Classification of SARS-CoV-2 Variants through The Epistatical Circos Plots with Convolutional Neural Networks

TL;DR

This work addresses automated SARS-CoV-2 variant classification by exploiting higher-order epistatic patterns inferred from genome sequences.Direct Coupling Analysis yields couplings between genomic loci and renders them as Circos plots, which are then treated as image data for CNN classifiers; the approach benefits from transfer learning to maximize performance.Using a dataset of 1,984 Circos images across Alpha, Beta, Gamma, Delta, and Omicron, the authors demonstrate near-perfect discrimination and provide GradCAM-based interpretability of the learned features.The study highlights the potential of combining population-genetic epistasis inference with deep learning for real-time genomic surveillance and extends naturally to other rapidly evolving pathogens.

Abstract

The COVID-19 pandemic has profoundly affected global health, driven by the remarkable transmissibility and mutational adaptability of the SARS-CoV-2 virus. Five major variants of concern, Alpha, Beta, Gamma, Delta, and Omicron, have been identified. By August 2022, over 12.95 million full-length SARS-CoV-2 genome sequences had been deposited in the Global Initiative on Sharing Avian Influenza Data (GISAID) database, offering an unprecedented opportunity to investigate viral evolution and epistatic interactions. Recent advances in epistatic inference, exemplified by Direct Coupling Analysis (DCA) (Zeng et al., Phys. Rev. E, 2022), have generated numerous Circos plots illustrating genetic inter-dependencies. In this study, we constructed a dataset of 1,984 Circos plots and developed a convolutional neural network (CNN) framework to classify and identify the corresponding genomic variants. The CNN effectively captured complex epistatic features, achieving an accuracy of 99.26\%. These findings demonstrate that CNN-based models can serve as powerful tools for exploring higher-order genetic dependencies, providing deeper insights into the evolutionary dynamics and adaptive mechanisms of SARS-CoV-2.

Figures (7)

• Figure 1: Temporal distribution of SARS-CoV-2 genome sequences by variants. The main panel shows the daily number of sequences collected from the GISAID database between December 2019 and August 2022, classified into the major variants of concern (Alpha, Beta, Gamma, Delta, and Omicron) and other minor lineages (Else). Distinct temporal waves correspond to the global emergence and dominance of each variant: Alpha surged in early 2021, Delta dominated mid-2021, and Omicron replaced all previous variants after late 2021. The inset highlights the early coexistence of Alpha, Beta, and Gamma variants. The sequencing volume declines after mid-2022, reflecting reduced global sequencing activity and fewer new submissions to GISAID.
• Figure 2: Effect of filtering threshold on the number of surviving loci for four variant categories. With the threshold $\phi$ increases from 0.90 to 0.99, the number of surviving loci rises steadily, reflecting the progressive removal of highly conserved sites. Among all variants, Omicron consistently retains the largest number of loci, followed by Delta, Else, and Alpha, indicating higher overall sequence variability in Omicron and Delta compared to Alpha. A near-exponential growth in surviving loci is observed when $\phi > 0.96$, demonstrating the sensitivity of the dataset size to the chosen conservation threshold.
• Figure 3: The graph is an image of the epistasis of data sets, with the line segments indicating the top 200 significant pairwise epistasis between the coding region motifs in the data set for this date. The colored lines indicate the top 50, and the gray lines indicate the top 51 to 200. Red lines indicate short-distance connections (distance less than or equal to 3 bp); blue lines indicate longer distances. Threshold $\phi$=95%,(a) Alpha;(b) Delta;(c) Omicron;(d) Else.
• Figure 4: Training loss and validation accuracy curves of the DenseNet121 model over 100 epochs. The training loss decreases steadily while the validation accuracy remains consistently high, indicating stable convergence and strong generalization without overfitting.
• Figure 5: Confusion matrix of the DenseNet121 model with transfer learning setting for predicting predicting SARS-CoV-2 variant classes from DCA-derived epistatic interaction images. Values represent normalized prediction frequencies for the four variant categories (Alpha, Delta, Omicron, and Else). The strong diagonal dominance indicates near-perfect classification performance, with only minor reciprocal confusion between Alpha and Delta and perfect separation for Omicron and the composite Else class. This matrix highlights the discriminability of lineage-specific epistatic structures captured by the model.