Genomic and Epigenomic Landscapes of Adult De Novo Acute Myeloid Leukemia

PATIENTS

We selected samples from 200 adults with de novo AML to represent the major morphologic and cytogenetic subtypes of AML.^8,15,16 The characteristics of these patients are fully described in Table 1, and in Tables S1 and S2, Figure S1, and the Materials section in the Supplementary Appendix, available with the full text of this article at NEJM.org. A video describing AML and this study is also available at NEJM.org.

ANALYTIC PLATFORMS

We performed whole-genome sequencing of the primary tumor and matched normal skin samples from 50 patients (with data from 24 of these patients reported previously¹⁷) and exome capture and sequencing for another 150 paired samples of AML tumor and skin (see Table S3 in the Supplementary Appendix for coverage data for the 200 samples).

All 200 patients who were selected for this study were enrolled in a single-institution tissue-banking protocol approved by the human studies committee at Washington University. Written informed consent for whole-genome sequencing was obtained from all study participants.

The samples, which were banked between November 2001 and March 2010, were selected from a set of more than 400 samples to reflect a real-world distribution of subtypes. Sample inventory and quality issues also had to be considered in the selection process, since the samples were analyzed on several different platforms. We identified candidate somatic variants using several algorithms (see the Methods section in the Supplementary Appendix), and all the variants for the 200 samples were verified with the use of hybridization capture–based methods and deep digital sequencing.¹⁸ We performed RNA-expression profiling on the Affymetrix U133 Plus 2 platform for 197 samples, RNA sequencing for 179 samples, microRNA (miRNA) sequencing for 194 samples, Illumina Infinium Human Methylation 450 BeadChip profiling for 192 samples, and Affymetrix SNP Array 6.0 for both tumor and normal skin samples from all 200 patients. Data sets were not completed for all samples on all platforms because of assay failures and availability and quality issues for some samples. The complete list of data sets is provided in Table S4 in the Supplementary Appendix. All data sets are available through the Cancer Genome Atlas (TCGA) data portal (https://tcga-data.nci.nih.gov/tcga).

COPY-NUMBER ALTERATIONS

In most de novo AML samples with a cytogenetic profile indicating intermediate or favorable risk, we detected very few copy-number events on high-resolution SNP arrays, as reported previously³ (Fig. S2 and Table S5 in the Supplementary Appendix). Unfavorable-risk samples had chromosomal copy-number alterations that were verified on the SNP arrays, and many had additional, cytogenetically cryptic events. No samples contained evidence of chromothripsis (a single genomic event that results in focal losses and rearrangements in multiple genomic regions).¹⁹ Details of these analyses are presented in the Materials section in the Supplementary Appendix.

MUTATIONS IN CODING SEQUENCES

We discovered 2315 somatic single-nucleotide variants (SNVs) and 270 small insertions and deletions (indels) in coding (tier 1) regions of the genome, with an average of 13 tier 1 mutations per sample (range, 0 to 51) (see box, and Table S6 in the Supplementary Appendix). We observed no recurring coding mutations in three samples, but all contained well-recognized fusion events that are known to initiate AML (NUP98-NSD1 in Patient 868231, MLL-MLLT3/AF9 in Patient 923966, and MLL-MLLT10/AF10 in Patient 558395). Three outlier samples contained 51, 36, and 35 tier 1 mutations; none of these samples contained mutations in known DNA-repair genes. Of the 2315 SNVs, 1539 (66%) were missense and 510 (22%) had no translational consequences. Small indels accounted for 270 of the 2585 validated mutations (10%); of these, 191 (71%) caused frameshifts.

Samples were stratified into 10 groups on the basis of the presence or absence of known recurring fusion events, cytogenetic-risk profile, or the presence or absence of TP53 mutations (which were strongly associated with an unfavorable cytogenetic risk) (Fig. 1A). We observed significant differences in the numbers of recurrent tier 1 mutations in some of these groups. Eleven samples had MLL fusions; this group had the fewest recurrent tier 1 mutations, with a mean of 2.09, as compared with a mean of 5.24 for all 200 samples (P = 0.002 after correction for multiple comparisons). This finding suggests that MLL fusions require fewer cooperating mutations than other AML-initiating events. Similarly, 20 samples containing PML-RARA fusions had fewer recurrent tier 1 mutations (mean, 3.25; P=0.001). We observed a higher mean number of recurrent tier 1 mutations in 7 samples containing either RUNX1-RUNX1T1 fusions (mean value, 7.85; P = 0.04) and in 13 samples with a combination of a high-risk cytogenetic profile and a TP53 mutation (mean, 7.00; P = 0.049). Larger sample sets will be required to confirm these observations.

A total of 260 genes had somatic mutations in at least 2 of the 200 samples; in 154 of these genes, more than one mutation was nonsynonymous. An additional 1623 genes were found to have a validated tier 1 mutation in one sample. Using the significantly mutated gene (SMG) test in the Mutational Significance in Cancer (MuSiC) suite of tools,²⁰ we identified 23 genes with a higher-than-expected mutation prevalence (false discovery rate, <0.05), including genes that are well established as being relevant to AML pathogenesis (e.g., DNMT3A, FLT3, NPM1, IDH1, IDH2, and CEBPA), along with genes that have only recently been implicated in AML pathogenesis, including U2AF1, EZH2, SMC1A, and SMC3 (Fig. 1B, and Table S7 in the Supplementary Appendix).

We also identified and verified all variants in noncoding regions in the 50 sample pairs that we analyzed using whole-genome sequencing. After the exclusion of 1 tumor sample, from Patient 817156, that had a high level of AML tumor cells (36%) in the skin sample (Table S1 in the Supplementary Appendix), the median number of non-coding mutations in tumor samples was 394, ranging from 68 to 1298. There was a strong correlation between the number of coding and noncoding mutations in each genome (Pearson’s correlation coefficient, 0.78), suggesting that most of the mutations were randomly distributed throughout each genome. Most mutations in AML genomes are probably background events that occurred in hematopoietic stem cells before the initiating event occurred; the clonal expansion of these cells captures their mutational history, as reportedly previously.¹⁷ This is also the reason why nearly all mutations in AML genomes are present in nearly all the cells in each sample.¹⁷ The results of an analysis of recurrently mutated regions in tiers 2 and 3 (nongenic regions¹¹) are presented in Table S8 in the Supplementary Appendix, as are data for mitochondrial variants (Table S9 in the Supplementary Appendix); the relevance of these events to pathogenesis is unclear.

Using deep digital sequencing, we verified all tier 2 and 3 variants that were discovered with the use of whole-genome sequencing. This provided a large number of sites for variant allele frequency (VAF)–based cluster analysis, which allowed us to define the clonal composition of each tumor.^17,18 More than half the tumors contained both a founding clone (the clone with the highest VAF values) and at least one subclone; we were able to identify as many as three independent subclones in one tumor sample (Fig. 1C, and Table S1 in the Supplementary Appendix). The coverage provided by whole-genome sequencing in this study (mean, 30.54×) limited the power to detect small sub-clones with VAFs of less than 10% (Fig. S3A in the Supplementary Appendix). Exome sequencing produced a higher level of coverage for the targeted sequences (mean, 167.50×), slightly increasing our ability to detect mutations with VAFs of less than 10%. However, the difference between the number of tier 1 mutations detected with whole-genome sequencing (14.5 per sample) and the number detected with exome sequencing (12.7 per sample) was not significant (P = 0.17) (Table S6 in the Supplementary Appendix). The mutational spectrum of all validated SNVs for all 200 samples is shown in Figure S3B in the Supplementary Appendix; transitions were the most common mutation type, as reported previously.^17,18

EXPRESSION OF MUTANT ALLELES

Analysis of RNA sequences revealed allelic bias for mutations in several genes. We observed increased or exclusive expression of the mutant DNMT3A, RUNX1, PHF6, and TP53 in several cases (Fig. S4A through S4F in the Supplementary Appendix). Loss of heterozygosity or partial uniparental disomy explained the enrichment of mutant allele expression in most samples; epigenetic modifications (e.g., altered patterns of DNA or histone methylation) may be responsible for the rest.

miRNA VARIANTS

We identified somatic SNVs in miRNA genes in 7 of 200 samples (4%) (Table S10 in the Supplementary Appendix). Of these 7 samples, 4 had mutations in miR-142 that were localized to the seed region of the mature strand (a sequence from the 3′ region called miR-142-3p) and were likely to affect messenger RNA (mRNA) target specificity (Fig. S5A in the Supplementary Appendix). The expression of miR-142-3p and miR-142-5p (a sequence from the 5′ region) is thought to be restricted primarily to the hematopoietic compartment.^21,22 Data from miRNA sequencing showed that miR-142 was highly expressed in AML samples (Fig. S5B in the Supplementary Appendix). The mutated miR-142 alleles were expressed at levels similar to those of the nonmutated allele in all samples (Fig. S5C in the Supplementary Appendix). Although miR-142 mutations have not previously been identified, several reports have linked aberrant expression of miR-142-3p with hematologic cancers, including precursor B-cell²³ and T-cell^24,25 acute lymphoblastic leukemia and AML.²⁶ The remaining mutations in miRNA genes were localized to precursor miRNAs (in 3 samples) or to a nonseed region of a mature miRNA (in 1 sample).

GERMLINE VARIANTS

We identified all the variants predicted to cause mRNA truncation that were found in both the skin and tumor samples (see the Results section, Table S11, and Fig. S6 in the Supplementary Appendix). However, very few of the genes with truncation variants were found to be expressed in most AML samples (including the samples carrying the variants themselves), suggesting that most of the inherited truncating variants were probably irrelevant for pathogenesis.

FUNCTIONAL CATEGORIZATION OF MUTATED GENES

We used complementary approaches to identify combinations of mutations in the samples; for this analysis, we considered only nonsynonymous mutations (Fig. 2). Using the HotNet algorithm,²⁷ we identified six subnetworks of a genome-scale protein–protein interaction network that had significant mutations (P<0.001). These included portions of known pathways and protein complexes, including the cohesin complex¹⁷ (Fig. S7 in the Supplementary Appendix).

We grouped mutations into larger sets or pathways and examined patterns of mutual exclusivity and co-occurrence between these groups. Of 200 samples, 199 (>99%) contained at least one mutation in one of nine categories that were defined according to biologic function and that have a putative role in AML pathogenesis: transcription-factor fusions (18% of cases), the gene encoding nucleophosmin (NPM1) (27%), tumor-suppressor genes (16%), DNA-methylation–related genes (44%), activated signaling genes (59%), chromatin-modifying genes (30%), myeloid transcription-factor genes (22%), cohesin-complex genes (13%), and spliceosome-complex genes (14%) (Fig. 2, and Fig. S8 in the Supplementary Appendix). For all genes that had mutations in only a small number of samples, the pathophysiological relevance of the events will require further validation. FLT3 mutations were identified in 56 samples, and an additional 62 samples were found to have mutations in genes encoding other kinases, phosphatases, or RAS family proteins (Fig. S9 in the Supplementary Appendix). However, most of these genes contained mutations in only 1 to 3 samples (with the exception of KIT, KRAS, NRAS, and PTPN11). In total, 59% of samples had a mutation in a gene encoding a signaling protein.

To more fully assess patterns of mutual exclusivity and co-occurrence between sets of genes, we applied Dendrix++ (see the Materials section in the Supplementary Appendix) to the mutation matrix. Dendrix++ identified three sets of genes with the strongest patterns of mutual exclusivity (groups A, B, and C) (Fig. 2). The most statistically significant set included the transcription-factor fusion genes, NPM1, RUNX1, TP53, and CEBPA (Fig. S10 in the Supplementary Appendix). The second most significant set showed exclusivity between mutations in FLT3 and in genes encoding other tyrosine kinases, serine–threonine kinases, protein tyrosine phosphatases, and RAS family proteins (Fig. S9 in the Supplementary Appendix). The third set included mutations in ASXL1 and in genes encoding components of the cohesin complex, other myeloid transcription factors, and other epigenetic modifiers (Fig. S10 in the Supplementary Appendix). Mutations within the spliceosome gene set, genes encoding other epigenetic modifiers, and genes encoding myeloid transcription factors also showed considerable mutual exclusivity within each set (Fig. S11 in the Supplementary Appendix).

Among the pairwise relationships between mutations in the AML samples, the most prominent was the significant co-occurrence between mutations in FLT3, DNMT3A, and NPM1 (Fig. S8 in the Supplementary Appendix). In particular, many samples had mutations both in NPM1 and DNMT3A or in NPM1 and FLT3. The likelihood that these mutations occurred together by chance is extremely small (P<6.3×10⁻⁷ for NPM1 and DNMT3A and P<1.9×10⁻⁶ for NPM1 and FLT3). This observation, combined with the strong association between samples having concurrent mutations in NPM1, FLT3, and DNMT3A and distinct clusters in mRNA, miRNA, and DNA methylation, suggests that samples with mutations in all three genes represent a novel subtype of AML. Furthermore, we observed relationships of strong mutual exclusivity. For example, PML-RARA, MYH11-CBFB, and MLL-containing fusion genes were mutually exclusive of mutations in NPM1 and DNMT3A (P<0.007, P<0.04, and P<0.04, respectively), and RUNX1 and TP53 mutations were mutually exclusive of FLT3 and NPM1 mutations. Table S12 in the Supplementary Appendix contains a full list of genes or gene sets with significant co-occurrences or exclusivity among the samples.

GENE FUSIONS

De novo assembly of RNA-sequencing data²⁸ for 179 AML samples identified 118 gene fusions in 80 samples (mean, 1.5 per sample), of which 71 were distinct events (Fig. 3A)²⁹; 99 samples had no detected fusions (Table S13 in the Supplementary Appendix). The range of fusions per sample was 0 to 8. The 74 in-frame fusions included many previously described, recurrent events, including PML-RARA, MYH11-CBFB, RUNX1-RUNX1T1, BCR-ABL1, PICALM-MLLT10/AF10, NUP98-NSD1, and multiple fusions involving MLL (Fig. 3B). We identified 15 new fusion events that maintained an open reading frame. Although none of them were recurrent in this cohort, several of the genes in the fusions were mutated or translocated in other samples of AML that we analyzed (i.e., MLLT10/AF10, NF1, GRID1, PPP2R1B, XIAP, ATP1B4, WSB1, KIAA0999, TBX15, and LRRC37B). An additional 42 gene fusions were out-of-frame (Fig. 3C), creating a truncated upstream gene or potential haploinsufficiency for both partner genes, with many that were mutated or translocated in other AML samples (including RUNX1, DNMT3B, MLLT10/AF10, NSD1, EDIL3, SCARB1, XIAP1, PPP2R1B, FOXP1, KSR2, MLL3, and CUL1). One out-of-frame fusion (GAS6-FAM70B) was detected in three AML samples, and one of its fusion partners (FAM70B P233L) was mutated in another sample. Most of the newly described fusion events in these samples were not detected by means of routine cytogenetic studies (Table S1 in the Supplementary Appendix).

GENE-EXPRESSION ANALYSIS

Unsupervised non-negative matrix factorization (NMF) consensus clustering (i.e., clustering data with inputs consisting only of gene or miRNA abundance information) suggested an optimum of seven RNA-sequencing groups and five miRNA-sequencing groups (Fig. S12 in the Supplementary Appendix). Associations between these groups and overall survival are shown in Fig. S13 in the Supplementary Appendix; the groups enriched for samples containing PML-RARA had better outcomes than other groups, as expected. The RNA-sequencing groups were highly concordant with a microarray data set for the same 178 samples (Fig. S14 in the Supplementary Appendix).

We used one-sided Fisher’s exact tests (corrected for multiple testing) to identify significant associations (P<0.05) between specific RNA-sequencing and miRNA-sequencing groups and covariates (Fig. 4). RNA-sequencing group 4 was associated with subtype M1 (AML with minimal maturation) in the French–American–British (FAB) classification of acute leukemias, group 3 with FAB subtype M3 (acute promyelocytic leukemia), group 5 with FAB subtype M4 (acute myelomonocytic leukemia), and group 7 with FAB subtype M5 (acute monoblastic or monocytic leukemia) (Fig. 4A). Concordance between gene-expression groups and FAB subtypes was similar to that previously reported for microarray data^30,31 and showed that some expression signatures were strongly correlated with the stage of myeloid differentiation of the AML sample.

For the miRNA-sequencing data, group 5 was associated with FAB subtype M3, and groups 2, 3, and 5 were associated with unfavorable, intermediate, and favorable cytogenetic risk categories, respectively. Group 3 was strongly associated with mutations in NPM1, DNMT3A, FLT3, and genes encoding the cohesin complex; miR-10a was expressed at high levels in this group, an observation that is consistent with reports correlating high miR-10a expression and NPM1 mutations^32,33 (Fig. 4B). Levels of miR-424 were relatively low in this group, making miR-424 the second-most discriminatory miRNA — an observation that is consistent with the findings in a previous study.³⁴ The data also confirmed that miR-196b, miR-130a, and let-7b were discriminatory in this group.³⁵ Additional comparisons with published sets of expression data are provided in the Materials section in the Supplementary Appendix.

DNA-METHYLATION ANALYSIS

Unsupervised analysis of changes in DNA methylation revealed significant differences among subsets of samples, particularly in CpG-sparse regions of the genome (Fig. 5).³⁶ Samples with IDH1 and IDH2 mutations showed extensive gains of methylation relative to CD34+CD38− cells obtained from healthy donors (Table S14 in the Supplementary Appendix), whereas some samples with MLL fusions or co-occurring NPM1, DNMT3A, and FLT3 mutations were associated with extensive loss of DNA methylation, as compared with normal CD34+CD38− cells. Specific patterns of methylation gain and loss distinguished samples with CEBPA mutations, as well as samples with PML-RARA, RUNX1-RUNX1T1, or MYH11-CBFB fusions. Significant changes in DNA methylation were identified across AML samples at 160,519 CpG loci (42% of sites tested), with 67% resulting in a gain of methylation and 33% resulting in a loss (see the Methods section and Table S15 in the Supplementary Appendix). Samples with triple mutations in NPM1, DNMT3A, and FLT3 showed methylation losses at 328 of 382 differentially methylated regions larger than 1 kb (86%), as compared with CD34+CD38− cells from healthy donors. Although both intergenic and genic regions were affected, approximately 71% of these changes were in coding regions (Table S16 in the Supplementary Appendix).

We also assessed the relationships between gene expression and DNA methylation in pairs of data types by identifying groups in one data type that were enriched in samples from a group in the other data type (Fig. S15A through S15E in the Supplementary Appendix). Clusters containing samples of acute promyelocytic leukemia were strongly concordant for mRNA, miRNA, and CpG-sparse DNA methylation (Fig. 5, and Fig. S15 in the Supplementary Appendix). In mRNA groups 1, 2, 5, and 6, there was a preferential association with one miRNA-sequencing group each, suggesting that regulatory relationships must exist between these groups (Fig. S15A in the Supplementary Appendix). Associations in CpG-sparse DNA methylation groups were more significant for miRNA groups than for mRNA groups, suggesting the existence of previously unrecognized epigenetic regulatory pathways (Fig. S15B through S15E in the Supplementary Appendix). Genes defining the RNA- and miRNA-sequencing groups are shown in Figure S16 in the Supplementary Appendix; a very striking set of small RNA genes within an imprinted locus on chromosome 14 were found to be consistently dysregulated in acute promyelocytic leukemia (Fig. S17 in the Supplementary Appendix).³⁷