The GIST is coming

“It is startling that here we are in 2024, recognizing how critical our immune system is for protecting us against infections and major diseases…yet we have no informative clinical test to get at it.  We desperately need a means for assessing our immune system.” Eric Topol July 2024

 

“The paper provides a first proof-of-concept, a stepping stone, towards a full immunome.” Eric Topol  February 2025

Topol first wrote about the need for an immunome—a way to assess all immune system components together—last July. He noted other studies that were beginning to grasp the “immunogen”, but all involved techniques that were too expensive to be used in the clinic. Not so with this latest technique, which offers the potential of an inexpensive way to chart the immunome—hence Topol’s title, “The First Diagnostic Immunome”.

Topol called the study, “Disease diagnostics using machine learning of B cell and T cell receptor sequences,” which appeared in the Science journal, “extraordinary,” and said it was co-authored by a who’s who list of immunologists and rheumatologists.

The Mal-ID process

The study assessed the receptors or antennae found on T and B cells. Because these receptors track and need to keep up with every pathogen, vaccination, autoimmune response, and toxic exposure our cells are exposed to, they provide a living encyclopedia of our immune history.

To respond to the many factors that our immune systems are exposed to, though, these receptors must be extremely nimble. That means they quickly rearrange their genes, or mutate, in order keep up with the bombardment of insults to which our immune systems are exposed. That complexity has made sequencing them difficult – if not impossible – to assess until now. Not surprisingly, it’s artificial intelligence that’s making the impossible possible.

Topol noted that Sarah Teichmann, at the University of Cambridge, stated, “This is a one-shot sequencing approach that captures everything that your immune system has been exposed to.”

A T-cell.

The researchers, led by Maxim Zaslavsky, used a process called Mal-ID (Machine Learning for Immunological Diagnosis). The study, which included nearly 600 individuals with a variety of conditions (hospitalized with acute COVID, autoimmune Type 1 diabetes, HIV, 37 – a recent flu shot) and healthy controls, sequenced the two parts of B and T-cells (the heavy chain of the B cell receptor and the beta chain of the T cell receptor) that play a critical role in recognizing immune insults (antigens) and binding to them (so that the immune system can respond) in tens of millions of T and B-cells.

The authors used three different analytic models to assess the data. They found that T cell receptor analysis was best at predicting lupus and Type 1 diabetes while B cell receptor sequences were best at identifying HIV, SARS-CoV-2 infection or recent influenza vaccination.

Combining the different models together to create an “ensemble model” was most effective. The study results shot off the charts with its ability to essentially classify everyone correctly ((AUROC) of 0.999 (1.0 = perfect)). That’s an unheard of diagnostic effectiveness.

The test’s ability to accurately assess disease specificity and sensitivity was lower (93% were accurately assessed as healthy or not, and 87.5% were accurately determined to have the correct disease). While some healthy people were assessed to have an autoimmune disease, Topol noted that these people could be vulnerable to an autoimmune condition.

Topol’s talk with the lead authors found that they’ve done sequencing in many more people with autoimmune conditions and that the tests have provided illuminating insights into therapies. Plus, when they become available, the tests should be inexpensively done at scale in batches with quick turnarounds.

The ability of this technique to subcategorize disease states at the molecular level could be particularly helpful in heterogeneous diseases like ME/CFS, FM, and long COVID. One of the co-authors – Boyd – noted that even autoimmune diseases have subsets that are significantly different at a biological or molecular level. Boyd told Science Daily:

“Mal-ID could help us identify subcategories of particular conditions that could give us clues to what sort of treatment would be most helpful for someone’s disease state.”

We’ve already seen that although Gulf War Illness and ME/CFS look the same in some ways, they are very different creatures indeed. A similar revolution in botanical classification took place with the advent of genetic analyses. Some plants put into families based on their appearance turned out not to be related at all, and plant classification was fundamentally changed.

Topol also noted that the use of a large language AI model called ESM-2 played a critical role in the study. While the study was a bit “contrived” because the authors used a limited number of conditions to test the model on, the authors stated that much progress had been achieved since the paper had been written.

The AI-trained system’s ability to identify molecular subsets within diseases could be a big boon in diseases like ME/CFS/FM and long-term COVID.

Topol envisioned integrating the T and B cell findings with other immune system findings (autoantibodies, virus and pathogen exposure, flow cytometry, T cell functional assays, interferon, HLA typing, etc) to create a full picture of the immunome. This seems eminently possible given the dramatic uptick in our ability to assess large amounts of data. The limiting factor doesn’t appear to be on the analytic side but on our ability to gather all the data; i.e., to get the funding to gather as much data as possible from as many people as possible.

Indeed, Topol stated that this new technique was perfect for understanding mysterious, complicated diseases like long COVID (and thanks to Mark Camenzind’s question) to ME/CFS and post-infectious diseases like post-treatment Lyme Disease (PTLD). Topol stated that not only that “diagnosing and following individuals with Long Covid is an exemplar for how this could be applied” but that it could be used to testing the effectiveness of treatments.

Applications to ME/CFS and other Post-Infectious Illnesses

When asked, Topol said the approach was “perfect for ME/CFS”.

When ME/CFS advocate Mark Camenzind asked if this approach could be applied to ME/CFS, Topol called it “definitely, a perfect application for ME/CFS”, and responded similarly to a question regarding Post-Lyme Disease Treatment Syndrome (PLDTS). The approach could also be used to uncover what’s gone wrong in post-vaccination syndrome and who is at risk for it. Mark Camenzind noted that the study originated at Stanford and that Paul Utz – who is studying ME/CFS and is the co-leader of the RECOVER long COVID program at Stanford – was a co-author.

This technology will take some time to become available, but it shows that we live in exciting times. With artificial intelligence enabling medical research in so many ways, who knows what’s around the corner?

The study was heavily funded by the NIH (20 grants (!))) as well as other grants and philanthropic support. Topol – who has been quite critical of the NIH – nevertheless decried the recent changes occurring at it, the CDC, the FDA and the Dept. of Health and Human Services. Calling it “dark times for biomedicine”, Topol – citing the “marked acceleration of progress” that is occurring, due in part to AI – said he had not lost optimism.