Development and Validation of a Cytokine-Based Predictive Model for Acute GvHD and Composite Outcomes in ATG-Based Haploidentical Hematopoietic Stem Cell Transplantation.

backgroundAcute graft-versus-host disease (aGvHD) stands as a critical complication following haploidentical hematopoietic stem cell transplantation (haplo-HSCT). Most existing predictive models, predominantly derived from HLA-matched donor cohorts, have been utilized for nonrelapse mortality (NRM) prediction; however, their utility in predicting aGvHD risk specifically in haplo-HSCT recipients receiving antithymocyte globulin (ATG)-based prophylaxis warrants further validation.

methodsA total of 280 patients undergoing ATG-based haplo-HSCT were retrospectively analyzed across two medical centers, split into training, internal test, and external validation cohorts. We first evaluated the predictive accuracy of the previously established Mount Sinai Acute GvHD International Consortium (MAGIC) algorithm for aGvHD, steroid-refractory aGvHD (SR-aGvHD). Subsequently, plasma concentrations of candidate cytokines (ST2, REG3α, Elafin, and TNFRI), selected a priori for their links to epithelial injury and inflammatory signaling in GvHD, were assessed for their predictive and causal relationships with aGvHD using logistic regression, weighted average area under the curve (wAUC), Mendelian randomization (MR), and restricted cubic spline (RCS) analyses. A new predictive model (the HAG model) was constructed based on identified key cytokines and validated across multicenter cohorts. MR analyses utilized external genome-wide association (GWAS) datasets to validate the reliability of identified cytokines. A visual interface for the model was created using R Shiny.

resultsMAGIC algorithm remains effective in the ATG-based haplo-HSCT setting for predicting aGvHD, achieving AUC values of 0.693 (training), 0.658 (internal test), and 0.622 (external validation). Among candidate cytokines, a combination of ST2, REG3α, and Elafin (the HAG model) demonstrated the highest predictive accuracy. MR analysis leveraging external GWAS data supported potential causal associations of ST2 (OR = 1.280, p = 0.004), REG3α (OR = 1.300, p = 0.012), and Elafin (OR = 1.209, p = 0.039) with aGvHD risk, providing complementary biological support for their selection as candidate biomarkers. The HAG model displayed good discrimination for aGvHD (AUC = 0.636-0.701) and SR-aGvHD (AUC = 0.666-0.779). Integration of clinical factors further enhanced prediction (HAG-C model, wAUC from 0.682 to 0.701).

conclusionThe HAG model, incorporating ST2, REG3α, and Elafin, provides clinically meaningful prediction of aGvHD and related clinical outcomes in ATG-based haplo-HSCT recipients, and may serve as a mechanistically informed tool for risk stratification and clinical management.