Predicting the trend of SARS-CoV-2 mutation frequencies using historical data

Area of Science:

• Epidemiology
• Computational Biology
• Machine Learning

Background:

• Predicting SARS-CoV-2 (Severe Acute Respiratory Syndrome Coronavirus 2) mutations is crucial for effective disease control.
• Viral evolution necessitates advanced methods for forecasting mutation patterns to guide diagnostics, therapeutics, and vaccines.

Purpose of the Study:

• To develop a real-time model for forecasting SARS-CoV-2 mutation surges.
• To improve the accuracy and robustness of predicting future viral mutation trajectories.

Main Methods:

• Transformed temporal prediction into a supervised learning framework using a sliding window approach.
• Modeled the first-order derivative of mutation frequency and predicted future values using machine learning models (Random Forest, XGBoost, SVM, Neural Networks).
• Validated the model using synthetic mutation patterns and real-world data from the USA and UK.

Main Results:

• Achieved low prediction error (0.1%-1%) for 30- and 80-day forecasts.
• Demonstrated significantly higher accuracy compared to traditional time-series models, with lower Mean Absolute Error (MAE) and Mean Squared Error (MSE).
• Successfully predicted unseen mutation patterns and showed superior performance on USA and UK datasets for a 2025 timeframe.

Conclusions:

• The developed model offers a robust and accurate method for forecasting SARS-CoV-2 mutations.
• The approach shows potential for application in other infectious disease forecasting and general prediction tasks.
• A GitHub package is available for enhanced accessibility and utility of the methodology.