A novel predictive approach for GVHD after allogeneic SCT based on clinical variables and cytokine gene polymorphisms

Abstract

Despite considerable advances in our understanding of the pathophysiology of graft-versus-host disease (GVHD), its prediction remains unresolved and depends mainly on clinical data. The aim of this study is to build a predictive model based on clinical variables and cytokine gene polymorphism for predicting acute GVHD (aGVHD) and chronic GVHD (cGVHD) from the analysis of a large cohort of HLA-identical sibling donor allogeneic stem cell transplant (allo-SCT) patients. A total of 25 SNPs in 12 cytokine genes were evaluated in 509 patients. Data were analyzed using a linear regression model and the least absolute shrinkage and selection operator (LASSO). The statistical model was constructed by randomly selecting 85% of cases (training set), and the predictive ability was confirmed based on the remaining 15% of cases (test set). Models including clinical and genetic variables (CG-M) predicted severe aGVHD significantly better than models including only clinical variables (C-M) or only genetic variables (G-M). For grades 3-4 aGVHD, the correct classification rates (CCR1) were: 100% for CG-M, 88% for G-M, and 50% for C-M. On the other hand, CG-M and G-M predicted extensive cGVHD better than C-M (CCR1: 80% vs. 66.7%, respectively). A risk score was calculated based on LASSO multivariate analyses. It was able to correctly stratify patients who developed grades 3-4 aGVHD (P < .001) and extensive cGVHD (P < .001). The novel predictive models proposed here improve the prediction of severe GVHD after allo-SCT. This approach could facilitate personalized risk-adapted clinical management of patients undergoing allo-SCT.

Graphical abstract

Figure 1.

Box plots of CCRs for patients who develop (CCR1) and do not develop (CCR0) GVHD/NRM as predicted with the different models. The AUC and the number of variables used are shown in each case. The predictive ability of each model, built using 85% of the samples (training set), was computed with the remaining 15% of the samples (test set). To evaluate the performance and predictive ability of each model, training and testing samples were randomly selected and the procedure repeated 100 times. The distribution of the CCR and AUC over the 100 iterations is shown by means of box plots. CCRs for the development of aGVHD, cGVHD, and NRM obtained using the predictive model including only clinical variables (upper panels), the model including only genetic variables (middle panels), and the model including both clinical and genetic variables (lower panels) are shown. Number of clinical/recipient SNPs/donor SNPs variables is indicated in parenthesis.

Figure 2.

Stratification of the whole cohort of patients (n = 263) according to the risk of developing acute GVHD. Risk was calculated using the proposed predictive model including clinical variables (upper panels), genetic variables (middle panels), or both clinical and genetic variables (lower panels). Cumulative incidence curves are shown for the development of grade 2 to 4 aGVHD (left panels) and grade 3 to 4 aGVHD (right panels). The cutoff used was 0.28 for grade 2 to 4 aGVHD and 0.11 for grade 3 to 4 aGVHD.

Figure 3.

Stratification of the whole cohort of patients (n = 201) according to the risk of developing chronic GVHD. Risk was calculated using the proposed predictive model including clinical variables (upper panels), genetic variables (middle panels), or both clinical and genetic variables (lower panels). Because the time of onset after transplantation was not available to build cumulative incidence curves, bar charts are shown for cGVHD (right panels) and extensive cGVHD (left panels). The cutoff used was 0.53 for cGVHD and 0.3 for extensive cGVHD.