Explainable convolutional neural network model provides an alternative genome-wide association perspective on mutations in SARS-CoV-2.

Identifying informative genomic features in SARS-CoV-2 can help clarify patterns of viral evolution. In this study, we developed an explainable convolutional neural network (CNN) model to classify SARS-CoV-2 genomic sequences into the WHO-designated Variants of Concern (VOCs), Alpha, Beta, Gamma, Delta, and Omicron. Using a balanced dataset of genomes, the classification CNN achieved 99.96% accuracy on the held-out test set. To interpret the model's predictions, we applied SHapley Additive exPlanations (SHAP) to estimate the contribution of each nucleotide position to VOC-label prediction and compared aggregated attributions with a chi-square GWAS baseline applied to the same categorical labels. SHAP prioritized several lineage-associated sites in Spike, including C23525T (S: H655Y) and A21801C (S: D80A), and also highlighted ORF8, ORF9, and intergenic positions that were not detected in the chi-square GWAS baseline. Based on the comparison between the CNN and GWAS, 23.8%-32.4% of top-ranked positions overlapped, with the shared subset enriched in Spike. We interpret these results as evidence that explainable deep learning can complement site-wise association analysis. This work therefore serves as a proof-of-concept that convolutional neural network modeling with post hoc attribution can provide an alternative genome-wide association perspective on mutations in SARS-CoV-2.