Abstract
The assessment of early treatment response based on minimal residual disease (MRD) detection is a powerful prognostic indicator in childhood acute lymphoblastic leukemia (ALL). To identify genes whose expression is associated with poorer early response and to define gene expression signatures predictive of MRD findings, we correlated gene expression profiles of diagnostic bone marrow blasts in 236 children with ALL enrolled in St. Jude Total Therapy XIIIA-XV protocols with MRD results obtained at days 19 and 46 of remission induction treatment. The dataset consisted of 46 T-lineage ALL and 190 B-lineage ALL; the latter included 10 BCR-ABL, 11 E2A-PBX1, 12 MLL rearranged, 49 TEL-AML1, 46 hyperdiploid >50 chromosomes (HD>50) karyotype, 3 BCR-ABL plus HD>50, and 59 other cases. RNA expression profiles were obtained using Affymetrix U133A gene chips; MRD was assessed by a flow cytometric assay that allows the identification of one leukemic cell among 10,000 normal bone marrow cells or greater and is applicable to approximately 95% of patients. We used a general linear model to eliminate the possible confounding influence of genetic subtypes known to be associated with treatment response. Then, we applied a t-test with the P value threshold of 0.006, determined by the profile information criterion for large-scale multiple tests. By this criterion, 279 probe sets were associated with MRD at day 19 (estimated false-discovery rate [FDR] 0.42) and 606 probe sets with MRD at day 46 (estimated FDR 0.17); 41 probe sets were associated with MRD at both time points. The expression of CASP8A2 (FLASH, CED-4), which encodes a key mediator of apoptosis and participates in glucocorticoid signaling, was significantly lower in cases with MRD at both time points. In a cluster analysis using the probe sets associated with MRD, the capacity to predict results of the MRD assay was limited. For example, only 69% of MRD-negative and 81% of MRD-positive results at day 19 were correctly classified. Similar results were obtained using the day 46 data. We also determined whether MRD status could be predicted by an unsupervised cluster analysis of all 236 cases with 17,269 probe sets (after removing transcripts not expressed in any of the samples). Although there was a strong association of cluster formation with lineage and genetic subtypes, there was no significant association with MRD status at days 19 or 46. Moreover, there was no significant association with MRD status in analyses limited to a series of 66 ‘standard-risk’ B-lineage ALL cases (excluding those with BCR-ABL, TEL-AML1, MLL, hypodiploid <45 chromosomes or HD>50), or to cases of each individual genetic subtype. In conclusion, leukemic cells at diagnosis express genes that are associated with MRD. Although gene expression profiles can accurately identify leukemia cell lineage and genotype, they cannot accurately predict MRD status, probably owing to the multifactorial nature of treatment response, which is influenced not only by cellular drug resistance but also by clinical and pharmacologic variables of the host.
Bayesian approach for predicting responses to therapy from high-dimensional time-course gene expression profiles
