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

Rapid growth of protein sequences generated through next-generation sequencing technologies lacks structural annotations. Accurately predicting the structural class of these large sequences is a challenging, yet crucial task. Over the years, using machine learning (ML) models, protein structural-class prediction methods were developed. Despite their strengths, these methods often lack generalization ability, efficient feature engineering, and interpretability. In this study, we have developed SCOP-KHNET (SCOP K-mer Hybrid Network), which utilizes the power of multiple ML models simultaneously with an optimized weighted strategy together with the K-mer sliding window approach. SCOP-KHNET predicts the four structural classes of proteins from only the sequence information and achieves 99.7% accuracy in the independent dataset of SCOP 2.08, which is significantly better than all the existing methods. This method also outperforms the large language model ESM-2 to predict protein structural classes. Moreover, kernel Shapley additive explanations (KernelSHAP) are used to identify important features that offer crucial structural insights.