The Genomic Landscape of Core-Binding Factor Acute Myeloid Leukemias

Abstract

Acute myeloid leukemia (AML) comprises a heterogeneous group of leukemias frequently defined by recurrent cytogenetic abnormalities, including rearrangements involving subunits of the core-binding factor (CBF) transcriptional complex. To better understand the genomic landscape of CBF-AMLs, we analyzed both pediatric (n=87) and adult (n=78) samples, including cases with RUNX1-RUNX1T1 (n=85) or CBFB-MYH11 (n=80) rearrangements, by whole-genome or whole-exome sequencing. In addition to previously reported somatic mutations in the Ras signaling pathway, we identified recurrent stabilizing mutations in CCND2, suggesting a recurrent and previously unappreciated cooperating pathway in CBF-AML. Outside of signaling alterations, RUNX1-RUNX1T1 and CBFB-MYH11 AMLs demonstrated a remarkably different spectrum of cooperating mutations as RUNX1-RUNX1T1 cases harbored recurrent somatic mutations in DHX15 and ZBTB7A, as well as an enrichment of somatic mutations in epigenetic regulators, including ASXL2, and in components of the cohesin complex. This detailed analysis provides insights into the pathogenesis and development of CBF-AML, while highlighting dramatic differences in the landscape of cooperating mutations between these related AML subtypes.

CBF-AMLs, which account for approximately 30% of pediatric and 15% of adult AMLs, contain chromosomal translocations or inversions that target the transcription factors RUNX1 (AML1) or CBFB. Normally RUNX1 and CBFB heterodimerize to bind DNA and recruit lineage defining transcription factors to regulate hematopoietic differentiation¹. The fusion products found in CBF-AMLs [RUNX1-RUNX1T1/t(8;21)(q22;q22) or CBFB-MYH11/inv(16)(p13.1q22) or t(16;16)(p13.1;q22)] block myeloid differentiation but are not sufficient to induce leukemia, as leukemia development in murine models of both fusions requires secondary mutations, including those in the Ras pathway^2–5.

While initial studies focused on validating known mutations, especially those in the Ras pathway^6–17, and recent profiling studies focused on limited gene sets^18–20 or had limited sample numbers^21–23, we sought to comprehensively understand genetic variations contributing to CBF-AML development by analyzing 165 cases by whole-genome sequencing (WGS, n=17) or whole-exome sequencing (WES, n=148) (Table 1 and Supplemental Table 1). In this cohort, the age distribution of patients with RUNX1-RUNX1T1 (n=85) or CBFB-MYH11 (n=80) rearrangements were comparable, as were the outcomes (Supplemental Figure 1). An average of 9.86 ± 6.16 (± s.d.) somatic mutations with functional consequences were identified per case, with more mutations in RUNX1-RUNX1T1 cases (11.86 ± 6.40 vs. 7.74 ± 5.13 for CBFB-MYH11, p <0.0001) and adult cases (12.56 ± 6.55 vs. 7.44 ± 4.62 for pediatric, p <0.0001) (Supplemental Figure 2 and Supplemental Tables 2–7), while the spectrum of mutations for both CBF-AMLs subtypes was similar to the previously reported signature for AML²⁴ (Supplemental Figure 3). Subclonal variants were identified in all CBF-AMLs analyzed by WGS followed by capture validation at > 1200× coverage, consistent with other AMLs^25,26 (Supplemental Figure 4). Few copy number alterations (CNAs) or additional structural variations (SVs) (Supplemental Figure 5 and Supplemental Tables 8 and 9) were present. Consistent with previous reports^14,27,28, 66.1% of the cases harbored activating mutations in NRAS, KIT, FLT3, KRAS, PTPN11 and/or loss-of-function mutations in NF1 (Figure 1); cumulatively these mutations were not associated with outcome (Supplemental Figure 6). NRAS was the most frequently mutated gene in CBF-AMLs, yet was not associated with outcome (p=0.638). Not only were NRAS mutations more common in CBFB-MYH11 AMLs (p=0.032), but the spectrum of mutations was also different as codon 61 mutations were more common in CBFB-MYH11 cases (p=0.0023, Supplemental Figure 7). Additionally, there was an enrichment of KIT exon 17 mutations (p=0.0051) in the RUNX1-RUNX1T1 cohort^13,15, which were associated with an inferior outcome (p=0.0097, See Supplementary Note and Supplemental Figures 6, 8, and 9).

Table 1.

Characteristics CBF-AML cohort

Subtype	Total	Pediatric	Adult	WGS	WES	RNAseq
RUNX1-RUNX1T1	85	44	41	7	78	27
CBFB-MYH11	80	43	37	10	70	9
Total	165	87	78	17	148	36

WES, whole exome sequencing; WGS, whole genome sequencing

Figure 1. Mutational landscape of CBF-AML.

(a) Mutational data for 165 CBF-AML cases sequenced either by whole genome (n=17) or whole exome sequencing (n=148). Signaling, epigenetic, and cohesin genes are grouped into functional groups. Cytogenetic abnormalities and patient age group (adult or pediatric) are shown along the bottom of the figure. Mutations in both epigenetic (p=4.3E–10) and cohesin (p=2.2E–16) genes are significantly enriched in RUNX1-RUNX1T1 AML (Fisher’s Exact test). (b) The frequency of recurrently mutated genes (n>3) separated by CBF-AML fusion type is shown. Of the 10 FLT3 mutations, 4 are internal tandem duplications (ITD), 5 are located in the tyrosine kinase domain, and 1 is classified as neither ITD nor TKD.

MGA, a negative regulator of MYC signaling²⁹, was also recurrently mutated in CBF-AMLs (Figure 2). The MGA mutations in RUNX1-RUNX1T1 AMLs are all predicted to be deleterious (including the splice site variant) and would eliminate the helix-loop-helix domain and binding of MGA to MAX²⁹ (Figure 2). Similar loss-of-function mutations in MGA have also recently identifed in lung cancer, chronic lymphocytic leukemia, ETV6-RUNX1 acute lymphoblastic leukemia, NK/T cell lymphomas and in a single case of RUNX1-RUNX1T1 AML in the AML TCGA study^30–34. We also found three cases with MYC mutations and seven cases (five RUNX1-RUNX1T1 and two CBFB-MYH11) with mutations in cyclin D2 (CCND2), a known downstream target of MYC³⁵. The CCND2 mutations (Figure 2 and Supplemental Figure 10) all surround a conserved phosphorylation site (Thr280) that regulates ubiquitination of Lys270 by FBXL2³⁶ and degradation by the proteosome³⁷. Germline and somatic mutations in this region of CCND2 have been identified in patients with Megalencephaly-Polymicrogyria-Polydactyly-Hydrocephalus (MPPH) syndrome³⁸, but have been uncommon in myeloid malignancies^25,39. Much like the germline mutations in MPPH, the CCND2 mutations identified here led to increased stability of CCND2 (Figure 2 and Supplemental Figure 10). Collectively, these data suggest that stabilization of CCND2 is a previously unidentified and potentially targetable mutation in CBF-AML leukemias.

Figure 2. Recurrent mutations in CCND2 and MGA.

Domain structure and the localization of mutations are shown for (a) CCND2 and MGA. (b) Representative western blot (of three independent experiments) of HEK293T cells expressing wild-type, P281R, or T282* CCND2 treated with cyclohexamide (CHX) and harvested at the indicated time points. Data show the expected increase in the levels of mutant CCND2 protein. GAPDH serves as a loading control.

Recurrent p. R222G mutations in DHX15, a member of the DExD/H-box family of RNA helicases whose yeast homolog (Prp43) participates in spliceosome function and ribosomal biogenesis^40–43, were also identified exclusively in the RUNX1-RUNX1T1 cohort. Interestingly, ribosomal biogenesis is also affected by RUNX1 deficiency⁴⁴. The R222G missense mutation lies in the back of the RecA1 domain along the translocation site and, therefore, is unlikey to affect its ATPase activity or RNA binding ability (Supplemental Figure 11). We used RNAi-mediated knockdown of DHX15 to mimic a complete loss of function and observed differential expression of splicing genes and genes involved in ribosomal biogenessis (Supplemental Tables 10–12) as well as an increase in the number of alternative splicing events upon loss of DHX15 protein. A similar increase in alternative splicing events was obtained when the R222G mutant was overexpressed in vitro. These data led us to examine proteins with differential binding capabilities between the wildtype and mutant DHX15 protein via mass spectroscopy and co-immunoprecipitation (Supplemental Table 13). We found that TFIP11, a component of the splicing machinery, exhibits reduced binding to the DHX15 R222G mutant (Figure 3). While common in RUNX1-RUNX1T1 cases, to date this mutation has not been seen in other myeloid neoplasms; surprisingly other myeloid neoplasms harbor mutations in other spliceosome components (e.g. SF3B1, U2AF1)⁴⁵, which were not observed in CBF-AMLs.

Figure 3.

DHX15 is recurrently mutated in RUNX1-RUNX1T1 AML. (a) Domain structure and the localization of mutations for DHX15. (b) Increased numbers of alternative splicing events were observed upon DHX15 knockdown (red bars) or overexpression of the R222G mutant (green bars) compared to overexpression of wildtype DHX15 (blue bars). (c) siRNA mediated knockdown of DHX15 leads to an enrichment of differentially regulated genes associated with splicing and ribosomal biogenesis. (d) Western blot showing the effectiveness of the DHX15 knockdown. Equal amounts of protein were loaded for each sample. An asterisk indicates a non-specific band also used as a loading control. (e) Co-immunoprecipitation of TFIP11 with DHX15 demonstrates reduced binding of TFIP11 to the R222G mutant form of DHX15.

A striking and statistically significant enrichment for putative loss-of-function mutations in chromatin modifying genes (ASXL2, EZH2, KDM6A, EED, SETD2, KMT2D, KMT2C, and CREBBP) was observed in the RUNX1-RUNX1T1 cases (44% vs. 4% in CBFB-MYH11 cases, p=4.3E–10, Figure 1, Supplemental Figure 12). Consistent with previous reports, frameshift mutations in ASXL2 were common and found exclusively in RUNX1-RUNX1T1 CBF-AMLs^46,47. These mutations were identified in both pediatric and adult cases, and were not associated with outcome (Supplemental Figure 13). These truncating mutations would eliminate the PHD protein-protein interacting domain, and are similar to mutations in ASXL1 that inhibit myeloid differentiation and induce a myelodysplastic syndrome (MDS)-like disease in mice^48,49. Whether ASXL2 modulates PRC2 function, like ASXL1, however, is currently unknown⁴⁸. In vitro studies using two separate ASXL2 mutations suggested that ASXL2 may play a role in myeloid differentiation (Supplemental Figure 14). We did not identify mutations in ASXL1 in the de novo leukemias, despite adequate sequence coverage in our WES cohort (Supplemental Figure 15). Many of the other mutated epigenetic regulators, such as KMT2C^50,51, KDM6A⁵², and CREBBP⁵³, functionally interact in complexes that alter enhancer states⁵⁴, suggesting that enhancer dysregulation may be an important cooperating event with RUNX1-RUNX1T1. As such, recent data generated from Lowe and colleagues observed that disruption of Kmt2c in a RUNX1-RUNX1T1 model was associated with resistant disease⁵⁵. Interestingly, of the 4 patients with KMT2C mutations in our cohort, 3 relapsed in <12 months and the 4^th patient had residual disease after one course of remission induction therapy.

Mutations in the cohesin complex (n=17) were exclusively identified in RUNX1-RUNX1T1 AMLs (Supplemental Figure 16). SMC1A, SMC3 and RAD21 encode members of the complex responsible for sister chromatid cohesin during mitosis and post-replicative DNA repair⁵⁶. Mutations in these genes have been reported in AML^19,57,58 and germline mutations are some of the underlying causes of the cohesinopathies^56,59. The seventeen cases with cohesin mutations lack evidence of aneuploidy or an increase rate of DNA mutations (Supplemental Tables 1 and 2) and mutation status has no effect on outcome (Supplemental Figure 17). We speculate that these mutations, similar to those in the epigenetic regulators mentioned above, alter transcriptional programs by dysregulating enhancer function⁵⁰. We also identified recurrent mutations in ZBTB7A/Pokemon (Supplemental Figure 18). The putative loss-of-function mutations identified here are consistent with recent reports of ZBTB7A acting as a tumor suppressor in RUNX1-RUNX1T1 AML^22,60

Transcriptome analysis was performed on the 36 (27 RUNX1-RUNX1T1 and 9 CBFB-MYH11) cases for which material was available. Differential expression analysis generated data consistent with previous array-based studies^61,62 (Supplemental Figure 19 and Supplemental Tables 14–16), yet mutation specific differences could not be definitively determined due to our relatively small sample size. We were able to detect the alternatively spliced 9A isoform of RUNX1T1 in every RUNX1-RUNX1T1 case, ranging in abundance from 13% to 64% (Supplemental Figure 20)⁶³.

Relapse material was characterized by WES for 8 samples in the cohort, demonstrating dynamic patterns of clonal evolution during disease progression including loss, retention or gain of somatic mutations and copy number alterations (Figure 4A, Supplemental Figure 21 and Supplemental Tables 17 and 18). For example, multiclonal KIT mutations were present at diagnosis in one case, with expansion of only one subclonal population at relapse (Figure 4B), whereas another case retained a KRAS mutation while losing a KIT mutation at relapse.

Figure 4. Mutant allele frequency (MAF) at diagnosis and relapse in CBF-AMLs.

a) The mutant allele frequencies of mutations in the indicated genes are shown for the eight samples with relapse material available. D – diagnosis, R – relapse. Darker color indicates higher MAF.

Despite sharing a common molecular alteration involving a component of the core binding factor transcription complex, AMLs with RUNX1-RUNX1T1 and CBFB-MYH11 alterations have a remarkably different spectrum of cooperating mutations (Supplemental Figure 22). Signaling mutations are common in both subtypes, yet RUNX1-RUNX1T1 cases have different types of signaling alterations that include CCND2. Mutations involving epigenetic regulators and the cohesin complex, as well as mutations in DHX15 and ZBTB7A, are also frequently observed in RUNX1-RUNX1T1 cases. While suggestive of a functional role in RUNX1-RUNX1T1 leukemogenesis, future studies will be required to evaluate their role in CBF-AML pathogenesis.

METHODS

Subject cohorts and sample detail

Tumor and germline samples from pediatric CBF-AMLs from the St. Jude Children’s Research Hospital tissue resource core facility and adult CBF-AMLs from the University Hospital of Ulm were obtained with informed consent using a protocol approved by the St. Jude Children’s Research Hospital institutional review board. Detailed clinicopathological and sequencing information are provided in Supplementary Tables 1–9. The study cohort comprised 165 leukemias (85 RUNX1-RUNX1T1 and 80 CBFB-MYH11 with matched germline samples) in 2 cohorts: a cohort for whole genome sequencing (n=17; 7 RUNX1-RUNX1T1 and 10 CBFB-MYH11, all pediatric cases) and a cohort for whole exome sequencing (n=148, 78 RUNX1-RUNX1T1 and 70 CBFB-MYH11). Matched germline DNA was acquired at remission from either bone marrow or peripheral blood or was flow sorted from leukemic samples.

Whole-genome and whole-exome sequencing and analysis

Whole-genome and whole-exome sequencing were performed as previously described^64,65. Paired-end sequencing was performed using Illumina HiSeq platform with a 100-bp read length. Whole-genome sequencing mapping, coverage and quality assessment, SNV and indel detection, tier annotation for sequence mutations and prediction of the deleterious effects of missense mutations have been described previously^64,65. Mapping statistics and coverage for each tumor is summarized in Supplemental Tables 3 and 4. SNVs discovered by whole genome sequencing were classified as tier 1, tier 2, tier 3 or tier 4 as previously described^64,65. All tier 1–3 SNVs were validated by a custom capture platform, as well as all coding indels and SVs (Supplemental Table 5). The overall validation rate was 83%, with a median validation rate of 87% per sample. All recurrent SNVs (Supplemental Table 6) in the whole-exome sequencing cohort were validated by either amplicon sequencing using the MiSeq platform or Sanger sequencing, as previously described^64,65. Briefly, PCR primers were designed to flank the putative variant using Primer3. Amplicon sizes ranged from approximately 400bp to 600b. PCR was performed using AmpliTaq Gold 360 master mix (Applied BioSystems), 400 nM of each primer (IDT) and 20 ng of repli-G whole genome amplified DNA (QIAGEN) in a 25 ul reaction volume. Thermo cycling was performed using the following parameters: 95°C for 10 min., 95°C for 30 sec., 65°C for 30 sec., 72°C for 1 min. for 35 cycles, 72°C for 7 min. All amplicons were quality checked on a 2% agarose E-gel (Invitrogen). Pooled amplicons were used to create DNA libraries using the Nextera XT kit (Illumina) following the manufacturer’s instructions. Libraries were normalized and sequenced on an Illumina MiSeq using a 2 × 150 paired-end version 2 sequencing kit. Sanger Sequencing was also performed on a subset of amplicons. Primer pairs and SNV targets used in validation are included in Supplemental Table 18. All validated SNVs are summarized in Supplemental Table 7. CNAs that were identified by WGS data were analyzed by evaluating the difference in read depth between each tumor and its matching normal (germline) sample using CONSERTING⁶⁶ (Supplemental Table 8). Other CNAs were detected by Affymetrix SNP6.0 arrays, as previously described^28,29. Structural variations in whole genome sequencing data were analyzed using CREST and were annotated as previously described^64,67(Supplemental Table 9). To assess tumor contamination of the germline sample we evaluated the existence of SVs supporting the RUNX1-RUNX1T1 and CBFB-MYH11 fusions in germline samples using CREST⁶⁷. The level of tumor in normal contamination is calculated as number of reads supporting fusion / number of total reads and was under the limit of detection in each sample (Supplemental Table 3).

Transcriptome sequencing

For library construction, 2–5 μg of total RNA was extracted from tumor samples by using Qiagen RNeasy Mini kits according to the manufacturer’s instructions. RNA concentration was measured by using a NanoDrop 100 Spectrophotometer (Thermo Scientific). RNA integrity was measured by using an Agilent Technologies 2100 Bioanalyzer Lab-on-a-chip system. Total RNA was treated with DNAse I (Invitrogen) and enriched for poly A containing mRNA using oligo dT beads (Dynabeads, Invitrogen). The cDNA synthesis used random hexamers and the Superscript Double-Stranded cDNA Synthesis kit (Invitrogen). Paired-end reads from mRNA-seq were aligned to the following 4 database files by using a Burrows-Wheeler Aligner (0.5.5): (i) human NCBI Build 37 reference sequence, (ii) RefSeq, (iii) a sequence file representing all possible combinations of non-sequential pairs in RefSeq exons, and (iv) AceView flat file downloaded from UCSC, representing transcripts constructed from human expressed sequence tag (EST). The final BAM (compressed binary version of the Sequence Alignment/Map [SAM] format) file was constructed by selecting the best alignment in the four databases.

CNA detection using WES data

Samtools⁶⁸ mpileup command was used to generate an mpileup file from matched normal and tumor BAM files with duplicates removed. VarScan2⁶⁹ was then used to take the mpileup file to call somatic CNAs after adjusting for normal/tumor sample read coverage depth and GC content. Circular Binary Segmentation algorithm⁷⁰ implemented in the DNAcopy R package was used to identify the candidate CNAs for each sample. B-allele frequency info for all high quality dbSNPs heterozygous in the germline sample was also used to assess allele imbalance.

Accession numbers

Genomic data have been deposited at the European Genome-phenome Archive (EGA), which is hosted by the European Bioinformatics Institute (EBI), under EGAS00001000349.

Lineage-negative enrichment and differentiation assays of murine hematopoietic cells

Experiments were approved by the St. Jude Children’s Research Hospital Institutional Animal Care and Use Committee. Briefly, bone marrow cells were harvested from 8-wk old C57BL/6 mice. Lineage-negative cells were purified by magnetic separation and cultured for 48hr in IMDM/20% FBS supplemented with penicillin-streptomycin, L-glutamine, recombinant mouse IL-3 (10 ng/ml), IL-6 (20 ng/ml), and SCF (50 ng/ml) (Peprotech). Cells were infected on RetroNectin-coated plates for 48hr (Takara Bio Inc.) with MSCV-IRES-mCherry retrovirus expressing wild-type or mutant Flag-tagged ASXL2. Transduced mCherry⁺ cells were obtained by fluorescence-activated cell sorting. For flow analysis, cells in culture were stained with anti-mouse-Gr-1 (RB6-8C5), anti-mouse-Mac1 (M1/70), anti-mouse-cKit (2B8), and anti-mouse-Sca1 (E13-161.7) (BD Biosciences) on day 14 post transduction. The data was analyzed using FlowJo software.

Mutagenesis

pCMV6 constructs containing wild-type human CCND2 and DHX15 were obtained from Origene. CCND2 mutations (P281R and T282*) and DHX15 (R222G) were introduced using the Quickchange II XL kit (Agilent) per manufacturer’s instructions using the primer pairs shown below:

T282*_F	5′-atcccgcacgtctcaaggggtgctggcttg-3′
T282*_R	5′-caagccagcaccccttgagacgtgcgggat-3′
P281R_F	5′-ccgcacgtctgtacgggtgctggcttg-3′
P281R_R	5′-caagccagcacccgtacagacgtgcgg-3′
R222G_F	5′-ctactgcagtcttcaaatccaatggagtaaccaacttcc-3′

Transient transfection of CCND2 mutants

HEK293T (ATCC) cell lines were transiently transfected using FuGene HD (Promega) as per the manufacturer’s instructions. Cells were incubated for 24 hr, and then subjected to 50 μg/mL cyclohexamide and harvested at the indicated time points in NP-40 lysis buffer (150 mM sodium chloride, 1% NP-40 and 50 mM Tris, pH 8.0) supplemented with Halt Protease and Phosphatase Inhibitor Cocktail (ThermoScientific). Western blotting of 293T cells expressing CCND2 constructs was performed using 10 μg of protein lysates prepared using NuPAGE LDS buffer (Life Technology) electrophoresed through 4–12% NuPAGE Bis-TRIS gels under reducing conditions. After transferring to PVDF membrane (iBlot, Life Technology), membranes were probed with total CCND2 (Abcam, ab3085) and GAPDH (Santa Cruz Biotechnology, 6C5) antibodies using the iBind system (Life Technology). Quantification of western blots was completed using the ImageJ software.

Modeling of DHX15

The structure of PRP43 bound to ADP was obtained from the Protein Data Bank⁷¹ (PDB: 2XAU⁷²). Mutations and models were generated using Pymol⁷³.

DHX15 knockdown, overexpression, and mass spectrometry

HEK293T cells were transfected with a pool of siRNAs targeting DHX15 (Supplemental Table 19) using Lipofectamine RNAiMax (ThermoFisher Scientific) according to manufacturer instructions. Cells were transfected again after 24 hours and harvested at day 4. For overexpression experiments, HEK293T cells were transfected with pCMV6-DHX15 or pCMV6-DHX15 R222G. RNA was extracted after 48 hours with Trizol and libraries prepared as described above. Pull down experiments for mass spectrometry were done using the Pierce c-Myc Tag IP/Co-IP Kit (Thermo Scientific). The MS analysis was performed according to the optimized platform as previously reported⁷⁴. Proteins in the gel band were in-gel digested by trypsin. Resulting peptides were loaded on a nanoscale reverse-phase column, and eluted by a gradient (~30 min). Eluted peptides were detected by an inline LTQ Orbitrap Elite mass spectrometer (Thermo Scientific). Acquired data were searched against the protein database using the Sequest algorithm and filtered to reduce the protein false discovery rate to below 1%. For binary comparison, statistical analysis was also performed based on the G-test. FDR was derived according to biological replicates and null hypothesis⁷⁵.

Western blots and co-immunoprecipitations

Co-immunoprecipitations (CO-IP) and immunoblots (WB) used standard techniques as described previously⁷⁶. Briefly, mouse embryonic fibroblasts were transduced to stably coexpress human TFIP11, 2× myc tagged with either wild type human DHX15, 2× HA tagged or R222G mutant, 2× HA tagged. Cells were lysed in NP-40 buffer (0.5% NP-40, 20 mM Tris, pH 7.5, 137 mM NaCl, 2 mM EDTA and 10% glycerol, 100 ug/mL RNase A). Agarose-conjugated anti-HA (Roche) or anti-myc (Cell Signaling Technology) antibodies were used for CO-IP. Horse radish peroxidase (HRP)-conjugated ani- HA (Roche) or anti-myc (Cell Signaling Technology) antibodies were used for WB.

Gene set enrichment analysis

Array data from Ross et al⁶² was used for Limma⁷⁷ analysis to create a list of differentially expressed genes between cases of RUNX1-RUNX1T1 and CBFB-MYH11 CBF leukemia (Supplemental Table 16). The resulting gene list was truncated at a false discovery rate of 0.001 and the remaining genes were used to create a gene set that represented the gene signature of each fusion. The transcriptome data was then used for gene set enrichment analysis (GSEA)⁷⁸.

Mutational spectrum analysis

We used the published computational framework for mutation signature identification^24,79. The algorithm deciphers the minimal set of mutational signatures that optimally explains the proportion of each mutation type found in each catalogue and then estimates the contribution of each signature to each catalogue. We used somatic mutations called from the whole exome sequencing diagnostic samples of RUNX1-RUNX1T1 (70 samples, 1595 mutations) and CBFB-MYH11 (66 samples, 1024 mutations).

Statistical Methods

Enrichment in the number of cases harboring mutations in epigenetic and cohesin genes in the RUNX1-RUNX1T1 vs. CBFB-MYH11 cohorts was calculated using the Fisher’s Exact test. Differences in the expression level of CCND2 were calculated using the Student’s t-test. The association of a mutation with event free survival (EFS) or overall survival (OS) was performed using Cox proportional hazard model; survival curves were also compared by log rank test. Event-free survival (EFS) was defined as the time elapsed from on study date to induction failure, withdrawal, relapse, secondary malignancy, or death, with those living and event-free at last follow-up censored. Overall survival (OS) was defined as the time elapsed from on study date to death with those still living at last follow-up considered censored. Outcome data was available for 159 out of 165 CBF-AML patients.