Gene expression signatures predictive of early response and outcome in high-risk childhood acute lymphoblastic leukemia: A Children's Oncology Group Study [corrected]

Abstract

Purpose: To identify children with acute lymphoblastic leukemia (ALL) at initial diagnosis who are at risk for inferior response to therapy by using molecular signatures.

Patients and methods: Gene expression profiles were generated from bone marrow blasts at initial diagnosis from a cohort of 99 children with National Cancer Institute-defined high-risk ALL who were treated uniformly on the Children's Oncology Group (COG) 1961 study. For prediction of early response, genes that correlated to marrow status on day 7 were identified on a training set and were validated on a test set. An additional signature was correlated with long-term outcome, and the predictive models were validated on three large, independent patient cohorts. Results We identified a 24-probe set signature that was highly predictive of day 7 marrow status on the test set (P = .0061). Pathways were identified that may play a role in early blast regression. We have also identified a 47-probe set signature (which represents 41 unique genes) that was predictive of long-term outcome in our data set as well as three large independent data sets of patients with childhood ALL who were treated on different protocols. However, we did not find sufficient evidence for the added significance of these genes and the derived predictive models when other known prognostic features, such as age, WBC, and karyotype, were included in a multivariate analysis.

Conclusion: Genes and pathways that play a role in early blast regression may identify patients who may be at risk for inferior responses to treatment. A fully validated predictive gene expression signature was defined for high-risk ALL that provided insight into the biologic mechanisms of treatment failure.

Fig 1.

Receiver operating characteristic (ROC) curve of the predicted score of early response on the test set. The model that comprised 24 probe sets that were derived from the training set was used for the prediction of early response on the test set. The ROC accuracy (ie, the area under the curve) is A = 0.77, which is significantly larger than that of noninformative prediction (P = .006). By using a threshold of C = 0.4 (determined in training set), the model correctly predicted 21 of 28 patient cases (success rate, 0.75). RER, rapid early responder; SER, slow early responder.

Fig 2.

Genes differentially expressed in patients that remained in complete continuous remission v in those that relapsed. Heatmap of the 47–probe set signature that was predictive of outcome (which represented 41 unique genes).