Integrating evolutionary and compositional features with ML and DL for robust and interpretable druggable protein prediction.

Druggable proteins are essential targets in contemporary therapeutic discovery, and their precise identification is necessary for the progression of rational drug design. Traditional biochemical screening techniques are costly, laborious, and protracted, whereas current computational models typically attain only moderate efficacy and seldom offer robust statistical validation or biological interpretability. To mitigate these limitations, we offer a hybrid computational approach that integrates evolutionary and compositional information by merging Average Block-based Position-Specific Scoring Matrix (AB-PSSM) characteristics with Dipeptide Composition (DPC) into a 600-dimensional representation. The hybrid feature space was assessed utilizing three machine learning algorithms (support vector machine, random forest, and XGBoost) alongside three deep learning architectures (CapsBiLSTM, ResCapsNetPlus, and ResNet1D) following a rigorous five-fold out-of-fold cross-validation protocol to guarantee fairness and reproducibility. A thorough assessment revealed that the hybrid attributes significantly surpassed individual descriptors, with the SVM and CapsBiLSTM classifiers attaining classification accuracies of 90% and ROC-AUC and PR-AUC values surpassing 95%. By integrating complementary evolutionary and compositional descriptors within a rigorously controlled out-of-fold validation framework, the proposed method improves predictive stability compared with previously reported sequence-based and ensemble learning predictors. Furthermore, unlike many existing approaches that primarily rely on accuracy-based evaluation, this study incorporates statistical significance testing and interpretable feature attribution, enabling both reliable generalization performance and biological interpretability. The robustness and reliability were further validated by statistical analyses, including DeLong's test, McNemar's test, and bootstrap confidence intervals, demonstrating that the observed enhancements were substantial and consistent. Furthermore, interpretability investigations employing SHAP feature attribution and t-SNE visualization underscored biologically significant sequence descriptors as pivotal factors influencing druggability, thus mitigating the "black box" constraint inherent in numerous deep learning methodologies. This study presents a statistically validated, interpretable, and high-performing framework for predicting druggable proteins, demonstrating competitive performance compared with existing machine learning and ensemble-based predictors while providing improved statistical reliability and interpretability, thereby establishing a clear basis for future applications in precision medicine and drug discovery.