MARSNet: A convolutional attention residual shrinkage network for RNA-protein binding site prediction.

RNA-binding proteins (RBPs) regulate post-transcriptional gene expression by recognizing specific RNA elements, making accurate identification of RBP binding sites essential for understanding gene regulation and disease mechanisms. However, experimental profiling is costly and often noisy, motivating robust sequence-based computational prediction. We propose MARSNet, a convolutional attention residual shrinkage network for RNA-protein binding site prediction. MARSNet adopts sequential multi-scale window encoding to capture both local motifs and broader contextual dependencies. For each window scale, a feature extractor integrates residual shrinkage blocks with lightweight attention modules. The residual shrinkage mechanism performs learnable soft-thresholding to suppress low-magnitude activations induced by background noise, significantly improving robustness, especially on data-scarce or heterogeneous datasets. Predictions from different window scales are then aggregated by a weighted fusion strategy to form the final output. Extensive experiments on the RBP-24 benchmark demonstrate that MARSNet achieves an average AUROC of 0.957, AP of 0.875, and MCC of 0.821, outperforming classic baselines and remaining competitive with recent methods such as PIONet and RMDNet. Notably, MARSNet exhibits superior performance on small-scale datasets, highlighting its stability. Furthermore, interpretability analyses, including motif discovery and an in silico mutagenesis case study on the TARDBP gene, confirm that MARSNet captures biologically meaningful binding determinants and is sensitive to pathogenic mutations. The source code is available at https://github.com/HNUBioinformatics/MARSNet.