Improved HCV genotyping based on sequence encoding and feature integration

Abstract

Genomic sequence classification is an important task for understanding viral genetic variability and supporting genotype identification, particularly for viruses such as the Hepatitis C Virus (HCV). In this study, we evaluate the effectiveness of different sequence encoding techniques, supplementary genomic descriptors, and class-balancing strategies for HCV genotype classification under a stricter CD-HIT-EST cluster-aware evaluation protocol designed to reduce sequence-identity leakage. Four encoding methods were investigated, namely Frequency Chaos Game Representation (FCGR), k-mer encoding, one-hot encoding, and label encoding. These representations were evaluated using multiple machine learning models, including XGBoost, Random Forest, Multi-Layer Perceptron (MLP), and K-Nearest Neighbors (KNN). To address class imbalance, SMOTE was applied, while GC content and GC skew were included as supplementary compositional descriptors. Overall, FCGR and k-mer encoding outperformed one-hot and label encoding. The best result was achieved using k-mer encoding with XGBoost and SMOTE, obtaining a macro F1-score of 0.890, accuracy of 0.922, macro precision of 0.866, macro recall of 0.925, macro specificity of 0.990, and MCC of 0.893 under the CD-HIT-EST cluster-aware evaluation setting. The statistical analysis showed that SMOTE had a clear positive effect on classification performance, particularly for the Random Forest model, while GC-based descriptors provided additional interpretability but did not show a statistically significant improvement in the k-mer-based experiment. Overall, this study demonstrates the importance of selecting appropriate sequence encodings and class-balancing strategies for machine-learning-based HCV genotype classification. The findings provide a comparative basis for future work on sequence-based classification tasks in virology and computational genomics.