Acute myeloid leukemias with reciprocal rearrangements can be distinguished by specific gene expression profiles

Selection and Characterization of Leukemia Samples.

We selected bone marrow samples from 37 AML patients representing four morphological and three underlying cytogenetic subgroups. All samples were newly diagnosed de novo AML and were characterized by cytomorphology, cytogenetics, fluorescence in situ hybridization, and molecular genetics using standard procedures (1, 3, 18–24). Samples used for gene expression analyses had been lysed immediately, frozen, and stored at −80°C from 1 to 34 months. The targets for GeneChip analyses were prepared according to the current Expression Analysis Technical Manual (Affymetrix, Santa Clara, CA). Detailed procedures are described in additional Methods, which are published as supporting information on the PNAS web site, www.pnas.org.

Class Separation by Principal Component Analysis.

Potential clusters corresponding to the genetic subgroups were visualized with a two-step approach. The data were scaled from each array to a target intensity value of 50 (Affymetrix MICROARRAY SUITE 4.0.1) to be able to perform interarray comparisons. All data were permutated 100 cycles by using the multiclass response parameter of the Significance Analysis of Microarrays algorithm (SAM) (25) (http://www-stat.stanford.edu/∼tibs/SAM/index.html). The total set of 12,600 genes was reduced to the significant differentially expressed genes. In a second step, the reduced set of genes was prepared for principal component analysis and analyzed with j-express (26) (http://www.molmine.com/). For visualization in a two-dimensional plot we chose the first two principal components as they captured most of the variation in the original data set.

Class Prediction by Weighted Voting.

We adapted a previously described method to reduce the number of candidate genes that could distinguish between the three different cytogenetic AML subgroups (13). Briefly, to avoid division by zero or negative numbers as occurs because of the expression algorithm (Affymetrix MICROARRAY SUITE 4.0.1) we set all average fluorescence intensities of 1 or less to 1. Then, gene expression levels were log-transformed. Performing pairwise comparisons (A vs. B), for each gene g P(g,c) values and votes [defined by: P(g,c) = (m1(g) − m2(g))/(s1(g) + s2(g))] were calculated based on mean expression levels (m) and standard deviations (s) in the respective cytogenetic subgroup. Subsequently, votes were summed and prediction strength values reflected the margin of victory in the direction of either cytogenetic group A or B of the pairwise comparison. Prediction strength values range between 0 and 1, values >0.45 demonstrate significance (according to the permutation test). The relevance of selected genes was assessed by performing leave-one-out cross-validation. Only those genes that were contained in all cross-validation classifiers were considered important. To determine a random association between genes we performed a permutation test (100 cycles). Because the number of informative genes, which are required to discriminate between samples, is unknown, we applied this method for different numbers of informative genes (range: 2 to 200). The minimal set of genes that provided optimal classification accuracy together with the highest prediction strength was selected to avoid overfitting. To visualize the identified genes and check their suitability for class separation a hierarchical cluster analysis was performed by using j-express (26) (cluster method: average linkage; distance metric: euclidean). The accuracy of this class prediction model was validated on an independent test set of five cases of AML not fulfilling the cRNA high-quality criterion.

Multiple-Tree Classifier.

As basic units in this classifier, classification trees are used (27–29). The optimal number of trees has been determined to be 15 (data not shown). Class votes of these trees are aggregated by a vote-by-majority rule. The classifier was fed with gene expression intensity values from a set of 973 genes that had been chosen based on their r statistic:

where μ_i refers to the class averages, μ̄ to the overall average, σ_i, to the within-class standard deviation, and summation is carried out over all k classes. The threshold was set to r > 0.75. Classification trees were constructed as follows: tree building was performed while restricting trees to contain no more than n-1 nodes to discriminate between n classes. The C5.0 algorithm was used (28). The variables (gene expression intensities) used for tree construction were eliminated from the data set, and a new tree was calculated based on the truncated data set. This procedure was iterated until the predetermined number of trees had been reached. The accuracy of the multiple-tree classifier was estimated by 10-fold cross-validation (30) and on an independent test set of data from five bone marrow aspirates, where the quality of the corresponding cRNA preparation was slightly lower than the high-quality standards required for the training set.

Characterization of Leukemia Samples.

We investigated 37 AML cases representing three defined cytogenetic aberrations corresponding to four FAB subtypes: t(8;21)(q22;q22)/AML M2 (n = 9), t(15;17)(q22;q12)/AML M3 or AML M3v (n = 10, n = 8), and inv(16)(p13q22)/AML M4eo (n = 10). All cases were characterized by cytomorphology, cytogenetics, fluorescence in situ hybridization, and reverse transcription–PCR (see Fig. 4, which is published as supporting information on the PNAS web site). All cases with AML and t(8;21) had AML M2, all with AML and inv(16) had AML M4eo, 10 cases with AML and t(15;17) had AML M3, and eight cases with AML and t(15;17) had AML M3v. All patients showed these balanced abnormalities as the sole karyotype change. Using fluorescence in situ hybridization analysis, more than 65% of cells demonstrated the specific signal constellation. The respective fusion transcripts were detected by reverse transcription–PCR in all samples. The median age of all patients was 53 years (range, 19–82 years; male/female = 15:22) and did not differ between the respective groups. AML subtypes M3 and M3v both carry the same chromosomal aberration but differ in morphological aspects like nuclear configuration, granulation, and clinical aspects like white blood cell count. The median white blood cell count was 20,000/μl (range, 800–168,000/μl) and was strikingly lower in patients with AML M3 as compared with all other patients (median, 6,200 vs. 36,500/μl, P = 0.0002).

Microarray Analyses.

The gene expression profiles of 37 AML samples were evaluated. Thirty-two hybridization cocktails demonstrated high-quality cRNA characteristics (Test3 probe arrays: 3′/5′ ratio of glyceraldehyde-3-phosphate dehydrogenase probe sets ≤3.0) and were selected for building class prediction models: t(8;21)/AML M2 (n = 7), t(15;17)/AML M3 or M3v (n = 9, n = 7), and inv(16)/AML M4eo (n = 9). Five cases were primarily excluded (3′/5′ ratios ranging between 3.9 and 5.4) and were used for subsequent validations of the class prediction models: t(8;21)/AML M2 (n = 2), t(15;17)/AML M3 or M3v (n = 1, n = 1), and inv(16)/AML M4eo (n = 1).

Class Separation by Principal Component Analysis.

To visualize clusters corresponding to the three underlying genetic subgroups we applied a two-step approach. Based on a permutation test (100 permutations) we correlated our expression data to the three different cytogenetic parameters (25). We obtained 1,000 significant genes. By principal component analysis we were able to clearly separate the three distinct chromosomal aberrations t(8;21), t(15;17), and inv(16) (Fig. 1) (26). These data suggest that genetically defined AML subtypes can be specified and identified based on their gene expression profiles.

Class Prediction by Weighted Voting.

To identify the genes that enable the accurate discrimination of these subgroups, we applied the data analysis methodology introduced by Golub et al. (13). We selected the minimal set of genes that provided optimal classification accuracy together with the highest prediction strength to avoid overfitting. Thirteen genes were sufficient to separate these AML subtypes with high precision (Table 1). GenBank accession numbers and detailed descriptions of the genes are given in Table 2.

All 32 clinical samples could be assigned to their corresponding cytogenetic subtype with best accuracy in leave-one-out cross-validation (1.0). Prediction strength values ranged from 0.91 to 0.98 (Table 1). To illustrate these results we applied hierarchical clustering (31). The resulting dendrogram clearly demonstrates the capacity of this subset of genes to separate all AML cases according to their cytogenetic aberration (Fig. 2). This finding demonstrates that class prediction of a chromosomal aberration in AML is feasible solely based on gene expression data.

For external validation, we tested whether primarily excluded samples could also be accurately assigned to their specific cytogenetic category. Despite their nonoptimal cRNA quality, all five cases were correctly classified with high prediction strength (0.76, 1.00, 1.00, 1.00, 1.00).

Class Prediction by Multiple-Tree Models.

As a second and independent methodological approach we developed a multiple-tree classifier to separate the three genetically defined subtypes based on the expression level of a minimal set of genes. In short, we computed classification trees to discriminate between the different AML subclasses. To avoid overfitting of a singular tree model, we computed a multiple-tree model by using an iteratively reduced set of genes. For each tree, we used only those genes that have not been used by the previously computed classification tree. The procedure is stopped when a predetermined number of trees has been reached. For this study, the optimal number of trees was calculated to be 15. The votes of the 15 trees were aggregated by a vote-by-majority rule. Equal votes for two of the three classes were counted as misclassification.

The classifier used the expression values of 29 genes (MYH11 was identified twice by two different probe sets; Table 2) to discriminate between three classes, namely samples displaying t(15;17), t(8;21), and inv(16) (Fig. 3). The accuracy on the training set (n = 32) was 100% and on the independent test set (n = 5) 100%. The average accuracy in 10-fold cross-validation was 94%.

In summary, we identified 36 genes by using two independent methodologies for class prediction in AML (Table 2). Six genes were described in both calculations, seven were found exclusively in the minimal set according to Golub et al. (13), and another 23 genes by using multiple-tree classifiers.

Correlation of Phenotype and Gene Expression Profile.

We were able to demonstrate striking correlations between genotype and gene expression profiles in three genetically defined subgroups of AML. In addition, we answered the question of whether the cytogenetically identical AML with t(15;17) but appearing with two different phenotypes, AML M3 or AML M3v (see Fig. 4), can also be separated by different gene expression patterns. We used 100-fold permutation of M3 (n = 10) and M3v (n = 8) data followed by principal component analysis and hierarchical cluster analysis based on 82 informative genes (data not shown). Separation into the corresponding two morphologically defined FAB subtypes M3 and M3v was possible in all cases (see Fig. 5, which is published as supporting information on the PNAS web site) and suggests also a close correlation between phenotype and gene expression profile.