Benchmarking machine learning robustness in Covid-19 genome sequence classification.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• The COVID-19 pandemic generated millions of SARS-CoV-2 genome sequences.
• Traditional methods struggle to analyze this vast amount of genomic data.
• Machine learning (ML) offers alternative approaches for viral sequence analysis.

Purpose of the Study:

• To establish a framework for benchmarking the robustness of ML models used in genomic analysis.
• To simulate biological sequences with errors mimicking common sequencing platforms.
• To assess the performance of various ML models under simulated sequencing noise.

Main Methods:

• Perturbing SARS-CoV-2 genome sequences to simulate sequencing errors from platforms like Illumina and PacBio.
• Experimenting with a diverse range of ML models.
• Evaluating model robustness and accuracy based on simulated noise levels and embedding methods.

Main Results:

• Different simulation-based approaches exhibit varying levels of robustness and accuracy for ML models.
• The effectiveness of ML models is dependent on specific embedding methods and noise simulation types.
• Certain perturbation budgets enhance model performance more than others.

Conclusions:

• A novel benchmarking framework for ML models in genomics has been developed.
• This framework helps researchers assess ML model reliability for analyzing SARS-CoV-2 data.
• The findings can improve understanding of viral evolution and aid in preventing future pandemics.