Gene Expression Profiling of Paired Pre- and Post-Prednisolone (PRED) BM Samples from Childhood ALL Identifies Robust Signatures for PRED Response and Eventual Outcome.

Early response to therapy is the most important prognostic factor for childhood ALL. CCG investigators have shown that Day-7 and Day-14 BM blast counts were prognostically important although there is great inter-observer variability. BFM group have shown that day 8 prednisolone (PRED) response is highly predictive of the treatment outcome. While gene expression profiling (GEP) of diagnostic marrow can discern a pattern of PRED sensitivity as determined by in vitro MTT assay, the accuracy was low at only 70%. We hypothesized that changes in global GEP after therapy have a higher likelihood to predict response as the signatures of sensitivity and resistance may be unmasked during the therapy. We prospectively studied the changes in GEP using Affymetrix HG-U133A or Plus 2 chips on paired BM samples before and after 7-day course of PRED and one dose IT MTX in 58 patients with newly diagnosed or relapsed ALL. Unsupervised hierarchical clustering revealed that pre- and post- PRED samples in the patients still tended to cluster together, indicating that expression profiles of molecular subgroups were still most important. To remove intrinsic influence of molecular subtypes and identify potential signatures independent of genetic abnormalities, we subtracted Day-0 GEP from its paired Day-8 profile and retained probe sets with significant changes (≥ 10-fold). To avoid the ambiguity of variation in BM blast counting at Day-8, we divided the samples into a stringently reproducible group where “Good” PRED response was defined as that Day-8 blast count in PBL < 10⁹/L and BM lymphoblasts ≤ 30% (n=16). “Poor” response was when Day 8 PBL ≥ 10⁹/L (n=11). This stringently reproducible group (n=27) formed the training group to help define a distinct signature while the rest (n=31 pairs) were used as a blinded test set. 54 and 19 discriminating genes were identified by 2 independent statistical methods respectively, and an integrated predictor model was constructed based on shortlisted entries. This model predicted the PRED response with 100% accuracy for the training set using the leave-one-out cross validation but was less accurate in predicting the BM blast count in blinded test set. But intriguingly, in the blinded test set, this model predicted correctly 19 out of 21 reliable “Good” PRED responses are in CCR (91%), while among 8 predicted as “Poor” responses, only 2 are in CCR (25%). This suggests that as gene expression profiling as early as day 8 of PRED could discern the beginning of leukaemia cell death even before morphological changes are discernable and is highly correlated to eventual outcome. In conclusion, we have shown that analyses on the relative changes of gene expression profile can identify real genetic signatures indicating the sensitivity to PRED administration which is highly correlated with outcome.