Dynamic Prediction of Non-Neutral SARS-Cov-2 Variants Using Incremental Machine Learning.

Area of Science:

• Genomics
• Computational Biology
• Epidemiology

Background:

• Emerging SARS-CoV-2 variants pose significant public health challenges.
• Accurate and timely classification of new lineages is crucial for pandemic surveillance.
• Traditional methods may struggle to keep pace with the rapid evolution of viral lineages.

Purpose of the Study:

• To evaluate the efficacy of Incremental Machine Learning (IML) for predicting SARS-CoV-2 lineage classification.
• To dynamically distinguish between neutral and non-neutral (variants of concern/interest) SARS-CoV-2 lineages.
• To assess the performance of an IML model using Spike protein sequences.

Main Methods:

• Utilized Spike protein sequences from the GISAID database.
• Derived k-mer features (amino acid subsequences) from the primary sequences.
• Implemented a Logistic Regression Incremental Learner, tested monthly from February 2020 to October 2021.

Main Results:

• Achieved an average balanced accuracy of 0.72 ± 0.2, increasing to 0.78 ± 0.16 in the last 12 months.
• Successfully identified variants like Alpha, Beta, Gamma, Eta, Kappa, and Delta as non-neutral with a mean recall of approximately 90%.
• Demonstrated the model's ability to adapt and improve performance over time with new data.

Conclusions:

• Incremental learning is a valuable tool for real-time pandemic surveillance.
• The IML approach provides a dynamic method for classifying emerging SARS-CoV-2 variants.
• This methodology supports timely identification of potentially concerning viral lineages.