A Molecular Representation Learning Model Based on Multidimensional Joint and Cross-Learning for Drug–Drug Interaction Prediction

Drug-drug interactions (DDIs) present significant challenges within clinical pharmacology, as they can impact therapeutic outcomes, especially given the growing prevalence of polypharmacy. Traditional methods for the clinical validation of DDIs typically exhibit inefficiency and high cost, underscoring the necessity for more advanced computational methodologies. Although deep learning-based methods have improved DDI prediction performance, current approaches often face challenges in extracting and integrating multidimensional molecular features and capturing molecular reaction patterns. To overcome these limitations, we propose a Multidimensional Joint and Cross-learning (MDJCL) model that effectively integrates 1D, 2D, and 3D molecular features of drugs. A cross-attention fusion module aggregates multidimensional features while minimizing information loss, and a molecular-pair reaction module pinpoints potential interaction sites. Experimental results on benchmark data sets demonstrate that MDJCL consistently outperforms state-of-the-art models. Ablation studies reveal that each module contributes distinctively to the overall enhancement of evaluation metrics. These results validate the effectiveness of multidimensional feature integration and cross learning mechanisms in enhancing DDI prediction, offering a reliable tool for clinical decision-making and precision medicine.