An Integrative Hybrid Machine Learning Framework with Metaheuristic Optimization for Drug-Drug Interaction Prediction Using Molecular Embeddings and Clinical Knowledge

Predicting drug-drug interactions (DDIs) is essential for ensuring patient safety, particularly in chronic conditions such as Type ۲ Diabetes Mellitus (T۲DM), where polypharmacy increases interaction risk. While many machine learning approaches have been proposed, several rely on complex architectures or insufficient integration of pharmacological knowledge, limiting interpretability and clinical applicability. We propose a leakage-free and interpretable framework that balances predictive performance with clinical relevance. The approach integrates two complementary SMILES-based molecular embeddings: Mol۲Vec, capturing fragment-level chemical patterns, and SMILES-BERT, encoding contextual structural information. To incorporate domain knowledge independently of interaction labels, we introduce a rule-based clinical score (RBScore) that encodes shared enzymes, targets, therapeutic classes, and side-effect similarity. The fused features are classified using a compact multilayer perceptron optimized through a three-stage RSmpl-ACO-PSO metaheuristic strategy, enabling efficient exploration of both discrete and continuous hyperparameter spaces. Evaluation on DrugBank and a T۲DM-specific subset under random, drug-level cold-start, and scaffold-based splits demonstrates consistent performance (DrugBank: ROC-AUC = ۰.۹۱۱, PR-AUC = ۰.۸۶۷; T۲DM: ROC-AUC = ۰.۹۰۲, PR-AUC = ۰.۸۵۹). Bootstrap confidence intervals and SHAP-based analysis further confirm robustness and interpretability.

کلیدواژه ها:

Drug-Drug Interaction Prediction ، Molecular Embedding ، Metaheuristic Optimization