Classification of SARS-CoV-2 Variants Through the Epistatic Circos Plots with Convolutional Neural Networks

Area of Science:

• Genomics
• Virology
• Computational Biology

Background:

• The COVID-19 pandemic, caused by SARS-CoV-2, is characterized by the virus's high transmissibility and rapid mutation.
• Several SARS-CoV-2 variants of concern (VOCs), including Alpha, Beta, Gamma, Delta, and Omicron, have emerged, posing significant public health challenges.

Purpose of the Study:

• To develop an integrative computational framework for classifying SARS-CoV-2 variants.
• To characterize lineage-specific epistatic signatures using genomic sequence data.

Main Methods:

• Combined direct coupling analysis (DCA) to infer pairwise mutational couplings.
• Utilized Circos-based visualization to transform coupling data into images.
• Employed convolutional neural networks (CNNs) for variant classification based on visual genomic signatures.

Main Results:

• The integrative framework achieved robust classification of SARS-CoV-2 variants (Alpha, Delta, Omicron, Else).
• The best CNN model demonstrated high performance with a weighted-average F1 score and an AUC close to 1.
• Temporal holdout analyses confirmed the framework's sustained predictive accuracy over evolutionary time, achieving 87.85% weighted-average F1 score.

Conclusions:

• The developed framework effectively leverages genomic data and machine learning for accurate SARS-CoV-2 variant classification.
• This approach provides a valuable tool for monitoring viral evolution and supporting public health strategies during pandemics.
• The study highlights the potential of integrating bioinformatic analyses with deep learning for rapid and reliable pathogen surveillance.