Explainable convolutional neural network model provides an alternative genome-wide association perspective on...

Area of Science:

• Genomics
• Virology
• Computational Biology

Background:

• Understanding SARS-CoV-2 evolution is crucial for public health.
• Identifying key genomic features aids in tracking viral variants.
• Traditional methods for genomic analysis have limitations.

Purpose of the Study:

• To develop an explainable deep learning model for classifying SARS-CoV-2 Variants of Concern (VOCs).
• To identify informative genomic features contributing to VOC classification.
• To compare deep learning-based attribution with traditional genome-wide association studies (GWAS).

Main Methods:

• Developed a convolutional neural network (CNN) model for SARS-CoV-2 genome classification.
• Utilized SHapley Additive exPlanations (SHAP) for model interpretability and feature attribution.
• Compared CNN-SHAP attributions with chi-square GWAS baseline.

Main Results:

• The CNN model achieved 99.96% accuracy in classifying VOCs (Alpha, Beta, Gamma, Delta, Omicron).
• SHAP analysis highlighted important genomic sites in Spike, ORF8, ORF9, and intergenic regions.
• Overlap between CNN-SHAP and GWAS top-ranked positions ranged from 23.8% to 32.4%, enriched in Spike.

Conclusions:

• Explainable deep learning, specifically CNNs with SHAP, offers a complementary perspective to traditional GWAS for analyzing viral mutations.
• This approach can uncover genomic features missed by conventional methods, aiding in understanding viral evolution.
• The study serves as a proof-of-concept for using deep learning in SARS-CoV-2 genomic association analysis.