A deep learning approach based on molecular graph features and residual blocks to predict interaction sites between...

CircRNAs are ubiquitously expressed across diverse tissues and cells, playing a pivotal role not only in protein-mediated biological processes but also in disease prevention and therapeutics. RNA-RBP interactions are critical for deciphering gene regulation, post-transcriptional RNA modifications, and protein synthesis. Although the computational methods involving sequence and structural data are widely adopted to predict RNA-RBP binding sites, existing approaches predominantly focus on linear RNAs, leaving circRNAs underexplored. Herein, MGFCRSites, a novel deep learning framework, is proposed to predict RBP binding sites on circRNAs by integrating molecular graph features with residual block structures. This method starts by encoding CircRNA molecular structures into ASCII-character-based representations. Then, molecular graphs are constructed in line with graph convolutional network principles. These graphs are processed through residual blocks to hierarchically extract discriminative features, prior to the binding site identification by a prediction module. Extensive experiments demonstrate that MGFCRSites performs well (AUC: 0.9663) across 37 benchmark datasets, outperforming existing methods by effectively capturing chemical structural patterns. To our knowledge, this is the first computational study that explicitly models CircRNA chemical molecular structures for binding site prediction. MGFCRSites provides a robust solution to the analysis of RBP-CircRNA interactions for exploring gene regulatory mechanisms and advancing disease research. The code of MGFCRSites is freely available at https://github.com/liuniannian717/MGFCRSites.git.