Predictive Modeling of Hematopoietic Stem Cell Transplant Outcomes: Discovering the Role of Immunogenetic Features through Machine Learning Models

Background: Allogeneic hematopoietic cell transplantation (allo-HCT) stands as a potentially curative therapy for various hematologic neoplasms. Its success hinges on a delicate balance: eradicating malignant cells (graft-versus-leukemia -GVL- effect) while minimizing complications such as relapse and graft-versus-host disease (GVHD). Over the past decades, various immunogenetic models have unraveled the role of human leukocyte antigen (HLA) compatibility and diversity in determining transplant outcomes. HLA evolutionary divergence (HED), a metric quantifying the genetic disparity in the antigen binding sites across HLA loci, has emerged as a predictor of response to immunotherapy in solid tumors, featuring its potential value also in transplant immunology. This study aims to harness machine learning (ML) tools to comprehensively evaluate the predictive power of HLA alleles, HED, and other transplant-related variables on key allo-HCT outcomes including relapse, acute GVHD (aGVHD), chronic GVHD (cGVHD), and overall survival (OS).

Methods: We accrued a large international cohort comprising patients receiving an allo-HCT in six major academic and clinical centers in USA and Europe. This dataset gathered time-dependent outcomes, transplant and disease-specific annotations and comprehensive HLA and related immunogenetic data. We employed multiple ML algorithms, notably Random Survival Forest (RSF) and Lasso/Elastic Net (EN) models to dissect the interplay between HLA supertypes, allele disparities, HED, and various clinical parameters.

Results: Overall, the study encompassed a diverse cohort of 2,192 adult patients [median age of 59 years (IQR 49-65) and male-to-female ratio of 1.3]. The disease spectrum included acute myeloid leukemia (N=1271), acute lymphoblastic leukemia (N=51), myeloproliferative neoplasms (N=219), myelodysplastic syndromes (N=590), lymphoma (N=37), bone marrow failure (N=15), plasma cell dyscrasia, and other lymphoproliferative disorders (N=9), along with a variety of donor types: matched related (N=492) and unrelated donors (N=1114), mismatched unrelated donors (N=97), and haploidentical donors (N=489). Also, graft sources included mobilized peripheral blood (N=1727), bone marrow (N=417), and cord blood stem cells (N=48). Additionally, both reduced-intensity (N=1671) and myeloablative conditioning regimens (N=521) were included, reflecting a broad range of clinical scenarios. Since an initial assessment of feature importance revealed that HLA supertypes did not have a significant relative impact on any of the outcomes, we excluded them from further analyses. HED values were calculated using established algorithms for the individual loci. The RSF models demonstrated strong discriminative abilities with relatively high area under the ROC curve (AUC) scores: 0.89 for OS, 0.91 for relapse, 0.80 for aGVHD, and 0.90 for cGVHD. Notably, HED values, particularly those in class II alleles, emerged as significant predictors. In this context, feature analysis showed that HED in DRB1 locus was crucial for relapse prediction while both HED in DQB1 and DRB1 were pivotal for acute GVHD, HED in DRB1 and B influenced chronic GVHD outcomes, and HED in A and DQB1 loci were critical for OS together with graft, donor and disease types. Subsequently, we explored Lasso/EN models to validate these associations and manage the dataset's high dimensionality and multicollinearity. These immunogenetic-integrated models achieved high accuracy in predicting relapse (86.3%) and chronic GVHD (95.4%), with moderate performance in predicting OS (73.9%) and aGVHD (63.7%). Specificity and sensitivity were particularly high for the cGvHD model (100%/84%) and moderately high for the other models (relapse: 96%/67%, aGvHD: 88%/28%, OS: 72%/75%).

Conclusion: This study demonstrates the effectiveness of advanced ML tools with detailed immunogenetic profiling to forecast key outcomes following allo-HCT. These results underpin the importance of integrating HLA-related metrics in clinical decision-making processes for the development of personalized therapeutic approaches. The successful use of ML tools such as RSF and regularization methods (Lasso and EN) sets a new benchmark for future predictive models in transplant medicine, focusing on optimizing patient care through customized risk assessment and tailored treatment strategies.