Over the decades, the success of a bone marrow transplant was gauged by one, painful wait; when the new immune system either accepts or rejects. When it comes to hematology, graft failure is the ghost in the machine: a very uncommon yet catastrophic complication in which the stem cells of the donor don’t grow, and a patient is left without any defenses and with very limited options.
The discussion at the 52 nd Annual Meeting of the European Society of Blood and Marrow Transplantation (EBMT 2026) in Madrid switched to the ability to predict this crisis rather than respond to it. An innovative study reported this week has demonstrated that Machine Learning (ML) can now see deeper than conventional HLA matching to select patients at high risk of graft failure even before the first cell is ever infused.
The Unmanifesto of Rejection
Conventional donor matching has been based upon a lock and key principle: the matching of the major HLA (human leukocyte antigen) markers between a donor and a recipient. And, as most clinicians can attest, even a 10 out of 10 match is not a sure-footed success.
The study that was mentioned in EBMT 2026 used the language models of proteins and the complex ML algorithms to serve more than 8,700 structural features of HLA molecules. This was not just a question of the composition of the genes being similar, but of the shape of the proteins and how the 3D form and structural diversity varied.
Important Advances in Foretelling:
- Multi-Locus Interactions: What the AI determined is that failure of the graft is not normally because of a single mismatched gene. Rather it is the effect of their interaction of various HLA regions (such as A, B, C and DR).
- Structural Breadth: A particular “structural breadth” of a HLA exon region was present in pairs that failed the grafts, and is not represented by traditional testing.
- Donor vs. Recipient Intrinsic Factors: The model effectively diffused factors that inherently belonged to the biology of the donor and those factors that were only problematic when combined with the unique environment of the recipient.
Feeding this high-dimensional data into a Random Forest model, or a form of machine learning that produces thousands of decision trees, researchers could predict results at a much higher accuracy than with the traditional EBMT risk scores with which they have measured the last two decades.
you may also like:
- Anthropic brings viral work tools on phone
- Claude AI Gets Smarter
- Bengaluru Goes Green with new Waste Plant
Trying to Humanize the Data: The Reason this is Important to Patients
Hiding behind each number in this research is a patient such as “Elena,” a 48-year-old survivor of lymphoma whose story was presented in the keynote of EBMT. Elena had a full-fledged donor match in paper, but she experienced primary graft failure which was a nightmare in clinical terms and she needed to undergo an emergency second transplant.
Previously, the doctors of Elena could only attribute her failure to bad luck. Had the structural incompatibility between her donor match weeks been revealed at all, they could have at least realized it with the new ML tools in Madrid.
The transition, observed one of the presenters, is to the Science of Precision, rather than the Art of Medicine. It is the difference between a preemptive approach and wait and see approach where a patient may have to save their lives.
A New Age: Multi-Outcome Prediction and SPRIGHT
The failure of grafts is not the only problem being addressed by AI. One more significant agenda of the 2026 congress was the SPRIGHT (Sickle Cell Predicting Outcomes of Hematopoietic Cell Transplantation) model. Although it is targeted at Sickle Cell Disease, the framework is a move towards multi outcome prediction.
These new tools show a dashboard of risks, rather than merely informing a doctor whether a transplant will work:
- Graft Failure (GF)
- There are acute and Chronic Graft-verified-Host Disease (GvHD).
- Overall Survival (OS)
- Event-Free Survival (EFS)
These models cease to be black boxes when they are used with SHAPley Additive Explanations values. In fact, they justify to the doctor why they tagged a patient as high-risk due to certain variables such as baseline hemoglobin levels or the actual dosage of conditioning therapy.
