Predictive Modeling of Post-Allogeneic Transplant Outcomes Using Machine Learning and Integrated Clinical and Immunogenetic Data: a Study from the SFGM-TC and a Multicenter US Consortium

Predictive Modeling of Post-Allogeneic Transplant Outcomes Using Machine Learning and Integrated Clinical and Immunogenetic Data: a Study from the SFGM-TC and a Multicenter US Consortium | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Predictive Modeling of Post-Allogeneic Transplant Outcomes Using Machine Learning and Integrated Clinical and Immunogenetic Data: a Study from the SFGM-TC and a Multicenter US Consortium Simona Pagliuca, Vincent Alcazer, Mélanie Gaudfrin, Nicole Raus, and 30 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9383323/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Allogeneic hematopoietic cell transplantation (allo-HCT) remains the only curative option for many hematologic malignancies, yet transplant-related toxicity and relapse continue to constrain long-term benefit. Prognostic tools based on clinical variables alone show limited transportability across centers. We hypothesized that integrating immunogenetic architecture, captured by HLA Evolutionary Divergence (HED), into machine-learning survival models would improve prediction of graft-versus-host disease–free/relapse-free survival (GRFS). We developed SMART, a time-dependent framework that estimates dynamic GRFS probabilities after allo-HCT. We analyzed 16,028 adults transplanted between 2010–2022. Model development used the French SFGM-TC registry (N = 13,979), split into training (N = 9,840) and held-out test (N = 4,139) sets. Random survival forests, XGBoost-Cox, and elastic-net Cox models were trained using 9 clinical predictors, with or without 10 recipient/donor locus-specific HED features (19 predictors total) and externally evaluated in 616 patients from five U.S. centers with complete data. Across algorithms, discrimination for this composite endpoint was modest (c-index < 0.60), but the addition of HED consistently improved performance and enabled reproducible stratification into low-, intermediate-, and high-risk groups based on the cumulative hazard score. Model-based simulations uncovered a non-linear (U-shaped) association between total HED and GRFS, with optimal outcomes at intermediate HED levels (~ 75th percentile). In haploidentical transplantation (N = 2,056), outcomes were maximized when donor and recipient HED were concordant (“match like with like”). In 9/10 mismatched unrelated donors (N = 1,326), HLA-B mismatches showed the greatest HED sensitivity. Integrating immunogenetics with clinical data improves GRFS risk modeling and supports HED as an actionable feature for donor selection and pre-transplant risk stratification. SMART is available for research use. Immunology Hematology Artificial Intelligence and Machine Learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Allogeneic hematopoietic cell transplantation (allo-HCT), the most paradigmatic form of cellular immunotherapy, is currently the solely curative treatment for many hematological disorders. Transplant-related complications, notably disease recurrence, graft-versus-host disease (GvHD) and infections continue to hamper the therapeutic index of this procedure, limiting its broader applicability 1 – 3 . Central to these complications is the complex interplay of donor-recipient alloreactivity and the balance between beneficial graft-versus-tumor (GvT) and harmful graft versus host (GvH) reactions 4 – 10 . Distinction between beneficial and deleterius alloreactivity remains the elusive "holy grail" of transplantation immunology essential for improving outcomes, expanding indications and implementation of innovative post-transplant immunomodulatory strategies 10 – 12 . Given the diversity of transplant platforms and patient populations across transplant centers and countries, traditional predictive models systematically fail to achieve generalizability, particularly when relying solely on clinical and demographic parameters 13 , 14 . This limitation underscores the clinical need to develop robust, accurate, and predictive tools capable of intergrate the interplay between the « beneficial » and the « deleterius » alloreactivity. To that end, computational biology and machine learning (ML) tools offer promising solutions to resolve complexity of datasets and uncover predictive patterns otherwise invisible to classical statistics 15 – 21 . Among biological features which could improve prediction of alloreactive complications in the context of all-HCT, the immunogenetic background, particularly the reciprocal human leukocyte antigen (HLA) configurations between donor and recipient, emerges as a critical determinant of post-transplant outcomes 22 . As robust measure of HLA heterogeneity, HLA Evolutionary Divergence (HED) concept has demonstrated a potential to effectively capture immunogenetic complexity in a readily interpretable manner, serving as a quantitative proxy for the degree of individual immunocompetence 23 – 29 . Given the limitations of models based solely on clinical variables to predict survival or mortality after allo-HCT 30 , 31 , we sought to develop a integrative computational framework to stratify the probability of GvHD-free remission as captured by GvHD-free/relapse-free survival (GRFS) 32 , 33 . For that purpose, we leveraged a very large, multicenter international cohort of patients with hematologic malignancies comprising 15,000 individuals, comprehensively annotated for clinical characteristics and HLA-related variables. We first quantified the contribution of individual immunogenetic interactions to clinical phenotypes and outcomes in a dedicated training cohort, and then trained GRFS prediction models to incorporate clinical and immunogenetic input using complementary ML-approaches. Model performance was subsequently evaluated in an internal hold-out test set and in an independent external validation cohort. Ultimately, our goal was to create an actionable risk stratification tool and to provide a rational basis for tailoring transplant platforms and post-transplant immunomodulatory strategies to individual patient–donor profiles. Methods Study design, cohorts, and eligibility We conducted a multicenter, international retrospective cohort study of 16,028 adults undergoing a first allogeneic hematopoietic cell transplant (allo-HCT) for hematologic malignancies between 2010–2022. Model development used the Societe Francophone de Greffe de Moelle et de Therapie Cellulaire (SFGM-TC) registry (N = 13,979), randomly split 70:30 into a training set (N = 9,840) and a held-out internal test set (N = 4,139) with stratification to preserve balance across key clinical and immunogenetic characteristics ( Supplementary methods ). External validation was performed in an independent US cohort (N = 1,931) from five centers (Cleveland Clinic, Johns Hopkins, Vanderbilt, University of Minnesota, and Karmanos Center), with 616 (32%) having complete data for all model predictors ( Supplementary methods , Table 1 , Table S1 and S2 ). Eligible patients were ≥ 18 years, received a first allo-HCT for a malignant hematological indication, had high-resolution HLA typing for A, B, C, DRB1, DQB1, and had available outcome data. Patients undergoing second transplants, transplanted for non-malignant diseases, or lacking required high-resolution HLA data were excluded. The study was approved by institutional review boards at all participating centers and was approved by the scientific committee of the SFGM-TC in april 2025. This study was declared to the Health Data Hub under the number 24237877. Table 1 Patient and Transplant Characteristics from SFGM-TC cohort Characteristic Training Cohort (n = 9,840) Test Cohort (n = 4,139) Patient Characteristics Patient age, years, median (IQR) 56.1 (44.5–63.5) 56.2 (44.3–63.3) Patient sex, n (%) Male 5,742 (58.4%) 2,426 (58.6%) Female 4,095 (41.6%) 1,712 (41.4%) Patient ABO blood group, n (%) A 4,042 (41.1%) 1,649 (39.8%) B 898 (9.1%) 382 (9.2%) AB 364 (3.7%) 152 (3.7%) O 3,998 (40.6%) 1,665 (40.2%) Patient CMV serostatus, n (%) Positive 5,338 (54.2%) 2,172 (52.5%) Negative 4,461 (45.3%) 1,950 (47.1%) Disease Characteristics Diagnosis, n (%) Acute myeloid leukemia 4,442 (45.1%) 1,885 (45.5%) Acute lymphoblastic leukemia 1,148 (11.7%) 474 (11.5%) Myelodysplastic syndrome 1,353 (13.8%) 561 (13.6%) Myeloproliferative neoplasm 1,150 (11.7%) 502 (12.1%) MDS/MPN overlap 73 (0.7%) 36 (0.9%) B-cell non-Hodgkin lymphoma 502 (5.1%) 187 (4.5%) T-cell non-Hodgkin lymphoma 392 (4.0%) 164 (4.0%) Hodgkin lymphoma 267 (2.7%) 103 (2.5%) CLL/PLL 223 (2.3%) 96 (2.3%) Multiple myeloma/PCL 290 (2.9%) 131 (3.2%) Disease status at transplant, n (%) Complete remission 6,262 (63.6%) 2,610 (63.1%) Active disease 3,232 (32.8%) 1,355 (32.7%) Donor Characteristics Donor age, years, median (IQR) 34.0 (26.0–45.0) 34.0 (26.0–45.0) Donor type, n (%) Matched sibling donor 1,933 (19.6%) 808 (19.5%) Matched unrelated donor 5,085 (51.7%) 2,165 (52.3%) Mismatched unrelated donor 1,248 (12.7%) 521 (12.6%) Haploidentical donor 1,451 (14.7%) 605 (14.6%) Cord blood 123 (1.2%) 40 (1.0%) Donor ABO blood group, n (%) A 3,802 (38.6%) 1,640 (39.6%) B 920 (9.3%) 367 (8.9%) AB 387 (3.9%) 151 (3.6%) O 4,202 (42.7%) 1,705 (41.2%) Donor CMV serostatus, n (%) Positive 4,204 (42.7%) 1,750 (42.3%) Negative 5,604 (57.0%) 2,374 (57.4%) Transplant Characteristics Stem cell source, n (%) Peripheral blood 8,426 (85.6%) 3,537 (85.5%) Bone marrow 1,291 (13.1%) 562 (13.6%) Cord blood 123 (1.2%) 40 (1.0%) Conditioning regimen, n (%) Myeloablative conditioning 3,453 (35.1%) 1,440 (34.8%) Reduced-intensity conditioning 6,387 (64.9%) 2,699 (65.2%) GvHD prophylaxis, n (%) CNI alone 1,418 (14.4%) 590 (14.3%) CNI + mycophenolate mofetil 5,091 (51.7%) 2,105 (50.9%) CNI + methotrexate 2,958 (30.1%) 1,294 (31.3%) Other 219 (2.2%) 88 (2.1%) In vivo T-cell depletion, n (%) ATG 6,238 (63.4%) 2,615 (63.2%) Post-transplant cyclophosphamide 1,513 (15.4%) 652 (15.8%) ATG + PTCy 209 (2.1%) 76 (1.8%) Other 21 (0.2%) 7 (0.2%) HLA Mismatches HLA-A mismatch, n (%) 0 7,913 (80.4%) 3,368 (81.4%) 1 1,686 (17.1%) 677 (16.4%) 2 241 (2.4%) 94 (2.3%) HLA-B mismatch, n (%) 0 8,061 (81.9%) 3,403 (82.2%) 1 1,552 (15.8%) 657 (15.9%) 2 227 (2.3%) 79 (1.9%) HLA-C mismatch, n (%) 0 7,977 (81.1%) 3,374 (81.5%) 1 1,600 (16.3%) 673 (16.3%) 2 263 (2.7%) 92 (2.2%) HLA-DRB1 mismatch, n (%) 0 8,150 (82.8%) 3,434 (83.0%) 1 1,453 (14.8%) 607 (14.7%) 2 237 (2.4%) 98 (2.4%) HLA-DQB1 mismatch, n (%) 0 8,274 (84.1%) 3,464 (83.7%) 1 1,348 (13.7%) 580 (14.0%) 2 218 (2.2%) 95 (2.3%) HLA-DPB1 mismatch, n (%) 0 5,620 (57.1%) 2,316 (56.0%) 1 2,229 (22.7%) 981 (23.7%) 2 1,991 (20.2%) 842 (20.3%) HLA typing and HLA evolutionary divergence (HED) High-resolution HLA typing (minimum 2-field/4-digit) was performed by accredited histocompatibility laboratories using sequence-based methods. HED was computed, as previously shown, at each locus using a Grantham distance–based approach quantifying physicochemical divergence across the antigen-binding domain between the two alleles carried by an individual 23 , 25 . HED was derived for A, B, C, DRB1, DQB1 (and DPB1 when available), separately for recipients and donors. Summary measures included Class I HED (A + B+C), Class II HED (DRB1 + DQB1), and Total HED (A + B+C+DRB1 + DQB1). Delta HED was defined as donor minus recipient total HED. Clinical variables and outcomes Pre-transplant variables included recipient age and sex, diagnosis, disease status at transplant (complete remission vs active disease), donor type, donor age, stem cell source, conditioning intensity, GVHD prophylaxis and in vivo T-cell depletion strategy. The primary endpoint was GRFS, defined as survival without grade III–IV acute GVHD, extensive chronic GVHD, relapse, or death; time was measured from transplant to the first component event, with censoring at last follow-up. Secondary endpoints included overall survival, relapse (with death as a competing event), non-relapse mortality (with relapse as a competing event), and acute/chronic GVHD analyzed using cumulative incidence methods with appropriate competing risks. Machine-learning model development and validation We evaluated three survival modeling approaches: random survival forests (RSF), XGBoost Cox survival, and penalized Cox regression (elastic net) with cross-validated hyperparameter tuning. For each method, we trained (i) a clinical model using 9 pre-transplant clinical predictors and (ii) a full model adding recipient and donor locus-specific HED (A, B, C, DRB1, DQB1; 10 HED features). Primary modeling used complete-case analysis (no imputation), yielding 9,196 patients for training and 3,826 for internal testing ( Supplementary methods ). Discrimination was assessed with Harrell’s c-index (95% CIs from 1000 bootstrap resamples) and time-dependent area under the curve (AUC) at 1, 2, and 5 years; calibration compared predicted vs observed risks across deciles. External validation applied the trained model to the US cohort without recalibration. For clinical interpretability, RSF-predicted cumulative hazard scores were used to define low/intermediate/high risk groups based on training-set tertiles, and the same cutpoints were applied unchanged to internal and external validation cohorts; Kaplan–Meier curves were compared by log-rank testing. Simulation analyses To visualize how HED profiles influence predicted GRFS risk under fixed clinical conditions, we simulated RSF predictions using a reference clinical profile (56-year-old male with acute myeloid leukemia [AML] in complete remission, receiving peripheral blood stem cells with reduced-intensity conditioning and calcineurin inhibitor [CNI] mychophenolate [MMF] prophylaxis) and systematically varied recipient/donor HED across observed ranges. Additional simulations examined matched donors, haploidentical donors, and 9/10 mismatched unrelated donors stratified by mismatched locus to explore locus-specific patterns. Statistical analysis Group comparisons used Wilcoxon rank-sum tests (continuous variables) and chi-square/Fisher’s exact tests (categorical variables). Survival was summarized with Kaplan–Meier methods; competing-risk endpoints were analyzed using Fine–Gray models. Analyses were performed in R (v4.2.0) using standard survival and machine-learning packages. Results Study cohorts In the French cohort, after exclusion of patients with missing data for model variables, 9,196 patients in training and 3,826 patients in test sets had complete data for all 19 predictors. The study cohorts used for the model development did not differ in clinical or immunogenetic variables (Fig. 1 A, Table 1 and S1). The median patient age was 56.1 years (IQR 44.5–63.5) in training and 56.2 years (IQR 44.3–63.3) in test cohorts. Males comprised 58 and 59% of training and test cohorts, respectively. Acute myeloid leukemia (AML) was the most common diagnosis (45% training, 45.5% test), followed by myelodysplastic syndrome (MDS, 14 vs.14%) and myeloproliferative neoplasms (MPN, 11 vs. 12%). Most patients (64% training, 63% test) were in complete remission at transplant. Donor types included matched unrelated donors (MUD, 52% training, 52% test), matched sibling donors (MSD, 20%, 19.5%), haploidentical donors (15% in both cohorts), mismatched unrelated donors (MMUD, 13% in both), and cord blood (1.0% in both cohorts). Missingness was assessed across all study variables, with an overall rate of 6% in training and test cohorts. Most variables were 98–100% complete, with notable exceptions: disease status (96%), T-cell depletion status (81%), and HLA-DPB1 alleles for recipients (74%) and donors (73%) ( Figure S1 ). The final dataset included clinical variables and HLA information encoded as recipient- and donor-specific HED at the locus level (A, B, C, DRB1, DQB1), as well as aggregated class I, class II, and total HED scores. Because HLA-DPB1 alleles were frequently missing, total HED was calculated using only loci A, B, C, DRB1, and DQB1. The external validation cohort comprised 1,931 patients from five US centers, with 616 (32%) having complete data for all model variables ( Table S2 ). After removal of cases with missing data, this cohort was enriched for myeloid diseases. Immunogenetic features across disease subgroups. First, we assessed whether immunogenetic, disease-specific patterns exist by examining the distributions of locus-specific, class I, class II, and total HED scores in recipients across ten disease subgroups, and we compared these with donor-related metrics (Figure S2[A,B] and S3). Notably, the Hodgkin (HL) lymphoma subgroup showed distinct per-locus distributions, with the lowest total and class II HED (driven primarily by DRB1 and DQB1) suggesting a potential immunogenetic predisposition related to specific class II HLA risk structures associated with the disease. When examining recipient–donor HED associations, we found strong correlations across all loci (R > 0.84, p < 2.2×10⁻¹⁶) except DPB1, which showed only a weak recipient–donor correlation (R = 0.38), Figure S4 . By donor subgroup, correlations in unrelated mismatched allo-HSCT ere very high for class II loci (R > 0.91) DQB1 and DRB but lower for class I (R ≤ 0.75), and very low for DPB1 ( Figure S5 ). In haploidentical transplants, correlations were uniformly weak across loci (R < 0.2). Cord blood was heterogeneous: higher for class II (R : 0.70–0.95) and lower for class I (R : 0.19–0.59), suggesting that cord blood units with class I mismatches are more frequently selected. DPB1 showed consistently weak correlations across donor groups (R < 0.3), with the notable exception of MSD, where not surprisingly the correlation was strong (R = 0.97), Figure S6. Exploring the impact of immunogenetic features on outcomes When we assessed the impact of specific immunogenetic features on outcomes, allele-enrichment analysis did not reveal any allele-specific effect on dichotomous clinical phenotypes (GRFS event vs. no event at 2y ; Table S3 ). However, several notable associations emerged when HLA loci were parameterized as HED features. Higher recipient HED at the DQB1 locus was an immunogenetic determinants of an increased risk of grade III–IV acute GvHD, whereas higher recipient HED at the B locus was associated with a lower risk of relapse and improved whole cohort OS ( Table S4 , Fig. 2 A). These associations displayed a time-varying effect, with an increasing HR over time (Fig. 2 B). Within the HLA-mismatched subset, the donor–recipient HED difference (Delta HED) did not differ across first-event categories, suggesting that a simple mismatch-derived summary metric together with binary outcomes, are unlikely to capture the post-transplant alloreactivity ( Figure S7 ). Machine learning modeling We next investigated the combinatorial impact of clinical and HED-related features using machine-learning models. When the SFGM-TC cohort was split in well-blanced training and a test set no differences in outcomes were observed ( Figure S8 ). We decided to study these clinical and immunogenetic interactions on GRFS, a composit endpoint expressing probability of being alive with/without an alloreactive complication. In the training set we applied principal component analysis (PCA) on all candidate predictors and projected patients onto the first 3 components, according to GRFS status (event vs event-free). PC1, PC2 and PC3 accounted for ~ 23%, 19% and 15% of the total variance, respectively, indicating that these axes captured just over half of the overall heterogeneity of the cohort (Fig. 3 A). However, in all pairwise projections (PC1–PC2, PC1–PC3 and PC2–PC3), patients with and without a GRFS event were widely intermingled, with no distinct clusters or clear separation between outcome groups. To assess the global contribution of immunogenetics to GRFS, without relying on linearity or other parametric assumptions, we implemented a comparative modeling framework. Specifically, we trained (i) a RSF, (ii) an XGBoost-based survival, and (iii) a Cox proportional hazards models with elastic-net regularization, and evaluated on an internal test set. For each modeling family, we specified two feature sets to isolate the incremental value of immunogenetics: i) a clinical model restricted to standard pre-transplant covariates routinely available at baseline (9 variables, Supplmentary methods ); and ii) a full model that extended the clinical backbone by adding the 19 variables (A, B, C, DRB1, DQB1, and HED metrics). This design allowed us to quantify, across distinct algorithmic assumptions (tree-based non-parametric, gradient-boosted, and penalized semi-parametric), whether inclusion of HED variables consistently improved prognostic performance and to compare the relative importance of clinical vs immunogenetic features within each framework. Nevertheless, discriminatory performance of all the models remained modest. Across all 3 models, addition of immunogenetic information (donor/recipient HED) to standard pre-transplant clinical covariates resulted in a small but generally consistent improvement in discrimination ( Tables S5–S6 , Fig. 3 B) while clinical predictors remained most important risk determinants (Fig. 3 C), donor and recipient HED, particularly at class II loci (DRB1 and DQB1), emerged among the most informative immunogenetic contributor (Fig. 3 D). RFS model validation Given comparable out-of-sample discrimination across modeling approaches, minimal evidence of overfitting, and the need for an interpretable prognostic tool, we selected the RSF model incorporating HED variables (RSF full) as the primary model for downstream analyses and validation. For each patient, we extracted the event risk score from the trained RSF model, defined as the predicted cumulative hazard. This continuous score captures the relative risk of experiencing the composite GRFS endpoint (grade III–IV acute GvHD, extensive chronic GvHD, relapse, or death from any cause). Patients were subsequently stratified into three risk categories—low (score 143)—using tertiles of the predicted risk score distribution in the training cohort. Optimized cutpoints ( Supplementary Methods ) were derived exclusively in the training set and then applied unchanged to the hold-out test set and the external validation cohort to ensure an unbiased assessment of performance. Across all three cohorts, Kaplan–Meier analyses demonstrated clear separation between risk groups, with statistically significant differences by log-rank testing (p < 0.001; Table S7, panel A , Fig. 4 A, B), supporting robust identification of three clinically meaningful risk strata, despite low C-indexes (Table S7, Panel B ). To assess the generalizability of the model across disease subtypes, we applied the same risk stratification to the separete subsets of patients with AML and MDS/MPN. Despite being trained on the full heterogeneous population, the model maintained robust discriminative ability within this disease-specific subgroup( Figure S9 ). HED contribution on RFS-based risk scores in matched setting. To characterize the combined effect of recipient and donor HED on GRFS outcomes, we performed a systematic simulation across the full range of total HED values from the French cohort (n = 13,979) using the trained RSF model. All clinical variables were held constant at reference values (56-year-old male with AML in complete remission, receiving peripheral blood stem cells from a MUD aged 35 years, with reduced-intensity conditioning and calcineurin inhibitor/mycophenolate mofetil prophylaxis). This analysis revealed a pronounced non-linear (U-shaped) relationship between total HED and GRFS risk, demonstrating a "Goldilocks effect" (Fig. 5 A). The optimal risk profile was observed at intermediate total HED levels of approximately 55 for both recipient and donor (corresponding to the 75th percentile of both HED distributions). Risk increased substantially with very low (Total HED = 0) or very high donor/recipient total HED (> 65)). Percentile-based analyses confirmed these findings (Fig. 5 B). Risk scores were elevated at the lowest HED percentile (P5: score = 71, total HED = 0, reflecting complete homozygosity), decreased to a minimum at P75 (score = 60, Total HED = 57), and increased again at the highest percentile (P95: score = 92, Total HED = 69). To assess the robustness of these findings, we repeated the simulation across three clinical profiles (all receiving matched donor transplantation) with distinct baseline GRFS risk levels: a higher-risk profile (65-year-old, MAC conditioning, not in remission), a lower-risk profile (45-year-old, MSD, complete remission), and a « standard risk » (reference) AML profile (56-year-old, MUD, RIC, complete remission). All three profiles exhibited the characteristic U-shaped HED-risk relationship, with nadir consistently at intermediate HED values (Fig. 5 C). For the reference clinical profile, predicted risk scores (RS) ranged from 60 (optimal, P75) to 92 (high HED, P95), with intermediate values for homozygous patients (71) and those at median HED (78). All simulated scenarios remained within the low-risk zone (score < 103), though the 32-point variation demonstrates that HED optimization provides meaningful risk modulation even within favorable clinical contexts (Fig. 5 D). HED contribution on RFS-based risk scores in mismatched setting. We then run anaologous analyses in the haploidentical subgroup. The optimal recipient/donor combination occurred at intermediate total HED levels (Fig. 6 A). Through a percentile based analysis, we quantified risk interactions across 49 recipient-donor combinations (P5-P95 for each). The optimal pairing was P75 recipient x P75 donor (RS = 79). Notably, the matrix revealed a distinct « match like with alike » pattern: recipients with low HED achieved optimal outcomes with low-HED donors (P5xP5: RS = 97), intermediate-HED recipients with intermediate donors (P50xP75; RS = 82 ; P75xP75; RS = 79) and high-HED recipients with high-HED donors (P90×P90 = 85; P95×P90 = 94). Mismatched pairings, particularly low-HED recipients with high-HED donors, produced the worst outcomes (P5×P95 RS = 119; P10×P95 : RS = 116). Donor selection effects were then examined across four representative recipient HED levels. For P5 recipients, risk decreased as donor HED decreased, with P5 donors being optimal (RS = 97). For P75 recipients, the lowest risk was also observed with matched P75 donors (RS = 79), while both lower and higher donor HED values increased risk. For P95 recipients, the best outcome was seen with P90 donors (RS = 97), whereas P95 donors paradoxically increased risk (Fig. 6 B). The potential risk reduction achievable through donor selection differed by recipient HED (Fig. 6 C,D), reaching 28 points for P25 recipients, 21 for P5, 20 for P50, and 10 even for P95 recipients. Overall, these findings indicate that donor HED is an actionable parameter in haploidentical transplantation, with the best strategy being to align donor HED with recipient HED level ( Table S7 ). We next assessed the impact of donor and recipient HED on RSF scores in the 9/10 mismatched unrelated donor setting (N = 1,326), restricting the analysis to transplants with a single HLA mismatch ( Figure S10 ). Using the same standardized AML patient profile, we simulated each mismatched subgroup while varying donor and recipient HED from P5 to P95 at the mismatched locus and keeping all other variables constant. HLA-B mismatch showed the greatest HED sensitivity, with RS varying from 73 to 97 (24.9-point difference), followed by HLA-DRB1 (23.2 points) and HLA-DQB1 (21.0 points), whereas HLA-A (16.6 points) and HLA-C (15.7 points) were less sensitive. For HLA-B, -DRB1, and -DQB1 mismatches, the lowest risk was consistently observed when both donor and recipient had high HED (P75) at the mismatched locus ( Figure S10 ). Implementation of a GRFS calculator To facilitate clinical implementation of these findings, we developed an interactive web-based calculator that integrates the trained RSF model with HED assessment. The tool accepts patient-specific clinical variables along with HLA typing data for both recipient and donor, used to automatically compute locus-specific HED values. The calculator outputs an individualized GRFS risk score with corresponding risk category and is freely available at https://huggingface.co/spaces/SPAGLIUCA/SMART ( Supplementary Appendix ). Discussion In this large, multicenter international study, we show that incorporating immunogenetic architecture information, captured by HED as a proxy of individual immunocompetence, into ML survival models yields modest but consistent, reproducible improvement in the prediction of GRFS after allo-HCT. The overall modest discrimination is expected for models originated from registry data. Neverthess, HED provided incremental prognostic signal across three distinct modeling families and enabled stable separation of patients into clinically interpretable risk strata in both internal and external cohorts, with compelling insight emerging from model-based simulations. Our results directly address a longstanding tension in transplantation immunology: the same donor–recipient immune interactions that underlie curative graft-versus-tumor effects can also fuel severe GvHD and immune-mediated morbidity. Disentangling beneficial from deleterious alloreactivity remains the elusive “holy grail” of the field, and classical risk models (largely anchored in clinical variables) have struggled to generalize across centers, platforms, and eras 34 . HED offers a biologically grounded, quantitative summary of HLA heterogeneity that moves beyond categorical matching and allele-level analyses, capturing the functional landscape of antigen presentation. In our data, locus-level HED features, particularly within class II, ranked among the most informative immunogenetic contributors, aligning with the central role of CD4-driven alloreactivity in GvHD biology and immune orchestration. Importantly, the time-varying associations observed in conventional Cox analyses reinforce that immunogenetic effects likely unfold dynamically, interacting with evolving immune reconstitution, infections, and post-transplant interventions. The non-linear “Goldilocks” effect observed in matched settings is biologically plausible and conceptually instructive. Very low HED (including complete homozygosity) may reflect a constrained peptide-binding breadth, potentially limiting immune surveillance, delaying immune reconstitution, or narrowing the repertoire of protective and antitumor responses. Conversely, very high HED may broaden antigenic space to a degree that amplifies alloreactive potential and inflammatory tissue injury, particularly in environments already primed by conditioning-related damage and microbial translocation. The stability of this U-shaped relationship across clinically distinct reference profiles suggests that HED is not merely a surrogate for baseline clinical risk but appears to act as an additive, orthogonal layer modulating outcome trajectories across transplant platforms. In haploidentical transplantation, where HLA disparity is structurally embedded, our simulations identified a striking “match similar with silmilar” principle: recipients achieved optimal predicted outcomes when paired with donors with comparable HED levels. This observation again suggests that the relative «immunogenetic distance» between donor and recipient (not only the presence of mismatch) may shape the balance between immune tolerance and immune aggression. One mechanistic hypothesis is that concordant HED profiles may harmonize the breadth of peptide presentation on both sides of the alloreactive interface, potentially reducing asymmetric immune recognition that could otherwise exacerbate GvH reactions or impair coordinated immune reconstitution. Conversely, discordant pairings (particularly low-HED recipients with high-HED donors) may create a steep gradient in antigen-presenting diversity, increasing the likelihood of broad donor T-cell activation against recipient targets without proportional gains in protective immunity. The 9/10 mismatched unrelated donor setting offered additional granularity, highlighting locus-dependent HED sensitivity in outcome risk prediction. HLA-B mismatches displayed the largest modulation of predicted risk across the HED spectrum, with substantial, yet directional, risk shifts dependent on donor and recipient divergence at the mismatched locus. That said, GRFS aggregates relapse, GvHD, and mortality into a single outcome, and different components may be differentially influenced by HED across loci and transplant platforms. Dissecting these competing forces will be essential to translate HED-informed strategies into practical donor selection rules tailored to the clinical priority (e.g., relapse prevention versus toxicity minimization). Several aspects of our results warrant cautious interpretation. First, c-index values remained below 0.60 across models, underscoring the inherent difficulty of predicting GRFS using baseline variables alone. GRFS is shaped not only by pre-transplant risk but also by post-transplant stochastic events, center-specific practices, infections, immune reconstitution kinetics, treatment adaptations, and evolving supportive care standards, many of which are not captured in registry-grade baseline datasets. In this context, a small but consistent improvement across algorithms is meaningful, particularly when it derives from biologically interpretable features and yields stable risk stratification across cohorts. Second, our external validation was necessarily restricted to complete cases, and missingness likely attenuated both model performance and the precision of HED-derived inferences. Third, although we intentionally avoided recalibration in the US cohort to provide a stringent generalizability test, center-level differences in practice and endpoint ascertainment may still influence absolute risk estimates. Notwithstanding these limitations, our study has notable strengths: the scale and heterogeneity of the dataset, the use of high-resolution HLA typing enabling locus-specific HED computation, the systematic comparison of complementary survival learners, and the application of a strict held-out internal test set alongside external validation. Importantly, we complemented “black-box” prediction with mechanistic-oriented simulations that transform model coefficients and interactions into clinically intuitive risk landscapes. This approach reframes immunogenetics from a static matching metric into a modifiable parameter that can be explored (and potentially optimized) at the time of donor selection. Looking forward, several directions emerge. Integrating additional immunogenetic layers (DPB1 when available, permissiveness metrics, KIR ligands, donor-specific antibodies) and disease/host biology (MRD status, cytogenetics/molecular risk, inflammatory markers) may raise performance. Such an rationally selected donor or donor/recipient immune profile invokes a definition of a transplant « immunome » as a sum of immunologic vectors determining alloreactivity. Furthermore, incorporating early post-transplant dynamics (e.g., immune reconstitution, viral reactivation, cytokine trajectories, and treatment intensification) could enable genuinely time-updated prediction that reflects the evolving biology of alloreactivity. Prospective studies are needed to test whether HED-guided donor selection can improve outcomes beyond standard matching, particularly in donor-rich settings where multiple acceptable donors exist. Ultimately, the goal is not to replace clinical judgment, but to equip it with a more faithful representation of immunogenetic diversity, helping clinicians navigate the narrow therapeutic corridor between graft-versus-tumor benefit and transplant-related toxicity. Authorship contributions SP and JM conceptualized the study, designed the analytical framework, and interpreted the data. SP developed the study synopsis, designed the study and statistical analysis plan, performed the biostatistical and bioinformatic analyses, created the visualizations, and wrote the manuscript. VA gave important inputs on the modeling design and provided patient data. MG contributed to data management. NR, AH, and RD coordinated data collection at the SFGM-TC registry level. AD, AK, LG, MJ, SKB, and FF performed data collection at the US center level. MR, IYA, EF, CEB, JBM, PC, CCL, XP, ML, JM, SN, FB, ED, NM, FM, MDA and MTR were involved in patient recruitment and performed data collection through the SFGM-TC registry. AA contributed to HLA typing quality control and immunogenetic data interpretation. MDA and MTR participated in study conception and interpretation of the data analysis, and provided important intellectual inputs. VV, TL, CG, and JM participated in the interpretation of the data analysis and provided important intellectual inputs and edited the manuscript. All authors reviewed and approved the final version of this manuscript. Declarations Authorship contributions SP and JM conceptualized the study, designed the analytical framework, and interpreted the data. SP developed the study synopsis, designed the study and statistical analysis plan, performed the biostatistical and bioinformatic analyses, created the visualizations, and wrote the manuscript. VA gave important inputs on the modeling design and provided patient data. MG contributed to data management. NR, AH, and RD coordinated data collection at the SFGM-TC registry level. AD, AK, LG, MJ, SKB, and FF performed data collection at the US center level. MR, IYA, EF, CEB, JBM, PC, CCL, XP, ML, JM, SN, FB, ED, NM, FM, MDA and MTR were involved in patient recruitment and performed data collection through the SFGM-TC registry. AA contributed to HLA typing quality control and immunogenetic data interpretation. MDA and MTR participated in study conception and interpretation of the data analysis, and provided important intellectual inputs. VV, TL, CG, and JM participated in the interpretation of the data analysis and provided important intellectual inputs and edited the manuscript. All authors reviewed and approved the final version of this manuscript. Conflict-of-interest disclosure SP has received travel expenses or honoraria for participation in advisory boards, symposia or other scientific events by Alexion, Novartis, Jazz Pharmaceutical, Sobi as well as research funding by JANSSEN HORIZON program. TLL is co-inventor on a patent application for using HED as a prognostic marker for immunotherapy success. Data sharing The data underlying this study were obtained from clinical registries and cannot be made publicly available because of institutional, ethical, and data protection restrictions. The authors are therefore unable to share individual-level data directly. Access to data from the SFGM-TC registry requires a formal request to the SFGM-TC study office. Requests regarding data from the US centers should be directed to Jaroslaw Maciejewski at [email protected] . The SMART model is hosted at https://huggingface.co/spaces/SPAGLIUCA/SMART. Acknowledgements This work was supported by SFGM-TC We acknowledge all participating centers. TLL was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – 437857095. SP received funding from Fondation ARC pour la Recherche sur le Cancer and Force Hemato. References Passweg JR, Baldomero H, Atlija M et al (2025) The 2023 EBMT report on hematopoietic cell transplantation and cellular therapies. Increased use of allogeneic HCT for myeloid malignancies and of CAR-T at the expense of autologous HCT. Bone Marrow Transpl 60(4):519–528. 10.1038/s41409-025-02524-2 Passweg JR, Baldomero H, Chabannon C et al Hematopoietic cell transplantation and cellular therapy survey of the EBMT: monitoring of activities and trends over 30 years. Bone Marrow Transpl. Published online 23 February 2021. 10.1038/s41409-021-01227-8 Trunk AD, Guo M, Budvytyte L et al (2024) Hematopoietic Stem-Cell Transplantation: Exploring the Latest Advances and Gaps in Disparities, Psychosocial and Symptom Management Interventions, and Chronic Graft-Versus-Host Disease Care. Am Soc Clin Oncol Educ Book 44(3):e432186. 10.1200/EDBK_432186 O’Neill AT, Chakraverty R (2021) Graft Versus Leukemia: Current Status and Future Perspectives. JCO 39(5):361–372. 10.1200/JCO.20.01801 Zeiser R, Blazar BR (2017) Acute Graft-versus-Host Disease — Biologic Process, Prevention, and Therapy. Longo DL, ed. N Engl J Med . ;377(22):2167–2179. 10.1056/NEJMra1609337 Malard F, Holler E, Sandmaier BM, Huang H, Mohty M (2023) Acute graft-versus-host disease. Nat Rev Dis Primers 9(1):27. 10.1038/s41572-023-00438-1 Pagliuca S, Michonneau D, Sicre de Fontbrune F et al (2019) Allogeneic reactivity-mediated endothelial cell complications after HSCT: a plea for consensual definitions. Blood Adv 3(15):2424–2435. 10.1182/bloodadvances.2019000143 Davies JK, Taussig D, Oakervee H et al (2013) Long-term survival with low toxicity after allogeneic transplantation for acute myeloid leukaemia and myelodysplasia using non-myeloablative conditioning without T cell depletion. Br J Haematol 162(4):525–529. 10.1111/bjh.12402 Pagliuca S, Prata PH, Xhaard A et al Long-term outcomes and risk factor analysis of steroid-refractory graft versus host disease after hematopoietic stem cell transplantation. Bone Marrow Transpl. Published online 25 June 2020. 10.1038/s41409-020-0977-3 Negrin RS (2015) Graft-versus-host disease versus graft-versus-leukemia. Hematol Am Soc Hematol Educ Program 2015:225–230. 10.1182/asheducation-2015.1.225 Van Bergen CAM, Van Luxemburg-Heijs SAP, De Wreede LC et al (2017) Selective graft-versus-leukemia depends on magnitude and diversity of the alloreactive T cell response. J Clin Invest 127(2):517–529. 10.1172/JCI86175 Lapoutge A, Gaudfrin M, Raus N et al (2025) Hallmarks of allogeneic hematopoietic cell transplant success: A study on immune signatures and predictive features of graft versus host disease-free remission states from the SFGM-TC registry and cryostem biobank. Blood 146(Supplement 1):2476–2476. 10.1182/blood-2025-2476 Shouval R, Fein JA, Shouval A et al (2019) External validation and comparison of multiple prognostic scores in allogeneic hematopoietic stem cell transplantation. Blood Adv 3(12):1881–1890. 10.1182/bloodadvances.2019032268 Saccardi R, Putter H, Eikema DJ et al (2023) Benchmarking of survival outcomes following Haematopoietic Stem Cell Transplantation (HSCT): an update of the ongoing project of the European Society for Blood and Marrow Transplantation (EBMT) and Joint Accreditation Committee of ISCT and EBMT (JACIE). Bone Marrow Transpl 58(6):659–666. 10.1038/s41409-023-01924-6 Awada H, Gurnari C, Durmaz A, Awada H, Pagliuca S, Visconte V (2022) Personalized Risk Schemes and Machine Learning to Empower Genomic Prognostication Models in Myelodysplastic Syndromes. Int J Mol Sci 23(5):2802. 10.3390/ijms23052802 Durmaz A, Gurnari C, Hershberger CE et al (2023) A multimodal analysis of genomic and RNA splicing features in myeloid malignancies. iScience 26(3):106238. 10.1016/j.isci.2023.106238 Kewan T, Durmaz A, Bahaj W et al (2023) Molecular patterns identify distinct subclasses of myeloid neoplasia. Nat Commun 14(1):3136. 10.1038/s41467-023-38515-4 Shouval R, Fein JA, Savani B, Mohty M, Nagler A (2021) Machine learning and artificial intelligence in haematology. Br J Haematol 192(2):239–250. 10.1111/bjh.16915 Shreve J, Meggendorfer M, Awada H et al (2019) A Personalized Prediction Model to Risk Stratify Patients with Acute Myeloid Leukemia (AML) Using Artificial Intelligence. Blood 134(Supplement1):2091–2091. 10.1182/blood-2019-128066 Romero M, Mosquera Orgueira A, Mejía Saldarriaga M (2024) How artificial intelligence revolutionizes the world of multiple myeloma. Front Hematol 3:1331109. 10.3389/frhem.2024.1331109 Alcazer V, Le Meur G, Roccon M et al (2024) Evaluation of a machine-learning model based on laboratory parameters for the prediction of acute leukaemia subtypes: a multicentre model development and validation study in France. Lancet Digit Health 6(5):e323–e333. 10.1016/S2589-7500(24)00044-X Arrieta-Bolaños E (2025) Differential Alloreactivity: Lessons Learned From a Singular HLA Locus. HLA 106(6):e70489. 10.1111/tan.70489 Pagliuca S, Gurnari C, Rubio MT, Visconte V, Lenz TL (2022) Individual HLA heterogeneity and its implications for cellular immune evasion in cancer and beyond. Front Immunol 13:944872. 10.3389/fimmu.2022.944872 Pierini F, Lenz TL (2018) Divergent Allele Advantage at Human MHC Genes: Signatures of Past and Ongoing Selection. Wilke C, ed. Molecular Biology and Evolution . ;35(9):2145–2158. 10.1093/molbev/msy116 Chowell D, Krishna C, Pierini F et al (2019) Evolutionary divergence of HLA class I genotype impacts efficacy of cancer immunotherapy. Nat Med 25(11):1715–1720. 10.1038/s41591-019-0639-4 Daull AM, Dubois V, Labussière-Wallet H et al (2022) Class I/Class II HLA Evolutionary Divergence Ratio Is an Independent Marker Associated With Disease-Free and Overall Survival After Allogeneic Hematopoietic Stem Cell Transplantation for Acute Myeloid Leukemia. Front Immunol 13:841470. 10.3389/fimmu.2022.841470 Pagliuca S, Gurnari C, Hong S et al (2021) Clinical and basic implications of dynamic T cell receptor clonotyping in hematopoietic cell transplantation. JCI Insight 6(13):e149080. 10.1172/jci.insight.149080 Crivello P, Arrieta-Bolaños E, He M et al (2023) Impact of the HLA Immunopeptidome on Survival of Leukemia Patients After Unrelated Donor Transplantation. J Clin Oncol 41(13):2416–2427. 10.1200/JCO.22.01229 Toulemonde C, Dubois V, Bouvier C et al (2025) HLA diversity is associated with TKI response and treatment-free remission in chronic myeloid leukemia. HemaSphere 9(11):e70261. 10.1002/hem3.70261 Shouval R, Labopin M, Unger R et al (2016) Prediction of Hematopoietic Stem Cell Transplantation Related Mortality- Lessons Learned from the In-Silico Approach: A European Society for Blood and Marrow Transplantation Acute Leukemia Working Party Data Mining Study. Bertolini F, ed. PLoS ONE . ;11(3):e0150637. 10.1371/journal.pone.0150637 Mussetti A, Rius-Sansalvador B, Moreno V et al (2024) Artificial intelligence methods to estimate overall mortality and non-relapse mortality following allogeneic HCT in the modern era: an EBMT-TCWP study. Bone Marrow Transpl 59(2):232–238. 10.1038/s41409-023-02147-5 Iwasaki M, Kanda J, Arai Y et al (2022) Establishment of a predictive model for GVHD-free, relapse-free survival after allogeneic HSCT using ensemble learning. Blood Adv 6(8):2618–2627. 10.1182/bloodadvances.2021005800 Holtan SG, DeFor TE, Lazaryan A et al (2015) Composite end point of graft-versus-host disease-free, relapse-free survival after allogeneic hematopoietic cell transplantation. Blood 125(8):1333–1338. 10.1182/blood-2014-10-609032 Negrin RS (2015) Graft-versus-host disease versus graft-versus-leukemia. Hematology 2015(1):225–230. 10.1182/asheducation-2015.1.225 Additional Declarations The authors declare potential competing interests as follows: SP has received travel expenses or honoraria for participation in advisory boards, symposia or other scientific events by Alexion, Novartis, Jazz Pharmaceutical, Sobi as well as research funding by JANSSEN HORIZON program. TLL is co-inventor on a patent application for using HED as a prognostic marker for immunotherapy success. Supplementary Files Supplementaryappendixnv.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9383323","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":621370936,"identity":"89ccf46d-eb5d-459b-9b20-687b796873b6","order_by":0,"name":"Simona Pagliuca","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+0lEQVRIiWNgGAWjYNCCAgkgwXyAGUlIjoAWA5AWtgRkLcaEtIAIHgPitMg38B78+MXAIrG//czHz4V7DjOYs589+JmHwSAfp/kH+JKlZQwkEmecyd0sPePZYQbLnrxkaaAWywacTuIxkJYAatnAkLuNmefA4foNB3LMmHkY/hjgdhiP8W+wFv43z0BaGAzOvwFpMcCpheEAj5nkB5AWiRw2iJYbOfi1GBzmS7MGBrLxjBvPjKV5DqQDtbwxlpxjgFuLfHvv4Zs/Kupk+/uTH37mOQDUfj7H8MObCjwOA7oBiNAAEw9uDUDAw8D4A10MU2QUjIJRMApGMgAAlKZMlI2x5k0AAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0003-4688-2478","institution":"Department of Hematology, CHRU de Nancy, Vandœuvre-lès-Nancy, France and UMR7365-CNRS, IMOPA, University of Lorraine, Vandœuvre-lès-Nancy, France","correspondingAuthor":true,"prefix":"","firstName":"Simona","middleName":"","lastName":"Pagliuca","suffix":""},{"id":621444056,"identity":"cc05203d-f361-4f10-a98d-bdda62ba0f5f","order_by":1,"name":"Vincent Alcazer","email":"","orcid":"","institution":"Service d’Hématologie Clinique, Centre Hospitalier Lyon Sud, Hospices Civils de Lyon, Pierre Bénite, France / Université Claude Bernard Lyon 1, CIRI INSERM U1111 - CNRS UMR5308","correspondingAuthor":false,"prefix":"","firstName":"Vincent","middleName":"","lastName":"Alcazer","suffix":""},{"id":621444057,"identity":"1096cb68-b077-4a4d-be06-63a45107c51e","order_by":2,"name":"Mélanie Gaudfrin","email":"","orcid":"","institution":"MR7365-CNRS, IMOPA, University of Lorraine, Vandœuvre-lès-Nancy, France","correspondingAuthor":false,"prefix":"","firstName":"Mélanie","middleName":"","lastName":"Gaudfrin","suffix":""},{"id":621444058,"identity":"133068d8-481a-4227-ab3c-9b4a97e8cc44","order_by":3,"name":"Nicole Raus","email":"","orcid":"","institution":"Société Francophone de Greffe de Moelle et de Thérapie Cellulaire (SFGM-TC) Centre Hospitalier Lyon Sud Pierre-Bénite France.","correspondingAuthor":false,"prefix":"","firstName":"Nicole","middleName":"","lastName":"Raus","suffix":""},{"id":621444059,"identity":"532e46a8-3741-45d0-a103-dcff87ad814f","order_by":4,"name":"Anne Huynh","email":"","orcid":"","institution":"Hematology unit, Institut Universitaire de Cancerologie de Toulouse, Oncopole, Toulouse, France.","correspondingAuthor":false,"prefix":"","firstName":"Anne","middleName":"","lastName":"Huynh","suffix":""},{"id":621444060,"identity":"70728a57-f99a-4280-aa91-107ed9813c5b","order_by":5,"name":"Raynier Devillier","email":"","orcid":"","institution":"Programme de Transplantation \u0026 Therapie Cellulaire, Marseille, France.","correspondingAuthor":false,"prefix":"","firstName":"Raynier","middleName":"","lastName":"Devillier","suffix":""},{"id":621444061,"identity":"2edfae72-023a-4c6b-89ae-6524d7a97621","order_by":6,"name":"Arda Dumaz","email":"","orcid":"","institution":"Department of Translational Hematology and Oncology Research, Cleveland Clinic, Cleveland, OH, USA.","correspondingAuthor":false,"prefix":"","firstName":"Arda","middleName":"","lastName":"Dumaz","suffix":""},{"id":621444062,"identity":"5f5889b0-3934-40a0-8de5-1be20ff2abff","order_by":7,"name":"Ashwin Kishtagari","email":"","orcid":"","institution":"Division of Hematology Oncology, Department of Medicine, Vanderbilt University Medical Center, Nashville, Tennessee, USA","correspondingAuthor":false,"prefix":"","firstName":"Ashwin","middleName":"","lastName":"Kishtagari","suffix":""},{"id":621444063,"identity":"9de0c38e-d9f4-431f-b397-3152a03e4619","order_by":8,"name":"Lukasz Gondek","email":"","orcid":"","institution":"Sidney Kimmel Comprehensive Cancer Center, Johns Hopkins University School of Medicine, Baltimore, MD, USA","correspondingAuthor":false,"prefix":"","firstName":"Lukasz","middleName":"","lastName":"Gondek","suffix":""},{"id":621444064,"identity":"9006b13e-8b3f-4fc4-9f9d-4976bc6a39c9","order_by":9,"name":"Mark Juckett","email":"","orcid":"","institution":"Department of Hematology, University of Minnesota, Minneapolis, USA","correspondingAuthor":false,"prefix":"","firstName":"Mark","middleName":"","lastName":"Juckett","suffix":""},{"id":621444065,"identity":"c32793c0-60ab-458f-a812-a1489f2e467e","order_by":10,"name":"Suresh Kumar Balasubramanian","email":"","orcid":"","institution":"Department of Oncology, Karmanos Cancer Institute/ Wayne State University, Detroit, MI, USA.","correspondingAuthor":false,"prefix":"","firstName":"Suresh","middleName":"Kumar","lastName":"Balasubramanian","suffix":""},{"id":621444066,"identity":"3ebe0f0b-e142-449d-bb40-400d25b325fd","order_by":11,"name":"Marie Robin","email":"","orcid":"","institution":"Hematology and Transplantation Unit, Saint Louis Hospital, Paris, France.","correspondingAuthor":false,"prefix":"","firstName":"Marie","middleName":"","lastName":"Robin","suffix":""},{"id":621444067,"identity":"87c7b04a-6a78-4bc6-acfd-50da8968e195","order_by":12,"name":"Ibrahim Yakoub-Agha","email":"","orcid":"","institution":"CHU de Lille, Univ Lille, INSERM U1286, Infinite, 59000, Lille, France.","correspondingAuthor":false,"prefix":"","firstName":"Ibrahim","middleName":"","lastName":"Yakoub-Agha","suffix":""},{"id":621444068,"identity":"f81836c5-1650-423d-80b5-e63a7dd620e8","order_by":13,"name":"Edouard Forcade","email":"","orcid":"","institution":"Service d'Hématologie Clinique et Thérapie Cellulaire, CHU Bordeaux, F-33000, Bordeaux, France.","correspondingAuthor":false,"prefix":"","firstName":"Edouard","middleName":"","lastName":"Forcade","suffix":""},{"id":621444069,"identity":"f05b1213-2a8f-4029-93d1-81727013e623","order_by":14,"name":"Claude Eric Bulabois","email":"","orcid":"","institution":"CHU Grenoble Alpes-Université Grenoble Alpes, Grenoble, France.","correspondingAuthor":false,"prefix":"","firstName":"Claude","middleName":"Eric","lastName":"Bulabois","suffix":""},{"id":621444070,"identity":"903d7418-9d39-46f8-aa26-1e84433ddff3","order_by":15,"name":"Jean-Baptiste Méar","email":"","orcid":"","institution":"Centre Hospitalier Universitaire de Rennes, Rennes, France.","correspondingAuthor":false,"prefix":"","firstName":"Jean-Baptiste","middleName":"","lastName":"Méar","suffix":""},{"id":621444071,"identity":"9d1ff7c5-ee8f-47e2-9463-666649555577","order_by":16,"name":"Patrice Chevallier","email":"","orcid":"","institution":"Hematology Department, CHU Hotel-Dieu, Nantes, France.","correspondingAuthor":false,"prefix":"","firstName":"Patrice","middleName":"","lastName":"Chevallier","suffix":""},{"id":621444072,"identity":"a6167141-c183-4af6-a3d9-ca60e81d332a","order_by":17,"name":"Cristina Castilla Llorente","email":"","orcid":"","institution":"Gustave Roussy Cancer Campus Grand Paris, Hematology Department, Villejuif, France.","correspondingAuthor":false,"prefix":"","firstName":"Cristina","middleName":"Castilla","lastName":"Llorente","suffix":""},{"id":621444073,"identity":"47ec2791-4157-4a18-86de-1b7155633a7f","order_by":18,"name":"Xavier Poiré","email":"","orcid":"","institution":"Department of Hematology, Institut Roi Albert II, Cliniques Universitaires St-Luc, Brussels, Belgium.","correspondingAuthor":false,"prefix":"","firstName":"Xavier","middleName":"","lastName":"Poiré","suffix":""},{"id":621444074,"identity":"b543ee87-3dbc-4d80-889c-fbe99ad718d6","order_by":19,"name":"Michaël Loschi","email":"","orcid":"","institution":"Department of Hematology, Hôpital L'Archet, Centre Hospitalier Universitaire de Nice, Nice, France.","correspondingAuthor":false,"prefix":"","firstName":"Michaël","middleName":"","lastName":"Loschi","suffix":""},{"id":621444075,"identity":"2bff82e9-778e-44ab-980d-a9fa95d94896","order_by":20,"name":"Johan Maertens","email":"","orcid":"","institution":"Department of Haematology, University Hospitals Leuven, Leuven, Belgium.","correspondingAuthor":false,"prefix":"","firstName":"Johan","middleName":"","lastName":"Maertens","suffix":""},{"id":621444076,"identity":"c7bbf241-9ae6-4648-b888-8a6b5e2817a7","order_by":21,"name":"Stéphanie Nguyen","email":"","orcid":"","institution":"Department of Hematology, Centre Hospitalier Universitaire Pitié-Salpêtrière, Assistance Publique-Hôpitaux de Paris, Paris, France.","correspondingAuthor":false,"prefix":"","firstName":"Stéphanie","middleName":"","lastName":"Nguyen","suffix":""},{"id":621444077,"identity":"efb9e20c-869d-4974-b8ba-dce73191ae35","order_by":22,"name":"Frédéric Baron","email":"","orcid":"","institution":"Department of Hematology, CHU Sart-Tilman University of Liege, Liege, Belgium.","correspondingAuthor":false,"prefix":"","firstName":"Frédéric","middleName":"","lastName":"Baron","suffix":""},{"id":621444078,"identity":"9b838388-d316-41e7-908b-4d31b2a73b7a","order_by":23,"name":"Etienne Daguindau","email":"","orcid":"","institution":"Department of Hematology, Jean Minjoz University Hospital, Besancon, France.","correspondingAuthor":false,"prefix":"","firstName":"Etienne","middleName":"","lastName":"Daguindau","suffix":""},{"id":621444079,"identity":"c63b10e8-46c8-4740-b86a-9cae4eaa4fee","order_by":24,"name":"Natacha Maillard","email":"","orcid":"","institution":"Hematology unit, Poitiers university hospital, Poitiers, France.","correspondingAuthor":false,"prefix":"","firstName":"Natacha","middleName":"","lastName":"Maillard","suffix":""},{"id":621444080,"identity":"24df482d-a337-424e-aa2b-bfd11e596641","order_by":25,"name":"Florent Malard","email":"","orcid":"","institution":"Department of Clinical Hematology and Cellular Therapy, Saint-Antoine Hospital, AP-HP, Sorbonne University, INSERM UMR 938, Centre de Recherche Saint-Antoine, Paris, France.","correspondingAuthor":false,"prefix":"","firstName":"Florent","middleName":"","lastName":"Malard","suffix":""},{"id":621444081,"identity":"d308df84-07e0-4007-8eff-d6ba61693502","order_by":26,"name":"Alice Aarnink","email":"","orcid":"","institution":"HLA and Histocompatibility Laboratory, CHRU de Nancy, Vandœuvre-lès-Nancy, France.","correspondingAuthor":false,"prefix":"","firstName":"Alice","middleName":"","lastName":"Aarnink","suffix":""},{"id":621444082,"identity":"437b87fc-38f5-4872-a3e5-58c3409661d3","order_by":27,"name":"Maud D’Aveni-Piney","email":"","orcid":"","institution":"Department of Hematology, CHRU de Nancy, Vandœuvre-lès-Nancy, France and UMR7365-CNRS, IMOPA, University of Lorraine, Vandœuvre-lès-Nancy, France","correspondingAuthor":false,"prefix":"","firstName":"Maud","middleName":"","lastName":"D’Aveni-Piney","suffix":""},{"id":621444083,"identity":"3ca9bb52-c9ec-40dc-9969-aa35991fcf97","order_by":28,"name":"Marie Thérèse Rubio","email":"","orcid":"","institution":"Department of Hematology, CHRU de Nancy, Vandœuvre-lès-Nancy, France and UMR7365-CNRS, IMOPA, University of Lorraine, Vandœuvre-lès-Nancy, France","correspondingAuthor":false,"prefix":"","firstName":"Marie","middleName":"Thérèse","lastName":"Rubio","suffix":""},{"id":621444084,"identity":"3398732d-6a29-44f6-991f-1c9ede1e101e","order_by":29,"name":"Francesca Ferraro","email":"","orcid":"","institution":"Division of Oncology, Department of Medicine, Washington University School of Medicine (WUSM), St. Louis, Missouri, USA.","correspondingAuthor":false,"prefix":"","firstName":"Francesca","middleName":"","lastName":"Ferraro","suffix":""},{"id":621444085,"identity":"d172f92c-f46c-4edd-b6d6-b4ef396e487b","order_by":30,"name":"Valeria Visconte","email":"","orcid":"","institution":"Department of Translational Hematology and Oncology Research, Cleveland Clinic, Cleveland, OH, USA","correspondingAuthor":false,"prefix":"","firstName":"Valeria","middleName":"","lastName":"Visconte","suffix":""},{"id":621444086,"identity":"2ecd0a68-9024-4393-a93e-53e355cd40d5","order_by":31,"name":"Tobias Lenz","email":"","orcid":"","institution":"Research Unit for Evolutionary Immunogenomics, Department of Biology, Universität Hamburg, Hamburg, Germany.","correspondingAuthor":false,"prefix":"","firstName":"Tobias","middleName":"","lastName":"Lenz","suffix":""},{"id":621444087,"identity":"8ac825d3-b317-4b9c-ba21-eb747d7972ad","order_by":32,"name":"Carmelo Gurnari","email":"","orcid":"","institution":"Department of Biomedicine and Prevention, University of Rome Tor Vergata, Rome, Italy and Department of Translational Hematology and Oncology Research, Cleveland Clinic, Cleveland, OH, USA","correspondingAuthor":false,"prefix":"","firstName":"Carmelo","middleName":"","lastName":"Gurnari","suffix":""},{"id":621444088,"identity":"49aeb911-b6e9-4b01-a000-86811f74f690","order_by":33,"name":"Jaroslaw Maciejewski","email":"","orcid":"","institution":"Department of Translational Hematology and Oncology Research, Cleveland Clinic, Cleveland, OH, USA","correspondingAuthor":false,"prefix":"","firstName":"Jaroslaw","middleName":"","lastName":"Maciejewski","suffix":""}],"badges":[],"createdAt":"2026-04-10 22:42:49","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":true,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9383323/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9383323/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107489541,"identity":"98e6c41a-04b9-4e88-8360-a6b8c2360d48","added_by":"auto","created_at":"2026-04-22 02:48:05","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":755372,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStudy design and overview of the SMART framework\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFrom the top left to the botton right : \u003cstrong\u003eI)\u003c/strong\u003eCohort assembly and split strategy. Model development was performed in the French SFGM-TC registry, with a training cohort (N=9,840) and an independent hold-out test cohort (N=4,139). External validation used a multicenter U.S. cohort (N=1,931) from five transplant centers (University of Minnesota, Vanderbilt University, Cleveland Clinic, Johns Hopkins University, and Karmanos Cancer Institute). \u003cstrong\u003eII)\u003c/strong\u003e End-to-end modeling pipeline. Clinical annotations and recipient/donor HLA genotypes were used to compute locus-level HED features (derived from polymorphic HLA exons; class I exons 2–3 and class II exon 2), which were combined with clinical covariates to train survival models (random survival forests [RSF], XGBoost survival, and elastic-net Cox). \u003cstrong\u003eIII) \u003c/strong\u003e\u0026nbsp;Model interpretation and performance assessment, including feature discrimination/importance, inter-model comparison of discrimination over time, risk-group stratification of patients, and risk simulation across covariate profiles. \u003cstrong\u003eIV) \u003c/strong\u003e\u0026nbsp;Translation to practice through the SMART-app : a user-facing GRFS calculator that accepts clinical variables and HLA inputs to generate individualized, dynamically updated GRFS risk estimates.\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9383323/v1/777e5d144995e318a4ddd5fd.jpeg"},{"id":107377545,"identity":"7f7b13dc-ba76-4800-9ab8-16af20cbbec4","added_by":"auto","created_at":"2026-04-21 01:30:41","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":756065,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAssociation between locus-specific HLA evolutionary divergence (HED) and transplant outcomes in multivariable survival models.\u003c/strong\u003e\u003cbr\u003e\n \u003cstrong\u003e(A)\u003c/strong\u003e Forest plots showing hazard ratios (HRs) and 95% confidence intervals from multivariable models adjusted for clinical covariates, evaluating the association of recipient HED at each HLA locus (HED_A, HED_B, HED_C, HED_DRB1, HED_DQB1) and aggregate indices (HED_Class_I_mean, HED_Class_II_mean, HED_Tot) with overall survival (OS), relapse, grade III–IV acute GvHD (aGvHD III–IV), and extensive chronic GvHD (cGvHD_ext). Point estimates represent HR per one-unit increase in HED; the vertical dashed line indicates HR=1. Colors denote statistical significance (NS, p\u0026lt;0.05, p\u0026lt;0.01).\u003cbr\u003e\n \u003cstrong\u003e(B)\u003c/strong\u003e Time-varying effects for selected HED predictors, displayed as estimated HR (solid line) with 95% confidence band (shaded area) across days after allo-HCT. Panels show HED_DQB1 for grade III–IV aGvHD, and HED_B for OS and relapse. The horizontal dashed line indicates HR=1; rugs along the x-axis indicate the distribution of observed event times.\u003c/p\u003e","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9383323/v1/61902a9026a097ca9ce2d747.jpeg"},{"id":107377546,"identity":"673450d4-f6b6-4284-98ef-f0dc80a184e9","added_by":"auto","created_at":"2026-04-21 01:30:41","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1186485,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eModel discrimination, calibration proxies, and feature importance for GRFS prediction with and without HED.\u003c/strong\u003e\u003cbr\u003e\n \u003cstrong\u003e(A)\u003c/strong\u003eLow-dimensional projection of the covariate space for the GRFS endpoint, with patients colored by occurrence of a GRFS event versus no event, illustrating overlap between outcome groups and the need for multivariable, non-linear risk modeling.\u003cbr\u003e\n \u003cstrong\u003e(B)\u003c/strong\u003e Overall discrimination (Harrell’s C-index) in the training set and the independent hold-out test set for each model family—random survival forest (RSF), XGBoost survival (XGB), and elastic-net Cox (Cox-EN) — trained using clinical variables only (“Clinical”, 9 variables) or the full feature set including HED (“Full”, 19 variables).\u003cbr\u003e\n \u003cstrong\u003e(C)\u003c/strong\u003eTime-dependent AUC in the SFGM-TC hold-out test cohort across prediction horizons, comparing clinical-only (dashed lines) versus full models including HED (solid lines) within each model family. Points and error bars represent estimates with uncertainty at each horizon.\u003cbr\u003e\n \u003cstrong\u003e(D)\u003c/strong\u003e Relative feature importance within each model (scaled so the top feature equals 1), highlighting the contribution of clinical predictors (red) and HED-derived immunogenetic features (yellow) across Cox-EN, RSF, and XGBoost models.\u003c/p\u003e","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9383323/v1/3a678cd7fd3ba41236e5eb89.jpeg"},{"id":107489540,"identity":"329fadb9-55db-4538-9e4c-8c0e8e9b06bf","added_by":"auto","created_at":"2026-04-22 02:48:04","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":2501199,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eRisk stratification by the RSF full-model GRFS risk score across development and validation cohorts.\u003c/strong\u003e\u003cbr\u003e\n \u003cstrong\u003e(A)\u003c/strong\u003eKaplan–Meier curves for graft-versus-host disease–free/relapse-free survival (GRFS) after allo-HCT, stratified into low-, intermediate-, and high-risk groups defined by the random survival forest (RSF) full-model risk score. Cut points were derived in the training cohort and applied unchanged to the hold-out test cohort and the external validation cohort. Numbers at risk are shown below each plot; between-group differences were assessed by the log-rank test.\u003cbr\u003e\n \u003cstrong\u003e(B)\u003c/strong\u003eDistribution of the continuous RSF risk score in each cohort, colored by assigned risk group. Vertical dashed lines indicate the predefined thresholds (low \u0026lt;103, intermediate 103–143, high ≥143). Insets report the number and proportion of patients in each risk category in the training, hold-out test, and external validation cohorts.\u003c/p\u003e","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9383323/v1/1c4e46f870e1df3ab8fd8757.jpeg"},{"id":107377548,"identity":"319e1224-324c-4c54-b711-9b42bf43992e","added_by":"auto","created_at":"2026-04-21 01:30:41","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":594121,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSimulated impact of recipient–donor HED interactions on RSF GRFS risk score in the matched setting.\u003c/strong\u003e\u003cbr\u003e\n \u003cstrong\u003e(A)\u003c/strong\u003eTwo-way partial dependence surface showing the joint effect of recipient and donor total HED on the RSF risk score. Colors indicate the predicted risk score (higher values = higher risk). The dashed contour delineates the boundaries between low and intermediate risk region, and the “optimal” point marks the combination associated with the lowest predicted risk within the observed range (corresponding to the 75° percentile for both donor and recipient HED).\u003cbr\u003e\n \u003cstrong\u003e(B)\u003c/strong\u003eOne-dimensional partial dependence of total HED (recipient = donor; sum across five loci) illustrating a non-linear, U-shaped relationship with the RSF risk score. Points indicate selected percentiles of the HED distribution; the shaded band denotes the low-risk region and the “optimal” percentile corresponds to the minimum predicted risk.\u003cbr\u003e\n \u003cstrong\u003e(C)\u003c/strong\u003ePredicted RSF risk score across total HED for representative clinical profiles, demonstrating that the magnitude and shape of the HED effect vary by baseline clinical risk (higher-risk, lower-risk, and typical AML reference profiles). Horizontal dashed lines indicate the predefined risk-score thresholds used for categorical stratification.\u003cbr\u003e\n \u003cstrong\u003e(D)\u003c/strong\u003eExample patient-level simulation for a reference clinical profile, showing the change in RSF risk score at illustrative HED levels (homozygous/very low, median, optimal, and high). Dashed horizontal lines indicate the low- and high-risk cutoffs, highlighting how immunogenetic diversity can potentially shift an otherwise fixed clinical profile across risk strata.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-9383323/v1/a6e5f79cf701dfa785946d02.png"},{"id":107377549,"identity":"1bf49d29-5a1b-408e-99b3-4db151b616f1","added_by":"auto","created_at":"2026-04-21 01:30:41","extension":"jpeg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":775532,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSimulated impact of recipient–donor HED interactions on RSF GRFS risk score in the haploidentical setting.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A)\u003c/strong\u003eTwo-way partial dependence surface illustrating the joint association of recipient and donor total HED with the RSF risk score in haploidentical transplantation. Colors represent the predicted risk score (higher values = higher risk). The dashed contour outlines the separation between low and intermediate region, and the “optimal” marker indicates the recipient–donor HED combination associated with the lowest predicted risk within the observed range.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(B)\u003c/strong\u003e Discretized interaction map showing predicted RSF risk scores across recipient and donor HED percentiles (total HED). Each cell reports the predicted risk score for the corresponding percentile pairing; the boxed cell highlights the percentile combination yielding the lowest predicted risk (both donor and recipinet HED around 75° percentile).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(C)\u003c/strong\u003e Predicted RSF risk score across donor HED percentiles stratified by recipient HED percentile, demonstrating that the donor HED associated with minimal predicted risk depends on the recipient’s immunogenetic diversity. The horizontal dashed line denotes the predefined risk-score threshold used for categorical risk stratification.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(D)\u003c/strong\u003e Estimated risk reduction achievable through donor selection, summarized for recipient HED percentiles. For each recipient percentile, dots indicate predicted risk at the least favorable versus most favorable donor HED percentile, and vertical segments represent the magnitude of improvement (annotated as Δ risk score). Together, these simulations suggest that, in haploidentical donor choice, optimizing donor HED relative to recipient HED may shift patients across clinically relevant risk-score thresholds.\u003c/p\u003e","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9383323/v1/e4bc20e43cc9a49c4203fe46.jpeg"},{"id":107490159,"identity":"b750e2db-c3fa-49a3-83a2-a869bfdce17f","added_by":"auto","created_at":"2026-04-22 02:50:53","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":7353625,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9383323/v1/1e91c6e1-8bcc-4a5f-89de-1b9bd34de35f.pdf"},{"id":107377544,"identity":"296f51bd-b358-48c8-b391-6fc6105f5301","added_by":"auto","created_at":"2026-04-21 01:30:41","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":8868961,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementaryappendixnv.docx","url":"https://assets-eu.researchsquare.com/files/rs-9383323/v1/9403fa61f2f1d2689979f37e.docx"}],"financialInterests":"The authors declare potential competing interests as follows: SP has received travel expenses or honoraria for participation in advisory boards, symposia or other scientific events by Alexion, Novartis, Jazz Pharmaceutical, Sobi as well as research funding by JANSSEN HORIZON program. TLL is co-inventor on a patent application for using HED as a prognostic marker for immunotherapy success. ","formattedTitle":"\u003cp\u003ePredictive Modeling of Post-Allogeneic Transplant Outcomes Using Machine Learning and Integrated Clinical and Immunogenetic Data: a Study from the SFGM-TC and a Multicenter US Consortium\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAllogeneic hematopoietic cell transplantation (allo-HCT), the most paradigmatic form of cellular immunotherapy, is currently the solely curative treatment for many hematological disorders. Transplant-related complications, notably disease recurrence, graft-versus-host disease (GvHD) and infections continue to hamper the therapeutic index of this procedure, limiting its broader applicability\u003csup\u003e\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Central to these complications is the complex interplay of donor-recipient alloreactivity and the balance between beneficial graft-versus-tumor (GvT) and harmful graft versus host (GvH) reactions\u003csup\u003e\u003cspan additionalcitationids=\"CR5 CR6 CR7 CR8 CR9\" citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Distinction between beneficial and deleterius alloreactivity remains the elusive \"holy grail\" of transplantation immunology essential for improving outcomes, expanding indications and implementation of innovative post-transplant immunomodulatory strategies\u003csup\u003e\u003cspan additionalcitationids=\"CR11\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Given the diversity of transplant platforms and patient populations across transplant centers and countries, traditional predictive models systematically fail to achieve generalizability, particularly when relying solely on clinical and demographic parameters\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. This limitation underscores the clinical need to develop robust, accurate, and predictive tools capable of intergrate the interplay between the \u0026laquo; beneficial \u0026raquo; and the \u0026laquo; deleterius \u0026raquo; alloreactivity. To that end, computational biology and machine learning (ML) tools offer promising solutions to resolve complexity of datasets and uncover predictive patterns otherwise invisible to classical statistics\u003csup\u003e\u003cspan additionalcitationids=\"CR16 CR17 CR18 CR19 CR20\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eAmong biological features which could improve prediction of alloreactive complications in the context of all-HCT, the immunogenetic background, particularly the reciprocal human leukocyte antigen (HLA) configurations between donor and recipient, emerges as a critical determinant of post-transplant outcomes\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. As robust measure of HLA heterogeneity, HLA Evolutionary Divergence (HED) concept has demonstrated a potential to effectively capture immunogenetic complexity in a readily interpretable manner, serving as a quantitative proxy for the degree of individual immunocompetence\u003csup\u003e\u003cspan additionalcitationids=\"CR24 CR25 CR26 CR27 CR28\" citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eGiven the limitations of models based solely on clinical variables to predict survival or mortality after allo-HCT\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e,\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e, we sought to develop a integrative computational framework to stratify the probability of GvHD-free remission as captured by GvHD-free/relapse-free survival (GRFS)\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e,\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. For that purpose, we leveraged a very large, multicenter international cohort of patients with hematologic malignancies comprising 15,000 individuals, comprehensively annotated for clinical characteristics and HLA-related variables. We first quantified the contribution of individual immunogenetic interactions to clinical phenotypes and outcomes in a dedicated training cohort, and then trained GRFS prediction models to incorporate clinical and immunogenetic input using complementary ML-approaches. Model performance was subsequently evaluated in an internal hold-out test set and in an independent external validation cohort. Ultimately, our goal was to create an actionable risk stratification tool and to provide a rational basis for tailoring transplant platforms and post-transplant immunomodulatory strategies to individual patient\u0026ndash;donor profiles.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eStudy design, cohorts, and eligibility\u003c/p\u003e \u003cp\u003eWe conducted a multicenter, international retrospective cohort study of 16,028 adults undergoing a first allogeneic hematopoietic cell transplant (allo-HCT) for hematologic malignancies between 2010\u0026ndash;2022. Model development used the Societe Francophone de Greffe de Moelle et de Therapie Cellulaire (SFGM-TC) registry (N\u0026thinsp;=\u0026thinsp;13,979), randomly split 70:30 into a training set (N\u0026thinsp;=\u0026thinsp;9,840) and a held-out internal test set (N\u0026thinsp;=\u0026thinsp;4,139) with stratification to preserve balance across key clinical and immunogenetic characteristics (\u003cb\u003eSupplementary methods\u003c/b\u003e). External validation was performed in an independent US cohort (N\u0026thinsp;=\u0026thinsp;1,931) from five centers (Cleveland Clinic, Johns Hopkins, Vanderbilt, University of Minnesota, and Karmanos Center), with 616 (32%) having complete data for all model predictors (\u003cb\u003eSupplementary methods\u003c/b\u003e, Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, \u003cb\u003eTable S1 and S2\u003c/b\u003e). Eligible patients were \u0026ge;\u0026thinsp;18 years, received a first allo-HCT for a malignant hematological indication, had high-resolution HLA typing for A, B, C, DRB1, DQB1, and had available outcome data. Patients undergoing second transplants, transplanted for non-malignant diseases, or lacking required high-resolution HLA data were excluded. The study was approved by institutional review boards at all participating centers and was approved by the scientific committee of the SFGM-TC in april 2025. This study was declared to the Health Data Hub under the number 24237877.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePatient and Transplant Characteristics from SFGM-TC cohort\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003e Characteristic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining Cohort\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;9,840)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTest Cohort\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;4,139)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePatient Characteristics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePatient age, years, median (IQR)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e56.1 (44.5\u0026ndash;63.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e56.2 (44.3\u0026ndash;63.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePatient sex, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5,742 (58.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2,426 (58.6%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4,095 (41.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,712 (41.4%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePatient ABO blood group, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4,042 (41.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,649 (39.8%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e898 (9.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e382 (9.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e364 (3.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e152 (3.7%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eO\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3,998 (40.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,665 (40.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePatient CMV serostatus, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5,338 (54.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2,172 (52.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4,461 (45.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,950 (47.1%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDisease Characteristics\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiagnosis, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAcute myeloid leukemia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4,442 (45.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,885 (45.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAcute lymphoblastic leukemia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,148 (11.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e474 (11.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMyelodysplastic syndrome\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,353 (13.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e561 (13.6%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMyeloproliferative neoplasm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,150 (11.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e502 (12.1%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMDS/MPN overlap\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e73 (0.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e36 (0.9%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eB-cell non-Hodgkin lymphoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e502 (5.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e187 (4.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eT-cell non-Hodgkin lymphoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e392 (4.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e164 (4.0%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHodgkin lymphoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e267 (2.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e103 (2.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCLL/PLL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e223 (2.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e96 (2.3%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMultiple myeloma/PCL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e290 (2.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e131 (3.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDisease status at transplant, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComplete remission\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e6,262 (63.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2,610 (63.1%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eActive disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3,232 (32.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,355 (32.7%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDonor Characteristics\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDonor age, years, median (IQR)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e34.0 (26.0\u0026ndash;45.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e34.0 (26.0\u0026ndash;45.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDonor type, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMatched sibling donor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,933 (19.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e808 (19.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMatched unrelated donor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5,085 (51.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2,165 (52.3%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMismatched unrelated donor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,248 (12.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e521 (12.6%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHaploidentical donor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,451 (14.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e605 (14.6%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCord blood\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e123 (1.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e40 (1.0%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDonor ABO blood group, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3,802 (38.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,640 (39.6%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e920 (9.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e367 (8.9%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e387 (3.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e151 (3.6%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eO\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4,202 (42.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,705 (41.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDonor CMV serostatus, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4,204 (42.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,750 (42.3%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5,604 (57.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2,374 (57.4%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTransplant Characteristics\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStem cell source, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePeripheral blood\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8,426 (85.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3,537 (85.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBone marrow\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,291 (13.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e562 (13.6%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCord blood\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e123 (1.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e40 (1.0%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConditioning regimen, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMyeloablative conditioning\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3,453 (35.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,440 (34.8%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eReduced-intensity conditioning\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e6,387 (64.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2,699 (65.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGvHD prophylaxis, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCNI alone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,418 (14.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e590 (14.3%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCNI\u0026thinsp;+\u0026thinsp;mycophenolate mofetil\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5,091 (51.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2,105 (50.9%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCNI\u0026thinsp;+\u0026thinsp;methotrexate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2,958 (30.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,294 (31.3%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e219 (2.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e88 (2.1%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIn vivo T-cell depletion, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eATG\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e6,238 (63.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2,615 (63.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePost-transplant cyclophosphamide\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,513 (15.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e652 (15.8%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eATG\u0026thinsp;+\u0026thinsp;PTCy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e209 (2.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e76 (1.8%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e21 (0.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e7 (0.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eHLA Mismatches\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHLA-A mismatch, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7,913 (80.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3,368 (81.4%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,686 (17.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e677 (16.4%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e241 (2.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e94 (2.3%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHLA-B mismatch, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8,061 (81.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3,403 (82.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,552 (15.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e657 (15.9%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e227 (2.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e79 (1.9%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHLA-C mismatch, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7,977 (81.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3,374 (81.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,600 (16.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e673 (16.3%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e263 (2.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e92 (2.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHLA-DRB1 mismatch, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8,150 (82.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3,434 (83.0%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,453 (14.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e607 (14.7%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e237 (2.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e98 (2.4%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHLA-DQB1 mismatch, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8,274 (84.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3,464 (83.7%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,348 (13.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e580 (14.0%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e218 (2.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e95 (2.3%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHLA-DPB1 mismatch, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5,620 (57.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2,316 (56.0%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2,229 (22.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e981 (23.7%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,991 (20.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e842 (20.3%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eHLA typing and HLA evolutionary divergence (HED)\u003c/p\u003e \u003cp\u003eHigh-resolution HLA typing (minimum 2-field/4-digit) was performed by accredited histocompatibility laboratories using sequence-based methods. HED was computed, as previously shown, at each locus using a Grantham distance\u0026ndash;based approach quantifying physicochemical divergence across the antigen-binding domain between the two alleles carried by an individual\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e,\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. HED was derived for A, B, C, DRB1, DQB1 (and DPB1 when available), separately for recipients and donors. Summary measures included Class I HED (A\u0026thinsp;+\u0026thinsp;B+C), Class II HED (DRB1\u0026thinsp;+\u0026thinsp;DQB1), and Total HED (A\u0026thinsp;+\u0026thinsp;B+C+DRB1\u0026thinsp;+\u0026thinsp;DQB1). Delta HED was defined as donor minus recipient total HED.\u003c/p\u003e \u003cp\u003eClinical variables and outcomes\u003c/p\u003e \u003cp\u003ePre-transplant variables included recipient age and sex, diagnosis, disease status at transplant (complete remission vs active disease), donor type, donor age, stem cell source, conditioning intensity, GVHD prophylaxis and in vivo T-cell depletion strategy.\u003c/p\u003e \u003cp\u003eThe primary endpoint was GRFS, defined as survival without grade III\u0026ndash;IV acute GVHD, extensive chronic GVHD, relapse, or death; time was measured from transplant to the first component event, with censoring at last follow-up. Secondary endpoints included overall survival, relapse (with death as a competing event), non-relapse mortality (with relapse as a competing event), and acute/chronic GVHD analyzed using cumulative incidence methods with appropriate competing risks.\u003c/p\u003e \u003cp\u003eMachine-learning model development and validation\u003c/p\u003e \u003cp\u003eWe evaluated three survival modeling approaches: random survival forests (RSF), XGBoost Cox survival, and penalized Cox regression (elastic net) with cross-validated hyperparameter tuning. For each method, we trained (i) a clinical model using 9 pre-transplant clinical predictors and (ii) a full model adding recipient and donor locus-specific HED (A, B, C, DRB1, DQB1; 10 HED features). Primary modeling used complete-case analysis (no imputation), yielding 9,196 patients for training and 3,826 for internal testing (\u003cb\u003eSupplementary methods\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eDiscrimination was assessed with Harrell\u0026rsquo;s c-index (95% CIs from 1000 bootstrap resamples) and time-dependent area under the curve (AUC) at 1, 2, and 5 years; calibration compared predicted vs observed risks across deciles. External validation applied the trained model to the US cohort without recalibration. For clinical interpretability, RSF-predicted cumulative hazard scores were used to define low/intermediate/high risk groups based on training-set tertiles, and the same cutpoints were applied unchanged to internal and external validation cohorts; Kaplan\u0026ndash;Meier curves were compared by log-rank testing.\u003c/p\u003e \u003cp\u003eSimulation analyses\u003c/p\u003e \u003cp\u003eTo visualize how HED profiles influence predicted GRFS risk under fixed clinical conditions, we simulated RSF predictions using a reference clinical profile (56-year-old male with acute myeloid leukemia [AML] in complete remission, receiving peripheral blood stem cells with reduced-intensity conditioning and calcineurin inhibitor [CNI] mychophenolate [MMF] prophylaxis) and systematically varied recipient/donor HED across observed ranges. Additional simulations examined matched donors, haploidentical donors, and 9/10 mismatched unrelated donors stratified by mismatched locus to explore locus-specific patterns.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eGroup comparisons used Wilcoxon rank-sum tests (continuous variables) and chi-square/Fisher\u0026rsquo;s exact tests (categorical variables). Survival was summarized with Kaplan\u0026ndash;Meier methods; competing-risk endpoints were analyzed using Fine\u0026ndash;Gray models. Analyses were performed in R (v4.2.0) using standard survival and machine-learning packages.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eStudy cohorts\u003c/p\u003e \u003cp\u003eIn the French cohort, after exclusion of patients with missing data for model variables, 9,196 patients in training and 3,826 patients in test sets had complete data for all 19 predictors. The study cohorts used for the model development did not differ in clinical or immunogenetic variables (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA, Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and S1). The median patient age was 56.1 years (IQR 44.5\u0026ndash;63.5) in training and 56.2 years (IQR 44.3\u0026ndash;63.3) in test cohorts. Males comprised 58 and 59% of training and test cohorts, respectively. Acute myeloid leukemia (AML) was the most common diagnosis (45% training, 45.5% test), followed by myelodysplastic syndrome (MDS, 14 vs.14%) and myeloproliferative neoplasms (MPN, 11 vs. 12%). Most patients (64% training, 63% test) were in complete remission at transplant. Donor types included matched unrelated donors (MUD, 52% training, 52% test), matched sibling donors (MSD, 20%, 19.5%), haploidentical donors (15% in both cohorts), mismatched unrelated donors (MMUD, 13% in both), and cord blood (1.0% in both cohorts).\u003c/p\u003e \u003cp\u003eMissingness was assessed across all study variables, with an overall rate of 6% in training and test cohorts. Most variables were 98\u0026ndash;100% complete, with notable exceptions: disease status (96%), T-cell depletion status (81%), and HLA-DPB1 alleles for recipients (74%) and donors (73%) (\u003cb\u003eFigure S1\u003c/b\u003e). The final dataset included clinical variables and HLA information encoded as recipient- and donor-specific HED at the locus level (A, B, C, DRB1, DQB1), as well as aggregated class I, class II, and total HED scores. Because HLA-DPB1 alleles were frequently missing, total HED was calculated using only loci A, B, C, DRB1, and DQB1.\u003c/p\u003e \u003cp\u003eThe external validation cohort comprised 1,931 patients from five US centers, with 616 (32%) having complete data for all model variables (\u003cb\u003eTable S2\u003c/b\u003e). After removal of cases with missing data, this cohort was enriched for myeloid diseases.\u003c/p\u003e \u003cp\u003eImmunogenetic features across disease subgroups.\u003c/p\u003e \u003cp\u003eFirst, we assessed whether immunogenetic, disease-specific patterns exist by examining the distributions of locus-specific, class I, class II, and total HED scores in recipients across ten disease subgroups, and we compared these with donor-related metrics \u003cb\u003e(Figure S2[A,B] and S3).\u003c/b\u003e Notably, the Hodgkin (HL) lymphoma subgroup showed distinct per-locus distributions, with the lowest total and class II HED (driven primarily by DRB1 and DQB1) suggesting a potential immunogenetic predisposition related to specific class II HLA risk structures associated with the disease.\u003c/p\u003e \u003cp\u003eWhen examining recipient\u0026ndash;donor HED associations, we found strong correlations across all loci (R\u0026thinsp;\u0026gt;\u0026thinsp;0.84, p\u0026thinsp;\u0026lt;\u0026thinsp;2.2\u0026times;10⁻\u0026sup1;⁶) except DPB1, which showed only a weak recipient\u0026ndash;donor correlation (R\u0026thinsp;=\u0026thinsp;0.38), \u003cb\u003eFigure S4\u003c/b\u003e. By donor subgroup, correlations in unrelated mismatched allo-HSCT ere very high for class II loci (R\u0026thinsp;\u0026gt;\u0026thinsp;0.91) DQB1 and DRB but lower for class I (R\u0026thinsp;\u0026le;\u0026thinsp;0.75), and very low for DPB1 (\u003cb\u003eFigure S5\u003c/b\u003e). In haploidentical transplants, correlations were uniformly weak across loci (R\u0026thinsp;\u0026lt;\u0026thinsp;0.2). Cord blood was heterogeneous: higher for class II (R : 0.70\u0026ndash;0.95) and lower for class I (R : 0.19\u0026ndash;0.59), suggesting that cord blood units with class I mismatches are more frequently selected. DPB1 showed consistently weak correlations across donor groups (R\u0026thinsp;\u0026lt;\u0026thinsp;0.3), with the notable exception of MSD, where not surprisingly the correlation was strong (R\u0026thinsp;=\u0026thinsp;0.97), \u003cb\u003eFigure S6.\u003c/b\u003e\u003c/p\u003e \u003cp\u003eExploring the impact of immunogenetic features on outcomes\u003c/p\u003e \u003cp\u003eWhen we assessed the impact of specific immunogenetic features on outcomes, allele-enrichment analysis did not reveal any allele-specific effect on dichotomous clinical phenotypes (GRFS event vs. no event at 2y ; \u003cb\u003eTable S3\u003c/b\u003e). However, several notable associations emerged when HLA loci were parameterized as HED features. Higher recipient HED at the DQB1 locus was an immunogenetic determinants of an increased risk of grade III\u0026ndash;IV acute GvHD, whereas higher recipient HED at the B locus was associated with a lower risk of relapse and improved whole cohort OS (\u003cb\u003eTable S4\u003c/b\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA). These associations displayed a time-varying effect, with an increasing HR over time (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB). Within the HLA-mismatched subset, the donor\u0026ndash;recipient HED difference (Delta HED) did not differ across first-event categories, suggesting that a simple mismatch-derived summary metric together with binary outcomes, are unlikely to capture the post-transplant alloreactivity (\u003cb\u003eFigure S7\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eMachine learning modeling\u003c/p\u003e \u003cp\u003eWe next investigated the combinatorial impact of clinical and HED-related features using machine-learning models. When the SFGM-TC cohort was split in well-blanced training and a test set no differences in outcomes were observed (\u003cb\u003eFigure S8\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eWe decided to study these clinical and immunogenetic interactions on GRFS, a composit endpoint expressing probability of being alive with/without an alloreactive complication. In the training set we applied principal component analysis (PCA) on all candidate predictors and projected patients onto the first 3 components, according to GRFS status (event vs event-free). PC1, PC2 and PC3 accounted for ~\u0026thinsp;23%, 19% and 15% of the total variance, respectively, indicating that these axes captured just over half of the overall heterogeneity of the cohort (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA). However, in all pairwise projections (PC1\u0026ndash;PC2, PC1\u0026ndash;PC3 and PC2\u0026ndash;PC3), patients with and without a GRFS event were widely intermingled, with no distinct clusters or clear separation between outcome groups.\u003c/p\u003e \u003cp\u003eTo assess the global contribution of immunogenetics to GRFS, without relying on linearity or other parametric assumptions, we implemented a comparative modeling framework. Specifically, we trained (i) a RSF, (ii) an XGBoost-based survival, and (iii) a Cox proportional hazards models with elastic-net regularization, and evaluated on an internal test set. For each modeling family, we specified two feature sets to isolate the incremental value of immunogenetics: i) a clinical model restricted to standard pre-transplant covariates routinely available at baseline (9 variables, \u003cb\u003eSupplmentary methods\u003c/b\u003e); and ii) a full model that extended the clinical backbone by adding the 19 variables (A, B, C, DRB1, DQB1, and HED metrics). This design allowed us to quantify, across distinct algorithmic assumptions (tree-based non-parametric, gradient-boosted, and penalized semi-parametric), whether inclusion of HED variables consistently improved prognostic performance and to compare the relative importance of clinical vs immunogenetic features within each framework. Nevertheless, discriminatory performance of all the models remained modest. Across all 3 models, addition of immunogenetic information (donor/recipient HED) to standard pre-transplant clinical covariates resulted in a small but generally consistent improvement in discrimination (\u003cb\u003eTables S5\u0026ndash;S6\u003c/b\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB) while clinical predictors remained most important risk determinants (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC), donor and recipient HED, particularly at class II loci (DRB1 and DQB1), emerged among the most informative immunogenetic contributor (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eD).\u003c/p\u003e \u003cp\u003eRFS model validation\u003c/p\u003e \u003cp\u003eGiven comparable out-of-sample discrimination across modeling approaches, minimal evidence of overfitting, and the need for an interpretable prognostic tool, we selected the RSF model incorporating HED variables (RSF full) as the primary model for downstream analyses and validation. For each patient, we extracted the event risk score from the trained RSF model, defined as the predicted cumulative hazard. This continuous score captures the relative risk of experiencing the composite GRFS endpoint (grade III\u0026ndash;IV acute GvHD, extensive chronic GvHD, relapse, or death from any cause). Patients were subsequently stratified into three risk categories\u0026mdash;low (score\u0026thinsp;\u0026lt;\u0026thinsp;103), intermediate (score 103\u0026ndash;143), and high (score\u0026thinsp;\u0026gt;\u0026thinsp;143)\u0026mdash;using tertiles of the predicted risk score distribution in the training cohort. Optimized cutpoints (\u003cb\u003eSupplementary Methods\u003c/b\u003e) were derived exclusively in the training set and then applied unchanged to the hold-out test set and the external validation cohort to ensure an unbiased assessment of performance. Across all three cohorts, Kaplan\u0026ndash;Meier analyses demonstrated clear separation between risk groups, with statistically significant differences by log-rank testing (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001; \u003cb\u003eTable S7, panel A\u003c/b\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA, B), supporting robust identification of three clinically meaningful risk strata, despite low C-indexes \u003cb\u003e(Table S7, Panel B\u003c/b\u003e). To assess the generalizability of the model across disease subtypes, we applied the same risk stratification to the separete subsets of patients with AML and MDS/MPN. Despite being trained on the full heterogeneous population, the model maintained robust discriminative ability within this disease-specific subgroup(\u003cb\u003eFigure S9\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eHED contribution on RFS-based risk scores in matched setting.\u003c/p\u003e \u003cp\u003eTo characterize the combined effect of recipient and donor HED on GRFS outcomes, we performed a systematic simulation across the full range of total HED values from the French cohort (n\u0026thinsp;=\u0026thinsp;13,979) using the trained RSF model. All clinical variables were held constant at reference values (56-year-old male with AML in complete remission, receiving peripheral blood stem cells from a MUD aged 35 years, with reduced-intensity conditioning and calcineurin inhibitor/mycophenolate mofetil prophylaxis). This analysis revealed a pronounced non-linear (U-shaped) relationship between total HED and GRFS risk, demonstrating a \"Goldilocks effect\" (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eA). The optimal risk profile was observed at intermediate total HED levels of approximately 55 for both recipient and donor (corresponding to the 75th percentile of both HED distributions). Risk increased substantially with very low (Total HED\u0026thinsp;=\u0026thinsp;0) or very high donor/recipient total HED (\u0026gt;\u0026thinsp;65)). Percentile-based analyses confirmed these findings (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eB). Risk scores were elevated at the lowest HED percentile (P5: score\u0026thinsp;=\u0026thinsp;71, total HED\u0026thinsp;=\u0026thinsp;0, reflecting complete homozygosity), decreased to a minimum at P75 (score\u0026thinsp;=\u0026thinsp;60, Total HED\u0026thinsp;=\u0026thinsp;57), and increased again at the highest percentile (P95: score\u0026thinsp;=\u0026thinsp;92, Total HED\u0026thinsp;=\u0026thinsp;69).\u003c/p\u003e \u003cp\u003eTo assess the robustness of these findings, we repeated the simulation across three clinical profiles (all receiving matched donor transplantation) with distinct baseline GRFS risk levels: a higher-risk profile (65-year-old, MAC conditioning, not in remission), a lower-risk profile (45-year-old, MSD, complete remission), and a \u0026laquo; standard risk \u0026raquo; (reference) AML profile (56-year-old, MUD, RIC, complete remission). All three profiles exhibited the characteristic U-shaped HED-risk relationship, with nadir consistently at intermediate HED values (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC). For the reference clinical profile, predicted risk scores (RS) ranged from 60 (optimal, P75) to 92 (high HED, P95), with intermediate values for homozygous patients (71) and those at median HED (78). All simulated scenarios remained within the low-risk zone (score\u0026thinsp;\u0026lt;\u0026thinsp;103), though the 32-point variation demonstrates that HED optimization provides meaningful risk modulation even within favorable clinical contexts (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eD).\u003c/p\u003e \u003cp\u003eHED contribution on RFS-based risk scores in mismatched setting.\u003c/p\u003e \u003cp\u003eWe then run anaologous analyses in the haploidentical subgroup. The optimal recipient/donor combination occurred at intermediate total HED levels (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eA). Through a percentile based analysis, we quantified risk interactions across 49 recipient-donor combinations (P5-P95 for each). The optimal pairing was P75 recipient x P75 donor (RS\u0026thinsp;=\u0026thinsp;79). Notably, the matrix revealed a distinct \u0026laquo; \u003cem\u003ematch like with alike\u003c/em\u003e \u0026raquo; pattern: recipients with low HED achieved optimal outcomes with low-HED donors (P5xP5: RS\u0026thinsp;=\u0026thinsp;97), intermediate-HED recipients with intermediate donors (P50xP75; RS\u0026thinsp;=\u0026thinsp;82 ; P75xP75; RS\u0026thinsp;=\u0026thinsp;79) and high-HED recipients with high-HED donors (P90\u0026times;P90\u0026thinsp;=\u0026thinsp;85; P95\u0026times;P90\u0026thinsp;=\u0026thinsp;94). Mismatched pairings, particularly low-HED recipients with high-HED donors, produced the worst outcomes (P5\u0026times;P95 RS\u0026thinsp;=\u0026thinsp;119; P10\u0026times;P95 : RS\u0026thinsp;=\u0026thinsp;116).\u003c/p\u003e \u003cp\u003eDonor selection effects were then examined across four representative recipient HED levels. For P5 recipients, risk decreased as donor HED decreased, with P5 donors being optimal (RS\u0026thinsp;=\u0026thinsp;97). For P75 recipients, the lowest risk was also observed with matched P75 donors (RS\u0026thinsp;=\u0026thinsp;79), while both lower and higher donor HED values increased risk. For P95 recipients, the best outcome was seen with P90 donors (RS\u0026thinsp;=\u0026thinsp;97), whereas P95 donors paradoxically increased risk (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eB). The potential risk reduction achievable through donor selection differed by recipient HED (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eC,D), reaching 28 points for P25 recipients, 21 for P5, 20 for P50, and 10 even for P95 recipients. Overall, these findings indicate that donor HED is an actionable parameter in haploidentical transplantation, with the best strategy being to align donor HED with recipient HED level (\u003cb\u003eTable S7\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eWe next assessed the impact of donor and recipient HED on RSF scores in the 9/10 mismatched unrelated donor setting (N\u0026thinsp;=\u0026thinsp;1,326), restricting the analysis to transplants with a single HLA mismatch (\u003cb\u003eFigure S10\u003c/b\u003e). Using the same standardized AML patient profile, we simulated each mismatched subgroup while varying donor and recipient HED from P5 to P95 at the mismatched locus and keeping all other variables constant. HLA-B mismatch showed the greatest HED sensitivity, with RS varying from 73 to 97 (24.9-point difference), followed by HLA-DRB1 (23.2 points) and HLA-DQB1 (21.0 points), whereas HLA-A (16.6 points) and HLA-C (15.7 points) were less sensitive. For HLA-B, -DRB1, and -DQB1 mismatches, the lowest risk was consistently observed when both donor and recipient had high HED (P75) at the mismatched locus (\u003cb\u003eFigure S10\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eImplementation of a GRFS calculator\u003c/p\u003e \u003cp\u003eTo facilitate clinical implementation of these findings, we developed an interactive web-based calculator that integrates the trained RSF model with HED assessment. The tool accepts patient-specific clinical variables along with HLA typing data for both recipient and donor, used to automatically compute locus-specific HED values. The calculator outputs an individualized GRFS risk score with corresponding risk category and is freely available at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://huggingface.co/spaces/SPAGLIUCA/SMART\u003c/span\u003e\u003cspan address=\"https://huggingface.co/spaces/SPAGLIUCA/SMART\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (\u003cb\u003eSupplementary Appendix\u003c/b\u003e).\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this large, multicenter international study, we show that incorporating immunogenetic architecture information, captured by HED as a proxy of individual immunocompetence, into ML survival models yields modest but consistent, reproducible improvement in the prediction of GRFS after allo-HCT. The overall modest discrimination is expected for models originated from registry data. Neverthess, HED provided incremental prognostic signal across three distinct modeling families and enabled stable separation of patients into clinically interpretable risk strata in both internal and external cohorts, with compelling insight emerging from model-based simulations.\u003c/p\u003e \u003cp\u003eOur results directly address a longstanding tension in transplantation immunology: the same donor\u0026ndash;recipient immune interactions that underlie curative graft-versus-tumor effects can also fuel severe GvHD and immune-mediated morbidity. Disentangling beneficial from deleterious alloreactivity remains the elusive \u0026ldquo;holy grail\u0026rdquo; of the field, and classical risk models (largely anchored in clinical variables) have struggled to generalize across centers, platforms, and eras\u003csup\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e. HED offers a biologically grounded, quantitative summary of HLA heterogeneity that moves beyond categorical matching and allele-level analyses, capturing the functional landscape of antigen presentation. In our data, locus-level HED features, particularly within class II, ranked among the most informative immunogenetic contributors, aligning with the central role of CD4-driven alloreactivity in GvHD biology and immune orchestration. Importantly, the time-varying associations observed in conventional Cox analyses reinforce that immunogenetic effects likely unfold dynamically, interacting with evolving immune reconstitution, infections, and post-transplant interventions. The non-linear \u0026ldquo;Goldilocks\u0026rdquo; effect observed in matched settings is biologically plausible and conceptually instructive. Very low HED (including complete homozygosity) may reflect a constrained peptide-binding breadth, potentially limiting immune surveillance, delaying immune reconstitution, or narrowing the repertoire of protective and antitumor responses. Conversely, very high HED may broaden antigenic space to a degree that amplifies alloreactive potential and inflammatory tissue injury, particularly in environments already primed by conditioning-related damage and microbial translocation. The stability of this U-shaped relationship across clinically distinct reference profiles suggests that HED is not merely a surrogate for baseline clinical risk but appears to act as an additive, orthogonal layer modulating outcome trajectories across transplant platforms.\u003c/p\u003e \u003cp\u003eIn haploidentical transplantation, where HLA disparity is structurally embedded, our simulations identified a striking \u0026ldquo;match similar with silmilar\u0026rdquo; principle: recipients achieved optimal predicted outcomes when paired with donors with comparable HED levels. This observation again suggests that the relative \u0026laquo;immunogenetic distance\u0026raquo; between donor and recipient (not only the presence of mismatch) may shape the balance between immune tolerance and immune aggression. One mechanistic hypothesis is that concordant HED profiles may harmonize the breadth of peptide presentation on both sides of the alloreactive interface, potentially reducing asymmetric immune recognition that could otherwise exacerbate GvH reactions or impair coordinated immune reconstitution. Conversely, discordant pairings (particularly low-HED recipients with high-HED donors) may create a steep gradient in antigen-presenting diversity, increasing the likelihood of broad donor T-cell activation against recipient targets without proportional gains in protective immunity.\u003c/p\u003e \u003cp\u003eThe 9/10 mismatched unrelated donor setting offered additional granularity, highlighting locus-dependent HED sensitivity in outcome risk prediction. HLA-B mismatches displayed the largest modulation of predicted risk across the HED spectrum, with substantial, yet directional, risk shifts dependent on donor and recipient divergence at the mismatched locus. That said, GRFS aggregates relapse, GvHD, and mortality into a single outcome, and different components may be differentially influenced by HED across loci and transplant platforms. Dissecting these competing forces will be essential to translate HED-informed strategies into practical donor selection rules tailored to the clinical priority (e.g., relapse prevention versus toxicity minimization).\u003c/p\u003e \u003cp\u003eSeveral aspects of our results warrant cautious interpretation. First, c-index values remained below 0.60 across models, underscoring the inherent difficulty of predicting GRFS using baseline variables alone. GRFS is shaped not only by pre-transplant risk but also by post-transplant stochastic events, center-specific practices, infections, immune reconstitution kinetics, treatment adaptations, and evolving supportive care standards, many of which are not captured in registry-grade baseline datasets. In this context, a small but consistent improvement across algorithms is meaningful, particularly when it derives from biologically interpretable features and yields stable risk stratification across cohorts. Second, our external validation was necessarily restricted to complete cases, and missingness likely attenuated both model performance and the precision of HED-derived inferences. Third, although we intentionally avoided recalibration in the US cohort to provide a stringent generalizability test, center-level differences in practice and endpoint ascertainment may still influence absolute risk estimates.\u003c/p\u003e \u003cp\u003eNotwithstanding these limitations, our study has notable strengths: the scale and heterogeneity of the dataset, the use of high-resolution HLA typing enabling locus-specific HED computation, the systematic comparison of complementary survival learners, and the application of a strict held-out internal test set alongside external validation. Importantly, we complemented \u0026ldquo;black-box\u0026rdquo; prediction with mechanistic-oriented simulations that transform model coefficients and interactions into clinically intuitive risk landscapes. This approach reframes immunogenetics from a static matching metric into a modifiable parameter that can be explored (and potentially optimized) at the time of donor selection.\u003c/p\u003e \u003cp\u003eLooking forward, several directions emerge. Integrating additional immunogenetic layers (DPB1 when available, permissiveness metrics, KIR ligands, donor-specific antibodies) and disease/host biology (MRD status, cytogenetics/molecular risk, inflammatory markers) may raise performance. Such an rationally selected donor or donor/recipient immune profile invokes a definition of a transplant \u0026laquo; immunome \u0026raquo; as a sum of immunologic vectors determining alloreactivity. Furthermore, incorporating early post-transplant dynamics (e.g., immune reconstitution, viral reactivation, cytokine trajectories, and treatment intensification) could enable genuinely time-updated prediction that reflects the evolving biology of alloreactivity. Prospective studies are needed to test whether HED-guided donor selection can improve outcomes beyond standard matching, particularly in donor-rich settings where multiple acceptable donors exist. Ultimately, the goal is not to replace clinical judgment, but to equip it with a more faithful representation of immunogenetic diversity, helping clinicians navigate the narrow therapeutic corridor between graft-versus-tumor benefit and transplant-related toxicity.\u003c/p\u003e \u003cp\u003eAuthorship contributions\u003c/p\u003e \u003cp\u003eSP and JM conceptualized the study, designed the analytical framework, and interpreted the data. SP developed the study synopsis, designed the study and statistical analysis plan, performed the biostatistical and bioinformatic analyses, created the visualizations, and wrote the manuscript. VA gave important inputs on the modeling design and provided patient data. MG contributed to data management. NR, AH, and RD coordinated data collection at the SFGM-TC registry level. AD, AK, LG, MJ, SKB, and FF performed data collection at the US center level. MR, IYA, EF, CEB, JBM, PC, CCL, XP, ML, JM, SN, FB, ED, NM, FM, MDA and MTR were involved in patient recruitment and performed data collection through the SFGM-TC registry. AA contributed to HLA typing quality control and immunogenetic data interpretation. MDA and MTR participated in study conception and interpretation of the data analysis, and provided important intellectual inputs. VV, TL, CG, and JM participated in the interpretation of the data analysis and provided important intellectual inputs and edited the manuscript. All authors reviewed and approved the final version of this manuscript.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAuthorship contributions\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSP and JM conceptualized the study, designed the analytical framework, and interpreted the data. SP developed the study synopsis, designed the study and statistical analysis plan, performed the biostatistical and bioinformatic analyses, created the visualizations, and wrote the manuscript. VA gave important inputs on the modeling design and provided patient data.\u0026nbsp;MG contributed to data management. NR, AH, and RD coordinated data collection at the SFGM-TC registry level. AD, AK, LG, MJ, SKB, and FF performed data collection at the US center level. MR, IYA, EF, CEB, JBM, PC, CCL, XP, ML, JM, SN, FB, ED, NM, FM, MDA and MTR were involved in patient recruitment and performed data collection through the SFGM-TC registry. AA contributed to HLA typing quality control and immunogenetic data interpretation. MDA and MTR participated in study conception and interpretation of the data analysis, and provided important intellectual inputs. VV, TL, CG, and JM participated in the interpretation of the data analysis and provided important intellectual inputs and edited the manuscript. All authors reviewed and approved the final version of this manuscript.\u003c/p\u003e\n\u003cp\u003eConflict-of-interest disclosure\u003c/p\u003e\n\u003cp\u003eSP has received travel expenses or honoraria for participation in advisory boards, symposia or other scientific events by Alexion, Novartis, Jazz Pharmaceutical, Sobi as well as research funding by JANSSEN HORIZON program. TLL is co-inventor on a patent application for using HED as a prognostic marker for immunotherapy success.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eData sharing\u003c/p\u003e\n\u003cp\u003eThe data underlying this study were obtained from clinical registries and cannot be made publicly available because of institutional, ethical, and data protection restrictions. The authors are therefore unable to share individual-level data directly. Access to data from the SFGM-TC registry requires a formal request to the SFGM-TC study office. Requests regarding data from the US centers should be directed to Jaroslaw Maciejewski at [email protected]. The SMART model is hosted at https://huggingface.co/spaces/SPAGLIUCA/SMART.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eThis work was supported by SFGM-TC We acknowledge all participating centers. TLL was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) \u0026ndash; 437857095. SP received funding from Fondation ARC pour la Recherche sur le Cancer and Force Hemato. \u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003ePassweg JR, Baldomero H, Atlija M et al (2025) The 2023 EBMT report on hematopoietic cell transplantation and cellular therapies. Increased use of allogeneic HCT for myeloid malignancies and of CAR-T at the expense of autologous HCT. Bone Marrow Transpl 60(4):519\u0026ndash;528. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41409-025-02524-2\u003c/span\u003e\u003cspan address=\"10.1038/s41409-025-02524-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePassweg JR, Baldomero H, Chabannon C et al Hematopoietic cell transplantation and cellular therapy survey of the EBMT: monitoring of activities and trends over 30 years. Bone Marrow Transpl. Published online 23 February 2021. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41409-021-01227-8\u003c/span\u003e\u003cspan address=\"10.1038/s41409-021-01227-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTrunk AD, Guo M, Budvytyte L et al (2024) Hematopoietic Stem-Cell Transplantation: Exploring the Latest Advances and Gaps in Disparities, Psychosocial and Symptom Management Interventions, and Chronic Graft-Versus-Host Disease Care. Am Soc Clin Oncol Educ Book 44(3):e432186. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1200/EDBK_432186\u003c/span\u003e\u003cspan address=\"10.1200/EDBK_432186\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eO\u0026rsquo;Neill AT, Chakraverty R (2021) Graft Versus Leukemia: Current Status and Future Perspectives. JCO 39(5):361\u0026ndash;372. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1200/JCO.20.01801\u003c/span\u003e\u003cspan address=\"10.1200/JCO.20.01801\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZeiser R, Blazar BR (2017) Acute Graft-versus-Host Disease \u0026mdash; Biologic Process, Prevention, and Therapy. Longo DL, ed. \u003cem\u003eN Engl J Med\u003c/em\u003e. ;377(22):2167\u0026ndash;2179. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1056/NEJMra1609337\u003c/span\u003e\u003cspan address=\"10.1056/NEJMra1609337\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMalard F, Holler E, Sandmaier BM, Huang H, Mohty M (2023) Acute graft-versus-host disease. Nat Rev Dis Primers 9(1):27. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41572-023-00438-1\u003c/span\u003e\u003cspan address=\"10.1038/s41572-023-00438-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePagliuca S, Michonneau D, Sicre de Fontbrune F et al (2019) Allogeneic reactivity-mediated endothelial cell complications after HSCT: a plea for consensual definitions. Blood Adv 3(15):2424\u0026ndash;2435. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1182/bloodadvances.2019000143\u003c/span\u003e\u003cspan address=\"10.1182/bloodadvances.2019000143\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDavies JK, Taussig D, Oakervee H et al (2013) Long-term survival with low toxicity after allogeneic transplantation for acute myeloid leukaemia and myelodysplasia using non-myeloablative conditioning without T cell depletion. Br J Haematol 162(4):525\u0026ndash;529. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/bjh.12402\u003c/span\u003e\u003cspan address=\"10.1111/bjh.12402\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePagliuca S, Prata PH, Xhaard A et al Long-term outcomes and risk factor analysis of steroid-refractory graft versus host disease after hematopoietic stem cell transplantation. Bone Marrow Transpl. Published online 25 June 2020. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41409-020-0977-3\u003c/span\u003e\u003cspan address=\"10.1038/s41409-020-0977-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNegrin RS (2015) Graft-versus-host disease versus graft-versus-leukemia. Hematol Am Soc Hematol Educ Program 2015:225\u0026ndash;230. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1182/asheducation-2015.1.225\u003c/span\u003e\u003cspan address=\"10.1182/asheducation-2015.1.225\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Bergen CAM, Van Luxemburg-Heijs SAP, De Wreede LC et al (2017) Selective graft-versus-leukemia depends on magnitude and diversity of the alloreactive T cell response. J Clin Invest 127(2):517\u0026ndash;529. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1172/JCI86175\u003c/span\u003e\u003cspan address=\"10.1172/JCI86175\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLapoutge A, Gaudfrin M, Raus N et al (2025) Hallmarks of allogeneic hematopoietic cell transplant success: A study on immune signatures and predictive features of graft versus host disease-free remission states from the SFGM-TC registry and cryostem biobank. Blood 146(Supplement 1):2476\u0026ndash;2476. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1182/blood-2025-2476\u003c/span\u003e\u003cspan address=\"10.1182/blood-2025-2476\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShouval R, Fein JA, Shouval A et al (2019) External validation and comparison of multiple prognostic scores in allogeneic hematopoietic stem cell transplantation. Blood Adv 3(12):1881\u0026ndash;1890. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1182/bloodadvances.2019032268\u003c/span\u003e\u003cspan address=\"10.1182/bloodadvances.2019032268\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaccardi R, Putter H, Eikema DJ et al (2023) Benchmarking of survival outcomes following Haematopoietic Stem Cell Transplantation (HSCT): an update of the ongoing project of the European Society for Blood and Marrow Transplantation (EBMT) and Joint Accreditation Committee of ISCT and EBMT (JACIE). Bone Marrow Transpl 58(6):659\u0026ndash;666. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41409-023-01924-6\u003c/span\u003e\u003cspan address=\"10.1038/s41409-023-01924-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAwada H, Gurnari C, Durmaz A, Awada H, Pagliuca S, Visconte V (2022) Personalized Risk Schemes and Machine Learning to Empower Genomic Prognostication Models in Myelodysplastic Syndromes. Int J Mol Sci 23(5):2802. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/ijms23052802\u003c/span\u003e\u003cspan address=\"10.3390/ijms23052802\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDurmaz A, Gurnari C, Hershberger CE et al (2023) A multimodal analysis of genomic and RNA splicing features in myeloid malignancies. iScience 26(3):106238. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.isci.2023.106238\u003c/span\u003e\u003cspan address=\"10.1016/j.isci.2023.106238\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKewan T, Durmaz A, Bahaj W et al (2023) Molecular patterns identify distinct subclasses of myeloid neoplasia. Nat Commun 14(1):3136. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-023-38515-4\u003c/span\u003e\u003cspan address=\"10.1038/s41467-023-38515-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShouval R, Fein JA, Savani B, Mohty M, Nagler A (2021) Machine learning and artificial intelligence in haematology. Br J Haematol 192(2):239\u0026ndash;250. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/bjh.16915\u003c/span\u003e\u003cspan address=\"10.1111/bjh.16915\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShreve J, Meggendorfer M, Awada H et al (2019) A Personalized Prediction Model to Risk Stratify Patients with Acute Myeloid Leukemia (AML) Using Artificial Intelligence. Blood 134(Supplement1):2091\u0026ndash;2091. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1182/blood-2019-128066\u003c/span\u003e\u003cspan address=\"10.1182/blood-2019-128066\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRomero M, Mosquera Orgueira A, Mej\u0026iacute;a Saldarriaga M (2024) How artificial intelligence revolutionizes the world of multiple myeloma. Front Hematol 3:1331109. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/frhem.2024.1331109\u003c/span\u003e\u003cspan address=\"10.3389/frhem.2024.1331109\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlcazer V, Le Meur G, Roccon M et al (2024) Evaluation of a machine-learning model based on laboratory parameters for the prediction of acute leukaemia subtypes: a multicentre model development and validation study in France. Lancet Digit Health 6(5):e323\u0026ndash;e333. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S2589-7500(24)00044-X\u003c/span\u003e\u003cspan address=\"10.1016/S2589-7500(24)00044-X\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArrieta-Bola\u0026ntilde;os E (2025) Differential Alloreactivity: Lessons Learned From a Singular HLA Locus. HLA 106(6):e70489. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/tan.70489\u003c/span\u003e\u003cspan address=\"10.1111/tan.70489\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePagliuca S, Gurnari C, Rubio MT, Visconte V, Lenz TL (2022) Individual HLA heterogeneity and its implications for cellular immune evasion in cancer and beyond. Front Immunol 13:944872. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fimmu.2022.944872\u003c/span\u003e\u003cspan address=\"10.3389/fimmu.2022.944872\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePierini F, Lenz TL (2018) Divergent Allele Advantage at Human MHC Genes: Signatures of Past and Ongoing Selection. Wilke C, ed. \u003cem\u003eMolecular Biology and Evolution\u003c/em\u003e. ;35(9):2145\u0026ndash;2158. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/molbev/msy116\u003c/span\u003e\u003cspan address=\"10.1093/molbev/msy116\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChowell D, Krishna C, Pierini F et al (2019) Evolutionary divergence of HLA class I genotype impacts efficacy of cancer immunotherapy. Nat Med 25(11):1715\u0026ndash;1720. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-019-0639-4\u003c/span\u003e\u003cspan address=\"10.1038/s41591-019-0639-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDaull AM, Dubois V, Labussi\u0026egrave;re-Wallet H et al (2022) Class I/Class II HLA Evolutionary Divergence Ratio Is an Independent Marker Associated With Disease-Free and Overall Survival After Allogeneic Hematopoietic Stem Cell Transplantation for Acute Myeloid Leukemia. Front Immunol 13:841470. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fimmu.2022.841470\u003c/span\u003e\u003cspan address=\"10.3389/fimmu.2022.841470\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePagliuca S, Gurnari C, Hong S et al (2021) Clinical and basic implications of dynamic T cell receptor clonotyping in hematopoietic cell transplantation. JCI Insight 6(13):e149080. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1172/jci.insight.149080\u003c/span\u003e\u003cspan address=\"10.1172/jci.insight.149080\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCrivello P, Arrieta-Bola\u0026ntilde;os E, He M et al (2023) Impact of the HLA Immunopeptidome on Survival of Leukemia Patients After Unrelated Donor Transplantation. J Clin Oncol 41(13):2416\u0026ndash;2427. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1200/JCO.22.01229\u003c/span\u003e\u003cspan address=\"10.1200/JCO.22.01229\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eToulemonde C, Dubois V, Bouvier C et al (2025) HLA diversity is associated with TKI response and treatment-free remission in chronic myeloid leukemia. HemaSphere 9(11):e70261. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/hem3.70261\u003c/span\u003e\u003cspan address=\"10.1002/hem3.70261\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShouval R, Labopin M, Unger R et al (2016) Prediction of Hematopoietic Stem Cell Transplantation Related Mortality- Lessons Learned from the In-Silico Approach: A European Society for Blood and Marrow Transplantation Acute Leukemia Working Party Data Mining Study. Bertolini F, ed. \u003cem\u003ePLoS ONE\u003c/em\u003e. ;11(3):e0150637. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pone.0150637\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0150637\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMussetti A, Rius-Sansalvador B, Moreno V et al (2024) Artificial intelligence methods to estimate overall mortality and non-relapse mortality following allogeneic HCT in the modern era: an EBMT-TCWP study. Bone Marrow Transpl 59(2):232\u0026ndash;238. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41409-023-02147-5\u003c/span\u003e\u003cspan address=\"10.1038/s41409-023-02147-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIwasaki M, Kanda J, Arai Y et al (2022) Establishment of a predictive model for GVHD-free, relapse-free survival after allogeneic HSCT using ensemble learning. Blood Adv 6(8):2618\u0026ndash;2627. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1182/bloodadvances.2021005800\u003c/span\u003e\u003cspan address=\"10.1182/bloodadvances.2021005800\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHoltan SG, DeFor TE, Lazaryan A et al (2015) Composite end point of graft-versus-host disease-free, relapse-free survival after allogeneic hematopoietic cell transplantation. Blood 125(8):1333\u0026ndash;1338. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1182/blood-2014-10-609032\u003c/span\u003e\u003cspan address=\"10.1182/blood-2014-10-609032\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNegrin RS (2015) Graft-versus-host disease versus graft-versus-leukemia. Hematology 2015(1):225\u0026ndash;230. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1182/asheducation-2015.1.225\u003c/span\u003e\u003cspan address=\"10.1182/asheducation-2015.1.225\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"CHRU de Nancy","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-9383323/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9383323/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAllogeneic hematopoietic cell transplantation (allo-HCT) remains the only curative option for many hematologic malignancies, yet transplant-related toxicity and relapse continue to constrain long-term benefit. Prognostic tools based on clinical variables alone show limited transportability across centers. We hypothesized that integrating immunogenetic architecture, captured by HLA Evolutionary Divergence (HED), into machine-learning survival models would improve prediction of graft-versus-host disease\u0026ndash;free/relapse-free survival (GRFS).\u003c/p\u003e \u003cp\u003eWe developed SMART, a time-dependent framework that estimates dynamic GRFS probabilities after allo-HCT. We analyzed 16,028 adults transplanted between 2010\u0026ndash;2022. Model development used the French SFGM-TC registry (N\u0026thinsp;=\u0026thinsp;13,979), split into training (N\u0026thinsp;=\u0026thinsp;9,840) and held-out test (N\u0026thinsp;=\u0026thinsp;4,139) sets. Random survival forests, XGBoost-Cox, and elastic-net Cox models were trained using 9 clinical predictors, with or without 10 recipient/donor locus-specific HED features (19 predictors total) and externally evaluated in 616 patients from five U.S. centers with complete data.\u003c/p\u003e \u003cp\u003eAcross algorithms, discrimination for this composite endpoint was modest (c-index\u0026thinsp;\u0026lt;\u0026thinsp;0.60), but the addition of HED consistently improved performance and enabled reproducible stratification into low-, intermediate-, and high-risk groups based on the cumulative hazard score. Model-based simulations uncovered a non-linear (U-shaped) association between total HED and GRFS, with optimal outcomes at intermediate HED levels (~\u0026thinsp;75th percentile). In haploidentical transplantation (N\u0026thinsp;=\u0026thinsp;2,056), outcomes were maximized when donor and recipient HED were concordant (\u0026ldquo;match like with like\u0026rdquo;). In 9/10 mismatched unrelated donors (N\u0026thinsp;=\u0026thinsp;1,326), HLA-B mismatches showed the greatest HED sensitivity.\u003c/p\u003e \u003cp\u003eIntegrating immunogenetics with clinical data improves GRFS risk modeling and supports HED as an actionable feature for donor selection and pre-transplant risk stratification. SMART is available for research use.\u003c/p\u003e","manuscriptTitle":"Predictive Modeling of Post-Allogeneic Transplant Outcomes Using Machine Learning and Integrated Clinical and Immunogenetic Data: a Study from the SFGM-TC and a Multicenter US Consortium","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-21 01:30:34","doi":"10.21203/rs.3.rs-9383323/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"263beb3b-7d4a-443a-822c-60c61ac24736","owner":[],"postedDate":"April 21st, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":66136551,"name":"Immunology"},{"id":66136552,"name":"Hematology"},{"id":66136553,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2026-04-21T01:30:35+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-21 01:30:34","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9383323","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9383323","identity":"rs-9383323","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}