Latest News

Metabolomics And Machine Learning Create ‘Signatures’ Of Viral Infections

By Deborah Borfitz


September 9, 2021 | Over 18,000 different metabolites have been identified in humans that could potentially be altered by a respiratory virus infection, according to Benjamin Pinsky, M.D., Ph.D., medical director of the clinical virology laboratory at Stanford Health Care, speaking on advances in metabolomics at the recent Next Generation Dx Summit. The changes can be detected by the coupling of liquid chromatography with mass spectrometry to produce host metabolic signatures.

Mass spectrometry can be used to look at nucleic acids (DNA and RNA) and other small molecules of proteins and metabolic intermediates, he says. While proteomics can reveal important information about small molecules such as viral antigens and antibodies, metabolomics can identify “all the other things” within a cell.

This has led to development of a reference metabolome database for understanding the enormously complex pathways of human metabolism, Pinsky says, providing a way to distinguish individuals who are infected from those who are not.

While sample preparation for liquid chromatography is relatively straightforward, data acquisition is “complicated at the moment,” he says. The process involves separating samples into its individual parts, which are then measured by mass spectrometry. 

The analytical technique identifies peaks in mass over charge that machine learning models turn into distribution patterns, Pinsky explains. The goal is to identify the pattern that best distinguishes between infected and noninfected individuals for a given virus.

Typically, mass spectrometry involves an upfront separation step to move molecules in a liquid solution into the gas phase, he continues. That step is unnecessary when the technique is being paired with liquid chromatography.

Data analysis looks at retention time, mass over charge, and peak intensity, Pinsky says, which correlates with the concentration of an analyte. “What’s tricky is that a metabolite can be charged or uncharged, so you need to separate them out.”

This required the use of a new, two-column chromatography method, developed at a different Stanford lab, to uncover distinct diagnostic signatures, he says.

Easy And Accurate

For biomarker discovery, Pinsky reports on use of liquid chromatography quadrupole time-of-flight mass spectrometry (LC/Q-TOF) for untargeted analysis of thousands of compounds in clinical samples to learn which ones make up the signature of a disease.

A triple quadrupole mass spectrometer, common in clinical labs, is used for targeted analysis of preselected compounds, he says. It works as a discriminator to look for a “limited signature.”

The metabolomics approach has several advantages as a diagnostic methodology, Pinsky continues. It can be performed directly from the primary specimen with minimal sample volume, the reagents are inexpensive, testing can be completed in two to three minutes, and the technique would be well suited to primary-care settings “as far-fetched as that may sound.”

Airport instruments that look for explosives residue are “mini-mass spectrometers,” says Pinsky. And the molecules being detected are just as small as the ones needing detection for his proposed metabolomics approach.

In a recently published study in EBioMedicine (DOI: 10.1016/j.ebiom.2021.103546), Pinsky and his Stanford colleagues demonstrated the feasibility and accuracy of an untargeted metabolomics approach from nasopharyngeal samples combined with machine learning for the identification of distinct metabolic signatures for the diagnosis of influenza infection.

Findings were based on an analysis of 236 samples—half positive and half negative and all age- and sex-matched—and used two machine learning methods (gradient boosted decision trees and random forests, variants of regression logistics). Researchers used ROC analysis (showing the trade-off between sensitivity and specificity) and the SHAP (SHapley Additive exPlanations) method to quantify the impact of each of the top 20 differentiating ion feature on the models.

Independent validation of the biomarker signature was performed in a prospective cohort of 96 symptomatic individuals, he reports. The metabolite most strongly associated with differential classification was pyroglutamic acid.

From Flu To COVID

The Stanford team has since become interested in developing a metabolomic signature for SARS-CoV-2, says Pinsky. Among the questions to be answered are what level of detection performance will be achievable, if COVID-19 and influenza signatures will overlap, and if the SARS-CoV-2 signature will be commutable to other patient populations.

Metabolomics more commonly peers into plasma, he notes, which Stanford researchers also tried. In a 262-patient study, the top-20 identified amino acids did a good job of differentiating SARS-CoV-2 from other viral infections, he says.

Pyroglutamic acid again seemed important as a distinguishing feature and, at least in positive patients, arginine—which is involved in nasal regulation. More features are needed to generate a disease signature for COVID-19.

Ultimately, the research team is interested in using their metabolomics approach to monitor severity of illness as well as predict the presence of disease, Pinsky says. “We can only speculate now on the mechanism of action [linking implicated metabolites].”

Load more comments