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

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• RNA-binding proteins (RBPs) are crucial regulators of gene expression.
• Accurate identification of RBP binding sites is vital for understanding gene regulation and disease.
• Experimental methods for RBP binding site profiling are often costly and prone to noise.

Purpose of the Study:

• To develop a robust and accurate computational method for predicting RNA-protein binding sites.
• To improve the prediction of RBP binding sites, particularly on challenging datasets.
• To provide insights into the mechanisms of RBP binding.

Main Methods:

• MARSNet, a convolutional attention residual shrinkage network, was developed for RNA-protein binding site prediction.
• The network utilizes sequential multi-scale window encoding to capture local motifs and contextual dependencies.
• A residual shrinkage mechanism was employed to enhance robustness by suppressing noise-induced activations.

Main Results:

• MARSNet achieved high performance on the RBP-24 benchmark, with AUROC of 0.957, AP of 0.875, and MCC of 0.821.
• The method demonstrated superior performance on small-scale and heterogeneous datasets, indicating its stability.
• Interpretability analyses confirmed MARSNet's ability to identify biologically relevant binding determinants and sensitivity to mutations.

Conclusions:

• MARSNet is a highly effective computational tool for predicting RNA-protein binding sites.
• The network's robustness and accuracy offer significant advantages over existing methods, especially in data-scarce scenarios.
• MARSNet provides a valuable resource for research in gene regulation, disease mechanisms, and drug discovery.