Enhancing predictive modeling with XGBoost-engineered probabilities and deep neural networks: A hybrid approach for...

Novel antineoplastic therapies focus on the inhibition of various protein kinases, creating an important need for new molecules capable of regulating their abnormal activity. Here, we introduce a specific hybrid Machine Learning (ML) architecture that combines an eXtreme Gradient Boosting (XGBoost) as a base model, and a deep neural network (DNN) to improve prediction accuracy in kinase inhibition QSARs. The approach utilizes XGBoost to process structured data features, while employing the DNN to refine and calibrate the probability estimates. The predictive probabilities from XGBoost are incorporated as new input features for a DNN to enhance classification performance. For the prediction of kinase inhibition, 40 reliable QSAR models were trained on large datasets (559-5675 compounds and ∼ 400-500 descriptors). As a result, the approach achieved a marked improvement in accuracy and robustness compared to the standalone XGBoost algorithm and various descriptors-based state-of-the-art methods, including Random Forest (RF) and Support Vector Machine (SVM). The accuracy enhancement reaches up to 14% for JAK2, BRAF and TRK-α, 13% for VEGFR2 and PIK3Cδ, and ranges between 5 and 12% for 30 other kinase datasets. Performance is thus analyzed through the discussion of several evaluation metrics. The proposed method will be useful for building accurate prediction models, either specific to the neoplastic drug design or QSAR modeling in general. Models constructed in this study are available at GitHub.