GGI-DDI: Identification for key molecular substructures by granule learning to interpret predicted drug-drug interactions

Deep learning-based approaches have achieved promising performance in predicting drug-drug interactions (DDIs). Nevertheless, a significant drawback of these approaches is their limited interpretability, hindering their practical applicability for clinicians. Based on current research findings, drug interactions frequently arise from specific substructures or functional groups present in drugs. To enhance the interpretability of deep learning models, we propose a novel end-to-end learning approach that employs granular computing to identify pivotal substructures instead of using conventional atom-based or predefined molecular fingerprint methods to predict DDIs. We refer to this model as GGI-DDI (Granule-Granule Interaction for Drug-Drug Interaction). In this method, drugs are granulated into a set of coarser granules that represent the key substructures or functional groups of drugs. Subsequently, the detection of DDIs occurs through the examination of interactions among these granules, aligning more closely with human cognitive patterns. We conducted thorough experiments on the TWOSIDES dataset, and the results show that GGI-DDI achieved impeccable accuracy in predicting DDIs. We compared GGI-DDI to state-of-the-art baseline models including DDIMDL, GoGNN, DNN, STNN-DDI and GMPNN-CS, GGI-DDI almost consistently outperforms the baselines across all metrics in terms of Accuracy (Acc), Area under the receiver operating characteristic (Auc), Area under precision recall curve (Aupr) and Precision (Pre) in both transductive and inductive scenarios. Finally, we provide case studies to illustrate how GGI-DDI can effectively reveal important substructure pairs across drugs about a specific DDI type, offering insights into the underlying mechanism of these interactions. We find that this interpretability can serve as valuable guidance in the advancement of novel drug development and poly-drug therapy strategies.

Automated summary

Deep learning-based models have limited interpretability in predicting drug-drug interactions (DDIs). We propose a novel approach that uses granular computing to identify key substructures and achieves high accuracy in predicting DDIs.