DGSite: A Well-Calibrated Deep Learning Framework for Protein-RNA Binding Site Prediction

Area of Science:

• Computational Biology
• Molecular Biology
• Bioinformatics

Background:

• Protein-RNA interactions are crucial for biological processes.
• Existing computational models for protein-RNA binding site prediction face challenges like feature integration, class imbalance, and prediction calibration.

Purpose of the Study:

• To develop a novel deep learning framework, DGSite, for accurate protein-RNA binding site prediction.
• To address limitations in existing models, including insufficient feature integration and poor prediction calibration.

Main Methods:

• Constructed a comprehensive dataset (Train-1086 and Test-107).
• Developed DGSite using a simplified Deformable Attention Transformer for long-range dependencies and Graph Attention Networks for local features.
• Introduced Adaptive Bias Loss (ABL) and a hybrid ABL+Focal Loss (FL) to mitigate class imbalance and model uncertainty.

Main Results:

• DGSite achieved high performance on multiple datasets.
• The ABL+FL loss function balanced precision and recall while significantly reducing model uncertainty.
• Predictions were well-calibrated and more trustworthy, with minimized false positives.

Conclusions:

• DGSite is a robust, high-performance, and well-calibrated computational tool for protein-RNA binding site prediction.
• The novel loss function enhances prediction accuracy and reliability.
• The model provides interpretable insights for biological applications.