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

Area of Science:

• Computational biology and bioinformatics
• Drug discovery and precision medicine

Background:

• Identifying druggable proteins is crucial for rational drug design.
• Traditional methods are costly and time-consuming.
• Existing computational models often lack accuracy, statistical validation, and interpretability.

Purpose of the Study:

• To develop a hybrid computational approach for predicting druggable proteins.
• To integrate evolutionary and compositional protein features for improved prediction.
• To enhance the statistical validation and biological interpretability of predictions.

Main Methods:

• Developed a 600-dimensional feature space by merging Average Block-based Position-Specific Scoring Matrix (AB-PSSM) and Dipeptide Composition (DPC).
• Evaluated the hybrid features using machine learning (SVM, Random Forest, XGBoost) and deep learning (CapsBiLSTM, ResCapsNetPlus, ResNet1D) models.
• Employed rigorous five-fold out-of-fold cross-validation, statistical significance testing (DeLong's, McNemar's tests), and feature attribution (SHAP, t-SNE) for validation and interpretability.

Main Results:

• The hybrid feature space significantly outperformed individual descriptors.
• SVM and CapsBiLSTM models achieved over 90% accuracy and >95% ROC-AUC and PR-AUC.
• The method demonstrated improved predictive stability and provided interpretable insights into key sequence descriptors influencing druggability.

Conclusions:

• The proposed hybrid approach offers a statistically validated, interpretable, and high-performing framework for predicting druggable proteins.
• This method surpasses existing sequence-based and ensemble predictors in reliability and interpretability.
• It provides a robust foundation for applications in precision medicine and accelerating drug discovery.