Microarray-Based Classification of a Consecutive Series of 121 Childhood

Microarray-based classiﬁcation of a consecutive series of 121 childhood acute leukemias: prediction of leukemic and genetic subtype as well as of minimal residual disease status

A Andersson1, C Ritz2, D Lindgren1, P Ede´n2, C Lassen1, J Heldrup3, T Olofsson4,JRa˚de5, M Fontes5, A Porwit-MacDonald6, M Behrendtz7,MHo¨glund1, B Johansson1 and T Fioretos1

1Department of Clinical Genetics, Lund University Hospital, Lund, Sweden; 2Department of Complex System Division, Theoretical Physics, Lund University, Lund, Sweden; 3Department of Pediatrics, Lund University Hospital, Lund, Sweden; 4Department of Hematology, Lund University Hospital, Lund, Sweden; 5Center for Mathematical Sciences, Lund University, Lund, Sweden; 6Department of Pathology, Karolinska Hospital and Institute, Stockholm, Sweden and 7Department of Pediatrics, Linko¨ping University Hospital, Linko¨ping, Sweden

Gene expression analyses were performed on 121 consecutive notype, white blood cell (WBC) count, central nervous system childhood leukemias (87 B-lineage acute lymphoblastic leuke- (CNS) involvement, response to therapy and genetic findings.6,7 mias (ALLs), 11 T-cell ALLs and 23 acute myeloid leukemias Despite the quite dramatic progress in treatment, risk- (AMLs)), investigated during an 8-year period at a single center. stratification and biological understanding of childhood leuke- The supervised learning algorithm k-nearest neighbor was 8 utilized to build gene expression predictors that could classify mias over recent decades, further improvements are still the ALLs/AMLs according to clinically important subtypes with needed. Refined risk assessment will hopefully result in the high accuracy. Validation experiments in an independent data identification of individual patients who may benefit from either set verified the high prediction accuracies of our classifiers. more or less intensive treatment regimens. Moreover, increased B-lineage ALLs with uncharacteristic cytogenetic aberrations understanding of the underlying biology of the various ALL and or with a normal karyotype displayed heterogeneous gene expression profiles, resulting in low prediction accuracies. AML subtypes will hopefully result in the development of novel Minimal residual disease status (MRD) in T-cell ALLs with a therapies. In this context, cDNA and oligonucleotide micro- high (40.1%) MRD at day 29 could be classified with 100% arrays have emerged as a promising tool for identifying both accuracy already at the time of diagnosis. In pediatric clinically and biologically important features. So far, only a few leukemias with uncharacteristic cytogenetic aberrations or with large-scale (4100 cases) gene expression studies of childhood a normal karyotype, unsupervised analysis identified two novel leukemias have been reported.9–13 In a previous study, we subgroups: one consisting mainly of cases remaining in complete remission (CR) and one containing a few patients in performed global gene expression profiling of a consecutive CR and all but one of the patients who relapsed. This study of a series of 121 pediatric acute leukemias of different lineages consecutive series of childhood leukemias confirms and (B-lineage ALLs, T-cell ALLs and AMLs) and six normal healthy extends further previous reports demonstrating that global bone marrows, and compared genetic subtype-specific gene gene expression profiling provides a valuable tool for genetic signatures with an extensive set of fractionated normal and clinical classification of childhood leukemias. 9 Leukemia (2007) 21, 1198–1203. doi:10.1038/sj.leu.2404688; hematopoietic cell subpopulations. Herein, we used supervised published online 5 April 2007 learning algorithms to build predictors capable of classifying the Keywords: pediatric leukemia; gene expression profiling; leukemias according to lineages and genetic changes with high supervised classification; ALL; AML accuracy. For the T-cell ALLs, we could classify samples with high tumor load at day 29, already at diagnosis. Among cases with uncharacteristic cytogenetic changes or with a normal karyotype, unsupervised hierarchical clustering analysis (HCA) identified two novel subgroups, one containing mainly patients Introduction in complete remission (CR) and one mainly patients who eventually relapsed. Validation experiments of prediction of Pediatric acute leukemias constitute a heterogeneous disease genetic subtypes in an external data set verified the high entity, comprised of different genetic, morphologic and im- accuracy of our classifier. munophenotypic subgroups with variable clinical features.1,2 The most common subtype is acute lymphoblastic leukemia (ALL) with an incidence of four cases per 100 000 per year; Materials and methods acute myeloid leukemia (AML) is less frequent with only 0.7 3,4 cases per 100 000 per year. The prognosis has improved Patient material significantly over the last few years, with the 7-year event-free Bone marrow (BM; n ¼ 108) or peripheral blood (PB; n ¼ 13) survival for ALL and AML now approaching 80 and 50%, samples from 121 children with ALL (87 B-lineage and 11 T-cell) respectively.3,5 This improved outcome has mainly been or AML (n ¼ 23) were obtained at the time of diagnosis. In achieved through stratifying individual patients to different risk addition, six normal bone marrows (NBMs) were obtained from and treatment groups based on, for example, age, immunophe- healthy donors. Details of the studied cohort have been described elsewhere.9 The leukemias were diagnosed and Correspondence: Dr A Andersson, Department of Clinical Genetics, treated at either Lund University Hospital (n ¼ 89) or Linko¨ping University Hospital, SE-221 85 Lund, Sweden. E-mail: [email protected] University Hospital (n ¼ 32), representing approximately 70% of Received 1 August 2006; revised 6 March 2007; accepted 8 March all childhood ALLs and AMLs diagnosed at these two hospitals 2007; published online 5 April 2007 during the study period (1997–2004). Two different treatment Array-based classification of pediatric AML and ALL A Andersson et al 1199 protocols – NOPHO (Nordic Society of Paediatric Haematology of the analysis is given in Supplementary Information in the and Oncology) ALL-923 and NOPHO ALL-2000 – were used expanded Materials and methods section. during the study period. Protocols were changed on 1 January HCA was performed in TMEV21 and principal component 2001 (time periods (1992–2001 and from 2001 onwards). Risk analysis (PCA) was applied using software developed at the classification during both treatment periods was based on age, Department of Mathematics, Lund, Sweden. Significance WBC count, immunophenotypic and karyotypic features, analysis of microarrays (SAM)21 was used to find genes dividing the ALLs into standard-risk/standard-intensive (SR/SI; associated with the putative novel groups in cases with no age 2–10 years, WBC o10 Â 109/l and no high-risk/intensive specific genetic change. (HR/I) features), intermediate-risk/intermediate-intensive (IR/II; 1–2 or 410 years, WBC 10–50 Â 109/l and no HR/I features), 9 Supervised classification HR/I (WBC 450 Â 10 /l, CNS or testicular leukemia, t(9;22), The k-nearest neighbors (k-NNs) algorithm22 was used for t(4;11), T-cell immunophenotype), very high-risk/extra-intensive supervised classification, where the class of a test sample is (VHR/EI; lymphomatous leukemia, together with other HR/I decided by the majority class among its k-NNs. A cross- criteria). The AMLs were treated according to the NOPHO- 5 validation procedure was used to select the number of neighbors AML93 protocol. The treatment protocols and the risk and genes used for classification. From each ranked list, the stratification for NOPHO ALL-92 have been reported pre- 3 classifier was evaluated in a leave-one-out cross-testing proce- viously. The study was reviewed and approved by the Research dure (for details, see the expanded method section in Ethics Committees of Lund and Linko¨ping Universities, Sweden. Supplementary Information). All cases were analyzed cytogenetically. In addition, the ALLs were screened for MLL rearrangements using Southern blot or fluorescence in situ hybridization analyses (FISH) and for the Validation experiments in an external data set presence of the fusion genes BCR/ABL1 (t(9;22), TCF3/PBX1 To validate the k-NN method to build predictors, the largest (t(1;19)) and ETV6/RUNX1 (t(12;21)) by the use of reverse childhood ALL data set reported to date was used.11 In this transcription-PCR. In addition, FISH investigations were per- analysis, our B-lineage ALL data set was used as a training set formed on cases with either normal karyotypes (NKs) or without and the B-lineage ALLs from the Ross data set were used as a analyzable metaphases, using probe cocktails for the chromo- validation data set. To validate the T-cell ALL classifier for 14 somes commonly gained in high hyperdiploid ALLs. All prediction of MRD, the only data set published reporting both analyses were performed at the Department of Clinical MRD status and gene expression data on pediatric T-cell ALLs Genetics, Lund, Sweden. The genetic features are summarized was used.23 The training and validation data sets were generated in Supplementary Table 1. using different array platforms (cDNA and Affymetrix, respectively); therefore, data were matched using the Entrez gene Monitoring of minimal residual disease status numbers, mean-centered individually with respect to array 9 The ALLs from Lund (n ¼ 61) were monitored for minimal platform, and analyzed as previously described. In addition, 10,24,25 residual disease (MRD) status by quantitative PCR as described gene lists from three additional data sets were used. in van Dongen et al.15 The ALLs from Linko¨ping (n ¼ 16) were monitored using flow cytometry as described previously.16,17 MRD was available in 77 ALLs; for the remaining 10 ALLs, no Results MRD data were available owing to the lack of a patient-specific primer or the non-availability of MRD measurement. As In a previous investigation, using unsupervised methods, we showed that pediatric leukemias, to a large extent, segregate determined by dilution experiments, the sensitivity of the PCR 9 and the flow cytometric MRD detection was 1 cell/105 BM cells according to lineage (91%) and primary genetic change (85%). and 1 cell/104 BM cells, respectively. It has been shown that In the present study, we focused on the use of a supervised MRD measured by flow cytometry or PCR yield concordant learning algorithm, that is, k-NN, to construct gene expression results.16,18 The MRD status was translated to a scale from 1to 6, predictors that could stratify the patients into clinically where an MRD of 1 corresponds to o0.001%, 2 to 0.001– important subgroups (Supplementary Information). 0.01%, 3 to 0.01–0.1%, 4 to 0.1–1%, 5 to 1–10% and 6 to 10– First, we constructed a classifier that made it possible to 100% leukemic cells. For classification, cases were divided into determine to which leukemia subtype (B-lineage ALL, T-cell ALL two groups: ‘Low’ (MRD of 1–2) or ‘High’ (MRD of 3–6). and AML) a sample belonged. In this predictor, we also included a set of NBMs. The overall classification rate was 97%, with only four cases being misclassified (Figure 1a, Table 1, Gene expression profiling Supplementary Table 2), verifying that microarray analysis in RNA extraction, labeling, hybridization, scanning and post combination with supervised learning algorithms can predict the hybridization washing were performed as described pre- 9,19 type of leukemia for an individual sample with very high viously Samples were hybridized to 27K microarrays accuracy. (Swegene DNA Microarray Resource Center, Lund University, Second, we constructed and evaluated the performance of a Sweden) containing 25 648 clones representing 13 616 Unigene classifier capable of predicting the presence of certain genetic clusters and 11 645 Entrez gene entries, according to Unigene changes within each lineage. Among the B-lineage ALLs, a build 186. The primary data are deposited in Gene Expression sufficient number of samples for classification of molecular Omnibus (GSE7186). genetic change was obtained for t(1;19) (six cases), t(12;21) (20 cases), high hyperdiploidy (450 chromosomes, 29 cases), NK Microarray data analyses (seven cases) and a group of 20 cases with uncharacteristic The data were analyzed using the GenePix4.0 software (Axon cytogenetic changes, designated ‘others’. Because the group Instruments, Foster city, CA, USA), and the obtained data matrix ‘others’ was expected to be heterogeneous, no gene lists for was uploaded to the BioArray Software Environment (BASE)20 ‘others’ versus the rest were used in any of the classifiers. The and analyzed as previously described.9,19 A detailed description overall accuracy, which was 75%, was affected by the fact that

Leukemia Array-based classification of pediatric AML and ALL A Andersson et al 1200 a Prediction of relapse and MRD status 2 (16%) In a next step, we aimed at identifying a gene expression pattern that at the time of diagnosis could predict which patients would develop a relapse. Among the B-lineage ALLs, 70 patients in CR and seven patients who later relapsed (R) were available and 11 reporters (false discovery rate (FDR) ¼ 0) associated with relapse 1 (34%) AML 3 (7%) were identified using SAM (Supplementary Table 7). In B-lineage ALL agreement with a previous report,11 we could not, however, T-cell ALL NBM construct a predictor that could predict relapse in ALLs across Misclassified the genetic subtypes already at diagnosis. In the AMLs, 10 cases Leukemic subtype cases were in CR and seven developed a relapse during the study period; no signature across the genetic subtypes was identified b 2 (22%) for relapse. The leukemias were therefore subdivided according to genetic subtype and reclassified. In this analysis, only ALLs diagnosed before 2001 (n ¼ 23) were considered to allow for detection of relapses up to 5 years from diagnosis. For the MLL- positive AMLs, all cases irrespective of date of diagnosis were 1 (31%) t(12;21)[ETV6/RUNX1] included. Classifications were performed for cases with high 3 (5%) High hyperdiploidy hyperdiploidy (CR ¼ 13; R ¼ 3), 11q23/MLL rearrangements t(1;19)[TCF3/PBX1] (AMLs) (CR ¼ 4; R ¼ 3), and T-cell ALL (CR ¼ 4; R ¼ 3). No Misclassified case signatures at the time of diagnosis were identified that could Genetic subtype predict patients that later relapsed. In ALL, it has previously been reported that a high tumor load c 2 (8%) (40.1%) at day 29 of treatment increases the risk of having a relapse.17 We hypothesized that such patients would have a unique gene expression profile already at the time of diagnosis, but we could not identify a single set of genes that identified samples with a high MRD at day 29 across the various subtypes.

3 (6%) 1 (63%) Instead, we investigated if the MRD status could be predicted in T-cell ALLs and among the genetic subtypes in B-lineage ALL. Among the B-lineage ALLs, no gene expression profile was High > 0.1% Low < 0.01% identified that could predict a high MRD at day 29. In the T-cell T-ALL MRD status ALLs, however, we could identify samples with a high or a low MRD at day 29, already at the time of diagnosis, with an Figure 1 PCA showing classifiers generated using k-NN. (a) PCA of B- overall accuracy of 100% (Figure 1c, Table 1, Supplementary lineage ALL (blue), AML (red), T-cell ALL (pink), and NBM (gray), using Table 8). Cases with a high MRD showed enrichment for genes the combined gene lists used for classification. The samples misclassified are colored in white; two cases (one T-ALL and one B-lineage ALL) involved in the JAK/STAT cascade (e.g., JAK3, STAT2 and were classified as AMLs, one AML was classified as an NBM and one B- STAT5A) or with catalytic activity (e.g., DNMT3B, CYP2S1, lineage ALL was classified as an AML. (b) PCA of the different genetic CYP2C8, PTGIS and CYP4F12) among the upregulated genes. subtypes in B-lineage ALLs (TCF3/PBX1,pink;ETV6/RUNX1,blue;high For the downregulated genes, an enrichment of genes with hyperdiploidy, green) using the combined gene lists. As above, ubiquitin-specific protease activity was found. No specific misclassified samples are colored in white. (c) PCA of the genes used genetic abnormalities were found to be characteristic for cases for classification of T-cell ALLs with high or low MRD status. having a high MRD; in fact, this group was cytogenetically quite heterogeneous. samples in the subgroup ‘others’ and NK were misclassified in 69 and 57% of the cases, respectively (Supplementary Table 3). When the group ‘others’ was removed and the remaining Unsupervised analysis of leukemias without a specific samples reclassified, the accuracy reached 94%, with only four chromosomal change cases misclassified (Supplementary Table 4). When only cases In ALLs and AMLs without specific aberrations, including those with TCF3/PBX1, ETV6/RUNX1 and high hyperdiploidy were with an NK, treatment decisions rely solely on clinical and considered, the classification accuracy reached 98%, as immunophenotypic variables. In an attempt to find subgroups of demonstrated previously by us9 (Figure 1b, Table 1, Supple- such ALLs with a poor or a favorable prognosis, HCA was used to mentary Table 5). Following the approach taken by Ross et al.,11 group samples based on their similarity, resulting in segregation we also built a classifier in which cases designated ‘others’ of the patients into two major clusters (groups I (n ¼ 11) and II were combined with those with an NK, obtaining an accuracy (n ¼ 9)) (Figure 2a; for patient details, see Supplementary Table of 83%. 9). Interestingly, in group II, the diagnostic samples from two of Only two classes could be considered among the AMLs – those the three cases that eventually relapsed co-segregated, together with MLL rearrangements (10 cases) and those without (13 cases). with one case in relapse. We hypothesized that HCA identified The overall classification accuracy reached 96%, with only one two groups of patients that might differ with respect to prognosis. case being misclassified (Table 1, Supplementary Table 6). SAM21 identified 72 genes (FDR ¼ 0) associated with the groups Attempts to build classifiers were also made for several (Figure 2b, Supplementary Table 10). GO and pathway analyses clinical variables, such as sex, age and WBC, but apart from revealed a highly significant enrichment of cell cycle-related prediction of sex, which resulted in 100% classification genes among the low expressed genes in group II, indicating a accuracy, all other parameters resulted in low prediction biological difference with regard to cell proliferation among the accuracies. two groups (Supplementary Table 11).

Leukemia Array-based classiﬁcation of pediatric AML and ALL A Andersson et al 1201 Table 1 Classiﬁcation accuracies for the supervised analyses

Class No. of samples No. of misclassiﬁed Apparent accuracy Sensitivity Speciﬁcity samples (%) (%) (%)

Prediction of leukemic subtypes NBM 6 0 100 100 100 B-lineage ALL 87 1 99 99 98 T-cell ALL 11 1 91 91 99 AML 23 2 91 91 98

Prediction of genetic subtypes among the B-lineage ALLs TCF3/PBX1 6 0 100 100 100 High hyperdiploidy 29 0 100 100 100 ETV6/RUNX1 20 1 95 95 97

Prediction of genetic subtypes among the AMLs MLL 10 1 90 91 100 Other 13 0 100 100 92.4 Prediction of MRD status in T-cell ALLs T-cell ALL high MRD 5 0 100 100 100 T-cell ALL low MRD 4 0 100 100 100 Abbreviations: ALL, acute lymphoblastic leukemia; AML, acute myeloid leukemia; MRD, minimal residual disease; NBM, normal bone marrow. Apparent accuracy was determined by leave-one-out cross-validation. Sensitivity ¼ the number of positive samples correctly predicted/the number of true positives. Speciﬁcity ¼ the number of negative samples correctly predicted/the number of true negatives.

Validation experiments in external data sets uncharacteristic cytogenetic changes (‘others’), the overall To validate the predictive strength of the classifier, our data set prediction accuracy, however, dropped markedly to 75%. The and the data set from Ross et al.11 were used as training and reasons for the difficulties in the latter prediction are most likely validation sets, respectively. Only genetic subtypes present in manifold, including that ALLs with an NK or with uncharacter- both data sets were considered (for details, see Supplementary istic genetic changes are heterogeneous, making it difficult to Table 12). The overall classification accuracy reached 89%, detect a significant gene expression pattern. In an attempt to with nine cases being misclassified, of which eight belonged to identify novel subgroups in B-lineage ALLs, we used unsuper- the group ‘others’. If samples designated ‘others’ were removed, vised HCA to investigate the 20 cases with either an NK or the overall classification accuracy increased to 96% with only lacking specific genetics changes. Interestingly, two patient two TCF3/PBX1-positive cases being misclassified, both as being subgroups were identified and GO analyses revealed hyperdiploid. Using a novel method for normalization of cross- a significant enrichment of cell cycle-related genes being platform differences, we have previously obtained similar results upregulated in group I. This group mainly contained cases using this data set as a validation set and the Ross data as a remaining in CR (10 out of 11), whereas group II consisted of six training set.26 diagnostic cases that remained in CR, one relapse sample and To validate the MRD predictor in T-cell ALL, the only data set two samples from diagnostic cases that later developed a available today with MRD data on pediatric T-cell ALLs was relapse. No distinct cytogenetic pattern was found to be used.23 This data set consisted of three cases with low and nine characteristic for the two groups. Notably, however, three cases cases with high MRD. The predictor classified correctly 2/3 in group I harbored a dic(9;20), a rearrangement suggested to be cases with low MRD (67%), but only 4/9 cases with high MRD correlated with a good prognosis.27 Compared to group I, (44%). Furthermore, a previously published MRD predictor in B- samples in group II showed a low expression of cell cycle- lineage ALLs was applied on our B-lineage ALLs, but could not related genes, possibly indicating impaired cell proliferation predict MRD status in our data set.24 Finally, genes found to be among these samples, which could be coupled to a decreased associated with MRD in ALLs in a recently published paper were sensitivity towards drugs that are effective on proliferating also investigated in our data set, but as above, they could not cells. However, this analysis was carried out on a small group separate out our leukemias based on MRD.25 of patients and further studies are clearly needed to validate these findings. We also applied the 72-gene signature on the samples designated ‘others’ on the data set from Ross et al., Discussion but were not able to separate the novel subgroup described by Ross et al.11 Using gene expression profiling and supervised learning Among the AMLs, classification was carried out for patients algorithms, we have classified a consecutive series of 121 with or without MLL rearrangements. The predictor showed an pediatric acute leukemias based on lineage and genetic subtype overall classification accuracy of 96%, with only one case, a with high classification accuracies (97 and 98%, respectively). t(6;11), being misclassified, in line with a previous study.12 The high accuracy of the predictor was independently validated A major problem in clinical practice is to identify patients with on the largest data set on B-lineage ALL available to date.11 an increased risk of relapse. To detect such patients already at These results, together with other recent investigations also the time of diagnosis would most likely be beneficial, making it including 4100 childhood leukemias,9–13 clearly demonstrate possible to use a different treatment strategy already up-front. To that gene expression profiling can be used to classify diagnostic address this problem, we first aimed at identifying a gene samples of childhood leukemia into well-known genetic expression signature that could predict cases with high risk of subgroups. When including cases with an NK or cases with relapse across all leukemic subtypes. This approach however

Leukemia Array-based classification of pediatric AML and ALL A Andersson et al 1202 failed, which is in agreement with a previous report.10 Instead, a Group I Group II the data set was divided based on genetic subtype and classified according to relapse status. We could not predict relapses in high hyperdiploid ALLs, in MLL-positive AMLs or in T-cell ALLs. The reasons for this are most likely manifold, including small patient groups, few relapses and/or heterogeneous expression patterns associated with relapse. Yeoh et al.10 obtained significant

Relapse CR Relapse CR Relapsed sample CR CR CR CR CR CR CR CR CR CR CR CR CR CR Relapse Outcome classifiers for relapse in ALLs with high hyperdiploidy and Group T-cell ALLs. The predictors contained a few reporters (20 and 7, respectively) and when matched against our microarray plat- b form, very few remained that could be used for classification. −2.0 0.0 2.0 Another parameter that recently has been shown to be associated with risk of relapse is high MRD status at day 29 after treatment.17 We therefore tried to build a classifier that MYH10 MYBL2 could classify such patients already at the time of diagnosis; HMGB3 CDC2 however, no single set of genes was identified that could predict CCNB2 CCNB2 a high MRD at day 29 across the genetic subtypes. For the TOP2A TOP2A genetic subtypes in B-lineage ALLs, prediction of MRD was not KIF4A GGH successful, resulting in low prediction accuracies. Only three C20orf129 STK6 previous expression studies have tried to predict MRD status STK6 23–25 24 TTK among childhood ALLs. Cario et al. reported a classifier MULTIPLE CLUSTER PTTG1 of 52 genes in B-lineage ALL showing an accuracy of 92%. PTTG1 DKFZp762E1312 Thirty-two of these genes were represented on our microarray, MK167 KIF23 but they could not predict MRD status in our data set. We also CENPA MGC57827 applied the gene signature associated with MRD in ALLs UBE2C 25 Full-length cDNA clone identified by Flotho et al. on our ALLs, but again their genes PLK1 SLC35E3 Transcribed locus could not separate our leukemias based on MRD status. Among SLC35E3 TYMS the T-cell ALLs, we could, despite the small number of patients, CDCA7 RAD51 build a classifier that could predict with an accuracy of 100% CDC45L BIRC5 patients with a high MRD at day 29. To the best of our TKI RACGAP1 knowledge, no other studies are available that report classifiers GNB3 CIT capable of predicting MRD status in T-ALL. Although our TMPO RFC3 classifier was specifically designed to avoid over-fitting, using a MCM7 CBX5 combined cross-validation and cross-testing procedure, it needs NUCKS CDC7 to be validated before future clinical implementation. Because STMNI FIGNLI of the low incidence of T-cell ALL, with only a few newly Transcribed locus CKSIB diagnosed cases available in our laboratory, we searched for AURKB PBK external data that could serve as independent validation sets. CCNA2 CDK2 We were able to identify only one data set reporting MRD data Transcribed locus, 23 ESPL1 on T-cell ALL. Applying our predictor to this relatively small BUB1B ATAD2 data set resulted in poor classification accuracy and hence ATAD2 MULTIPLE CLUSTER future validation studies are clearly needed. Looking at the gene RAD54L EXO1 list discriminating high from low MRD, we identified JAK3,a KIF2C 28 DCCI gene known to have an important role in T-cell development, CDC25A ZWINT as highly expressed in cases with a high MRD. Several genes TPX2 CDCA8 known to become activated by JAK3 were also upregulated, for HSPD1 29 NEK2 example, STAT5A and STAT2. In addition, genes belonging to NEK2 CENPF the cytochrome P450 superfamily, involved in drug metabolism, KIF23 KIFI4 were highly expressed. It is tempting to speculate that these DEPDCI D2IS2056E genes are responsible for the slow treatment response in T-cell CDCA2 Transcribed locus ALL, although this remains to be elucidated. In contrast to MRD BUBI KNTC2 status, relapse could not be predicted among the T-cell ALLs. PLK4 CTNNAL1 The reason for this is presently unclear, but could be a reflection FLJ40629 KIF11 of genetic heterogeneity, the small sample size or different BM039 PRC1 biological mechanisms underlying slow treatment response CDC6 PCNA MTB and relapse. MLF11P MLF11P To validate the k-NN algorithm used for classifying genetic subtype among the B-lineage ALLs in the present study, the ALL Figure 2 HCA identifies a novel subgroup among leukemias with data set from Ross et al.11 was retrieved and used as a validation uncharacteristic genetic changes. (a) Unsupervised HCA using 4451 data set, resulting in maintained prediction strength. This reporters reveals segregation of samples into two groups using Pearson correlation and average linkage. (b) Supervised HCA of the genes strongly indicates that although different platforms were used identified using SAM analysis. GO analysis of these genes showed a to generate the microarray data, genes having a high predictive highly significant enrichment of cell cycle genes, which are highly strength in our data set were also differentially expressed in the expressed in group I. external data set.11 Thus, these genes seem to be excellent predictors of genetic subtypes of B-lineage ALLs.

Leukemia Array-based classification of pediatric AML and ALL A Andersson et al 1203 In conclusion, we have analyzed the gene expression 11 Ross ME, Zhou X, Song G, Shurtleff SA, Girtman K, Williams WK signatures of a consecutive series of 121 childhood leukemias. et al. Classification of pediatric acute lymphoblastic leukemia by Using k-NN, predictors were constructed that could predict with gene expression profiling. Blood 2003; 102: 2951–2959. high accuracy leukemic and genetic subtypes. This suggests that 12 Ross ME, Mahfouz R, Onciu M, Liu HC, Zhou X, Song G et al. Gene expression profiling of pediatric acute myelogenous microarray analysis is a promising tool for classification of leukemia. Blood 2004; 104: 3679–3687. individual patients into current risk groups based on lineage and 13 van Delft FW, Bellotti T, Luo Z, Jones LK, Patel N, Yiannikouris O genetic abnormality present at diagnosis. It is however, et al. Prospective gene expression analysis accurately subtypes important to bear in mind that leukemias with an NK or with acute leukaemia in children and establishes a commonality uncharacterized genetic changes are more difficult to classify, between hyperdiploidy and t(12;21) in acute lymphoblastic necessitating the continuous need for detailed cytogenetic and leukaemia. Br J Haematol 2005; 130: 26–35. 14 Paulsson K, Panagopoulos I, Knuutila S, Jee KJ, Garwicz S, Fioretos T molecular analyses. We also found two putatively novel et al. Formation of trisomies and their parental origin in hyperdiploid subgroups in B-lineage ALLs with variable genetic changes, childhood acute lymphoblastic leukemia. Blood 2003; 102: characterized by a differential gene expression signature of cell 3010–3015. cycle-related genes, and identified genes predicting a high MRD 15 van Dongen JJ, Seriu T, Panzer-Grumayer ER, Biondi A, Pongers- at day 29 in T-cell ALLs at the time of diagnosis. Hopefully, Willemse MJ, Corral L et al. Prognostic value of minimal residual additional genome-wide studies together with larger validation disease in acute lymphoblastic leukaemia in childhood. Lancet 1998; 352: 1731–1738. studies should bring forth novel and clinically valuable tools 16 Malec M, van der Velden VH, Bjo¨rklund E, Wijkhuijs JM, towards the goal to further improve current treatment stratifica- So¨derhall S, Mazur J et al. Analysis of minimal residual disease tion of childhood ALL. in childhood acute lymphoblastic leukemia: comparison between RQ-PCR analysis of Ig/TcR gene rearrangements and multicolor flow cytometric immunophenotyping. Leukemia 2004; 18: Acknowledgements 1630–1636. 17 Bjo¨rklund E, Mazur J, So¨derha¨ll S, Porwit-MacDonald A. Flow Grant support was provided by The Swedish Cancer Society, the cytometric follow-up of minimal residual disease in bone marrow gives prognostic information in children with acute lymphoblastic Swedish Children’s Cancer Foundation, the Medical Faculty of leukemia. Leukemia 2003; 17: 138–148. Lund University and the IngaBritt and Arne Lundberg foundation. 18 Neale GA, Coustan-Smith E, Stow P, Pan Q, Chen X, Pui CH et al. PE was supported by the Swedish Foundation for Strategic Comparative analysis of flow cytometry and polymerase Research through the Lund Center for Stem Cell Biology and Cell chain reaction for the detection of minimal residual disease in Therapy. childhood acute lymphoblastic leukemia. Leukemia 2004; 18: 934–938. 19 Andersson A, Edeń P, Lindgren D, Nilsson J, Lassen C, Heldrup J References et al. Gene expression profiling of leukemic cell lines reveals conserved molecular signatures among subtypes with specific 1 Hall GW. Childhood myeloid leukaemias. Best Pract Res Clin genetic aberrations. Leukemia 2005; 19: 1042–1050. Haematol 2001; 14: 573–591. 20 Saal LH, Troein C, Vallon-Christersson J, Gruvberger S, Borg A˚ , 2 Pui CH, Relling MV, Downing JR. Acute lymphoblastic leukemia. Peterson C. BioArray Software Environment (BASE): a platform for N Engl J Med 2004; 350: 1535–1548. comprehensive management and analysis of microarray data. 3 Gustafsson G, Schmiegelow K, Forestier E, Clausen N, Glomstein A, Genome Biol 2002; 3: SOFTWARE0003. Jonmundsson G et al. Improving outcome through two decades in 21 Saeed AI, Sharov V, White J, Li J, Liang W, Bhagabati N et al. TM4: childhood ALL in the Nordic countries: the impact of high-dose a free, open-source system for microarray data management and methotrexate in the reduction of CNS irradiation. Nordic Society of analysis. Biotechniques 2003; 34: 374–378. Pediatric Haematology and Oncology (NOPHO). Leukemia 2000; 22 Dudoit S, Fridlyand J. Introduction to classification in microarray 14: 2267–2275. experiments. In: Berrar DP, Dubitzky W, Granzow M (eds). A 4 Forestier E, Heim S, Blennow E, Borgstrom G, Holmgren G, Practical Approach to Microarray Data Analysis. Kluwer Academic Heinonen K et al. Cytogenetic abnormalities in childhood acute Publishers: London, 2002, pp 201–215. myeloid leukaemia: a Nordic series comprising all children 23 Willenbrock H, Juncker AS, Schmiegelow K, Knudsen S, Ryder LP. enrolled in the NOPHO-93-AML trial between 1993 and 2001. Prediction of immunophenotype, treatment response, and relapse Br J Haematol 2003; 121: 566–577. in childhood acute lymphoblastic leukemia using DNA micro- 5 Lie SO, Abrahamsson J, Clausen N, Forestier E, Hasle H, Hovi L arrays. Leukemia 2004; 18: 1270–1277. et al. Treatment stratification based on initial in vivo response in 24 Cario G, Stanulla M, Fine BM, Teuffel O, Neuhoff NV, Schrauder A acute myeloid leukaemia in children without Down’s syndrome: et al. Distinct gene expression profiles determine molecular results of NOPHO-AML trials. Br J Haematol 2003; 122: 217–225. treatment response in childhood acute lymphoblastic leukemia. 6 Grimwade D. The clinical significance of cytogenetic abnormal- Blood 2005; 105: 821–826. ities in acute myeloid leukaemia. Best Pract Res Clin Haematol 25 Flotho C, Coustan-Smith E, Pei D, Iwamoto S, Song G, Cheng C 2001; 14: 497–529. et al. Genes contributing to minimal residual disease in childhood 7 Johansson B, Mertens F, Mitelman F. Clinical and biological acute lymphoblastic leukemia: prognostic significance of importance of cytogenetic abnormalities in childhood and adult CASP8AP2. Blood 2006; 108: 1050–1057. acute lymphoblastic leukemia. Ann Med 2004; 36: 492–503. 26 Nilsson B, Andersson A, Johansson M, Fioretos T. Cross-platform 8 Ravindranath Y. Recent advances in pediatric acute lymphoblastic classification in microarray-based leukemia diagnostics. Haema- and myeloid leukemia. Curr Opin Oncol 2003; 15: 23–35. tologica 2006; 91: 821–824. 9 Andersson A, Olofsson T, Lindgren D, Nilsson B, Ritz C, EdeńP 27 Clark R, Byatt SA, Bennett CF, Brama M, Martineau M, et al. Molecular signatures in childhood acute leukemia and their Moorman AV et al. Monosomy 20 as a pointer to dicentric correlations to expression patterns in normal hematopoietic (9;20) in acute lymphoblastic leukemia. Leukemia 2000; 14: subpopulations. Proc Natl Acad Sci USA 2005; 102: 241–246. 19069–19074. 28 Eynon EE, Livak F, Kuida K, Schatz DG, Flavell RA. Distinct effects 10 Yeoh EJ, Ross ME, Shurtleff SA, Williams WK, Patel D, Mahfouz R of Jak3 signaling on alphabeta and gammadelta thymocyte et al. Classification, subtype discovery, and prediction of outcome development. J Immunol 1999; 162: 1448–1459. in pediatric acute lymphoblastic leukemia by gene expression 29 Ward AC, Touw I, Yoshimura A. The Jak–Stat pathway in normal profiling. Cancer Cell 2002; 1: 133–143. and perturbed hematopoiesis. Blood 2000; 95: 19–29.

Supplementary Information accompanies the paper on the Leukemia website (http://www.nature.com/leu)

Leukemia