparison of deep learning (DL) and classical machine learning (ML) techniques for modeling mTOR inhibitor activity. Unlike prior studies that focused on specific algorithms or limited descriptors, we benchmark a wide range of models, from traditional models like Random Forest, logistic regression, and SVM to modern algorithms like CNNs, GRUs, and LSTMs, on both descriptor-based features obtained by Dragon and descriptor-free inputs, including raw SMILES string, and Morgan fingerprints. This comprehensive analysis provides a robust foundation for using the best QSAR models specific to mTOR inhibition. Our findings revealed that while the Random Forest classifier achieved the highest accuracy among all models (0.9290 accuracy, 0.8940 F1-score, 0.9737 AUC), DL methods also demonstrated strong predictive capabilities, with nearly all models attaining an accuracy above 0.90. Among the DL models, CNN-QSAR using Morgan fingerprints achieved the highest accuracy (0.9271), F1-score (0.8950), and AUC (0.9696), demonstrating its effectiveness in capturing structural characteristics. The GRU-QSAR and LSTM-QSAR models, which utilized tokenized SMILES, achieved accuracies of 0.9002 and 0.9021, F1-scores of 0.8595 and 0.8603, and AUCs of 0.9270 and 0.9529, respectively, leveraging their ability to process sequential data.