Capsule enclosed coordinate attention based dual batch depthwise convolutional knowledge distillation model for...

Drug-drug interactions (DDIs) are a significant issue in drug discovery, impacting research efficiency and patient safety. Precise prediction of DDIs is important, particularly when drugs are co-administered. The combination of heterogeneous data sources that reflect drug relationships and properties can greatly enhance predictive accuracy. This paper proposes a new Capsule-enclosed Coordinate Attention-based Dual Batch Depthwise Convolutional Knowledge Distillation (CC-DBDKD) model for DDI prediction. The input data drawn from the DrugBank dataset is preprocessed with the RDKit to standardize SMILES strings into their canonical representations. Various techniques of molecular fingerprint generation, such as Extended Connectivity Fingerprints, MACCS keys, PubChem Fingerprints, 3D molecular fingerprints, and molecular dynamics fingerprints, are used to map drug chemical structures onto feature vectors. Drug similarities are subsequently calculated by the Tanimoto coefficient, and the Structural Similarity Profile (SSP) is calculated as an average of these fingerprint types. A lightweight model, CC-DBDKD, improves DDI prediction by introducing capsule networks to learn spatial hierarchies and complex drug relationships. Coordinate attention mechanisms improve feature extraction by attending to key interaction patterns. Adding dual-batch depthwise convolutional layers improves computational efficiency to support scalability with large datasets. In addition, knowledge distillation reinforces the model by mapping knowledge from a teacher model to a student model, enhancing accuracy and robustness. The proposed model realizes superior accuracy values of 0.987 and 0.989 and an F1-score of 0.986, which outshines other prevailing models like CNN, CNN-LSTM, Autoencoder, and D-CNN. The outcomes position the CC-DBDKD model as a strong and scalable instrument for accurate DDI prediction.