APPLICATION OF MACHINE LEARNING IN DRUG INTERACTION PREDICTION: SYSTEMATIC REVIEW

Abstract

Drug-drug interactions (DDIs) represent a significant challenge in modern pharmacotherapy, contributing to 17-23% of adverse drug reaction-related hospitalizations. Machine learning (ML) has emerged as a promising approach for computational DDI prediction, yet a comprehensive synthesis of methodologies, performance benchmarks, and clinical translation challenges remains lacking. This systematic review aims to identify and evaluate ML algorithms applied to DDI prediction, compare their performance across different datasets and validation strategies, analyze feature representation methods, and identify critical gaps impeding clinical deployment. Following PRISMA 2020 guidelines, we conducted a systematic search across five electronic databases (PubMed, IEEE Xplore, Scopus, Web of Science, Google Scholar) for studies published between January 2019 and March 2025. Dual independent screening and extraction were performed with quality assessment using adapted PROBAST criteria. Included studies were analyzed for algorithm types, feature representations, datasets, validation strategies, and performance metrics. From 1,285 initial records, 60 high-quality studies were included. Graph neural networks (GNNs) emerged as state-of-the-art methods (mean F1-score: 0.931 ± 0.024), significantly outperforming traditional ML (0.842 ± 0.038, p < 0.001) and deep neural networks (0.893 ± 0.031, p = 0.003). Multi-modal approaches integrating chemical structure, biological targets, and phenotypic data achieved highest performance (F1: 0.945-0.982). DrugBank was the most utilized dataset (63.3% of studies), though severe class imbalance (positive:negative ratio 1:20 to 1:50) posed significant challenges. Critical gaps identified include: cold-start problem (18.3% performance degradation for unseen drugs), interpretability issues (45% black-box models), and minimal real-world validation (only 6.7% used EHR data). A severe reproducibility crisis was evident, with only 11.7% of studies fully reproducible. While ML-based DDI prediction has achieved impressive benchmark performance, substantial challenges remain for clinical translation. Priority research directions include: developing explainable AI methods for biological validation, addressing cold-start generalization through meta-learning and transfer learning, conducting multi-center real-world validation studies, establishing standardized evaluation protocols, and implementing federated learning infrastructure for privacy-preserving collaboration. Community-wide efforts toward reproducibility, standardization, and responsible deployment are essential for translating computational advances into clinically impactful systems that enhance medication safety.