P432: RISK STRATIFICATION IN ACUTE MYELOID LEUKEMIA: EXTERNAL VALIDATION OF THE STELLAE-123 GENE EXPRESSION MODEL AND COMPARISON WITH CURRENT PROGNOSTIC SCORES

Abstract

Background: The identification of recurrently mutated genes and cytogenetic anomalies has provided high prognostic and therapeutic significance in patients diagnosed with AML. Currently, some of this information has been incorporated into risk stratification guidelines, such as those of the European Leukemia Net (ELN). However, these guidelines are based on cytogenetic analyses and a limited number of mutations, and they don’t consider the genomic complexity of AML. Recently, we presented a new prognostic score based on gene expression analysis (Stellae-123) which achieved high discriminative power in survival prediction of AML patients, particularly among those bearing high-risk mutations. It is necessary to evaluate this signature in new AML cohorts and to test its performance compared with standard risk stratification scores. Aims: To validate the prognostic precision and discriminative power of the Stellae-123 gene expression signature in external cohorts and to compare its performance with standard prognostic scores: ELN-2017 in adult patients and clinical risk score in pediatric patients. Methods: RNA-seq data from 2 adult cohorts and one pediatric AML cohort was retrieved. The BeatAML (N=334) and the AMLCG-2008 (N=199) cohorts consisted of adult AML patients, whereas the TARGET AML cohort (N=144) was composed solely of pediatric patients. Gene expression estimates (FPKM) were normalized between cohorts using ComBat. Common genes were retrieved, and we selected those included in the Stellae-123 predictor, reaching a total of 69 genes. Random forests were built to predict survival in the largest cohort (BeatAML). The precision of the predictors was evaluated using cross-validated time-dependent AUCs derived from cox models. In the particular case of BeatAML, gene expression survival models were based on out-of-bag predictions in order to reduce the risk of overfitting during the training phase of the model. Results: We initially built a survival predictor using random forests based on the expression of the 69 genes incorporated in Stellae-123. This predictor was trained in the BeatAML cohort, and external validation was performed on the remaining two cohorts. The model achieved c-indexes of 0.635, 0.645 and 0.598 in the BeatAML, AMLCG-2008 and TARGEL AML cohorts, respectively. We then evaluated the precision of this signature to predict survival at 6 months, 1 year and 2 years after diagnosis using cross-validated cox models. The results indicated that this model achieved greater precision than the ELN-2017 and the pediatric clinical risk score in the majority of evaluated time points (Figure 1). Since age is a variable deeply associated with survival in AML, we evaluated the performance of the models including this covariate. We observed clear improvements in AUCs for the BeatAML and TargetAML, but not for the AMLCG-2008 cohort, a finding which is probably related to the fact that these were fit adult patients recruited in a clinical trial. Notably, we also observed an improved performance of the gene expression signature plus age model compared with the ELN-2017 and pediatric clinical risk scores plus age models, particularly in the prediction of survival within the first year after diagnosis. Image:Summary/Conclusion: The Stellae-123 gene expression signature can predict overall survival in AML with greater precision than the ELN-2017 and the pediatric clinical risk score. Therefore, gene expression profiling emerges as a powerful tool to optimize patient risk stratification. There is a growing need to standardize these tests for clinical use.