SCOP-KHNET: An Interpretable Hybrid Network for Protein Structural Class Prediction.

Area of Science:

• Computational Biology and Bioinformatics
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Next-generation sequencing generates vast protein sequences lacking structural annotations.
• Accurate prediction of protein structural classes from sequence is essential but challenging.
• Existing machine learning (ML) methods often suffer from poor generalization, feature engineering limitations, and lack of interpretability.

Purpose of the Study:

• To develop an accurate and interpretable method for predicting protein structural classes solely from sequence information.
• To overcome the limitations of existing ML-based protein structure prediction techniques.

Main Methods:

• Developed SCOP-KHNET (SCOP K-mer Hybrid Network), integrating multiple ML models with an optimized weighted strategy.
• Employed the K-mer sliding window approach for sequence feature extraction.
• Utilized kernel Shapley additive explanations (KernelSHAP) for feature interpretability.

Main Results:

• SCOP-KHNET achieved 99.7% accuracy in predicting the four structural classes of proteins on an independent SCOP 2.08 dataset.
• Demonstrated superior performance compared to all existing protein structural-class prediction methods.
• Outperformed the large language model ESM-2 in predicting protein structural classes.

Conclusions:

• SCOP-KHNET offers a highly accurate and robust method for protein structural class prediction from sequence data.
• The method provides crucial structural insights through interpretable feature identification via KernelSHAP.
• This advancement addresses the critical need for efficient structural annotation of rapidly growing protein sequence data.