Exploring NTK Kernel And NLP based approach for Robust DNA Sequence Classification

This paper classifies viral sequences from five different viruses: Zika, Influenza, HIV, Ebola, and SARS-CoV-2. First, the preprocessing on viral sequences, a clean and high-quality dataset for precise analysis, gives a starting point for analysis. Following the preprocessing part, k-mers are extracted as critical features obtained from the sequences. Then, three different feature extraction methods are employed: TF-IDF vectorization, Bag of Words (BoW), and Word2Vec embeddings. Thereafter, each set of features is fed into a variety of algorithms developed under machine learning which includes SVM, Random Forest, Naive Bayes, and a customized SVM using the NTK. The primary purpose behind the present research work is to explore the capability of performance in SVM models in combination with NTK with respect to traditional classifiers within viral sequence classification. From the experiments, it can be seen that SVM with a linear kernel as well as with Random Forest are very accurate in all the methods of feature extraction, and, on average, SVM is giving 97%. Naive Bayes is relatively poorer and is taken here as a baseline. In this regard, the SVM-NTK classifier is notably good, and it had an average accuracy of 98% at all feature extraction techniques. The results clearly suggest that the quality of the inputs is highly sensitive to the accuracy of the model. The cross-validation scores for all the models were excellent, thereby proving stability and reproducibility in viral sequence classification. Results were successful in establishing the potential to use the combination of independently developed feature extraction techniques along with powerful classifiers for making an accurate viral sequence identification. Future work will include greater sophistication in neural network architecture and more advanced methods of sequence alignment to further improve the accuracy of classification.