Prognostic evaluation is a critical step in newly diagnosed patients with acute myeloid leukemia (AML) in order to identify those at high risk of relapse. For AML patients, cytogenetic abnormalities as well as gene mutations and/or hyper-expressions detected at diagnosis are the main prognostic factors guiding the initial treatment strategy in a risk-oriented manner 1234. Hypergranular acute promyelocytic leukemia (APL), as well as AMLs with either translocation t(8;21)(q22;q22) [t(8;21)-AMLs] or inversion inv(16)(p13q22)/t(16;16)(p13;q22) [inv(16)-AMLs], are well-defined entities associated with a favorable outcome 13. They can be distinguished from all other AML subtypes based on specific chromosomal alterations and fusion genes: PML/RAR-alpha fusion gene with reciprocal translocation t(15;17)(q24;q21) for APLs, AML1/ETO (also called RUNX1/RUNX1T1) fusion gene for t(8;21)-AMLs, and CBFB/MYH11 fusion gene for AMLs with either inv(16)(q21;q22) or balanced reciprocal translocation t(16;16)(q21;q22). Of note, in up to 15 percent of APLs, no translocation t(15;17)(q24;q21) is detected by conventional cytogenetics at diagnosis, despite PML/RARA fusion gene is detected using molecular assays 56. Similarly, cryptic t(8;21)(q22;q22) and inv(16)(q21;q22), undetected by conventional cytogenetics, have also been reported 789.

With samples containing a high leukemic blast load, microarray-based gene expression profiling (GEP) and class prediction analyses have demonstrated their ability to assign AML samples to one of these three well characterized favorable cytogenetic risk AML subtypes, with high accuracy and low error rates 101112131415161718. One of the largest class prediction analyses in AML achieved 100 percent classification accuracy with respect to APL, t(8;21)-AML, and inv(16)-AML subtypes, indeed 10. However, in the majority of those studies, the minimum percentage of leukemic cells within each sample is most often above 60 percent 101112131415161718, an arbitrary threshold that is significantly different from the one used by cytologists for the diagnosis of AML, i.e., excess of blasts greater or equal to 20 percent.

To our knowledge, the impact of the leukemic blast load and/or of the sample quality on microarray-derived class prediction results has not been specifically studied in AMLs. For centers wishing to integrate microarray-based GEP in a routine prognostic workflow for newly diagnosed AML patients, one critical issue is the ability to perform accurate class prediction analysis with suboptimal samples, i.e., containing as low as 20 percent blasts and even below, and/or not fulfilling all quality control criteria along their process.

In this study, GEP was first used to define a limited set of markers allowing to correctly classify 71 patients with either APL, t(8;21)-AML, inv(16)-AML or cytogenetically normal AML (NK-AML) based on samples containing at least 60 percent of leukemic blasts and characterized by good quality control criteria (training set including optimal samples). The classifiers derived from this first supervised analysis were then evaluated for their ability to assign to the correct class 111 suboptimal samples with low leukemic blast load (from 2 to 59 percent) and/or poor quality control criteria (test set including suboptimal samples), as well as duplicates of three AML cell lines.

The integration of microarray-derived data to the current workflow dealing with the prognostic evaluation of AML patients requires the technology to deliver informative data for the majority of samples, including those with suboptimal characteristics. The current study focused on this kind of samples - either with a low leukemic blast load, poor quality control criteria or both - and the ability to identify favorable-risk karyotype AMLs in such situations, using microarray-based GEP and a class prediction method based on GGA. Apart from the study of de Ridder et al, which addressed the impact of random, fixed and group-specific impurities on the results of differential gene expression analyses, using a limited set of samples and computer simulations, the influence of the leukemic blast load on class prediction accuracy based on GEP data has never been specifically studied so far 25. To evaluate the capacity of microarray-derived predictors to correctly assigned samples with a low leukemic blast load, besides using "real" AML samples with low blast percentage, i.e., less than 60 percent, a "dilution/mixture" approach was considered 2627. This strategy has already been successfully applied to assess the efficiency of various normalization methods applied to microarray datasets. It consisted in diluting labeled cRNAs of originally high blast load AML samples within pools of normal bone marrow labeled cRNAs. This approach was also considered as in some AML subtypes, such as t(8;21)-AMLs, more differentiated cells (not counted as blasts) are frequently of leukemic origin, leading to under-estimate the leukemic load of some "low blast content" samples. Consequently, it became possible to accurately control the percentage of leukemic blasts within the Test Set sample down to 2 percent.

Overall, considering the 111 Test Set samples with low leukemic blast load and/or suboptimal quality control criteria, whether being peripheral blood or bone marrow samples, the predictive capacity of the GGA-derived model with 10-marker classifiers was encouraging, with specificities ranging from 0.98 to 1.00, and sensitivities ranging from 0.91 to 1.00. Moreover, considering AML samples containing 20 to 56 percent blasts, the overall error rate was lower than 1.4 percent (1 misclassified NK-AML sample out of 72). Finally, APL, t(8;21)-AML, and inv(16)-AML samples containing as low as 2 percent blasts could be correctly classified, whether the classifiers mainly relied on a single marker, such as for inv(16)-AMLs and t(8;21)-AMLs, or on the overall set of 10 markers defining the classifier, as for APL samples. However, regarding APL, t(8;21)-AML and inv(16)-AML samples, the confidence level of the class assignment was correlated to the percentage of leukemic blasts within the studied samples. For NK-AML samples, such a finding was not observed, probably because NK-AMLs represent a more heterogeneous group of leukemias as compared to t(8;21)-AMLs, inv(16)-AMLs or APLs in term of oncogenic processes. This heterogeneity likely led to a higher variance of gene expression profiles within the NK-AML class, to the detriment of the interclass variance, hence influencing the class prediction confidence level more than the leukemic blast load. This hypothesis was supported by the fact that, within the NK-AML class, the best predictive capacity of the GGA was achieved for NPM1-mutated samples, in accordance with previous studies, and the worst one for samples with no NPM1 and FLT3 mutations 28293031.

Unexpectedly, the worst results of the class prediction analysis were achieved for inv(16)-AML Test set samples (Sensitivity, 0.91; negative predictive value, 0.96). Among the three misclassified inv(16)-AML samples and for which the presence of CBFB/MYH11 fusion gene was confirmed by FISH analysis, all contained less than 20 percent blasts and two were obtained at the time of relapse - from patients UPN7 and UPN37 -. Regarding these two samples, which respectively contained 5 and 14 percent blasts, they were characterized by a low expression level of most markers defining the inv(16)-AML specific signature; especially MYH11, which expression level was within the estimated background signal range. It is noteworthy that for one of those two patients, UPN37, the bone marrow sample obtained at diagnosis was correctly classified, even when diluted at 50 and 75 percent (blast content, 20 and 10 percent, respectively). For this patient, a clonal evolution of the disease was suspected as CD2, CD13, CD34 and CD117 expression values, assessed by flow cytometry on gated leukemic cells, and MYH11 expression level (microarray data) were significantly lower in the relapse sample as compare to the ones observed in the diagnostic sample. Verhaak et al in their study using Affymetrix GeneChips reported that MYH11 expression level was sufficient to identify inv(16)-AMLs 10. The present results suggest that the use of MYH11 as the sole marker for the inv(16)-AML class could lead to unstable results in a class prediction model, especially in low leukemic blast load samples or in case of a clonal evolution at relapse. Of note, the inv(16)-AML signature developed on the AMLProfiler™ kit http://www.skyline-diagnostics.com, which uses the Affymetrix™ platform, also includes a set of 19 markers 32.

As reported by Kohlmann et al., sub-optimal AML samples with poor quality control criteria, mainly because of low amounts of hybridized cRNAs, were all correctly classified, even though half of these contained less than 60 percent blasts (down to 9 percent for UPN90) 33. Studying the impact of RNA degradation on GEP analyses, similar results have been recently reported by Opitz et al on a different microarray platform 34. As in the present study, they showed that useful information could still be obtained from thermically degraded RNA; at least to a certain extent, and depending on the length of the mRNA molecules. Interestingly, even AML samples for which a third of the required cRNA amount was hybridized on BeadChips could be assigned to the correct class in the present study, suggesting that the biological variance between the AML classes was higher than the one related to technical variability or artifacts. This result could also be related to the fact that, during the sample processing, the TotalPrep RNA Amplification kit that was used, did not required any fragmentation step, which could further alter poor quality RNA or lower the already reduced cRNA amounts. These encouraging results, on a limited set of poor quality samples, could also be related to the classification method used, as the GGA-based class prediction strategy aimed at finding the most influential markers for each AML class (top markers with the highest inter-class variance and the lowest intra-class variance). This GGA-based class prediction method was indeed associated with the lowest error rate and the highest predictive accuracy when compared to other supervised methods, such as support vector machines, random forests, artificial neural networks, k-Nearest Neighbor or nearest shrunken centroids (also known as PAM) (data not shown).

Although, in most cases, these favorable prognostic chromosomal abnormalities and/or their related gene products can be detected by karyotype, FISH or PCR assay, they represent main targets for a microarray-based class prediction analysis to be identified before considering GEP as a useful tool in a routine workflow for prognostic assessment of AML patients. The fact that the AMLProfiler, a microarray-derived kit based on the Affymetrix™ platform, has been recently commercialized with the aim of achieving a molecular diagnosis for APLs, t(8;21)-AMLs and inv(16)-AMLs, respectively using 27, 31 and 19 markers, emphasizes this assumption 32.

Finally, regarding the classifiers associated with APLs, t(8;21)-AMLs or inv(16)-AMLs, 18 of the 28 markers identified in this study (64 percent) had already been reported in previous studies using various microarray technologies (additional file 1, Table S1). These findings (1) confirm that the Illumina bead-based technology is as reliable, robust and sensitive as other microarray technologies developed by commercial manufacturers or academic facilities, and (2) suggest that the GGA-derived class prediction approach is a highly efficient one as it required a limited set of ten markers per class to achieve accurate class assignments. Furthermore, this study has identified new AML markers that will need further studies to delineate their role in the leukemogenic events involved in APLs (CERCAM, COL23A1, LOC643201, LOXL4, SLC39A11), t(8;21)-AMLs (LOC440030, TNFRSF21), and inv(16)-AMLs (EFHC2, GPR12, MEGF10).

In more than 96 percent of the suboptimal cases, using a microarray-derived GEP approach and a GGA-based class prediction method, favorable cytogenetic risk AML samples with low leukemic blast load and/or poor quality control criteria could be correctly assign to the appropriate class with a limited set of markers, allowing to consider GEP as a useful tool in a routine workflow for prognostic assessment of AML patients.

APL: acute promyelocytic leukemia; AML: acute myeloid leukemia; GEP: gene expression profiling; GGA: grouping genetic algorithms; NK-AML: cytogenetically normal acute myeloid leukemia; RIN: RNA integrity number.

The authors declare that they have no competing interests.

PhG was the principal investigator and takes primary responsibility for the paper. DRB, OB and PhG contribute to the conception of design. DRB, PCL, AC, LB, FG, IL, MHB, AB, AST, MZ, YD, NI, OB, PhG contribute to the collection and assembly of data. DRB, OB, YD, MZ, NI, PhG contribute to the data analysis and interpretation. DRB, PhG contribute to manuscript writing. DRB, OB, PCL, AC, LB, FG, IL, MHB, AB, AST, MZ, YD, NI, PhG contribute to the drafting of the article, revise it and approved the final version. PhG coordinated the research.

PhG is the Head of Plateforme SNP, Transcriptome & Epigénomique, i.e., the genomic platform of Angers University Hospital, and is also an Associate Editor at BMC Genomics and a member of the American Society of Human Genetics.