Dear All, Over the past 6 months, I had the opportunity to speak with a number of researchers on ME/CFS and share my research work and methods on the use of Artificial Intelligence and Network Analysis to identify promising research targets. With the introduction of Large Language Models (LLMs), it became even clearer to me that we need a resource where we should collect what is known about ME/CFS : Which genes, metabolites, biological pathways appear to be more relevant? Which of these findings have been replicated? Which findings failed to be replicated? Without such resource, I fear that any use of AI will lead to incoherent results and therefore I wanted to ask if a website / database or other knowledge base exists which assimilates and categorises such findings from current and previous ME/CFS research. If such resource does not exist, I would welcome any thoughts on what is the best way to proceed in order to create it.
It isn’t very upto date, and there is some more dodgy stuff in there, but https://me-pedia.org is probably the closest to what your looking for. It also has an API. Otherwise, there are databases of ME research. Such as this one: https://www.meresearch.org.uk/research/other-resources/research-database/ Sorry if I’m not correctly understanding your question though.
Indepently of the use of AI this seems to be a good idea to keep track of replicated and non-replicated findings over the years and decades. Would be awesome if one could get one of the organisations involved.
Have you seen the mapMECFS project, @mariovitali? It describes itself as "an interactive data portal providing access to research results across many biological disciplines from studies that are focused on advancing our understanding of Myalgic Encephalomyelitis / Chronic Fatigue Syndrome (ME/CFS). The goal of mapMECFS is to enable exploration of complex experimental results and connect datasets to offer new opportunities for discovery for ME/CFS researchers." I happened upon it yesterday and thought that it looked useful. I've created an account today but haven't yet had a chance to look at the various datasets that have been uploaded.
Sam, Thank you for letting me know about this resource. From what I understand, to access information to this portal you need to be part of an academic or otherwise research institution (which I am not).
WeCrunchME may also be useful. Although the website seems to be down for me. “a volunteer project, mapping out the research, clinical, tech, and advocacy ecosystem around ME/CFS, Long Covid, and other infection-associated chronic conditions. Our aim is to help provide an information base to new and existing members, and so accelerate progress towards treatments and cures.”
This is over simplistic, but I think the answer is something like this: Replicated: – about 2/3 of people with ME report they had an infectious onset. – infectious mononucleosis leads to me in a significant minority of people who develop it in their teens/20s. – There is a reduced ventilatory threshold on day two of a two-day maximal exercise test. I have probably missed a few things, but that’s pretty much it. Not replicated: Everything else. For things like metabolites it’s more complicated and then they will be common findings. For genetics, it’s not;,nothing has yet been replicated, though DecodeME will report soonish and may change that. I also think you’re right that without good data, AI isn’t going to give us any useful insights. The fundamental problem is the state of research (which I’ve been following for 30 years). There’s been no attempt to pursue hypothesis to clear answers. Instead, promising avenues are simply abandoned without the publication of any further negative or positive results. I have no doubt that ME is a very hard problem to solve. But a bluebottle approach to research problems doesn’t help. PS for non-replication, try rituximab. The initial promising findings were ruled out by a more thorough later trial. Although this is a clinical finding, it also implies that B-cells are less promising of research.
Yes, you could well be right about this. I have no institutional affiliation and my registration is currently pending, so we'll see if they allow me, a layperson, access to the datasets they have.
I put "unaffiliated" under institution and gave a brief description of what I'd hope to use the data for when I applied a couple months ago, and my account was approved. I can't remember how long it took to be approved. Edit: I remembered I can check my emails for dates. It took four days to be approved.
@mariovitali Assuming you are considering running an LLM over a pre-specified set of high quality published journal articles? I've just started looking at LLMs running on my laptop, using pre-built models eg the Ollama models. I find that the 7B models run very smoothly while the 70B models are definitely slower but still useable. The answers look correct and convincing, but it can go well off course. It seems to completely fabricate journal references, but if you ignore that the underlying information I guess can represent a pointer. I don't think the LLMs are anywhere near ready to drive the direction of research yet. Ground truth needs things to be 100% accurate so we don't appear to be there yet with LLMs. Human researchers (the good ones) coming up with hypotheses and testing them still outperform the current state of the art of AI. As an example I asked meta's llama3-70B (the 70B means 70 billion tokens) which is a general not scientific model: "describe the rituximab studies in me/cfs by fluge and mella". Response — Let's dive into the details of the two open-label studies on rituximab in Myalgic Encephalomyelitis/Chronic Fatigue Syndrome (ME/CFS) conducted by Fluge and Mella: Fluge et al. (2011) Study design: Open-label, single-arm study Patient population: 15 patients with ME/CFS (Fukuda criteria) Rituximab regimen: 2 infusions of rituximab (500 mg/m²) 2 weeks apart Follow-up infusions every 6 months for up to 2 years Primary outcome measures: Fatigue severity (Fatigue Questionnaire, FQ) Pain severity (Visual Analog Scale, VAS) Cognitive function (Symbol Digit Modalities Test, SDMT) Results: Significant improvement in fatigue (FQ) at 3 months (p = 0.001) and 6 months (p < 0.001) compared to baseline Reduced pain severity (VAS) at 3 months (p = 0.01) and 6 months (p = 0.003) Improved cognitive function (SDMT) at 3 months (p = 0.02) and 6 months (p = 0.01) Adverse events: Mild to moderate infusion reactions in 4 patients No serious adverse events reported Mella et al. (2013) Study design: Open-label, single-arm study Patient population: 30 patients with ME/CFS (Fukuda criteria) Rituximab regimen: 2 infusions of rituximab (500 mg/m²) 2 weeks apart Follow-up infusions every 6 months for up to 1 year Primary outcome measures: Fatigue severity (Fatigue Questionnaire, FQ) Quality of life (Short Form-36, SF-36) Pain severity (Visual Analog Scale, VAS) Results: Significant improvement in fatigue (FQ) at 3 months (p < 0.001) and 6 months (p < 0.001) compared to baseline Improved quality of life (SF-36) at 3 months (p = 0.002) and 6 months (p = 0.004) Reduced pain severity (VAS) at 3 months (p = 0.01) and 6 months (p = 0.005) Adverse events: Mild to moderate infusion reactions in 5 patients No serious adverse events reported Common findings: Both studies reported significant improvements in fatigue, pain, and quality of life measures, as well as improved cognitive function, following rituximab treatment. The results suggest that rituximab may have a positive effect on ME/CFS symptoms, particularly fatigue and pain. Limitations: As open-label studies, these trials lacked a control group, which limits the ability to draw causal conclusions about the efficacy of rituximab in ME/CFS. Additionally, the small sample sizes and short follow-up periods may not be representative of the broader ME/CFS population or long-term treatment outcomes. These pioneering studies laid the groundwork for further research into the potential benefits and risks of rituximab in ME/CFS. So I then asked: "give correct references to the rituximab studies in me/cfs by fluge and mella". Response — Here are the correct references to the Rituximab studies in Myalgic Encephalomyelitis/Chronic Fatigue Syndrome (ME/CFS) conducted by Fluge and Mella: Fluge et al. (2011) Reference: Fluge Ø, Bruland O, Risa K, Storstein A, Kyte H, Strand EB, et al. Benefit from B-lymphocyte depletion using the anti-CD20 antibody rituximab in chronic fatigue syndrome. A pilot study. Biochem Pharmacol. 2011 Oct 15;82(8):715-23. doi: 10.1016/j.bcp.2011.07.003. Mella et al. (2013) Reference: Mella O, Fluge Ø, Risa K, Eide H, Wyller VB, Strand EB. Rituximab targets CD20+ B-cells in ME/CFS: a pilot study. PLoS One. 2013 Jul 23;8(7):e69793. doi: 10.1371/journal.pone.0069793.
The problem with consensus is that there will always be subjectivity and bias involved in determining it. Assessments of "relevance" are going to stem from agenda or other bias. There are things that have been studied heaps that are unlikely to go anywhere, but who is to say whether they go into the model or not? The model may be impartial but I don't see how the information going into it can be. You would need a large and well balanced consensus panel of people with different backgrounds, agendas, and whatever else, and they would have to review a mountain of work and also keep their collective assessment of the literature up to date. I don't know if this is feasible. This is a very good idea in principle though, the issue is just getting the input information right.
Jaime Seltzer updated the MEAction research summary in 2019. Maybe some useful information there. https://www.meaction.net/wp-content/uploads/2019/08/MEAction_2019_Research_Summary.pdf Original 2015 Summary http://www.meaction.net/wp-content/uploads/2015/05/ME2FCFS-RESEARCH-SUMMARY-Jamie-Seltzer.pdf
In my mind (right or wrong) one of the most replicated findings in ME and Long Covid is reduced cerebral blood flow. I've seen patient reports of using Clonidine and Guanfacine to try to address it with mixed reports, maybe the side effects due to effects on neurotransmitters and in the case of Clonidine, the short half life. Add also many reports of suspected low blood volume (but very very little real data on this) and patient response to saline / salt loading in the short term that could also affect cerebral blood flow, I think this is an area that could benefit from your tools. @MelbME also had thoughts on how PEM could possibly affect water balance and CO2 levels that could affect blood volume and CBF in other recent threads.
Among patients I think one of the most common themes is sensitivity to medications with recommendations to start low and go slow. What is tolerated by one person is not tolerated by another. I would be interested to know what the general literature states about this. Not hypothesis on MTHR and such, but actual facts. @mariovitali You are a bigger liver "pointer", and there is also MCAS, but what about other diseases. Is sensitivity to medications reported in other diseases, and if so, what diseases?
@SNT Gatchaman Thank you for your message, yes LLMs is a technology I started looking at and I have been trying to build a knowledge base in order to use it with them. However, it is not about LLMs only. During my discussions with Wenzhong Xiao's team I introduced an Information retrieval (IR) system that could be used to rank concepts based on a input query. If we have a "ground truth" that includes, symptoms, metabolites, genes and biological pathways then these can be fed to the IR system. An example is shown below where the IR system was used to rank differentiated genes found in 18 ME/CFS studies : So transferases, appear to get the highest relevance ranking. Obviously, if the input changes, so does the output. Needless to say I am not someone who is able to define the ground truth. It is also interesting to use different methods to see if some results converge. Even though Chat GPT-4o gave completely different results, let's see what Claude from Anthropic and its output suggests : So based on those two methods we see that Transferases are being selected (METTL3) and also dna methylation (appears in IR results although Claude selects a gene related to RNA methylation). This is a problem I always had : I am using the right input or not? And this is why I believe that trying to collect and list what we know can be useful. We can also assign certain weights according to the confidence of the finding and use it during analysis. @wigglethemouse Thank you and what you mention regarding medications is a key question. Again, during my discussions I stressed out that there should be screening on how medications under investigation are being metabolized when Randomized Control Trials are involved because of these sensitivites to medications.
