Deep phenotyping of post-infectious myalgic encephalomyelitis/chronic fatigue syndrome, 2024, Walitt et al

Just as a sanity check, since I'm noticing something counterintuitive. The worse the fatigue severity, the more Earnings from the EEFRT study.

I pulled up the raw data, and looked at Earnings vs some of the scores of the MFI-20 or MASQ surveys. (They aren't summed together in the raw data, they're split into things like "mental score", "physical score", "verbal memory".)



It does seem that way. I never really looked at the EEFRT portion too deeply, so I don't really remember what all the "Median_Trial_Time_Hard_Trials", "Total_Trials_Completed_Hard_Trials", etc actually means.

But it seems like the spreadsheet above is showing that the more severe the fatigue, the less time spent per hard trial, but the more hard trials total completed?

 

Couple of points @forestglip

1. I'm loving the embedded google sheets, that works really well for me.
2. I'm very impressed by your interrogation of the data and your commitment to extracting the true value from it. I notice you're not afraid to put in the effort.
3. I believe Benjamini Hochberg is for independent tests, some of those cognitive tests are probably not independent. So the BH correction could be a bit strong.
I haven't used it but there is a test that you can use to control false discovery rate under conditions of dependence, called Benjamini-Yekutieli. https://projecteuclid.org/journals/...-multiple-testing/10.1214/aos/1013699998.full

 

SF-36 seems like a better metric than fatigue for disability severity.

Here's the SF-36.

I might exclude this question and make a custom total since this seems irrelevant for the level of disability one is experiencing at the moment:

And it looks like the "RN Polysymptom Index" has a bunch of yes or no questions about PEM that I could add to the SF-36 score. These seven are the ones where not everyone put yes and there are no missing values:

Here is the what the spread of summed PEM answers looks like:


Maybe I should multiply the PEM score to be higher than the SF-36 score before adding them together since it's the main symptom of ME/CFS?

Edit: Missed one PEM question, it's the last one added above. Also updated the plot.

 

Yeah I was really excited to learn the forum could do that!

I have a ton of fun trying to treasure hunt in data. I'm not very good at it, but I can imagine the possibilities.

Thanks, you're probably right. I'll look at that. Though I didn't really care about multiple test correction with this massive dataset. With only 17 participants and 3000 tests, I think there's very little hope of much being significant after correction. I mainly just wanted to get them in order to see what kind of stuff is near the top for most correlated.

 

I agree with this approach. Worrying about absolute p values isn't as important as ranking things and seeing if those things show up near the top in other studies. That's when there's actual signal.

Another bit of info for anyone scrolling the mega list of correlates with severity, the acronym VM seems to stand for Vector Movements. It's about whether accelerometers attached to people's wrists and hips moved. Makes sense that more severe people stayed more still post-CPET.

 

Thanks I didn't know what VM was.

But weirdly, I think the correlation is the opposite. The higher the severity, the more movements. I think it's mainly because of the MASQ, which has a higher score so weighs more in the final score, and is more for cognitive dysfunction.

I wasn't sure if maybe I was reading the MASQ scores backwards, but it seems right. Here is the raw data for the mental fatigue scale from the MFI-20 and "visual/perceptual ability" from MASQ for both groups. ME/CFS is red.

It does look like higher values for both scales are worse because both are higher in the ME/CFS group:


And yet a worse visual/perceptual ability is correlated to higher wrist and hip movements:


Much less correlation of movement with the physical fatigue scale:


Also, it doesn't even seem like the ME/CFS group is moving their hip less than the healthy group at this time point, which is a bit weird.


----

Edit: And because it was bugging me if maybe I was wrong and higher MASQ scores mean less cognitive problems, I searched and found a paper that says:

I can't find any official instructions, but this study says higher equals worse. And in the NIH study, ME/CFS had a higher score for all MASQ domains.

So I can probably safely add it to the MFI-20 scores where higher also means worse there. So the correlations should be good in the previous post, they're just heavily weighted to cognitive complaints since that is all of MASQ and a portion of MFI.

But I'm going to do a new score without these surveys anyway.

 

Oh, I looked up the instructions for SF-36 and question 2 about how your health today compares to a year ago isn't actually included in any of the composite scores. Table 2 shows how to add them together, and it's the only question missing.

And I found a paper that says it's basically just used for validating the test. For people that say they improved over the past year, their total disability scores should on average be higher than those who didn't:

Garratt AM, Ruta DA, Abdalla MI, et al. SF 36 health survey questionnaire: II. Responsiveness to changes in health status in four common clinical conditions.. BMJ Quality & Safety 1994;3:186-192.

So the SF-36 has these eight domain scores:

What I'm thinking is to mainly use the PEM score in the post above, but since a lot of participants are tied, the SF-36 will just kind of act as a tie-breaker. I added up all the scores for the SF-36 domains above which gives a score around 200-500. I scaled them so that they are between 0 and 0.99. It's less than 1 because the PEM scores are separated by 1, and I don't want adding the SF-36 score to be able to change the order, only break ties. And since higher scores on SF-36 indicate better health/less disabled, I will subtract them from the PEM severity score. So people who are less disabled (higher SF-36 score) but same PEM severity will end up with a lower total "severity score".

So here is a plot of the PEM score and the transformation after subtracting the SF-36 score. The third column is after I converted them to equidistant ranked scores to visualize them better, since for Spearman correlation, only the order matters, not the spacing between them.


Amazingly, two participants are still tied. They had the same PEM and SF-36 scores (not all SF-36 domains were identical, but they happened to add up to the same number). I don't know if I'll just let them be tied or try to think of another tiebreaker.

Edit: If anyone has thoughts on this PEM + SF36 not being the optimal way to measure severity, I'd love to hear them. I don't want to create many sets of correlations from tweaking the severity score over and over as I think of improvements.

 

Hopefully this is the final product. I used the PEM - SF-36 metric as described in the last post, as well as in the Notes tab of the spreadsheet below.

I used Kendall's Tau test for correlations because I read that it is more robust to tied values (two participants having the same severity or lab result) than Spearman's rho. This may or may not be accurate, as I had trouble finding scientific papers specifically supporting or disputing this.

The thinking with the severity metric was first ranking by PEM score since that is probably the best available data for "severity" of ME/CFS. Since there are a lot of ties, I broke the ties based on their score on the SF-36 questionnaire, with the thinking that if they score worse on a general healthy survey, there is a good chance they have worse ME/CFS.

Still, they may have other comorbid conditions which are impacting their health or life which may cause lower scores on the SF-36 without necessarily meaning worse ME/CFS. It may be better to only use the PEM score, but there may be less power to find correlations as there are only 6 unique scores split across 17 participants, i.e. there are many ties.

Description of SF-36

All 8 of these subscales were added together. It may be better to only use some of these subscales.

A red/positive correlation indicates that as severity increases, the given test's value also increases. A blue/negative correlation indicates that as severity increases, the test's value decreases.

Link to browser version


Edit: I decided to add in a "pure" PEM score, since I want the correlations to be as "clean" as possible. It's the second tab of the spreadsheet. It's only based on the 8 yes or no questions mentioned that relate to PEM.

 

The PEM - SF-36 severity metric seems to correlate to physical fatigue from the MFI-20:


Not as much for mental fatigue:




Again weirdly one of the wrist movement counts after CPET was one of the most correlated, with more severe PEM meaning higher wrist movements. This time 27-43 hours post-CPET:


But compared to the healthy group, the ME/CFS group as a whole looks slightly lower, though it's not significant.

 

I decided to just rerun it only based on the PEM score. I want the score to match ME/CFS as well as possible, and I can't be sure the SF-36 is doing that. I added the new correlations as the second tab in the spreadsheet above.

One thing I've noticed so far is that pretty much everything from the "Seahorse Mitochondrial Function Baseline" dataset is negatively correlated with PEM severity.


Here are the descriptions of these tests:

I know pretty much nothing about these tests, so I hope someone else can chip in. But the only one that sounds "bad" to me is "Proton Leak", and that's the only one that is higher with worse PEM. The rest that sound like things that would be good, like "Spare Respiratory Capacity" are lower with worse PEM.

And then here we are with the same tests before and after a CPET. I guess they did a baseline twice? Because the baseline values in the CPET file don't match the values in the "baseline" file above.


Pretty much the same ones are correlated negatively or positively here.

What I'm noticing is that tests at baseline, 72 hour, and 48 hour seem to be the most correlated, but most of the tests at 24 hours are barely correlated. Maybe some temporary improvement immediately after exercise? Potentially a marker of the "adrenaline" effect?

ATP production has a negative correlation of at least -0.24 at all time points, including the other baseline value in the table above. Maybe this is a more stable marker of ME/CFS?

 

Looking at metrics between groups, it's definitely not that straightforward. While those with worse PEM tend to have lower ATP production at baseline, healthy people's ATP production seems to be lower than ME/CFS.


I quickly looked at the rest of the comparisons. ATP production is either the same or higher in the ME/CFS group at all time points. Pretty much all of the tests are no difference or the opposite of what I would expect if basing it on the correlations with PEM.

I've double, triple checked I did everything right. This plot of ATP from the baseline file is about as raw from the downloaded data files as I can get it:

It does look like a somewhat negative correlation where ATP production gets lower as PEM increases in the ME/CFS group, and yet the healthy group is lower on average, even though they have zero PEM.

Here's another, Spare Respiratory Capacity (%), which was most correlated in the first table above. Same pattern:


So it'll either take something quite unintuitive to explain this or it's nothing.

Edit: This makes me wonder. What if I include healthy controls and set all their PEM scores to zero and run the correlations with everyone?

I guess the way I have it now might be better because I have a "validation set", to see if the top correlations from the only ME/CFS group extend to healthy controls. Which was useful here for providing contrary evidence for mitochondrial function.

But maybe just use half the healthy controls in the correlations and half as validation?

 

I thought including half the healthy controls was a decent enough idea so I went ahead and did that. Everything the same, except half the controls are tied at a PEM score of zero and the other half are excluded for later validation.

Browser link


First impressions: Now all the top correlations are surveys. SF-36 questions, physical fatigue scale, etc. Which makes sense. Those surveys separate the two groups best, and some questions are bound to correlate with PEM.

Taking a peak at how the top correlation looks:

I added a bit of "jitter" to the dots. It just moves them a tiny bit at random so they aren't overlapping, since so many of them have the exact same value.

So that's question 15 of the SF-36:

The top 69 correlations are all surveys.

There's a few things from the heart monitor files near the top: "Frequency domain measures of heart rate variability collected using a 24 hour Holter monitor." But there are at least 631 measures related to HRV, so some are bound to be highly correlated.

Next category I see going down is CSF metabolomics. First one is at row 109. (You'll have to follow the browser link to see row numbers.) The metabolite is X-22162. Whatever that is.

But in any case, I'm taking a position of ignoring everything from the CSF metabolomics portion for now, for the reasons I gave in a previous post:

I think there was some sort of technical artifact making all metabolites downregulated in ME/CFS. Whether during sampling, during the actual lab measurements, maybe just something like all ME/CFS were lying down and all HV were sitting up during testing. Not sure, but artifact seems most likely. And if it's not an artifact, then all encompassing CSF metabolite downregulation is potentially a big finding. But I think that's the less likely option.

Next category I see is something from the blood labs: MCV (fL) (mean corpuscular volume or average size of red blood cells)

(The spacing on the x axis is arbitrary. The red dots could be a million units farther right and it'd be the same correlation since the spacing doesn't matter for Kendall's tau and there's no way for me to say how much "PEM severity" is between any two participants or between the groups.)

It looks interesting.

Next a couple things from the lipidomics study. A triglyceride and a diglyceride. (positive correlation)

Then I see something from the "Free living accelerometry" study where they wore an activity monitor at home for at least five days and during the exercise test: (Hip Moderate [2020 - 5998 cnts; 3-5.9 METs] Time (min) negative)

Peak VO2 during CPET is at row 179. (negative)

More lipidomics. (all positive)

Another from the accelerometry: (Hip Avg Wear Time METs, negative)

Negative for a hand grip metric.

There's something from CSF flow cytometry at row 277: CD4+ T cell subset Memory (%) (positive)
This CSF study doesn't seem skewed like CSF metabolomics. In this one, 49% of tests are higher in ME/CFS, 49% are lower, 2% the same. Seems more realistic.

Something from CSF catecholamine study at 283: concentration of DOPA
In this study, it's 8 catecholamines. Mostly lower in ME/CFS.

Next thing that looks interesting to me at 384: Lymphocyte NK CD56dim (cells/ul) (negative)

Oh fun, we got a stool metabolite at 472: Xylose (negative)

At 506, from clinical master labs: Triglycerides (mg/dL) (I assume in blood) (positive)

Another from CSF flow cytometry at 530: Lymphocyte NK cell (cells/ul) (negative)

At 544 from tilt catecholamine study: Plasma concentration of dopamine at the end of head-up tilt, in pg/mL (negative) Highly correlated at -0.72 but there's only data for 7 participants. Just to see what the correlations look like at row 544:


For reference, these are sorted by p value, and that last one at 544 I mentioned has an uncorrected p value of 0.02. And there are about 3300 total tests, though many are correlated to each other.

So yeah, might be something interesting in there. Since I have the other 11 healthy controls that weren't included I can eventually test to see if any of these correlations hold up.

Edit: I thought about it more and realized my logic wasn't logicing. I can't truly validate with only healthy controls in the validation set. I can do like a "half validation" by replacing the controls and checking the correlation, but it's possible that random variation in the ME/CFS group caused a non-real effect, and I can't check that. Should have done half of both groups from the start.

 

Okay, hopefully I'm done. Here's spreadsheet_final_v2(1)_20. (relevant comic)

I split the data into two datasets. Details of how I split are in the Notes tab. I tested correlations for each split and the full dataset. The correlations for the full dataset are the first tab, and the others are the next two.

On the first tab, I marked which tests had a p<0.15 in all three datasets. (0.15 was just an arbitrary cutoff I chose to get a decent number of tests past the cutoff.) It's the first column. It shows the higher of the two p values for that test from the two splits. Significance is also indicated by the color. The more green, the more significant.

I further colorcoded the tests that met this significance cutoff in the Lab Test column based on if they are a subjective survey or an objective test. Bluish color for survey and purple for objective. So the most interesting ones should be the purple tests in the first tab. And the greener the first column, the more significant in both splits.

Browser link


So what do we have that is an objective test that reaches p<0.15 in both data splits?

178 total tests

2 from accelerometry:

2 from microbiome "Alpha Diversity" file:

I don't know what these mean.

1 from before and after CPET datafile:

1 from blood flow cytometry:

1 from body composition:

7 from CPET:

2 from CSF catecholamine (but really just 1 since it's just different units):

4 from EEfRT:

1 from food records:

70 from the 5 heart rate variability files (some of these you'll have to look at the spreadsheet to see if it's from the LF, HF, or the other files):

14 from lipidomics (all increased with worse PEM except LPC):

69 from CSF metabolomics:

That's 15% of all tests in CSF metabolomics. Very over-represented compared to all the other studies here. All but one (X-23195) lower with worse PEM. Which, hey, if all ME/CFS tests were shifted down for some reason, and this X thing is still higher, then maybe that means it's very significant.

1 lower in stool metabolites:

3 lower in tilt catecholamine

Minutes is a measure of angle here.

--------

So yeah, those are the ones I'd say are maybe interesting just based on this study alone. But I think the first tab can be useful for comparing other study results. The first tab has correlations from the whole dataset, so it has the most power to find correlations. The correlation/significance data there can be used to compare to tests from other studies, even if the data splits didn't necessarily show high significance for the tests in question. The splits have very few participants, so it's likely some correlations were missed in those due to random variation.

 

ATP is also a key signalling molecule .
So if something is going awry, it's function can alter .
Most molecules multi task , but we tend to focus on their primary association

 

Since low fecal hypoxanthine was flagged above, I did a search.

Plots of fecal hypoxanthine from the NIH study:


Plots of hypoxanthine vs other metabolites mentioned in other studies below as lower. Propionate, uracil, and butyrate look like they might be positively correlated with hypoxanthine.


Also, one of the studies found higher fecal taurine in ankylosing spondylitis and rheumatoid arthritis. The plot for taurine in the NIH study is interesting. Everyone along the bottom has zero, there was just jitter added to the dots. Only three had taurine. Notice the dot in the top right corner who had the max possible PEM score.


Fecal hypoxanthine

The faecal metabolome in COVID-19 patients is altered and associated with clinical features and gut microbes, Lv et al, 2021

Longitudinal Multi-omics Reveals Subset-Specific Mechanisms Underlying Irritable Bowel Syndrome, 2020, Mars et al

Untargeted Metabolomics of Feces Reveals Diagnostic and Prognostic Biomarkers for Active Tuberculosis and Latent Tuberculosis Infection: Potential Application for Precise and Non-Invasive Identification, 2023, Luo et al

Characterizing the metabolomic signature of attention-deficit hyperactivity disorder in twins, 2023, Swann et al

Characterization of ankylosing spondylitis and rheumatoid arthritis using 1H NMR-based metabolomics of human fecal extracts, 2016, Shao et al

Blood and urine hypoxanthine

Untargeted Metabolomics and Quantitative Analysis of Tryptophan Metabolites in [ME] Patients and Healthy Volunteers, 2024, Abujrais+

Metabolic profiling reveals anomalous energy metabolism and oxidative stress pathways in CFS, 2015, Armstrong et al

Development of an expert system for the classification of myalgic encephalomyelitis/chronic fatigue syndrome 2024 Yagin et al

Post-Exertional Malaise Is Associated with Hypermetabolism, Hypoacetylation and Purine Metabolism Deregulation in ME/CFS Cases, 2019, McGregor et al

Purinergic signaling elements are correlated with coagulation players in peripheral blood and leukocyte samples from COVID-19 patients, 2022, Schultz et al

High-Intensity Interval Training Decreases Resting Urinary Hypoxanthine Concentration in Young Active Men—A Metabolomic Approach, 2019, Kistner et al

 

So low hypoxanthine in the stool has been seen in many quite different health conditions (acute COVID, TB, ADHD, IBS, ankylosing spondylitis, RA). One more if you include ME/CFS. So maybe it is a marker of poor health in general. The plot of hypoxanthine in the two groups of this study above shows much more variability in the healthy controls, overlapping with ME/CFS on the low end, but also going much higher. Maybe this indicates that there is a spectrum of healthiness in the controls, with some dealing with other health problems and some fairly healthy, but most or all ME/CFS would be classified as "unhealthy" based on this marker.

Edit: "Unhealthy" could just mean some specific subset of health conditions, not all.

Edit 2: Updated disease list.

 

The deep phenotyping study has the stool metabolomics statistics in Supplementary File 21. These are sorted by significance.



Tyrosine, phenylacetate, and threonine are the other ones they found were significant. (All increased according to the spreadsheet above.)

Edit: Unclear if they did multiple test correction, but probably not since I got similar p values before correction.

 

I don’t have the energy to follow but thank you for doing these literature searches @forestglip it is so important.

 

Thanks for your deep diving into all of those data @forestglip.
Could hypoxia be a common factor in all those hypoxanthine findings?

 

No idea. Hard to think of what could cause the same thing in so many conditions. My mind usually goes to confounders like less physical activity in many diseases, but I don't typically think of people with ADHD as being less active. (Edit: maybe poor diet both causing low hypoxanthine and being a risk factor for or result of many conditions?)

I haven't yet found any studies in humans where fecal hypoxanthine wasn't lower in the unhealthy group. (All the ones I found are listed above, and I'll add to the post if I find more.)