Abstract

Full Text

NEJM -- Bullinger L et al. Use of Gene-Expression Profiling to Identify Prognostic Subclasses in Adult Acute Myeloid Leukemia. 350(16):1605-1616. Supplementary Appendixes Article: Bullinger L et al. Use of Gene-Expression Profiling to Identify Prognostic Subclasses in Adult Acute Myeloid Leukemia. N Engl J Med 2004;350(16):1605-1616.

View Figure

Supplementary Appendix 1. AML Study Group Ulm Multicenter Treatment Trials HD98A (Panel A) and HD98B (Panel B).

In Panel A, patients received two courses of induction therapy of idarubicin, cytarabine, and etoposide (ICE), one consolidation cycle of high-dose cytarabine and mitoxantrone (HAM), and a risk-adapted late consolidation cycle based on cytogenetic findings. The asterisk indicates that the second cycle of ICE was given with all-trans-retinoic acid (A or ATRA) and idarubicin (IDA) in patients with acute promyelocytic leukemia with t(15;17). The dagger indicates that bone marrow (BM) was collected, if leukapheresis (LP) was not possible. Double daggers indicate that this abnormality was independent of additional chromosomal aberrations. In Panel B, two cycles of ICE and one cycle of HAM (either with or without high-dose cytarabine and etoposide) were followed by a randomized post-remission therapy: either intravenous (IV) consolidation or oral maintenance therapy with idarubicin and etoposide (IE). Paragraph marks indicate all-trans-retinoic acid with idarubicin (AIDA) intravenously, plus all-trans-retinoic acid plus high-dose cytarabine and mitoxantrone (A-HAM) intravenously, plus AIDA orally in acute promyelocytic leukemia (APL) with t(15;17). A-HAE denotes all-trans-retinoic acid plus high-dose cytarabine and etoposide, G-CSF granulocyte colony-stimulating factor, HSCT hematopoietic stem-cell transplantation, CR complete response, PR partial response, RD refractory disease, R randomization, IC intensive chemotherapy, AUTO autologous HSCT, and ALLO allogeneic HSCT.

Supplementary Appendix 2

To develop an outcome predictor in the training sample set, we used a SAM gene set (genes correlating with the survival time) to group the samples into two subgroups by means of k-means cluster analysis (which is computationally less intensive than hierarchical clustering). We next used Kaplan–Meier survival analysis to determine the prognostic relevance of the two subgroups (and to assign good-outcome and poor-outcome labels to each subgroup). We then used the PAM method to identify a 10-fold cross-validated gene-expression predictor for these cluster-defined outcome classes. The entire procedure was then iterated, with the use of fractionally fewer SAM genes each time, until a minimal Kaplan–Meier P value was obtained by means of a log-rank test (i.e., the best separation of samples into prognostic groups with the use of SAM) and a minimal cross-validation error rate (i.e., the best prediction of outcome group with the use of PAM) was achieved, represented in our analysis by a set of 1600 genes with the use of SAM and a prediction set of 2655 genes with the use of PAM. A more manageable prediction set of 133 unique genes (represented by 149 complementary DNAs) was selected near the point on the PAM shrinkage curve at which the cross-validation error started to rise steeply (but was still within 1 SE of the minimum) and was used for all subsequent analyses.

View Figure

Supplementary Appendix 3. Effect of the Use of Various Preclustering Data-Filtering Criteria.

In Panel A, to determine the robustness of our AML sample classes, we performed two-way hierarchical cluster analysis using 12 different sets of genes, obtained by varying preclustering data-filtering criteria as shown. The number of genes remaining in each set after data filtering is also indicated. Panels B and C show examples of two hierarchical cluster-derived dendrograms produced with the use of the lowest-stringency filtering criteria (18,291 genes) in Panel B and the highest-stringency filtering criteria (1435 genes) in Panel C. For the 2 sample dendrograms shown, as well as for the 10 other sample dendrograms generated from the filtering criteria shown in Panel A, the majority of AML samples with a normal karyotype segregated into two main groups (I and II), with membership largely preserved within each group. Sample labels are color-coded to indicate cytogenetic groups. Dendrogram branches are color-coded to indicate subgroups I and II with a predominantly normal karyotype.

View Figure

Supplementary Appendix 4. Supervised Analysis with the Use of the Multiclass Significance Analysis of Microarrays (SAM) Method.

Panel A shows the two-way hierarchical cluster of 116 AML samples (columns) and the 728 genes (rows) identified by multiclass SAM analysis to be significantly distinctly expressed among nine different cytogenetic groups. Mean-centered gene-expression ratios are depicted by a log-transformed (on a base 2 scale) pseudocolor scale (indicated). Panel B shows an enlarged sample dendrogram. Samples are color-coded according to cytogenetic groups as indicated. Panels C, D, E, F, G, H, I, and J show selected gene-expression “features” extracted from the cluster (locations indicated by vertical colored bars). Named genes (but not expressed-sequence tags) are indicated.

View Figure

Supplementary Appendix 5. Schematic Depiction of the Hypothetical Relationship between Prognostically Relevant Tumor Subclasses and Survival.

The overlapping probability-density curves suggest that the duration of survival is a poor surrogate for underlying tumor subclasses, illustrating the motivation for developing our approach to outcome prediction using gene- expression profiles.

Supplementary Appendix 6

In the training set, hierarchical cluster analysis across the 133 predictive genes (represented by 149 complementary DNAs) identified two major gene clusters, one more highly expressed in the good-outcome subgroup, and the other more highly expressed in the poor-outcome subgroup (Fig. 4A). Two major gene clusters were also identified in the test set (Fig. 4B). A chi-square test demonstrated a significant correlation between membership in the good-outcome and poor-outcome gene clusters in the training set and membership in the two major gene clusters in the test set (P<0.001), confirming the detection of good-outcome and poor-outcome gene signatures in the test set. Furthermore, the average (centroid) gene-expression signature across the 133 genes for the good-outcome and for the poor-outcome subgroups in the training set was highly correlated with the average gene-expression signature across the 133 genes for each of the two test sample subgroups defined by hierarchical clustering (Pearson correlation coefficients, 0.74 and 0.72, respectively), also confirming the presence of good-outcome and poor-outcome signatures in the test set.

View Figure

Supplementary Appendix 7. Outcome Prediction by Means of the Nearest Shrunken Centroid.

In Panel A, columns represent AML samples in the training set ordered according to k-means clustering; rows represent the 149 predictive complementary DNAs (cDNAs), ordered according to hierarchical clustering. Mean-centered gene-expression ratios are depicted by a log-transformed (on a base 2 scale) pseudocolor scale; gray denotes poorly measured data. Good-outcome and poor-outcome subgroups were identified by means of Kaplan–Meier analysis. In Panel B, columns represent AML samples in the test set, ordered according to the rank value of Pearson correlation to the nearest shrunken centroid of good-outcome or poor-outcome class in the training set (see text); rows represent the 149 predictive cDNAs, ordered according to hierarchical clustering. Vertical bar (left) indicates genes that were expressed in the good-outcome (blue) or poor-outcome (red) subgroup in the training set. Panel C shows Kaplan–Meier survival analysis in the two subgroups according to the defined nearest-shrunken-centroid method; P = 0.034 for the difference between subgroups by the log-rank test.


	Abstract
	Full Text

Comments and questions? Please contact us.