Predicting COVID-19 disease severity from SARS-CoV-2 spike protein sequence by mixed effects machine learning

Area of Science:

• Virology
• Epidemiology
• Computational Biology

Background:

• COVID-19 variants like Delta and Omicron present varying severe disease risks.
• Existing studies often lack sequence-level viral data or are limited in scope.
• Analyzing SARS-CoV-2 sequence data from GISAID for disease severity is challenging due to temporal and geographic biases.

Purpose of the Study:

• To develop robust and biologically meaningful genotype-patient status models for predicting COVID-19 severity.
• To address limitations in existing studies by incorporating sequence-level data and metadata.
• To evaluate machine learning models for predicting the impact of viral mutations on disease outcomes.

Main Methods:

• Utilized a subset of GISAID data with "patient status" metadata (mild/severe disease).
• Employed efficient mixed-effects machine learning with GPBoost, treating country as a random effect.
• Trained and validated models using temporally split GISAID data, including Omicron variants.

Main Results:

• GPBoost models demonstrated higher predictive accuracy for the impact of spike protein mutations on patient outcomes.
• Outperformed fixed-effect models such as XGBoost, LightGBM, random forests, and elastic net logistic regression.
• Highlighted the challenges posed by data biases and evolving pandemic responses in GISAID data.

Conclusions:

• Mixed-effects machine learning, specifically GPBoost, offers a more predictive approach to modeling COVID-19 variant severity using GISAID data.
• This method can better capture the influence of viral genetic changes on disease outcomes.
• Future research should consider advanced modeling techniques to overcome data limitations in viral surveillance.