Metabolomics-Based Machine Learning Diagnostics of Post-Acute Sequelae of SARS-CoV-2 Infection

Ethan Cai ¹, Valentina L Kouznetsova ^2,3,4, Igor F Tsigelny ^2,3,4,5

⁰REHS Program, San Diego Supercomputer Center, University of California San Diego, San Diego, CA 92093, USA.

Metabolites +

|

December 24, 2025

View abstract on PubMed

Summary

Researchers developed a machine-learning model to identify Post-Acute Sequelae of SARS-CoV-2 Infection (PASC), or Long COVID, by analyzing metabolites. The model shows promise in differentiating PASC from similar conditions, aiding faster diagnosis.

Area Of Science

Computational biology and bioinformatics
Infectious disease research
Metabolomics

Background

COVID-19 has caused millions of deaths globally and its long-term effects, known as Post-Acute Sequelae of SARS-CoV-2 Infection (PASC) or Long COVID, remain poorly understood.
PASC presents a significant public health challenge due to its widespread impact and complex symptomatology.
Distinguishing PASC from other fatiguing illnesses is crucial for accurate diagnosis and effective treatment.

Purpose Of The Study

To develop and validate a machine-learning model capable of distinguishing PASC from PASC-similar diseases based on metabolite profiles.
To identify unique metabolic signatures associated with PASC.
To explore the potential for accelerating PASC diagnosis and intervention.

Main Methods

A machine-learning model, specifically a multi-layer perceptron, was trained using molecular descriptors of PASC-dysregulated metabolites (p ≤ 0.05).
The model was tested on its ability to recognize PASC-dysregulated metabolites and differentiate PASC from myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS), Lyme disease, postural orthostatic tachycardia syndrome (POTS), and irritable bowel syndrome (IBS).
Pairwise model testing was conducted to elucidate similarities between PASC and other conditions, particularly fibromyalgia (FM).

Main Results

The machine-learning model achieved an AUC-ROC of 0.8991 in recognizing PASC-dysregulated metabolites in an independent testing set.
The model successfully differentiated PASC from ME/CFS, Lyme disease, POTS, and IBS.
The model was unable to differentiate fibromyalgia (FM) from PASC, indicating a significant metabolic similarity between these two conditions.

Conclusions

The developed machine-learning approach offers a novel method for PASC diagnosis, potentially reducing the need for lengthy exclusion processes.
The model's reliance on molecular descriptors allows for broad applicability across various metabolites.
This study highlights the metabolic overlap between FM and PASC and presents a promising tool for faster PASC diagnosis, potentially leading to improved patient outcomes.

Keywords:

IBS Long COVID ME/CFS PASC POTS fibromyalgia machine learning metabolites