AI-driven multi-omics modeling of Myalgic Encephalomyelitis / Chronic Fatigue Syndrome, 2025, Xiong et al.

The Independent:

Gut bacteria could help diagnose long Covid and chronic fatigue syndrome, researchers find​

Bacteria in the gut could help diagnose long Covid and chronic fatigue syndrome, researchers have found.

The debilitating condition, which can cause extreme tiredness, sleep problems, dizziness and brain fog, is often overlooked as there is no specific test to diagnose it. This means doctors have to simply rule out other illnesses.
But research published in the journal Nature Medicine has looked at gut bacteria, immune responses and metabolism to find a way of diagnosing the condition.

The findings, potentially relevant to long Covid due to its similarity with chronic fatigue syndrome, come from data on 249 individuals analysed using a new artificial intelligence (AI) platform that identifies disease biomarkers from stool, blood, and other routine lab tests.

“Our study achieved 90 per cent accuracy in distinguishing individuals with chronic fatigue syndrome, which is significant because doctors currently lack reliable biomarkers for diagnosis,” said study author Dr Derya Unutmaz, professor in immunology at the Jackson Laboratory in the US.

“Some physicians doubt it as a real disease due to the absence of clear laboratory markers, sometimes attributing it to psychological factors.”
It is estimated that 404,000 people in the UK have chronic fatigue syndrome or ME, according to Action for ME. About half of the 1.9 million people in the UK with long Covid are also thought to have symptoms that are similar to ME.

Although it is not yet known what causes chronic fatigue syndrome, there is evidence that certain infections, including but not limited to viruses, can trigger the illness.

This new research, led by Dr Julia Oh, formerly at the Jackson Laboratory and now a microbiologist and professor at Duke University in North Carolina, investigates how the gut microbiome – the bacteria in your gut – and immune system interact in a patient with chronic fatigue syndrome.

To conduct the study, researchers used data collected from the Bateman Horne Center, a leading ME/CFS, long Covid and fibromyalgia research centre in Salt Lake City, Utah. Dr Ruoyun Xiong, also a lead author on the study, developed a research tool called BioMapAI.
Click to expand...
This tool helped to compare gut bacteria, immune cells, blood test data, and clinical symptoms from 153 patients and 96 healthy participants over four years.

Researchers found analysing immune cells proved the most accurate in predicting how severe the participants’ symptoms were, but found data from gut bacteria helped predict participants’ emotional symptoms and sleep disturbances. They found those with chronic fatigue had lower levels of butyrate, a beneficial fatty acid produced in the gut, along with other nutrients essential for metabolism, inflammation control, and energy.
Click to expand...
annoying use of chronic fatigue as ever,
 
I'll bet you can also build "multi-omics" model for stock prices. The question is: can it predict? It's not that difficult to test, really; there is no need to say "it may predict". Publishing a paper about a model without actually testing it would be no more than a hand-waving. In this case, they seem to be proposing a hypothesis based on the prediction, without testing the prediction.

Based on BioMapAI’s predictions and subsequent network analyses, we propose that some of the disease-specific changes in ME/CFS arise from disrupted associations between the gut microbiome, immune system, and metabolome (Figure 5).
Click to expand...
 
Jonathan Edwards said:
I can't follow much of this sort of spin but they seem to be saying MAIT and gamna delta T cells are making IFN-G. Suits me and Jackie Cliff fine I would say. I doubt your, actual "gut flora' matters that much.
Click to expand...
I was surprised that they noted Granzyme A and not Granzyme B. Can you remember what Jackie found regarding Granzyme A vs B?

Background :
AI Overview said:
Granzyme A and Granzyme B are both serine proteases stored in the cytotoxic granules of T cells and natural killer (NK) cells. However, they induce cell death through different mechanisms. Granzyme B induces apoptosis by activating caspases, while granzyme A induces a caspase-independent cell death pathway characterized by single-stranded DNA damage and mitochondrial dysfunction.
Click to expand...
 
Jonathan Edwards said:
I can't follow much of this sort of spin but they seem to be saying MAIT and gamna delta T cells are making IFN-G. Suits me and Jackie Cliff fine I would say. I doubt your, actual "gut flora' matters that much.
Click to expand...
From the text:
Increased correlations between gut microbiome and mucosal/inflammatory immune modules, including CD8+ MAIT and IFN-γ+ CD4 memory cells, suggested an increased association with microbiome and inflammatory elements in ME/CFS
Click to expand...
We further validated these findings with two independent cohorts (Guo et al. and Raijmakers et al.). For example, increased tryptophan metabolism, associated with gastrointestinal issues, lost its negative association with Th22 cells, and gained correlations with γδ T cells and the secretion of IFN-γ and GzA from CD8 and CD8+ MAIT cells.
Click to expand...
In healthy controls, these microbial metabolites are associated with activity of mucosal immune cells, including Th17, Th22 and Treg cells. In ME/CFS, however, these regulatory networks break down, with heightened proinflammatory responses mediated by γδ T cells and CD8 MAIT cells producing IFN-γ and GzA, which in turn were associated with subjective health perception and social functioning.
Click to expand...
I don’t think there’s any strong association with IFN-gamma production and disease state here. You’ll find IFN-gamma+ T cell subsets via flow cytometry in every participant if you start with a high enough cell count.

Out of the dozens of increasingly fine-grained subsets they sorted, they had a couple of scattered weak associations, most of which were only apparent in long term and not short term disease. Even with the p-value corrections I would say this is some egregious fishing for correlation and doesn’t tell us much of anything about a possible role of IFN-g in ME/CFS.
 
Abstract of published paper (paragraph breaks added)

Myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS) is a chronic illness with a multifactorial etiology and heterogeneous symptomatology, posing major challenges for diagnosis and treatment. Here we present BioMapAI, a supervised deep neural network trained on a 4-year, longitudinal, multi-omics dataset from 249 participants, which integrates gut metagenomics, plasma metabolomics, immune cell profiling, blood laboratory data and detailed clinical symptoms. By simultaneously modeling these diverse data types to predict clinical severity, BioMapAI identifies disease- and symptom-specific biomarkers and classifies ME/CFS in both held-out and independent external cohorts.

Using an explainable AI approach, we construct a unique connectivity map spanning the microbiome, immune system and plasma metabolome in health and ME/CFS adjusted for age, gender and additional clinical factors. This map uncovers altered associations between microbial metabolism (for example, short-chain fatty acids, branched-chain amino acids, tryptophan, benzoate), plasma lipids and bile acids, and heightened inflammatory responses in mucosal and inflammatory T cell subsets (MAIT, γδT) secreting IFN-γ and GzA.

Overall, BioMapAI provides unprecedented systems-level insights into ME/CFS, refining existing hypotheses and hypothesizing unique mechanisms—specifically, how multi-omics dynamics are associated to the disease’s heterogeneous symptoms.
 
Finally: A diagnostic test for chronic fatigue syndrome

Specifically the team found that ME/CFS patients had lower levels of a fatty acid produced in the gut called butyrate as well as reduced levels of other nutrients that play a role in energy production and overall metabolism. The researchers also found heightened inflammatory responses in T cells (a type of white blood cell) known as MAIT.
 
Supplementary table 3 gives some more info about the predictions. Using all datasets (metabolomics, immune, microbiome, etc.) the BioMapAI had an accuracy of 72.5%. Precision was 0.71 meaning that approximately 7 out of 10 participants predicted to have ME/CFS actually had ME/CFS.

1753542745101.png

On the other cohorts BioMapAI had an accuracy between 58-72%, often because the dataset was smaller or only included microbiome and no immune data.
1753542907285.png
 
Rather than just predicting ME/CFS versus control, the model could also estimate symptom scores such as subscales of the SF-36. It seems that the immune data was most useful for prediction physical functioning and general health while info on gut microbiome was useful for gastrointestinal problems (and emotional well being and sleeping problems for some reason).

Figure 2.C

1753543054944.png
 
I think my main criticism is that this is a combination of:
1) small cohort (for a supervised analysis)
2) a large amount of associations done (some with very specific symptoms that would have been endorsed by only a fraction of the small cohort)
3) no attempt at feature selection to test the predictive power of the top factors and test them as separate hypotheses

All this with a machine learning method that is notorious for overfitting. The test cohort was ~30 people, and the independent cohort validations were still done with a large amount of features.

And I need to stress how much a predictive power of >0.8 for nearly all of those individual symptoms is remarkable—too good, really. But their main model barely performed better than other more standard methods.

So a very complicated model with several factors stacked in their favor for overfitting even with additional validation, and all that could be pulled out was a handful of metabolites and a reference to some T cell subsets that seems pulled out of nowhere.
 
Moved post
---------

https://newatlas.com/medical/diagnostic-fatigue-syndrome/

There is a question mark in the news title, but still, it implies that a diagnostic test for ME is just around the corner. I can't remember whether this study (Jackson Lab study about microbiome, metabolites, etc) has been discussed here, but I thought I'd point out that it has gotten news interest.
 
Last edited by a moderator:
All in all my sense is that this model is driven by the sheer amount of features included in a model with way more parameters than observations, such that validation is basically just testing whether it can pick something out of thousands of features to classify any new case.

The things that are driving prediction for any new data presented to the model are likely different sets of features every time. Which wouldn’t necessarily be an issue for most tasks that you would use a model like this for, like trying to target advertising to someone’s search history.

But ultimately the purpose of multi-omics in biology is to find relevant pathways in disease that would have been missed if you didn’t cast a wide net, or at least find clinically testable markers that can distinguish cases/controls with a manageable number of features.

I’d like to be more positive given this kind of analysis is exactly what I’m doing in my PhD. But I’m just not sure what this particular analysis gives us—it doesn’t tell any convincing biological story, despite their attempts to pull things together in the final figure.
 
You must log in or register to reply here.
Back
Top Bottom