Uncertainty Quantification and Temperature Scaling Calibration for Protein-RNA Binding Site Prediction

Area of Science:

• Computational biology
• Bioinformatics
• Deep learning applications

Background:

• Deep learning models often function as black boxes, raising concerns about prediction reliability and uncertainty.
• Protein-RNA interactions are vital, but existing computational methods for predicting binding sites lack uncertainty quantification (UQ).

Purpose of the Study:

• To develop a robust method for predicting protein-RNA binding sites with enhanced reliability.
• To quantify and reduce prediction uncertainty using UQ techniques.

Main Methods:

• Proposed MGCA (multiscale graph convolutional networks, convolutional neural networks, and attention) for improved feature extraction.
• Implemented UQ using Expected Calibration Error (ECE) and temperature scaling (TS).
• Introduced a novel split-bins screening method based on ECE to assess uncertainty reduction.

Main Results:

• MGCA achieved competitive results in protein-RNA binding site prediction.
• The split-bins method effectively reduced false positives (FP).
• Temperature scaling significantly decreased model ECE, improving prediction precision when combined with the split-bins method.

Conclusions:

• Temperature scaling is effective in reducing uncertainty for protein-RNA binding site prediction.
• Minimizing model uncertainty demonstrably enhances prediction quality and reliability.