Prediction of circular RNA-RNA binding protein binding sites based on structural feature and dynamic feature...

Circular RNA (circRNA)-RNA binding protein (RBP) interactions play critical roles in various diseases, and predicting their binding sites can elucidate regulatory mechanisms and identify potential therapeutic targets. Current deep learning methods for this task rarely integrate predictions across both whole-sequence and single-nucleotide resolutions, and they also fail to adequately address dynamic feature selection. To overcome these limitations, we present circGMST. A breast cancer-specific dataset comprising variable-length circRNA sequences of seven RBPs is constructed. For sequence encoding, we combine multi-scale sliding-window GC content with nucleotide-level structural features. We then design a Gated Multi-scale Fusion (GMF) block, which integrates Gated Linear Units (GLU) and multi-scale dilated convolution. Three GMF blocks are stacked to form a homogeneous encoder-deep processor-decoder framework for hierarchical feature learning, followed by fully connected layers for final prediction. Comparative and ablation experiments, along with visualization analyses, demonstrate the superior nucleotide-level predictive performance of circGMST. By employing a multi-scale window strategy and a soft-label assignment scheme, circGMST is successfully extended to fragment-level and sequence-level binding affinity prediction, confirming its architectural advantages. Furthermore, motif analysis and a case study show that circGMST can extract biologically relevant motifs and generate high-confidence interaction candidates, providing valuable leads for experimental validation. The datasets and source code of circGMST are available at https://github.com/gyj9811/circGMST.