Enhanced deep Convolutional Neural Network for SARS-CoV-2 variants classification

Rapid and scalable classification of SARS-CoV-2 genomes from spike-gene sequences can support real-time genomic surveillance in contexts where whole-genome data or high-end computing resources are limited.

We curated approximately 35,800 quality-filtered spike sequences spanning multiple clades and lineages and trained a hybrid CNN-BiLSTM model with standard regularization and class-imbalance handling. Model performance was benchmarked against Nextclade assignments and compared with classical machine-learning baselines.

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. Saliency and feature attribution analyses highlighted recurrent spike substitutions consistent with known variant-defining mutations, as well as additional uncharacterized motifs with potential biological relevance.

These findings demonstrate that spike-only deep models can provide rapid and accurate clade or variant classification, while also yielding interpretable feature importance. Such models complement phylogenetic approaches in settings with constrained resources and enable efficient triage of samples for confirmatory whole-genome analysis, supporting more timely genomic surveillance.