DeepInsulin-Net: A Deep Learning Model for Identifying Drug Interactions Leading to Specific Insulin-Related Adverse Events

Predicting clinical adverse effects resulting from drug-drug interactions is a critical research area for drug safety and patient health. Specifically, predicting adverse effects associated with insulin is crucial for clinical decision support systems and pharmacovigilance applications. This study proposes a deep learning-based model with high accuracy to predict adverse effects caused by drug interactions. In the literature, 17 different clinical side effects commonly associated with the hormone insulin have been identified. The properties of the drug molecules causing these interactions were calculated through MACCS, Morgan fingerprints and RDKit descriptors. These features are filtered by the variance thresholding method and optimized to improve classification performance. The model is built on a 1D CNN architecture that handles drug pairs as parallel inputs and a class weighting technique is used to eliminate class imbalance. Experimental results show that the model achieves 99.66% accuracy in training and 94.03% in validation, with training loss decreasing to 0.01 and validation loss stabilizing at 0.22. The ROC-AUC metric is above 0.99, indicating that the model can predict infrequent adverse events. The developed model provides a scalable, computationally efficient and highly reliable approach to predict the clinical consequences of drug interactions.