Enhanced Deep Convolutional Neural Network for SARS-CoV-2 Variants Classification

Olaitan I. Awe^1,2*Hesborn Obura^3,4Charles Ssemuyiga^5*Evans Mudibo^3,4Mike J. Mwanga^3,4

• ¹African Society for Bioinformatics and Computational Biology, Cape Town, South Africa
• ²Department of Computer Science, Faculty of Science, University of Ibadan, Ibadan, Oyo, Nigeria
• ³Department of Biochemistry and Biotechnology, School of Pure and Applied Sciences, Pwani University, Kilifi, Kilifi, Kenya
• ⁴Pwani University Bioscience Research Centre (PUBReC), Kilifi, Kenya
• ⁵Kampala International University, Kampala, Uganda

High-throughput sequencing enables the taxonomic classification of pathogens in clinical samples through alignment-based comparisons with reference databases. While effective, this approach is computationally intensive, particularly as genomic databases continue to grow. Machine learning offers a scalable alternative for rapid and accurate classification of viral genomes. In this study, we developed a hybrid deep learning model combining Convolutional Neural Networks and Bidirectional Long Short-Term Memory networks (CNN-BiLSTM) to classify five dominant SARS-CoV-2 variants of concern/VOCs (Omicron, Delta, Beta, Gamma, and Alpha) based on full-length spike gene sequences. The model was trained on 27,236 high-quality SARS-CoV-2 sequences retrieved from the GISAID database and evaluated on an independent validation set comprising 8,585 sequences. Across 10 experimental runs, the model achieved a mean training accuracy of 99.74% ± 0.11, a validation accuracy of 99.00% ± 0.00, and a test accuracy of 99.91% ± 0.03. In benchmarking against the molecular epidemiology tool Nextclade, our model demonstrated superior performance, correctly identifying 100% of Omicron sequences, compared to 34.95% achieved by Nextclade. Feature attribution analysis using saliency maps revealed biologically meaningful nucleotide regions corresponding to known variantdefining mutations and other nucleotide motifs, further supporting the interpretability of the model. This work provides a deep learning-driven alternative to classical alignment-based approaches for SARS-CoV-2 variant classification, with potential applications in real-time genomic surveillance, public health decision-making, and pandemic preparedness.