Detection of CEBPA Double Mutants in Acute Myeloid Leukemia Using a Custom Gene Expression Array

Abstract

Double (bi-allelic) mutations in the gene encoding the CCAAT/enhancer-binding protein-alpha (CEBPA) transcription factor have a favorable prognostic impact in acute myeloid leukemia (AML). Double mutations in CEBPA can be detected using various techniques, but it is a notoriously difficult gene to sequence due to its high GC-content. Here we developed a two-step gene expression classifier for accurate and standardized detection of CEBPA double mutations. The key feature of the two-step classifier is that it explicitly removes cases with low CEBPA expression, thereby excluding CEBPA hypermethylated cases that have similar gene expression profiles as a CEBPA double mutant, which would result in false-positive predictions. In the second step, we have developed a 55 gene signature to identity the true CEBPA double-mutation cases. This two-step classifier was tested on a cohort of 505 unselected AML cases, including 26 CEBPA double mutants, 12 CEBPA single mutants, and seven CEBPA promoter hypermethylated cases, on which its performance was estimated by a double-loop cross-validation protocol. The two-step classifier achieves a sensitivity of 96.2% (95% confidence interval [CI] 81.1 to 99.3) and specificity of 100.0% (95% CI 99.2 to 100.0). There are no false-positive detections. This two-step CEBPA double-mutation classifier has been incorporated on a microarray platform that can simultaneously detect other relevant molecular biomarkers, which allows for a standardized comprehensive diagnostic assay. In conclusion, gene expression profiling provides a reliable method for CEBPA double-mutation detection in patients with AML for clinical use.

Introduction

One of the hallmarks of acute myeloid leukemia (AML) is the accumulation of undifferentiated myeloid cells carrying specific genetic aberrations in the bone marrow (Löwenberg et al., 1999). The CCAAT/enhancer-binding protein-alpha gene (CEBPA) encodes a transcription factor that has been found to play a critical role in granulopoiesis and is mutated in 5%–14% of patients with AML (Pabst et al., 2001; Gombart et al., 2002; Preudhomme et al., 2002; Snaddon et al., 2003; van Waalwijk van Doorn-Khosrovani et al., 2003; Fröhling et al., 2004; Nerlov, 2004; Bienz et al., 2005). Two types of mutations in CEBPA are more frequently found in AML; first, out-of-frame mutations or mutations introducing a stop codon in the N-terminal part of the gene causing premature stops in translation (Zhang et al., 2004), and second, in-frame mutations in the C-terminal basic leucine zipper (bZIP) region of the gene, which appear to impair DNA binding and/or homo- and heterodimerization (Nerlov, 2004). Most AML cases with mutant CEBPA have two mutations, most probably bi-allelic (referred to as CEBPA dm, double mutant, here). The majority of these cases carry a combination of an N-terminal and C-terminal mutation (Nerlov, 2004; Leroy et al., 2005; Pabst and Mueller, 2007). A minority of CEBPA mutant AML cases are mono-allelic, with a single mutation (referred to as CEBPA sm, single mutant) (Nerlov, 2004; Leroy et al., 2005).

The WHO has defined AML with CEBPA mutations (single and double collectively) as a provisional favorable subtype of AML (Swerdlow et al., 2008). AML cases with a CEBPA mutation are predominantly found among the cytogenetically intermediate risk group, mainly those with normal karyotypes (Grimwade et al., 1998). However, AMLs carrying CEBPA double mutations rather than CEBPA single mutations represent a distinct prognostically favorable group of AML cases, and therefore, CEBPA double mutants provide a clinically useful marker for risk stratification of the leukemia (Pabst et al., 2009; Renneville et al., 2009; Wouters et al., 2009; Green et al., 2010; Taskesen et al., 2011). Finally, AML cases have been identified that are CEBPA wt, yet are hypermethylated (Figueroa et al., 2009). Not surprisingly, these AMLs display gene expression patterns that are very similar to those of CEBPA double mutants because a shared set of downstream transcripts are low either due to the bi-allelic mutations or the hypermethylated promoter status of CEBPA (Figueroa et al., 2009; Taskesen et al., 2011). However, CEBPA double-mutation AMLs and CEBPA hypermethylated AMLs are different as regard to the gene expression levels of CEBPA itself, which are generally higher in the double-mutant AML subtypes, but low to absent in the hypermethylated subtype.

Classification of CEBPA dm cases has been previously demonstrated using gene expression profiling (Wouters et al., 2009); however, misclassified cases were the CEBPA hypermethylated AMLs (Figueroa et al., 2009; Taskesen et al., 2011). Herein we have developed a two-step classifier, which explicitly distinguishes the CEBPA double mutants from the CEBPA hypermethylated. Accuracy of this two-step classifier was established by means of a double-loop cross validation on a cohort of 505 primary AML cases. The two-step classifier was then incorporated on a custom chip, along with detection capabilities for AML1-ETO, CBFB-MYH11, PML-RARA, other mutations such as NPM1 (Van Vliet et al., Submitted), as well as overexpression of prognostic genes EVI1 and BAALC (Brand et al., in preparation). Thus, this gene expression assay provides a standardized test for multiple molecular biomarkers relevant for AML at the time of diagnosis.

Materials and Methods

Custom in vitro diagnostic gene expression array

A customized Affymetrix (Santa Clara, CA) GeneChip microarray was designed (AMLprofiler^™). This custom chip includes a subset of the probe sets present on the Affymetrix HG-U133 Plus2 platform and a subset of probes designed for specific purposes. Its resulting .CEL file is automatically routed through proprietary software that reports 0 or 1 for CEBPA wt and CEBPA dm, respectively.

Datasets

We employed the blood or bone marrow specimens from 505 AML patients, who had been enrolled in the Dutch–Belgian Hematology–Oncology Cooperative group protocols -04, -04A, -29, and -42 (HOVON-SAKK, the Erasmus University Medical Center, Rotterdam), (Valk et al., 2004; Wouters et al., 2009). The 505 cases employed here are a subset from the 524 cases (Wouters et al., 2009) for which the hybridization cocktail was available. The 505 set contains all 26 CEBPA dm and 12 CEBPA sm cases from the Wouters et al. (2009) cohort (Table 1). Sample processing and quality control were performed as previously described (Wouters et al., 2009). Bi-allelic mutant, single mutant, and wild-type CEBPA annotations were confirmed in all cases by the entire CEBPA coding region investigation by denaturing high-performance liquid chromatography (dHPLC), analysis of selected regions by agarose gel analysis (van Waalwijk van Doorn-Khosrovani et al., 2003), and/or nucleotide sequencing (Wouters et al., 2009). CEBPA promoter hypermethylation was determined using the HELP assay (HpaII tiny fragment enrichment by ligation-mediated PCR), as previously described (Valk et al., 2004). Table 1 shows an overview of the CEBPA mutation status of all AML cases in the cohort.

Table 1.

Overview of Cases in the Cohort, and Distribution of CEBPA Mutations

		Training cohort		Above step 1 threshold
CEBPA status	Label	n	%	n	%
CEBPA WT	Neg	458	90.7%	48	60.8%
CEBPA Hypermethylated	Neg	9	1.8%	0	0.0%
CEBPA single mutant	Neg	12	2.4%	5	6.3%
CEBPA double mutant	Pos	26	5.1%	26	32.9%
Total		505	100.0%	79	100.0%

Hybridization cocktails were obtained for all cases (Verhaak et al., 2009), and were hybridized onto the AMLprofiler platform. All .CEL files were preprocessed using the MAS5 algorithm (scaling to 1500), and probe set intensities below 30 truncated to 30. Subsequently, geometric mean centering was applied per probe set relative to a subset of 244 AML cases.

AML profiler CEL files are available at the Gene Expression Omnibus (National Center for Biotechnology Information) under accession number GSE42194.

All patients provided written informed consent in accordance with the Declaration of Helsinki.

Double-loop cross validation

To build diagnostic classifiers from high throughput data, we used the double-loop cross-validation framework (Wessels et al., 2005). We adopted this methodology combined with forward filtering as the feature selector, the signal to noise ratio as a criterion to evaluate the individual genes, and ClaNC (Dabney, 2005) a simple classifier that is known to perform well on this type of data. The double-loop cross validation was executed with 100 repeats of 26-fold cross validation in the outer (validation) loop, and 10-fold cross validation in the inner loop. Learning curves were constructed for up to 100 genes, using the average of the sensitivity and specificity as criterion (reported as percentages, 50% is random classification, 100% is perfect classification) to be optimized. At all points, data splits were stratified with respect to the class prior probabilities. To optimize a classifier toward less FN or less FP cases, the prior probabilities were adjusted, such that the classifier boundary gets shifted.

Results

Development of a classifier for CEBPA dm

We set out to develop a classifier for detecting CEBPA dm AML cases in a cohort of 505 patients with AML (Table 1). For this, we consider the CEBPA dm as the positive group, and the remainder as negative (CEBPA wt, sm, or hypermethylated). This provides a six gene classifier with high performance, as shown in Figure 1 and Table 2. The misclassified cases are predominantly the hypermethylated cases (Fig. 1), which are consistent with their biological background, as the downstream gene expression profiles from a nonfunctional CEBPA protein due to the bi-allelic gene mutations may be equivalent to the absence of the protein due to the hypermethylated gene status. Although this classifier has a high accuracy, from a clinical utility point of view, false-positive cases are undesirable as they might be at risk of undertreatment given the favorable prognosis of CEBPA double mutants. Therefore, we investigated methods to exclude the hypermethylated cases before the application of a classifier.

FIG. 1.

CEBPA hypermethylated cases are often FP in CEBPA double-mutant classifiers. The scatter plot showing the output of a one-step classifier trained on the 505 cases as determined by the double-loop cross-validation procedure. Numbers indicate how often a case was misclassified in the 100 repeats (samples always correctly classified do not have a number printed). Color images available online at www.liebertpub.com/gtmb

Table 2.

Double-loop Cross-Validation Performances of the Classifiers for Predicting CEBPA Double Mutations

							95% CI			95% CI
Classifier	n	TP	FN	TN	FP	Sens	LL	UL	Spec	LL	UL
Overall classifier	505	24.79	1.21	474.70	4.30	95.3	80.0	99.1	99.1	97.8	99.6
Two-Step classifier	505	25.00	1.00	479.00	0.00	96.2	81.1	99.3	100.0	99.2	100.0

Sens, sensitivity; spec, specificity; CI, confidence interval; LL, lower limit; UL, upper limit.

Development of a two-step classifier for CEBPA dm

Since AML cases with hypermethylated CEBPA show reduced CEBPA gene expression levels, we propose a two-step classifier. In step 1, the CEBPA expression level is assessed in comparison with a threshold. In step 2, only the selected cases exceeding that threshold will subsequently be input into the classifier (Fig. 2).

FIG. 2.

Overview of the two-step classifier for predicting CEBPA dm.

The step 1 threshold was developed as follows. The expression levels for the different CEBPA mutation groups were compared (Fig. 3A). As expected, the hypermethylated CEBPA cases all have notably low CEBPA expression values. The CEBPA wt cases show a wide range of variable CEBPA expression levels, whereas the CEBPA sm cases have slightly elevated CEBPA expression. However, the CEBPA dm cases all share high CEBPA expression levels. The aim of the step 1 classifier is to classify all hypermethylated cases below the threshold and all CEBPA dm cases above the threshold. The threshold is determined by the intersection of the two fitted normal distributions on the group of CEBPA dm cases and CEBPA hypermethylated cases. At that point, the groups have an equal chance of correct classification (i.e., where the overall chance of misclassification is minimal), indicated by the red line in Figure 3A. A total of 79 cases exceeds this threshold, including all 26 CEBPA dm cases, 5 CEBPA sm cases, and 48 CEBPA wt cases (see Table 1).

FIG. 3.

(A) Step 1: Scatter plot indicating the distribution of the CEBPA expression level (y-axis) for all 505 cases split across the four groups of CEBPA status (x-axis). The red line indicates the threshold chosen for the first classifier. (B) Step 2: Scatter plot showing the classifier trained on the 79 cases that are above the threshold in (A), as determined by the double-loop cross-validation procedure. Numbers indicate how often a case was misclassified in the 100 repeats (samples always correctly classified do not have a number printed). Color images available online at www.liebertpub.com/gtmb

The step 2 classifier was developed using the 79 cases that exceed the threshold of the first step of the classifier. On these 79 cases, we trained a classifier without FPs and minimal FNs. This results in a 55-gene classifier with 0 FP and 1 FN (sensitivity 96.2%; specificity 100.0%; Fig. 3B). This classifier has a better sensitivity and specificity compared with the classifier trained on all 505 cases (see Table 2). The 55 probe sets used in the classifier are provided in Table 3.

Table 3.

55 Probe Sets Used in the Classifier

AffyID	Gene name	Ensembl gene	AffyID	Gene name	Ensembl gene
223095_at	MARVELD1	ENSG00000155254	204069_at	MEIS1	ENSG00000143995
200765_x_at	CTNNA1	ENSG00000044115	1559477_s_at	MEIS1	ENSG00000143995
200764_s_at	CTNNA1	ENSG00000044115	213844_at	HOXA5	ENSG00000106004
1555630_a_at	RAB34	ENSG00000109113	235521_at	HOXA3	ENSG00000105997
232227_at	LOC100505976	ENSG00000228401	205453_at	HOXB2	ENSG00000173917
1553183_at	UMODL1	ENSG00000177398	235753_at	HOXA7	ENSG00000122592
209191_at	TUBB6	ENSG00000176014	225368_at	HIPK2	ENSG00000064393
1556599_s_at	ARPP21	ENSG00000172995	204039_at	CEBPA	ENSG00000184771
210844_x_at	CTNNA1	ENSG00000044115	206940_s_at	POU4F1	ENSG00000152192
206622_at	TRH	ENSG00000170893	209686_at	S100B	ENSG00000160307
222423_at	NDFIP1	ENSG00000131507	204082_at	PBX3	ENSG00000167081
202252_at	RAB13	ENSG00000143545	209994_s_at	ABCB4	ENSG00000005471
215772_x_at	LOC283398	ENSG00000172340	225841_at	C1orf59	ENSG00000162639
217226_s_at	SFXN3	ENSG00000107819	224428_s_at	CDCA7	ENSG00000144354
205382_s_at	CFD	ENSG00000197766	211682_x_at	UGT2B10	ENSG00000109181
217853_at	TNS3	ENSG00000136205	206847_s_at	HOXA7	ENSG00000122592
241706_at	CPNE8	ENSG00000139117	228904_at	HOXB3	ENSG00000120093
213147_at	HOXA10	ENSG00000153807	216733_s_at	GATM	ENSG00000171766
209905_at	HOXA9	ENSG00000078399	202746_at	ITM2A	ENSG00000078596
210762_s_at	DLC1	ENSG00000164741	226065_at	PRICKLE1	ENSG00000139174
228365_at	CPNE8	ENSG00000139117	206772_at	PTH2R	ENSG00000144407
214651_s_at	HOXA9	ENSG00000078399	211341_at	POU4F1	ENSG00000152192
216667_at	ECRP/RNASE2	ENSG00000136315	206111_at	RNASE2	ENSG00000169385
225116_at	HIPK2	ENSG00000064393	203509_at	SORL1	ENSG00000137642
213150_at	HOXA10	ENSG00000153807	206589_at	GFI1	ENSG00000162676
224822_at	DLC1	ENSG00000164741	209360_s_at	RUNX1	ENSG00000159216
201841_s_at	HSPB1	ENSG00000106211	202718_at	IGFBP2	ENSG00000115457
217800_s_at	NDFIP1	ENSG00000131507

Discussion

Given the favorable prognostic value of CEBPA double mutants, accurate detection of CEBPA mutations at the time of AML diagnosis is important. Current methods include dHPLC followed by Sanger sequencing, direct sequencing, melting curves, etc. However, CEBPA is a notoriously difficult gene to sequence due to its high GC content. In addition, due to short sequence reads, most next generation sequencing (NGS) techniques cannot confirm whether two co-occurring mutations inactivate both alleles or reside on the same allele. Alternatively, a gene expression classifier for CEBPA dm in AML has been built previously (Wouters et al., 2009), but not one that explicitly prevented CEBPA promoter hypermethylated AML cases as false-positive CEBPA double mutants (Figueroa et al., 2009; Taskesen et al., 2011).

Given the favorable prognosis of CEBPA double mutants, FP cases might be at risk of undertreatment. CEBPA hypermethylated cases have a gene expression pattern that is overall very similar to CEBPA double-mutant cases (Figueroa et al., 2009; Taskesen et al., 2011), and therefore, they are more likely than others to end up as FP. Therefore, we have built a two-step classifier that first eliminates CEBPA hypermethylated cases based on their low CEBPA expression level, which is the only difference that distinguishes them from CEBPA double-mutant cases. Next, a classifier with unequal prior probability is employed such that there were zero FPs. This two-step classifier provides an accurate method for predicting CEBPA dm status in AML.

Validation of a classifier on an independent cohort is a common way to estimate its performance. However, such gene expression data measured using the custom Affymetrix GeneChip platform is currently unavailable. As an alternative, we employed the double-loop cross-validation procedure, which estimates the classifier's performance using multiple resamplings of the same data (generalization performance, Wessels et al., 2005). Finally, the two-step classifier presented here will also be validated in an ongoing prospective clinical trial.

In conclusion, the two-step classifier provides an accurate means to detect CEBPA dm status. Moreover, the CEBPA dm classifier has been included on the AML profiler Affymetrix platform that also detects the presence of t(8;21), t(15;17), inv(16)/t(16;16), NPM1 mutations, EVI1 overexpression, and BAALC overexpression, in a single standardized test. Together, this enables a more efficient analysis at diagnosis, with a single sample work up that is suitable for clinical use.

Acknowledgments

This research was performed within the framework of CTMM, the Center for Translational Molecular Medicine, project BioCHIP grant 03O-102.

Author Disclosure Statement

M.H.V.V. performed research, design of experiments, data analysis, data interpretation, and manuscript writing. P.B., L.D.Q., J.P.M.B., L.C.M.B., and H.V. performed research. B.L. and P.J.M.V. conceptuated the project, and were involved in research, design of the experiments, data interpretation, and manuscript writing. EHVB performed research, design of experiments, data interpretation, and manuscript writing. M.H.V.V., P.B., L.D.Q., J.P.M.B., L.C.M.B., H.V., B.L., P.J.M.V., and E.H.V.B. have declared ownership interests in Skyline Diagnostics.