A transformer-based graph local-global fusion model for interpretable circRNA-miRNA interaction prediction

Background Aberrant circRNA–miRNA interactions (CMIs) play a pivotal role in the pathogenesis of major diseases. Although several computational models have been proposed for CMI prediction with considerable results, their inter-pretability in elucidating the underlying mechanisms of CMI remains limited. Therefore, developing a highly interpretable CMI prediction model is of great significance for uncovering the molecular mechanisms and biological implications. Results We propose TransFusion, a transformer-based graph fusion model for predicting CMI. TransFusion focuses on molecular biological characteristics, extracting intrinsic semantic features using advanced natural language processing techniques, which are subsequently integrated into the network. By combining graph convolutional networks with graph transformers, the model captures both local and global structural information, facilitating the fusion of multi-source features. We visualized the process of molecule representation to explore the inter-pretability of TransFusion in capturing and distinguishing molecular features. Disease-related case studies validate the practicality of TransFusion from both computational and biological perspectives. Conclusions TransFusion is capable of revealing molecular interaction mechanisms , providing valuable insights for the design of targeted cancer therapies. To facilitate communication with relevant researchers, we developed an online prediction server (http://39.106.153.201/TransFusion/home.html) and open-sourced the code on GitHub (https://github.com/591286260/TransFusion).