Enhancing Nucleotide Pattern Recognition: A Hybrid Encoding Approach with Bi-LSTM and GRU

Gene sequence classification is based on the order of nucleotides present inside a DNA or RNA molecule. The nucleotide pattern recognition is challenging due to the complex structure and repetitive nature of gene sequences. Recent research on virus gene classification lacks a robust gene sequence classifier with parameter efficiency that can be evaluated on multiple datasets to produce efficient results. This work aims to utilize the computation power of RNN architectures Bi-LSTM and GRU to capture the sequential dependencies among nucleotides in viruses having variable length and complex pre-determined structure using a hybrid encoding approach by merging k-mer and label encoding techniques for data pre-processing. The results obtained for the RNN architectures were observed on 3 datasets with variable class sizes of 4, 6, and 14. The Bi-LSTM achieved accuracies of 96.11%, 98.82%, and 97.55% for datasets 1, 2, and 3, respectively. Similarly, the GRU architecture showed impressive results, with accuracies of 97.55%, 98.72%, and 97.54% across the same three datasets. The research outcome strongly suggests that, even though both GRU and LSTM produce nearly identical performance metrics values, the fine-tuned GRU network emerges as a better choice for classifying viruses’ gene sequences due to 1,685,774 trainable parameters, which are significantly less when compared to Bi-LSTM which had 2,012,046 trainable parameters. This work contributes to the nucleotide pattern recognition using RNN architectures, primarily GRUs, that offer improved computational efficiency without compromised accuracy across datasets, therefore achieving a balance between accuracy and computational efficiency across several datasets.