Acute myeloid leukemia (AML) is a genetically heterogeneous hematological malignancy with variable responses to chemotherapy. Although recurring cytogenetic abnormalities and gene mutations are important predictors of outcome, 50% to 70% of AMLs harbor normal or risk-indeterminate karyotypes. Therefore, identifying more effective biomarkers predictive of treatment success and failure is essential for informing tailored therapeutic decisions. We applied an artificial neural network (ANN)-based machine learning approach to a publicly available data set for a discovery cohort of 593 adults with nonpromyelocytic AML. ANN analysis identified a parsimonious 3-gene expression signature comprising CALCRL, CD109, and LSP1, which was predictive of event-free survival (EFS) and overall survival (OS). We computed a prognostic index (PI) using normalized gene-expression levels and β-values from subsequently created Cox proportional hazards models, coupled with clinically established prognosticators. Our 3-gene PI separated the adult patients in each European LeukemiaNet cytogenetic risk category into subgroups with different survival probabilities and identified patients with very high-risk features, such as those with a high PI and either FLT3 internal tandem duplication or nonmutated nucleophosmin 1. The PI remained significantly associated with poor EFS and OS after adjusting for established prognosticators, and its ability to stratify survival was validated in 3 independent adult cohorts (n = 905 subjects) and 1 cohort of childhood AML (n = 145 subjects). Further in silico analyses established that AML was the only tumor type among 39 distinct malignancies for which the concomitant upregulation of CALCRL, CD109, and LSP1 predicted survival. Therefore, our ANN-derived 3-gene signature refines the accuracy of patient stratification and the potential to significantly improve outcome prediction.
© 2019 by The American Society of Hematology.