Allogeneic hematopoietic stem cell transplantation is an important last resort therapy for hematological diseases. Unfortunately, the success of a large proportion of these transplants is limited by graft-versus-host disease (GVHD). Currently known risk factors for acute GVHD (histoincompatibility, sex mismatch, older patients, previous pregnancies) do not provide a precise estimate of individual patient risk and do not help for individualization of the therapy. Although GVHD may have beneficial graft-versus-tumor effects, observational studies have identified GVHD as the main cause of non-relapse mortality. Thus, early identification of those patients who will develop aGVHD may allow for individualized treatment, and also for the reduction of unnecessary treatment for those patients not at risk. Nowadays, however, there is no diagnostic method that allows prediction of aGVHD. Thus, the goal of our study was to reveal a gene expression profile that would predict the occurrence of aGVHD.Material and methods: Between May 2004 and August 2007, we collected blood samples from 89 recipients of myeloablative (N=71) and non-myeloablative (N=17) HLA-identical sibling allogeneic hematopoietic stem cell transplants at the time of successful engraftment (2–3 weeks after transplantation). Acute GVHD prophylaxis was cyclosporine plus short course of methotrexate and cyclosporine plus mycophenolate mofetil in myeloablative and non myeloablative transplants, respectively. Forty patients (45%) experienced acute GVHD (Glucksberg criteria). All the patients had acute GVHD before the first three months after transplantation (median 31 days; range 21–88 days). We isolated total RNA from the peripheral blood mononuclear cells, amplified it, labeled, and co-hybridized to the microarray slides containing probes for 22,000 genes. We used BRB ArrayTools for data analysis. The patients were divided into training and test groups, the former used to build a model discriminating patients with and without aGVHD and the latter - to test the model on independent samples. We selected the informative genes using “recursive feature elimination” method followed by six different multivariate classification algorithms (compound covariate predictor, diagonal linear discriminant analysis, 1- and 3-nearest neighbor predictor, nearest centroid predictor, support vector machine, Bayesian predictor) in order to establish a molecular classifier in the training set. Then we validated this new classifier in the test set of patients. Because multiple classification algorithms may give different results, we considered a result being robust and a patient being “classifiable” when at least five out of seven algorithms were concordant. The other patients were considered as “unclassifiable” because no reliable prediction regarding development of aGVHD could be made.Results: We found a molecular classifier comprised by 244 gene probes in the training set of samples which were selected based on the most accurate classification. In the test group of samples, we found that 80% of patients could be classified based on the concordance between classification methods as described above. For these patients, the classifier showed 78% of a predictive accuracy (75% of sensitivity and 80% of specificity).Conclusion: Our results show that molecular profiling is able to identify patients under high risk of acute GVHD at the time of engraftment. Establishing of a predictive diagnostic method for aGVHD is the first step of individualization of therapeutic strategy after hematopoetic stem cell transplantation.