Beyond Primary Sequences: Assessing Secondary Structure Contributions in Protein Classification

Protein classification is a fundamental task in computational biology, aiding in function prediction and disease analysis. This study uniquely investigates the complementary role of secondary structure information alongside primary sequences in protein categorization through machine learning and deep learning techniques. A comprehensive dataset of 203,591 protein samples from the top 20 protein classes was processed using TF-IDF and embedding-based feature extraction. Seven classifiers, including Random Forest, XGBoost, SVM, Neural Networks, LSTM, Bi-LSTM, and GRU, were evaluated using hold-out cross-validation. Results indicate that primary sequence-based classification achieved the highest accuracy (92%) using Random Forest with TF-IDF, while secondary structure-based classification was comparatively less effective (86%). Combining primary and secondary sequences yielded marginal improvements, suggesting that while secondary structure contributes to feature distribution and biological interpretability, it does not significantly enhance classification accuracy. Notably, oxidoreductase, ribosome, and virus proteins exhibited consistently high classification performance across all models, with 99% accuracy in both primary and combined sequence classification. SHAP analysis highlighted that secondary structure features (‘Secondary_hh’ and ‘Secondary_hhh’) contributed to virus protein classification, emphasizing their potential for structural interpretation.