Exploring metabolic anomalies in COVID-19 and post-COVID-19: a machine learning approach with explainable artificial...

Area of Science:

• Metabolomics and Computational Biology
• Infectious Diseases and Immunology
• Artificial Intelligence in Healthcare

Background:

• The COVID-19 pandemic presents diverse clinical outcomes and persistent symptoms (Long COVID).
• Metabolic reprogramming is increasingly implicated in the long-term effects of SARS-CoV-2 infection.
• Understanding these metabolic alterations is crucial for diagnosing and managing post-COVID conditions.

Purpose of the Study:

• To employ machine learning (ML) and explainable artificial intelligence (XAI) for analyzing metabolic alterations in COVID-19 and Post-COVID-19 patients.
• To compare the efficacy of ML/XAI methods against traditional statistical approaches in metabolomics.
• To identify specific metabolic signatures associated with COVID-19 and Long COVID.

Main Methods:

• Analysis of 111 identified metabolites from 142 COVID-19, 48 Post-COVID-19, and 38 control patients.
• Comparison of Principal Component Analysis (PCA) and Partial Least Squares Discriminant Analysis (PLS-DA) with eXtreme Gradient Boosting (XGBoost) enhanced by SHAP values.
• Utilized SHAP values for explainability to interpret ML model predictions.

Main Results:

• XGBoost with SHAP values demonstrated superior predictive performance compared to traditional methods.
• Identified distinct metabolomic subgroups within COVID-19 and Post-COVID-19 cohorts, indicating heterogeneous responses.
• Key metabolic signatures in Post-COVID-19 included taurine, glutamine, alpha-Ketoglutaric acid, and LysoPC a C16:0.

Conclusions:

• ML and XAI integration provides a powerful, fine-grained approach for metabolomics research.
• This methodology offers deeper insights into the metabolic basis of COVID-19 progression and Long COVID.
• The identified metabolic signatures may serve as potential biomarkers for Post-COVID-19 syndrome.