AttentionRBP: Integrating Multimodal Features via Sparse Cross- Attention for RNA-Binding Protein Prediction

Abstract

RNA-Binding Proteins (RBPs) play critical roles in post-transcriptional gene regulation by modulating RNA splicing, transport, stability, and translation. Dysfunction of RBPs is associated with various diseases, including cancer, neurodegenerative disorders, and viral infections. However, existing computational methods for RBP prediction often rely on single-modality features or direct feature concatenation, limiting their ability to capture the complex interactions between RNA and proteins. We propose AttentionRBP, a multimodal framework that integrates protein-language-model embeddings, structural information, and physicochemical properties for RBP prediction. To improve cross-modal feature fusion, a Dropout-based sparse cross-attention mechanism was designed to suppress noisy interactions and enhance representation robustness. Experimental results demonstrated that AttentionRBP outperformed several baseline methods in terms of Accuracy, MCC, AUC, and F1-score. Ablation studies confirmed the effectiveness of each modality and the contribution of the sparse cross-attention strategy. In addition, SHAP-based interpretability analysis indicated that cross-modal features contributed substantially to model predictions, while structural information complemented sequence-based embeddings. The proposed multimodal framework effectively integrates heterogeneous biological information, thereby improving the robustness and interpretability of RBP prediction. The sparse crossattention mechanism also reduces the influence of irrelevant cross-modal interactions during feature fusion. AttentionRBP provides an effective framework for multimodal RBP prediction and shows promising generalization performance on unseen data. A publicly accessible web server was also developed to support batch submission and multi-species prediction for efficient RBP identification.