Sachidananda Behera

India

An SNF-Integrated Hybrid Convolutional Neural Network for Drug–Drug Interaction Prediction in Cancer Patients

Sachidananda Behera1,2; Deepa Kumari3; Subhrakanta Panda3; Parool Gupta2,4; Jitendra Kumar Singh2
1. Department of Pharmacy, AISECT University, Hazaribagh, Jharkhand, India.
2. Department of Research and Clinical Trials, S. S. Hospital and Research Institute, Patna, India
3. CSIS Department, BITS-Pilani, Hyderabad Campus, Telangana, India
4. Buddha Institute of Pharmacy, GIDA, Gorakhpur, Uttar Pradesh, India

Abstract

Background

Drug–drug interactions (DDIs) represent a major concern in clinical practice, particularly in the context of polypharmacy, as they can lead to adverse drug reactions, compromised therapeutic efficacy, and increased healthcare burden. Accurate and early prediction of DDIs is therefore essential for ensuring medication safety. However, existing computational approaches often struggle to effectively integrate heterogeneous drug-related data and capture complex interaction patterns. In this study, we have developed a novel predictive framework termed Similarity Network Fusion–Hybrid Convolutional Neural Network (SNF–HCNN) for enhanced DDI predictions in Cancer patients under Chemotherapy.

Methods

We have developed a novel predictive framework, termed Similarity Network Fusion–Hybrid Convolutional Neural Network (SNF–HCNN), for robust DDI prediction. Drug-related data were integrated from DrugBank, PubChem, and SIDER, encompassing seven key drug features: targets, transporters, enzymes, chemical substructures, carriers, off-targets, and side effects. Feature-wise drug similarities were computed using the Jaccard similarity measure to construct individual similarity matrices. A similarity selection strategy was applied and then selected similarity matrices were then fused using Similarity Network Fusion (SNF) to generate a unified drug similarity network. The fused representation was fed into multiple hybrid convolutional neural network (HCNN) architectures, combining CNNs with Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM) classifiers. We cross-checked some datasets in Cancer patients (n=25).

Results

Experimental evaluation demonstrates that the proposed SNF–HCNN framework significantly outperforms conventional DDI prediction approaches. Among the tested models, CNN+LR achieved the highest accuracy (95.19%), followed by CNN+RF (94.45%) and CNN+SVM (93.65%). Notably, CNN+LR also exhibited superior performance across critical evaluation metrics, including precision, sensitivity, F1-score, and area under the ROC curve (AUC), indicating enhanced predictive reliability and robustness. We have also observed some interesting results from Cancer patients’ datasets.

Conclusions

The proposed SNF–HCNN framework effectively integrates heterogeneous drug features and leverages deep learning to achieve high-accuracy DDI prediction. By capturing complex drug similarity patterns and reducing feature redundancy, the model offers a scalable and reliable solution for DDI identification. This approach holds substantial promise for improving medication safety and supporting clinical decision-making for Cancer treatment in real-world healthcare settings.