Machine learning analysis of human leukocyte antigen structure identifies novel predictors of graft failure after haploidentical stem cell transplantation, providing potential advances in donor selection and risk stratification. The results of this study were presented as a research poster at the 52nd EBMT Annual Meeting.
Identifying risks of port failure using machine learning
Graft failure remains a life-threatening complication after haploidentical stem cell transplantation, and traditional matching methods cannot fully capture immunological incompatibility. In this study, researchers applied a protein language model to analyze the structural features of human leukocyte antigen molecules in 127 patients who experienced graft failure and underwent a second transplant.
From the initial set of 8,718 features, five structural features were significantly associated with graft failure (p<0.05). These features reflect complex interactions between class I and class II loci rather than single gene effects, highlighting the importance of overall structural compatibility. One characteristic is donor-specific and four represent donor-recipient interactions.
Structural biomarkers and donor selection
Compared to successful second transplants, failed first donors showed decreased structural diversity across specific loci spanning HLA-A to HLA-C (A_C_exon23, p=0.012; hazard ratio: 0.4987). In contrast, donor-recipient pairs associated with graft failure showed increased structural width across several multilocus regions, including A_B_C_DR_DQ_exon3 (p<0.001), A_DR_exon3 (p<0.001), C_DR_exon2 (p=0.003), and A_B_exon1-5 (p=0.033).
Five major structural biomarkers were identified when compared to a control cohort of 1,699 patients without graft failure. These include three recipient-specific features A_B_C_DR_DQ_exon3, C_DR_exon1-5, and A_B_C_exon5 and two donor-recipient interaction features A_B_C_exon1 and A_C_exon1, reinforcing the multifactorial nature of transplant compatibility.
A composite risk score derived from these five features showed moderate predictive performance with an area under the curve of 0.656, suggesting potential clinical utility in identifying high-risk donor-recipient combinations.
Implications for transplantation practice
These findings represent a new approach to understanding graft failure by integrating structural immunology and machine learning. This result suggests that subtle physicochemical and spatial properties of human leukocyte antigen molecules can influence transplantation outcomes beyond traditional allele-level matching.
This approach allows clinicians to identify high-risk donor or donor-recipient pairs prior to transplantation, potentially improving outcomes in the haploidentical setting. Although further validation in larger cohorts is required, this study is an important step toward more accurate and biologically informed donor selection strategies.
reference
Ma R et al. HLA structural profiles identified by protein language models predict graft failure in haploidentical transplants. Abstract B275. EBMT52n.d. Annual General Meeting; March 22-25, 2026.
