🧬 biology

Metabolomic reconstruction fidelity between people and between visits in COVID-19 cohorts

This study evaluates the fidelity of metabolomic reconstructions across individuals and visits in COVID-19 cohorts, finding that modeling within-person deviations improves reconstruction accuracy in the Karu cohort while highlighting the sensitivity of these results to statistical objectives and the lack of established clinical utility.

Original authors: Roberto Navarro Quiroz, Eloina Zarate Peñata, Cecilia Fernández-Ponce, Lisandro Pacheco Lugo, Leonardo Pacheco Londoño, Antonio Acosta-Hoyos, Nataly Galan Freyle, Andrea Jaruffe Pinilla, Katherine Esc

Published 2026-10-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Roberto Navarro Quiroz, Eloina Zarate Peñata, Cecilia Fernández-Ponce, Lisandro Pacheco Lugo, Leonardo Pacheco Londoño, Antonio Acosta-Hoyos, Nataly Galan Freyle, Andrea Jaruffe Pinilla, Katherine Escorcia Lindo, Ibeth Luna-Rodríguez, Elkin Navarro Quiroz

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

When doctors draw blood to understand a disease, they are often looking for a signal hidden within a storm of noise. The human body is a complex chemical factory, constantly producing and breaking down thousands of tiny molecules called metabolites. These chemicals tell a story about how a person is feeling, how their immune system is fighting, and how their cells are working. However, this story changes from person to person. One individual might naturally have higher levels of a specific chemical than another, simply because of their unique biology. At the same time, that same person's chemical levels can shift from day to day as they recover from an illness or react to their environment. The challenge for scientists is to figure out which parts of the data represent a person's permanent biological fingerprint and which parts represent the temporary fluctuations of their daily life. If researchers cannot distinguish between these two types of variation, they risk drawing the wrong conclusions about how a disease like COVID-19 affects the body.

A team of researchers set out to test how well they could separate these two layers of information using data from people who had been infected with COVID-19. They focused on a specific group of eight chemical signals in the blood, including methionine and its oxidized form, methionine sulfoxide. These molecules are part of the body's sulfur metabolism and are known to interact with each other. The scientists gathered blood samples from hundreds of participants across three different studies. Some of these people were tested multiple times over several days or weeks, allowing the researchers to see how their chemical levels changed over time. The goal was to see if they could create a simple, two-dimensional summary of this complex data that would accurately capture both the unique chemical signature of each person and the specific changes that happened to them during their illness.

To do this, the researchers treated the data in two different ways. First, they looked at the average chemical levels for each person, treating those averages as the person's unique identity. Second, they looked at how much a person's levels deviated from their own average during each visit, treating those shifts as the story of their recovery. They then tried to rebuild the original data using only two simplified directions, or coordinates, learned from these different perspectives. It is like trying to describe a three-dimensional object using only a flat shadow; the researchers wanted to know which method of casting that shadow would preserve the most detail when they tried to reconstruct the original shape. They tested this by training their models on most of the participants and then seeing how well those models could predict the data for a person they had never seen before.

The results showed that the answer depends heavily on which group of people they were studying. In one specific cohort of thirty-three people, the researchers found that learning from the daily fluctuations within each person was significantly better at reconstructing the data than learning from their average levels. When they used the daily changes to build their model, the error in reconstructing the data dropped by a measurable amount. Specifically, the loss in accuracy was reduced by 0.0670 units when using the daily changes compared to using the averages. This improvement was not uniform across all the chemicals; for some molecules like taurine and glutamic acid, the daily-change model worked much better, while for others, the improvement was smaller or even slightly worse. However, in the other two groups of people they studied, this clear advantage did not appear. The models built from the daily changes performed about the same as, or sometimes slightly worse than, the models built from the averages. This suggests that there is no single, universal rule for how to summarize this kind of biological data; the best approach depends on the specific population and the conditions under which the samples were collected.

The study also examined the relationship between methionine and methionine sulfoxide, a pair of chemicals that are chemically linked because one is the oxidized version of the other. The researchers found that these two molecules tended to rise and fall together in the blood samples, both within the same person over time and across different people in a separate, large dataset collected years before those people were diagnosed with COVID-19. This consistency suggests a strong biological connection between the two. However, the strength of this connection changed depending on how the data was processed. When the researchers applied a specific mathematical adjustment to account for differences in sample concentration, the strength of the link between the two chemicals decreased noticeably. This indicates that while the relationship is real, its apparent strength can be influenced by technical factors in how the blood samples were handled and measured.

Ultimately, the paper does not claim to have discovered a new cure or a definitive diagnostic tool. Instead, it provides a careful evaluation of how well current methods can summarize complex biological data. The researchers demonstrated that while it is possible to compress a large amount of chemical information into a simpler form, the accuracy of that compression varies. In some cases, focusing on how a person's body changes day-to-day offers a clearer picture than focusing on their long-term average. In other cases, the difference is negligible. The study highlights that the choice of how to analyze the data matters just as much as the data itself. By showing exactly where these methods succeed and where they fall short, the work helps future scientists understand the limits of what can be learned from blood tests and reminds them that the story of a disease is written in both the steady traits of a person and the shifting tides of their daily recovery.

Technical Summary: Metabolomic Reconstruction Fidelity in COVID-19 Cohorts

Problem Statement
The study addresses a fundamental question in multilevel metabolomics: when summarizing repeated measurements from a panel of metabolites into a low-dimensional representation (specifically two coordinates), does learning from differences between people (personal means) retain the fidelity of within-person deviations as well as learning directly from those deviations? The authors note that while both levels contain relationships between metabolites, these relationships need not be identical. A high correlation between personal means does not guarantee that deviations during follow-up are similarly correlated. The paper seeks to evaluate the "retrospective fidelity" of such summaries for measurements already acquired, without claiming to measure physiological importance, Shannon information, or clinical prognosis.

Methodology
The research performs an exploratory reanalysis of publicly available, processed metabolomic data from four distinct sources: two waves of the BEAT-COVID study, the Karu cohort, and the Cornell cohort (split into GIM1 and GIM2 batches). A separate cross-sectional analysis was conducted using the LEOCC biobank.

Data Scope: The analysis focuses on a common panel of eight analytes (C08): methionine, cystathionine, cysteine, glycine, serine, glutamic acid, methionine sulfoxide, and taurine. The study includes 224 participants with repeated measurements (581 samples) and 417 participants from LEOCC with single prediagnostic samples.
Statistical Framework: The authors decompose the total variation ( $C_T$ $C_{T}$ ) into between-person variation ( $C_B$ $C_{B}$ , differences in personal means) and within-person variation ( $C_W$ $C_{W}$ , deviations from personal means).
- Models: Three models were learned from standardized matrices: Model $B$ (from $C_B$ ), Model $W$ (from $C_W$ ), and Model $T$ (from $C_T$ ).
- Objective: The primary objective was to assess reconstruction fidelity. Models were trained on all participants except one (held-out), and the loss was measured as the squared residuals of the held-out person's data projected onto the two learned coordinates.
- Metrics: Loss was calculated for two objectives: reconstructing within-person deviations ( $L_W$ ) and reconstructing personal means ( $L_B$ ). The primary contrast was $L_W(B) - L_W(W)$ , where a positive value indicates that Model $W$ (learned from deviations) better reconstructs deviations than Model $B$ (learned from means).
Sensitivity and Validation: The study employed 1,000 bootstrap resamples per stratum to generate exploratory 2.5–97.5 percentiles. Sensitivity analyses included probabilistic quotient normalization (PQN), varying visit counts, and adjusting for covariates (age, sex, BMI) in the LEOCC cross-sectional analysis.
Complementary Analysis: The authors examined specific molecular pairs, particularly methionine and methionine sulfoxide, to see if previously prioritized relationships recurred in external resources with repeated measurements and in a prediagnostic cross-sectional context.

Key Results

Reconstruction Fidelity in Karu: The Karu cohort provided the clearest evidence of an advantage for Model $W$ $W$ . Learning from within-person deviations reduced the reconstruction loss for held-out deviations by 0.0670 squared standardised log units compared to learning from personal means ( $L_W(B) - L_W(W) = 0.0670$ $L_{W} (B) - L_{W} (W) = 0.0670$ ; 95% exploratory interval: 0.0347–0.1646). This represented a 30.2% relative reduction in loss with standardised logarithms.
- This advantage was unevenly distributed: taurine, methionine, and glutamic acid contributed to the loss reduction, while five other signals showed higher loss with Model $W$ .
- The cost of reconstructing personal means using Model $W$ was uncertain; the interval for the loss difference included zero.
Other Cohorts: In BEAT (waves 1 and 2) and Cornell (GIM1, GIM2, and combined), the advantage of Model $W$ was not established as clearly. Intervals for the loss differences often included zero, and in some cases, Model $T$ (total variation) showed slightly lower loss than Model $W$ . No clear between-cohort differences were demonstrated.
Molecular Relationships:
- Methionine–Methionine Sulfoxide: This pair showed a positive within-person correlation in all repeated-measures cohorts (Karu, BEAT) and a strong positive correlation in the cross-sectional LEOCC cohort (Pearson $r \approx 0.49$ ). This association remained robust across sensitivity analyses (PQN, visit restrictions, covariate adjustment).
- Other Pairs: Secondary pairs (serine–cystathionine, cystathionine–cysteine) did not show consistent patterns across cohorts or between within-person and cross-sectional contexts.
Scale Sensitivity: The relative magnitude of the advantage depended heavily on standardisation. Without standardisation, the relative loss reduction in Karu dropped to 3.3%, highlighting that the weighting of analytes significantly alters the objective.

Significance and Claims
The paper modestly frames its contribution as an evaluation of retrospective fidelity for measurements already acquired.

Primary Claim: The study demonstrates that for specific cohorts (notably Karu), learning a low-dimensional representation directly from within-person deviations yields a more faithful reconstruction of those deviations than learning from personal means. However, this is not a universal rule across all cohorts or scales.
Limitations: The authors explicitly state that metabolic function and clinical utility are not established by these results. The reconstruction loss does not measure physiological importance or prognostic value. The methionine–sulfoxide covariation is described as a complementary characterisation that does not validate the reconstruction gain for the full panel.
Methodological Stance: The work serves as a diagnostic tool to compare how different objectives (means vs. deviations) affect the summarization of a measured panel. It does not propose a new predictive model or claim that better reconstruction improves clinical decision-making. The findings are presented as descriptive and exploratory, with intervals serving as exploratory resampling descriptions rather than confirmatory hypothesis tests.

In conclusion, the study provides a rigorous, reproducible framework for assessing how well two-dimensional metabolomic summaries retain specific levels of variation, revealing that the optimal learning objective depends on the specific cohort, the scale of measurement, and the target of reconstruction.

Technical Summary: Metabolomic Reconstruction Fidelity in COVID-19 Cohorts

More like this