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

Area of Science:

• Genomics
• Virology
• Machine Learning

Background:

• Genomic surveillance of SARS-CoV-2 is crucial for tracking variants.
• Whole-genome sequencing can be resource-intensive.
• Spike gene sequences offer a potential alternative for rapid classification.

Purpose of the Study:

• To develop and evaluate a deep learning model for classifying SARS-CoV-2 variants using only spike gene sequences.
• To assess the model's performance against existing tools like Nextclade.
• To explore the interpretability of the model's classifications.

Main Methods:

• Curated ~35,800 quality-filtered SARS-CoV-2 spike sequences.
• Trained a hybrid Convolutional Neural Network-Bidirectional Long Short-Term Memory (CNN-BiLSTM) model.
• Benchmarked performance against Nextclade and classical machine learning models.

Main Results:

• The CNN-BiLSTM model achieved high accuracy (e.g., 99.91% test accuracy).
• The model correctly identified 100% of Omicron sequences, significantly outperforming Nextclade (34.95%).
• Saliency analyses identified known and novel mutation motifs.

Conclusions:

• Spike-only deep learning models offer rapid and accurate SARS-CoV-2 variant classification.
• These models are valuable complements to phylogenetic methods, especially in resource-constrained settings.
• The approach enables efficient sample triage for further genomic analysis and timely surveillance.