Metabolomic Classification of ME/CFS via Explainable Ensemble Learningand Pareto-Guided Feature Selection,2026, Yagin et.al

[mariovitali]

Abstract


Myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS) is a debilitating multisystem illness characterised by post-exertional malaise, non-restorative sleep, and cognitive impairment, yet no objective diagnostic biomarkers have been established. Untargeted plasma metabolomics provides a broad view of the biochemical disturbances underlying ME/CFS; however, the high dimensionality of omics datasets and the limited interpretability of conventional classifiers nevertheless hinder translation into clinical practice.

This study evaluates three ensemble classifiers—Explainable Boosting Machine (EBM), XG-Boost, and LightGBM—for binary ME/CFS classification using plasma metabolomic and lipidomic profiles from 197 participants (106 ME/CFS; 91 healthy controls; 888 features). Feature dimensionality was reduced using a Pareto-Guided Recursive Neural Network (PRNN) pipeline. Model performance was assessed via 50-repeat stratified hold-out validation. EBM achieved the highest accuracy (0.909; 95% CI: 0.868–0.949) and area under the receiver operating characteristic curve (AUC: 0.940; 95% CI: 0.909–0.983), with XGBoost and LightGBM performing comparably.

Interpretability analyses revealed that pairwise metabolite interaction terms—particularly proline & indole-3-lactate, tyrosine & N-acetylornithine, and maleic acid & arachidic acid—contributed the greatest discriminative signal. An ablation analysis comparing the full interaction-augmented EBM (AUC = 0.940) with a main-effects-only EBM (AUC = 0.882) confirmed that pairwise metabolite co-variation contributes additional discriminative value beyond individual metabolite levels, implicating amino acid catabolism, tryptophan–kynurenine pathway dysregulation, mitochondrial energy impairment, and lipid remodelling as central pathophysiological features.

Global and instance-level explanations jointly demonstrated population-level metabolic signatures alongside individual heterogeneity, highlighting the added clinical value of explainable artificial intelligence (XAI) in metabolomics. These findings support EBM-based metabolomic profiling as an internally validated approach for ME/CFS classification, subject to external validation, calibration assessment, and prospective testing.


Link to study : Link

 

[Jacob Deasy]

The performance of this model is quite impressive!

I have taken a cursory glance at the data preprocessing, splitting, and validation procedure. This looks reasonably robust to me, with correct hyperparameter optimisation only on the training set, which is nice to see. There is even a comment that the feature compression steps are done correctly on each train split - pretty good, and my only remaining skepticism would lie in checking the actual code.

One suggestion for improvement would be splitting based on feature clusters (rather than the random outer splits), as this is more robust, or splitting based on the site of the data subsets. I’m surprised the latter wasn’t an obvious grouping as it reflects real-world deployment.

The data is available upon request. Health permitting, I will have a play around with it.

 

[Jonathan Edwards]

Can anyone tell me what this result is supposed to mean?

 

[Jacob Deasy]

I can give it a go.

They use multiple different machine learning classification algorithms to classify ‘has ME/CFS’ vs ‘does not’. Some of these algorithms focus on making use of the fewest possible input metabolites because this, in theory, leads to a more interpretable model and an understanding of which inputs are important.

Their best performing model has an AUROC (denoted AUC) of 0.94. This is very good because AUROC can be interpreted as the probability that the model will rank a randomly chosen positive (ME/CFS) instance higher than a randomly chosen negative (control) instance. AUROC this high is great, but also does not reflect the true imbalance of lots of controls with few ME/CFS cases (which you would see if you just used this in the wild). For that, you should focus on another metric called AUPRC. I do not have the focus right now to check if they have somewhere in the paper.

As they note, in practice the model would need to be calibrated against other datasets. Also, in practice you would need to pick a cut off between 0 and 1, with a certain specificity and sensitivity, above which you would diagnose.

Is that the kind of understanding you were looking for?

 

[Jonathan Edwards]

I had thought they were subsetting but as you say it seems they are just comparing ME/CFS to normals. This sort of statistical analysis has been claimed to be useful for at least a decade now and I find it very hard to believe that the molecules they have picked out tell us anything useful. As faras I am aware the levels are within normal ranges for pretty much everyone with ME/CFS. The chances that anything useful will come out of combinations is historically negligible. And there are, as you say, a whole load of questions about what groups and sizes shoud really be compared to make this meaningful.

And even if some association is found it is likely downstream rather than, as they put it, 'underlying'.

 

[Utsikt]

For that, you should focus on another metric called AUPRC. I do not have the focus right now to check if they have somewhere in the paper.

See figure 2:

Precision–recall (PR) analysis (Figure 2) yielded a consistent picture. Area under the precision-recall curve (AUPRC) reached 0.886 for the control class and 0.950 for the ME/CFS class. The higher AUPRC for ME/CFS indicates that precision was maintained even at high recall thresholds—a clinically important property given that undetected ME/CFS cases carry greater harm than false-positive referrals.

—-

They note this limitation in the paper:

Before interpreting these findings biologically, we emphasise that the EBM pairwise terms quantify statistical co-variation between metabolites within the fitted classifier and do not, in themselves, denote direct biochemical or causal interactions; the mechanistic correspondences proposed in the following paragraphs are therefore offered as biologically motivated, hypothesis-generating interpretations rather than demonstrated molecular relationships.

But say this in the abstract:

An ablation analysis comparing the full interaction-augmented EBM (AUC = 0.940) with a main-effects-only EBM (AUC = 0.882) confirmed that pairwise metabolite co-variation contributes additional discriminative value beyond individual metabolite levels, implicating amino acid catabolism, tryptophan–kynurenine pathway dysregulation, mitochondrial energy impairment, and lipid remodelling as central pathophysiological features.

It might be that my interpretation of the word «implicating» is affected by English being my second language, but I think it’s far too strong of a word to be used to describe hypotheses.

And «central pathophysiological features» is completely unwarranted as JE says.

I can’t say I’m a fan of putting the methods after the results either.

The paper does not state clearly where it got its data from. It says it’s a de-identified omics data set that was previously described, but there is no citation at the end of that sentence.

Sixth, although cases and controls were frequency-matched by sex, age, race/ethnicity, and geographic/clinical site at enrolment, and participants with significant confounding conditions or immunomodulatory drug use were excluded at enrolment, non-metabolomic variables including body mass index, detailed medication use, collection site, season of collection, and comorbidities were not incorporated as explicit inputs into the machine learning models.

This makes me unsure about if they controlled for these confounders at all. There’s also nothing about the female reproductive cycle.

This decision was consistent with the primary objective of the study to develop and evaluate a metabolite- and lipid-driven explainable classification pipeline and with the established practice of restricting model inputs to the omics feature space in biomarker discovery analyses.

They say they didn’t include that data because they were primarily aiming to get a pipeline set up, but that brings into question why there were even doing interpretations at all. And it makes the statement from the abstract look even worse.

 

[forestglip]

The paper does not state clearly where it got its data from. It says it’s a de-identified omics data set that was previously described, but there is no citation at the end of that sentence.

It's cited at the end of the paragraph as ref 13. The cohort size of the following study is the same.

Deficient butyrate-producing capacity in the gut microbiome is associated with bacterial network disturbances and fatigue symptoms in ME/CFS (2023, Cell) (S4ME link)

 

[Utsikt]

It's cited at the end of the paragraph as ref 13. The cohort size of the following study is the same.

Deficient butyrate-producing capacity in the gut microbiome is associated with bacterial network disturbances and fatigue symptoms in ME/CFS (2023, Cell) (S4ME link)

Thank you. I see they also cite 38 there, is that the same? They really should have put 13 in the first sentence if that’s what they are using.

 

[forestglip]

I see they also cite 38 there, is that the same?

Oh yeah, the sample size is the same for that one too. It's this:

Metabolomic Evidence for Peroxisomal Dysfunction in Myalgic Encephalomyelitis/Chronic Fatigue Syndrome, 2022, Levine,Hornig,Lipkin et al

Using regression, Bayesian and enrichment analyses, we conducted targeted and untargeted metabolomic analysis of plasma from 106 ME/CFS cases and 91 frequency-matched healthy controls.

Study linked a couple posts up (ref 13: "Deficient butyrate-producing capacity..."):

geographically diverse cohort of 106 cases and 91 healthy controls

Current thread study:

(106 ME/CFS; 91 healthy controls; 888 features).