“AI has literally changed my life, being an ME/CFS patient for many years and now symptom-free since 2015. AI has identified well before several findings that have been identified at a later time by conventional research”. Efthymios Kalafatis
After his studies in computer science Efthymios Kalafatis got an MSc in knowledge extraction. He’s been working since 2000 in data mining and later machine learning and natural language processing consulting. He has a 2018 patent named “Machine Learning, Natural Language Processing, And Network Analysis-guided Discovery Related To Medical Research“.
In 2022 he was a co-author of a paper examining LongCOVID in Greece and is the principal investigator of a Ramsay Award grant from the Solve ME/CFS Initiative that will use data from the You+ME Patient Registry to identify symptom, treatment, activity, and life event patterns in ME/CFS and long COVID.
His onset of ME/CFS in 2004 began a long decline in his health. In 2012 he used machine learning and other tools to associate his personal factors (food, supplements, activity, etc.) with his good days and bad days. The lessons he learned from that experiment allowed him to stabilize his health.
In 2013, he began using artificial intelligence methods to attempt to make some headway in the ME/CFS. In part because ME/CFS is under-studied, his idea was to capitalize on the massive amount of data concerning diseases closely related to ME/CFS in order to learn more about ME/CFS.
To that end, he started by collecting via software the abstracts of all the published research regarding the symptoms found in ME/CFS such as fatigue, orthostatic intolerance, depression, anxiety, tinnitus, etc. He also collected all published research related to ME/CFS, fibromyalgia, Gulf War illness, post-treatment Lyme disease syndrome, post-finasteride syndrome, and post-Accutane Syndrome. (Efthymios chose those diseases because he read about many cases of patients of these syndromes ending up having ME/CFS).
The next step was to identify -via text analytics- the most frequently occurring potential biological factors such as genes, biological pathways (e.g. mevalonate pathway), and other biological entities (e.g. mitochondria) that existed within these abstracts but also adding medical concepts as these were identified to be important year after year by researchers.
That list of concepts prompted another search that resulted in published research for each of these topics. That resulted in millions of scientific abstracts being collected which resulted in the next text analysis finding associations between genes, biological pathways, and symptoms.
The first analysis involved a type of machine learning analysis called classification analysis during which all relevant symptoms were grouped. Machine learning algorithms were then used to identify potential key biological factors associated with those symptoms.
By 2014 the first AI results were out: they identified endoplasmic reticulum (ER) stress, oxidative stress and many liver-associated factors as important targets. The results were ahead of their time as ER stress was just recently identified by NIH researchers as a potentially important contributor to ME/CFS (https://pubmed.ncbi.nlm.nih.gov/37579159/).
Using the results from the artificial intelligence study Efthymios was able to, after 11 years, regain his health. See his recovery story here.
At the end of 2015, Efthymios sent an email to ME/CFS researchers proposing that liver injury (that does not always show up in tests) in conjunction with genetic polymorphisms (slight changes in genes) involved in a variety of areas (cytochrome P450 pathway, methylation, glucuronidation, choline metabolism, bile acid synthesis and enterohepatic circulation, oxidative stress and endoplasmic reticulum stress) resulted in uncontrolled endoplasmic reticulum and oxidative stress and the unfolded protein response.
A new development in artificial intelligence in 2015 that allowed any possible biological pathway, gene, or compound to be given a relevance score enabled Efthymios to take the next step. Now AI programs are better able to pluck out the key factors in a disease.
In 2017, Efthymios added an additional analytical method named “Network Analysis” which researchers such as Nancy Klimas and Derya Unutmaz have used in ME/CFS. His goal was to create a knowledge map that highlighted prominent factors in ME/CFS. He succeeded. His findings (which have not been published) are below.
The tightly correlated findings with their clear center that resulted (see below) suggested he was on the right track. Contrary to those who believe ME/CFS is a heterogeneous mess, Efthymios’ analysis suggested that core areas of high biological interest – or hotspots exist. (Red denotes highly important topics, orange less important, blue lower importance).
The Liver Shows Up Big Time
Of the five “hottest” or reddest areas two (hepatotoxicity -liver damage), steatohepatitis (fatty liver disease) involved the liver. The others involved oxidative stress – long known to be present in ME/CFS, vitamin K (blood coagulation, calcium metabolism), the urea cycle (ammonia – a possibly big player in ME/CFS), fmo3 – a liver enzyme which can be found in the endoplasmic reticulum and is responsible for the metabolism of several xenobiotics, and LXR (a liver receptor) which has been recently (August 2023) identified in a paper by Maureen Hanson’s group and was found downregulated in COVID19.
These results suggested that liver injury caused by viral infections, medications, or other environmental factors provides a link between syndromes like ME/CFS, fibromyalgia, Gulf War Illness, post-treatment Lyme disease syndrome, etc. It also proposed that concepts related to oxidative stress, the liver, bile acids (CYP7B1, CYP27A1), peroxisomal function, mitochondrial biogenesis, detoxification, phospholipid metabolism (choline_deficiency) had to be looked at.
I was surprised at how often the liver showed up and asked Efthymios about it. He said “I did not know what to expect really. I was surprised though that it was showing over and over again as new information was being fed into the AI system as years passed by. And I believe some other researchers may now be getting similar signals (note that normal liver enzymes do not rule out liver disease!)”
A Liver ME/CFS Connection
It turns out that the liver has shown up quite a bit recently in both ME/CFS and long COVID. The urea cycle which converts ammonia to urea in the liver cells, in particular the mitochondria of liver cells, has shown up in at least three metabolomic studies including two by the Hanson group as well as a gene expression study.
Germain’s 2018 “Metabolic profiling of a myalgic encephalomyelitis/chronic fatigue syndrome discovery cohort reveals disturbances in fatty acid and lipid metabolism” study found reductions in primary bile acids that were suggestive of liver damage. Plus the pathway with the highest “impact factor” was taurine – which is part of the primary bile acid biosynthesis pathway. The authors noted that lower levels of taurine – which was part of Efthymios’s recovery protocol – have been found in ME/CFS and that the supplement is used by many ME/CFS patients. Plus one of the most impactful pathways found – the glycerophospholipid metabolism – is dependent upon proper liver functioning.
The authors even went so far as to suggest that a panel for hepatotoxicity prediction might constitute a serum metabolic signature for ME/CFS. (The paper did note that the reductions could reflect toxicity from the use of prescription or over-the-counter drugs (high doses of acetaminophen/Tylenol can impact the liver)
In April 2022 the Morten Group at Oxford asked “Could an upset liver be a key clue in ME/CFS?” and stated their metabolic screen “found a consistent liver issue in mild, mod, and severe patients”. Chris Armstrong of the Open Medicine Foundation is currently engaged in a project to understand what’s going on with ammonia metabolism in ME/CFS. It turns out that when a dysfunctional liver fails to break down nitrogen, excess ammonia can produce neuroinflammation and encephalopathy. It was the attempt to understand what’s happening with ammonia that led to the itaconate shunt hypothesis. (A blog on the hypothesis is coming up.)
A Liver Long COVID Connection
With regard to long COVID a presentation at the Keystone International Long COVID conference proposed that liver dysregulation in long COVID may be, among other things, whacking the mitochondria. A recent conference report also found increased levels of “liver stiffness” months after the infection. The lead researcher Dr Firouzeh Heidari at Massachusetts General Hospital, said “Our study is part of emerging evidence that Covid-19 infection may lead to liver injury that lasts well after the acute illness.”
Endoplasmic Reticulum Stress and the Unfolded Protein Response
Efthymios’s model also highlighted endoplasmic reticulum (ER) stress and problems with the unfolded protein response (UPR). Proteins are folded into their correct shapes (shape is everything with proteins) in the endoplasmic reticulum (ER). Problems with the ER can cause proteins such as amyloids to gum up the works in the cell. An accumulation of misfolded proteins triggers the “unfolded protein response (UPR)” which then should degrade them. It can also trigger “autophagy” a process whereby the cell uses “autophagosomes” to digest the strangely shaped proteins.
Cells exposed to stressors like viral infections often trigger the unfolded protein response (UPR) and autophagy. Two studies, thus far suggest that problems cleaning up these misfolded proteins may exist in ME/CFS. Baraniuk’s 2005 paper found evidence of amyloid proteins in the cerebral spinal fluid of ME/CFS patients and evidence of impaired autophagy recently showed up in ME/CFS.
The recent WASF3 study highlighted the role that damaged endoplasmic reticulum may play in ME/CFS. Time will tell what role ER stress and the unfolded protein response plays in ME/CFS but they are another example of new findings (like microclots) that potentially open a wide variety of new treatment approaches.
Brian Vastag reported that the lead author is trying to get a clinical trial of an amyotrophic lateral sclerosis drug, Relyvrio, underway in ME/CFS and many different drugs (Dextromethorphan, Bromocriptine, Dantrolene, Verapamil, diltiazem, thalidomide, lenalidomide, pomalidomide, tamoxifen) can impact it.
Xanthine oxidase was another hotspot that popped up in Efthymios’s AI analysis. Xanthine oxidase, an enzyme, catalyzes the oxidation of hypoxanthine to xanthine. In his 2019 ME/CFS metabolomic study, McGregor found that hypoxanthine was the principal metabolite associated with post-exertional malaise. McGregor believed that the hypoxanthine increase was associated with the inhibition of protein synthesis in the muscles during exertion. His study suggested that ME/CFS patients’ cells were breaking down amino acids – a poor energy source – to fuel their mitochondria.
AI-Driven Approach Predicts Future Findings
Only time will tell what role the liver, the endoplasmic reticulum, etc. play in this disease but it is notable that Efthymios’s artificial intelligence/network analysis approach was able to pluck out potentially key themes (liver, endoplasmic stress, peroxisomal functioning, phospholipid metabolism, and xanthine oxidase metabolism) which showed up, in some cases, in ME/CFS study results years later.
Note that the clues were already there – buried in research papers on ME/CFS and allied diseases. They simply needed to be unearthed using artificial intelligence methods.
Note that Efthymios’ project took him up to 2017; i.e. six years of data have been produced since then. One wonders what new clues a 2023 AI project might produce given what we’ve learned so far, the emergence of long COVID, and AI’s apparently rapidly accelerating capabilities.
- About ten years ago, Efthymios Kalafatis, a data scientist, used artificial intelligence (AI) to plow through millions of documents and pluck out potentially key hotspots in ME/CFS and related diseases.
- He also developed a treatment protocol that allowed him to recover from ME/CFS. Efthymios stated “AI has literally changed my life, being an ME/CFS patient for many years and now symptom-free since 2015.”
- As AI methods developed over time Efthymios first used machine learning and then network analysis to sharpen his results. The results suggested that instead of being heterogenous illnesses that core hotspots existed not just in ME/CFS but in allied diseases as well.
- The liver, endoplasmic reticulum stress, the misfolded protein response, peroxisomal dysfunction, lipid issues were all highlighted. Most of these were not thought to play a role in ME/CFS at the time but recent studies suggest that all could play a core role.
- The liver, in particular, showed up. Efthymios’s analysis suggested that liver injury caused by viral infections, medications or other environmental factors provided a link between syndromes like ME/CFS, fibromyalgia, Gulf War Illness, post-treatment Lyme disease syndrome, etc.
- Since Efthymios’s analysis (which was never published) several studies have highlighted the liver’s possible role in ME/CFS. In April 2022 the Morten Group at Oxford asked “Could an upset liver be a key clue in ME/CFS?” and stated their metabolic screen “found a consistent liver issue in mild, mod, and severe patients”.
- Recently stressed-out endoplasmic reticulum – the organelles that fold proteins to the correct shape in our cells – were highlighted in an NIH-funded ME/CFS study. Similarly problems with lipid metabolism, the urea cycle and xanthine oxidase have popped up since Efthymios’s analysis in 2017.
- Only time will tell what role the liver, the endoplasmic reticulum, etc. play in this disease but it’s notable that Efthymios’s artificial intelligence/network analysis approach was able to pluck out potentially key themes (liver, endoplasmic stress, peroxisomal functioning, phospholipid metabolism, and xanthine oxidase metabolism) which showed up, in some cases, in ME/CFS study results years later.
- Note that the clues were already there – buried in research papers on ME/CFS and allied diseases. They simply needed to be unearthed using artificial intelligence methods.
- AI is limited in several ways. It can work through huge amounts of data more quickly and accurately than any human can but it cannot explain how it gets the results it does and cannot go beyond the data given it; i.e. it can help explain the past but cannot predict the future.
- Efthymios’s analysis was done in 2017 – six long years ago. One wonders what a 2023 analysis using the data generated since then in ME/CFS, long COVID, and other diseases would uncover and how it might help shorten the search for the answers to ME/CFS and allied diseases.
In 2018 Efthymios presented his AI-assisted research in a talk “Machine Learning and Network Analysis Guided Medical Research” at the Euromene Conference and was granted a patent application “Machine Learning, Natural Language Processing And Network Analysis-guided Discovery Related To Medical Research“.
The goal of natural language processing is to create a program or computer that is capable of fully “understanding” the contents of documents and can “accurately extract information and insights” from them. Being able to feed huge amounts of data from many studies into a computer, find associations and have the computer extract novel insights from them could obviously provide a major step forward in our understanding of disease.
One of Efthymios ‘ goals with his patented methodology is to use AI algorithms and Network Analysis to produce hypotheses on how ME/CFS is related to other diseases. Another goal is to create an ordered/ranked list of high-priority biological topics, symptoms, and environmental, or nutritional factors – which he has done.
It was fascinating stuff and with artificial intelligence suddenly becoming the talk of the town in 2023 I wanted to pick Efthymios’ brain about the possibilities it raises for understanding and finding treatments for ME/CFS and allied diseases. He said :
“AI has literally changed my life, being an ME/CFS patient for many years and now symptom-free since 2015. AI has identified well before several findings that have been identified at a later time by conventional research”.
How is AI different from, say, the sophisticated statistical programs used to assess complex study results? What makes it different from other data-driven approaches?
“First, some definitions: AI is a technology that attempts to achieve or even exceed the level of human intelligence. Machine Learning (ML) is a subset of AI where the computer “learns” from data given a specific problem (i.e. An algorithm “learns” from the data being presented and automatically identifies subgroups (clusters) of patients according to a number of metabolites).
ML has been heavily relying on statistical methods -no question about that- and there is this huge debate as to whether ML is no different than statistics. In statistics, we have a “tighter” set of rules to analyze and draw inferences from data”
AI has been so in the news lately – 2023 is truly the year that AI made it into the public consciousness. We know that AI can perform complex calculations and analyses faster and more accurately than humans, recognize patterns in large amounts of data that humans would never be able to pick up, etc. I wanted to know, though, what AI cannot do
In something called the “explainability problem,” Efthymios stated that AI is something of a black box in a way. It cannot explain to us why it’s drawing the conclusions it’s drawing. For instance, it cannot say why it asserts that “x” research target should be studied to explain “y” symptom. A technology named XAI (Explainable AI will eventually change this)
AI also cannot go beyond the data it’s given; i.e. if the data on say a gene – is not included in its program – it can never logically infer that manipulating that gene might be helpful. It can only say things about the data it already has.
We’ve been seeing studies using machine learning and other AI techniques for quite some time but this is the year that artificial intelligence showed up big time publicly. Why is that? Has something recently happened to move the field forward in a dramatic way?
I think that Chat-GPT’s widespread use is something to consider here. It showed everyone the immense potential of AI but also that we must proceed with caution”
How are researchers using artificial intelligence right now to understand diseases? Have any breakthroughs been produced?
“Researchers rely more and more on machine learning methods for diagnostic purposes mainly, for example identifying factors that differentiate patients versus controls.
A breakthrough here would be to be able to provide to an AI system like ChatGPT with everything we know about symptoms, metabolites, and any other ME/CFS findings and have it suggest the possible causes (I would happily contribute to implementing this!)
Another breakthrough would be for an AI system to automatically identify patient subtypes and indicate factors as to why specific interventions are extremely helpful to some patients and not helpful to others.”
You stated that software -using non-AI methods- can help us connect various research findings from ME/CFS research within seconds and speed up the knowledge extraction process. This new knowledge could then be forwarded to relevant research teams, enabling them to see the “bigger picture” of ME/CFS and guide resources to the most promising research targets.
How would this work and what would it do? If you were to embark on an AI-driven ME/CFS project right now what would you do differently?
“We see many theories and research teams working to identify key mechanisms of ME/CFS. Techniques like Information Retrieval, Natural Language Processing, and Network Analysis can help us identify – with unprecedented speed – which medical concepts may connect these theories with existing findings and other syndromes such as Ehlers-Danlos Syndrome, fibromyalgia, post-treatment Lyme Disease, long COVID, etc.
If I did a project now I would try to capture patient historical data as I believe that important insights may be found in conditions or events that occurred prior to the onset of ME/CFS. LongCOVID would also be included as it appears that many of its symptoms match those of ME/CFS, including post-exertional malaise. Causal AI would also be of interest to try to predict how different interventions change the outcome and why.
“Regarding any hidden connections: Post-Translational modifications -such as Ubiquitination-, apoptotic cell clearance mechanisms, and finally the potential role of CH25H (Cholesterol-25-Hydroxylase) in viral persistence mechanisms should be further investigated according to the methods I have been using”
(Note – “Causal machine learning (CML)… provides a complete toolset for investigating how a system would react to an intervention (e.g.\ outcome given a treatment).”)
Interested in the Role AI Could Play in Understanding and Treating ME/CFS and Allied diseases?
Please Support Health Rising and Keep the Blogs Coming
HEALTH RISING IS NOT A 501 (c) 3 NON-PROFIT