Accurate and fast clade assignment via deep learning and frequency chaos game representation

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• The COVID-19 pandemic led to extensive sequencing of SARS-CoV-2, creating vast genomic datasets.
• Managing and analyzing these large SARS-CoV-2 sequence datasets is challenging.
• Accurate and rapid classification of SARS-CoV-2 variants into clades is crucial for tracking viral evolution.

Purpose of the Study:

• To develop a novel, accurate, and fast method for SARS-CoV-2 variant clade assignment.
• To implement this method into a user-friendly tool named CouGaR-g.
• To identify specific genetic markers (k-mers) associated with different SARS-CoV-2 clades.

Main Methods:

• Utilizing frequency chaos game representation (FCGR) to visualize genomic sequences.
• Applying convolutional neural networks (CNNs) for deep learning-based classification.
• Training the model on a large dataset from the GISAID platform.
• Employing feature importance methods to identify significant k-mers.

Main Results:

• CouGaR-g achieved 96.29% overall accuracy in SARS-CoV-2 clade assignment on a GISAID subset.
• The tool significantly outperformed Covidex (77.12% accuracy), a random forest-based method.
• The method successfully identified k-mers corresponding to SARS-CoV-2 marker variants.

Conclusions:

• Combining FCGR and CNNs provides a highly accurate approach for SARS-CoV-2 variant classification.
• CouGaR-g offers comparable running times to existing tools but with superior accuracy.
• The identified k-mers offer insights into the genetic basis of SARS-CoV-2 variant differentiation.