Abstract

The use of risk adapted therapy has improved outcome for children with ALL and assessment of early response (ER) has been used to identify a subset of patients who benefit from augmented treatment. In spite of these advances, the majority of treatment failures occur in a “good risk” subset indicating the critical need for better predictors of outcome. We performed oligonucleotide microarrays on 99 B-precursor patients treated on CCG 1961 for high risk ALL to identify: 1) genetic pathways that play a role in early response to therapy (42 day 7 rapid (R)ER, 40 slow (S)ER) and 2) a signature that would predict event free survival (28 pts > 4 yr CCR, 31 relapses < 3 yrs). RNA was extracted and amplified before hybridization to Affymetrix U133 Plus 2.0 microarrays representing over 47,000 transcripts. Data were normalized, filtered and then analyzed by both unsupervised and supervised methods. We identified a robust signature that predicted SER at day 7 (79% of cases in a test set) but not RER thus limiting its utility in risk stratification. Furthermore, we also identified a highly significant set of genes that correlated with SER (M2/M3) at day 14. Importantly, expression of 47 probe sets were significantly different in patients who were in CCR for more than 4 years compared to cases that experienced a relapse within the first 3 years (FDR ≤0.05%). Logistic regression indicated that each of the 47 probe sets has significant prognostic value beyond that contained in the clinical covariates (including gender, age and WBC count). A major subset of these genes were involved in signal transduction (MAP3K3, YWHAZ, PCTK2, TGFBR1, PTPRE, IFNGR2, HSPA8). In a separate validation analysis, subsets of genes were selected using logistic regression with backward, forward and stepwise variable selection and the models proved to be significantly predictive of response in the independent expression data set of 220 patients generated previously using a smaller probe set with U95Av2 microarrays (Mosquera-Caro et al, Blood 2003, 102(11)). Interestingly, the genetic models superseded clinical co-variates such as age, WBC and gender. Finally, to assess further the predictive accuracy of these gene signatures, we randomly divided the data set into a training set (2/3 cases) that was used to develop a signature predictive of relapse that was tested for accuracy in the remaining cases (1/3 cases). This was repeated three times. On average, 77% of patients in CCR and 73% of patients who relapsed were accurately classified in the test set. With further validation these signatures of outcome (not early response) will allow for more precise identification of patients who are at very high risk for relapse and may provide a mechanistic understanding of why patients fail therapy.

Author notes

Corresponding author