Prognostic Mutation Constellations in AML and MDS

Abstract

Purpose of review

In the past decade numerous studies analyzing the genome and transcriptome of large cohorts of acute myeloid leukemia (AML) and myelodysplastic syndrome (MDS) patients have substantially improved our knowledge of the genetic landscape of these diseases with the identification of heterogeneous constellations of germline and somatic mutations with prognostic and therapeutic relevance. However, inclusion of integrated genetic data into classification schema is still far from a reality. The purpose of this review is to summarize recent insights into the prevalence, pathogenic role, clonal architecture, prognostic impact and therapeutic management of genetic alterations across the spectrum of myeloid malignancies.

Recent findings

Recent multiomic-studies, including analysis of genetic alterations at the single-cell resolution, have revealed a high heterogeneity of lesions in over 200 recurrently mutated genes affecting disease initiation, clonal evolution and clinical outcome. Artificial intelligence and specifically machine learning approaches have been applied to large cohorts of AML and MDS patients to define in an unbiased manner clinically meaningful disease patterns including disease classification, prognostication and therapeutic vulnerability, paving the way for future use in clinical practice.

Summary

Integration of genomic, transcriptomic, epigenomic and clinical data coupled to conventional and machine learning approaches will allow refined leukemia classification and risk prognostication and will identify novel therapeutic targets for these still high risk leukemia subtypes.

INTRODUCTION

Acute myeloid leukemia (AML) and myelodysplastic syndromes (MDS) are a heterogeneous group of diseases characterized by clonal expansion of undifferentiated myeloid precursors, impaired hematopoiesis and bone marrow failure. The threshold of 20% bone marrow blasts distinguishes AML from MDS on morphologic examination ^[1]. Approximately 30% of MDS cases progress to AML, the incidence of which varies across different MDS subtypes. Progression is generally associated with acquisition of driver mutations leading to clonal outgrowth ^[2]. In both diseases genetic alterations are progressively acquired overtime in hematopoietic stem cells that first gain a selective advantage (premalignant clone) and then fully transform in malignant clones when secondary mutations are acquired (Fig. 1). According to their biological function and clinical significance, most frequent mutations can be grouped into nine classes including NPM1, myeloid transcription factors, DNA methylation–related genes, epigenetic/chromatin modifiers, tumor suppressor genes, signal transduction pathways, RNA processing and splicing factors, cohesin complex and DNA repair. Mutations in DNA methylation, epigenetic modifiers and splicing machinery genes are typically acquired early, while mutations in transcriptional regulators, signal transduction pathways and chromosomal abnormalities may occur later. In addition to somatic mutations, germline mutations in several genes, such as DDX41 ^[3], RUNX1 ^[4,5], GATA2 ^[6], ANKRD26 ^[7], ETV6 ^[8], TP53 ^[9], CEBPA ^[10], SAMD9/SAMD9L ^[11-13], Fanconi anemia genes ^[14], or telomerase complex genes^[15] predispose to AML and MDS in both children and adults ^[16,17]. Although common targets of mutations are largely similar across age and between AML and MDS, their frequency, type and co-mutational patterns vary by age and disease subtype ^[18,19]. The main genetic similarities, differences and their impact on prognosis will be discussed in this review.

Figure 1. Simplified model of clonal expansion of hematopoietic stem and progenitor cells leading to MDS and AML.

An initiating driver mutation in a hematopoietic stem cell promotes the expansion of mutant stem cells and abnormal hematopoietic progenitor and precursor cells. These cells expand and become dominant. The occurrence of secondary mutations promotes a malignant cell transformation. According to cell morphology features, occurrence of dysplasia and percentage of bone marrow and peripheral blood blasts a diagnosis of AML or MDS is made. In MDS the acquisition of additional driver mutations or the emergence of preexisting ones leads to progression to AML in around 30% of cases.

CONSTELLATIONS OF GENETIC ALTERATIONS

Constellations of genetic alterations in AML

Acute myeloid leukemia includes a clinical and genetic heterogeneous group of hematopoietic malignancies that arise as a result of clonal expansion of undifferentiated myeloid precursors in the bone marrow due to genetic abnormalities that impair self-renewal, proliferation and differentiation. The prevalence of AML increases with age from 20% of all leukemias in childhood (with a peak in infancy) to 80% in older adults ^[20]. AML can present as either de novo or secondary disease (therapy related or post-antecedent hematologic disorder). Although intensified treatment regimens and risk-adapted patient stratification have improved overall survival, outcome is still dismal with overall 5-year survival of 35%-40% in patients aged <60 years and 5%-15% in older patients ^[20]. Since the first whole genome sequencing of a single AML patient in 2008 ^[21], large sequencing studies have considerably increased ^[19,22,23]. These studies have shown extensive genetic heterogeneity and important differences between the genomic landscapes of pediatric and adult AML. While cytogenetically normal AML represents ~40% of adult AML, it is only 15-20% of pediatric cases (Fig. 2A). In childhood, common gene fusions drive distinct pathogenetic pathways. Among those, KMT2A (also known as mixed lineage leukemia, MLL) rearrangements (KMT2Ar) are common in childhood AML (~15%), especially in infants (~50%), and, although partner dependent, are generally associated with unfavorable outcome ^[24,25]. The fusion RBM15- MKL1 arising from the t(1;22)(p13;q13) chromosomal rearrangement is a World Health Organization (WHO)-defined subtype of AML that occurs in around 10% of non-Down syndrome AML displaying megakaryoblastic differentiation (AMKL) ^[26,27], and is associated with intermediate clinical course ^[28]. Cryptic gene fusions, including NUP98-fusions (4-9% of pediatric AML) ^[18,29,30], CBFA2T3-GLIS2 (2% of all AML pediatric cases but 20–30% of pediatric AMKL cases) ^[26,27,31], and MNX1-ETV6 (~ 1% of all pediatric AML cases but 4-30% of AML in children less than 2 years old) ^[32] are almost exclusively found in pediatric AML and predict poor outcome ^[33-36]. Similar to MLL, NUP98, encoding a component of the nuclear pore complex, is rearranged to multiple different partner genes in AML. NUP98-KDM5A frequently co-occurs with RB1 gene deletion and is recurrent in acute erythroid leukemia (AEL, ~20% of pediatric cases), while cells with NUP98-NSD1 fusion frequently harbors FLT3 internal tandem duplications (ITD) or WT1 mutations ^[36-38]. Fusions affecting the core binding factor (CBFB-MYH11), the retinoid acid receptor (PML-RARA and variant RARA rearrangements) or the runt-related transcription factor 1 (RUNX1-RUNX1T1; AML1-ETO) disrupt transcription factors important for myeloid differentiation, occur at any age but with a peak in children and young adults, and are associated with a relatively favorable prognosis in both adults and children ^[39,40]. Co-occurring genetic alterations are observed and they significantly alter outcome, such as in the case of KIT mutations in core-binding factor leukemias ^[41-43]. Additional AML-associated cytogenetic abnormalities, more frequently found in adults than pediatric cases, with dismal outcome ^[44,45] include the DEK proto-oncogene fused to the nucleoporin gene 214 (NUP214) ^[46] (< 2%) and inv(3) or t(3;3) (1-4%) which is associated with elevated platelet counts, dysplastic megakaryocytes, multilineage dysplasia and MECOM overexpression ^[44,47-49]. With regard to mutations (Fig. 2B), alterations in transcription factors such as WT1 and GATA2 or in signaling genes (e.g. RAS and KIT) are more prevalent in pediatric cases compared to adult patients. Conversely, mutations in epigenetic modifiers, such as DNMT3A, TET2 and IDH1/2, in NPM1 and TP53 are more frequent in adult AML ^{[19,38,50-53]}.

Figure 2. Recurrent genetic alterations in AML and MDS.

A) Main cytogenetic abnormalities according to age in pediatric (left panel) and adult (right panel) AML. B) Prevalence of somatic mutations (single nucleotide variations, indels and duplications) in adult versus pediatric AML. Data are from landmark previously published studies ^{[18,19,23,38]}. C) Prevalence of somatic mutations (single nucleotide variations, indels and duplications) in adult versus pediatric MDS. Data are from landmark previously published studies ^[60,62,63]. Genes are grouped according to their biological annotation.

Although the complexity of the AML genetic landscape, the mutation status of only few genes is considered by the current WHO classification model ^[1]. These include mutations in NPM1, FLT3, CEBPA and a provisional entity represented by RUNX1 mutations. For the remaining subgroups, classification is still based on morphological and immunophenotypic criteria with most entities in this group reminding the old French American British (FAB) classification subtypes ^[1]. Both recurrent cytogenetic and molecular alterations are also considered for prognostication by two current scoring systems: the European LeukemiaNet^[54] and the Medical Research Council ^[55]. Here, mutations in TP53, ASXL1, DNMT3A and partial tandem duplication (PTD) of KMT2A are added as predictors of adverse outcome. However, except for cases with NPM1 mutations and co-occurring FLT3 ITDs, which are associated with intermediate prognosis compared to the favorable prognosis of NPM1 alone and adverse prognosis of FLT3 ITD alone, these models do not consider: i) functional consequences of distinct hotspot mutations; and ii) concurrent mutations that may dramatically alter the contribution of specific AML disease alleles to clinical outcomes. For example, NPM1 mutations preferentially occur with NRAS^G12/13 but not NRAS^Q61 and they generally are associated with favorable outcome ^[22]. In the case of NPM1-IDH2 mutational co-occurrence, IDH2^R140 is significantly associated with NPM1 mutations, while IDH2^R172 is mutually exclusive with NPM1 and other class-defining lesions. This hotspot is associated with a distinct gene expression, methylation and metabolic profiles ^[22]. Moreover, patients who have concurrent mutations of FLT3^ITD, DNMT3A, and NPM1 have a very poor clinical outcome, compared to patients with NPM1 and DNMT3A mutations without FLT3^ITD. In contrast, in pediatric AML, FLT3^ITD and NPM1 mutations co-occur in the absence of DNMT3A mutations in a group of subjects with superior outcomes. However, when FLT3 ITD co-occurs with NUP98-NSD1 or WT1 mutations the prognosis is poor ^[18].

Morphology-based criteria have led to several changes in leukemia classification. This is notable for AEL, especially the historic FAB M6a (AML, not otherwise specified, NOS, erythroid-subtype) that in the latest revision of the WHO classification ^[56] was merged to a hybrid group of MDS or AML, NOS non-erythroid subtype. We have recently shown that AEL is a genetically heterogeneous with six defined age-related genomic subgroups:(1) biallelic TP53 mutations, often with concomitant mutations of chromatin regulators, transcription factors and tumor suppressors, 32%; (2) NPM1 mutations, 12%; (3) KMT2A/MLL mutations or rearrangements, 11%; (4)- NUP98-rearrangements, 4%; (5) DDX41 mutations, 3%; and (6) an MDS-like group with mutations in chromatin regulators and splicing factors, 37%; and overall with marked differences in mutation frequency between AEL and non-AEL AML and MDS ^[36]. For example, AEL has much lower frequency of canonical genes mutated in AML such as FLT3 and NPM1 when compared to non-erythroid AML, but they are more common than in MDS. Conversely, MDS-associated mutations such as SF3B1 and ASXL1 are less frequent in AEL compared to MDS, but more common than in non-erythroid AML. Genomic subgroups but not morphological phenotypes were the strongest predictors of outcome, highlighting the importance of genomic features to properly diagnose and risk-stratify patients ^[36].

Constellations of genetic alterations in MDS

Myelodysplastic syndromes encompass a spectrum of myeloid neoplasms characterized by ineffective hematopoiesis, cytopenia, abnormal cell morphology and a high propensity of progression to AML in 30% of cases (sAML) ^[57,58]. MDS is uncommon in children but it increases markedly with age, with a median age of onset of 71-76 years old ^[59]. It can be de novo or related to prior use of cytotoxic chemotherapy and/or radiation (therapy-related MDS, t-MDS)^[1,2]. The latter is commonly associated with monosomies in chromosome 5 or 7, complex cytogenetics and poor outcome ^[2,60]. Although large DNA-sequencing studies have led to the identification of multiple recurrent mutations ^[60-65], genetic lesions are not used to define MDS subtypes in the WHO classification of myeloid neoplasms ^[1] nor in the current traditional scoring systems, including the international prognostic scoring system (IPSS) and the revised IPSS (IPSS-R) ^[66,67]. The only exceptions are isolated del(5q), which is associated with refractory anemia and normal to increased platelet counts with micromegakaryocytes ^[68], and SF3B1 mutations which have been included as a diagnostic criterion in MDS with ring sideroblasts (MDS-RS) ^[1,69-71]. Mutations in SF3B1 occur in 25% of all MDS cases but affect >80% of MDS-RS and are independent predictors of favorable outcomes ^[71,72]. Mutational targets in MDS overlap those in AML although with a different frequency and prognostic role and include those involved in RNA splicing, epigenetic modification, cohesin complex, transcription, DNA damage response, and signal transduction (Fig. 2C). RNA splicing is the most commonly mutated pathway in MDS, and six genes (SF3B1, TET2, SRSF2, ASXL1, DNMT3A, and RUNX1) are mutated in at least 10% of patients, while the others have a less occurrence and heterogenous patterns. In contrast to adult MDS, Ras/MAPK pathway mutations are common in pediatric MDS (45%), while mutations in RNA splicing genes are rare (2%) ^[73]. Mutation-interactions are not random, but several specific patterns have been described. For example, splicing mutations are almost mutually exclusive of each other, and, similarly, mutations in genes of cohesin complex, due to synthetic lethality mechanisms. By contrast, splicing mutations are significantly associated with mutations in epigenetic modifiers or transcription factors. For example, SRSF2 mutations are significantly associated with mutations in RUNX1, ASXL1, IDH2, CUX1, TET2, RUNX1 and STAG2. Although gene sequencing results are not included in the current prognostication guidelines, several mutation-prognosis associations have been described. Among those, the most concordant across different studies, include association between mutations in TP53, EZH2, ETV6, RUNX1, ASXL1, and SRSF2 and poor overall survival ^{[61-63,65,74,75]}. Notably, somatic mutations can predict overall survival independent of IPSS-R. However, genetic prognostic system may be improved by considering intra-patient co-mutational occurrence, type and number of mutation and variant allele frequency. For example, a recent study in over 3,000 MDS cases showed that MDS patients with monoallelic TP53 mutations did do not differ from TP53 wild-type patients in term of outcome and response to therapy, demonstrating the importance of allelic state in outcome ^[76**]. Two-thirds of TP53-mutated patients had biallelic targeting including more than one gene mutation, mutation and deletion, mutation and copy neutral loss of heterozygosity. These patients had association with genome instability, treatment resistance, disease progression and dismal outcomes, independently of IPSS-R ^[76**].

NOVEL GENETIC AND COMPUTATIONAL APPROACHES

Clonal architecture at the single cell resolution

Single-cell sequencing studies can dissect clonal architecture and identify rare populations important for pathogenesis and response to therapy (Fig. 3). Dr. Peter van Galen and colleagues ^[77] combined single-cell RNA sequencing (Fig. 3A) and genotyping in AML and normal bone marrow cells and identified six malignant cell phenotypes whose abundancy varied between patients and between subclones in the same tumor, as well as across 179 bulk AML profiles from the Cancer Genome Atlas (TCGA) ^[19] queried with cell-type- specific gene signatures. This analysis yielded seven AML groups with distinct cell-type compositions and associated to characteristic genetic lesions. For example, cells with FLT3 tyrosine kinase domain (TKD) mutations were enriched in AML with differentiation, whereas those with FLT3^ITD have higher abundances in primitive stem cells. Interestingly, NPM1 subgroups showed different phenotype according to co-occurrent lesions: a strong stem/progenitor phenotype when co-occurring with FLT3^ITD and a more differentiated monocyte- to dendritic cell–like signatures in FLT3^ITD-negative cases.

Figure 3. Schematic genomic analysis at the single-cell resolution.

A) Single-cell transcriptome sequencing can detect rare populations with distinct expression patterns and important for leukemogenesis. B) Single cell-DNA sequencing can dissect clonal intra patient genetic heterogeneity and identify co-occurrent or mutually exclusive aberrations. C) Representative fish plot graph showing emergence of mutations responsible for relapse. A schematic example is provided with data from [78**]. Each color represents an individual genetic clone. Abbreviations: sc GEX, single-cell gene expression; mut, mutation.

Single-cell mutation data have the power to unambiguously reveal co- occurrence and mutual exclusivity of driver mutations at the cellular level (Fig. 3B). For example, Dr. Morita and colleagues ^[78**] described that when multiple signaling pathway genes (e.g. KRAS, NRAS, FLT3) are present in the same patient, at the cellular level they often occur in mutually exclusive clones. A similar mutually exclusive relationship was observed among other functionally redundant mutations such as IDH1 and IDH2 or in the case of TET2 and IDH1. This may have important therapeutic implications since inhibition of one mutation may favor the expansion of the other clone. A notable group of mutations that display statistically significant mutual exclusivity in MDS and AML is that with mutations in splicing factors, although in very rare cases (< 1%) 2 concomitant splicing factor mutations may be present ^[79**]. By single-cell DNA-sequencing of patient cells with > 1 splicing mutation, Taylor et al ^[79**] showed that escape from this epistasis occurs when there is i) selection for less common alleles, such as SF3B1^non-K700E mutations, that have reduced effects on RNA splicing and/or binding compared with the most common alleles; ii) mutations that occur in cis with preservation of the wild-type allele ^[79**].

Progress in single-cell technologies allows to simultaneously explore genotype-phenotype correlations by analysis of single-cell DNA and cell surface proteins ^[78**,80**]. For example, cells with NPM1 or IDH mutations have been described to express lower levels of CD34 and HLA-DR, while cells with a single TP53 mutations have higher levels of CD34 and CD117, but double TP53 mutations have a monocytic immunophenotype ^[78**,80**].

Moreover, single-cell DNA sequencing studies can dissect the evolution of clonal architecture in response to different therapies (Fig. 3C) ^{[78**,81-83*]}. For example, a selection of a small subclone with FLT3 ^TKD (D835Y) was observed during azacitidine and sorafenib (a FLT3 inhibitor) treatment, which was associated with relapse. Similarly, selection of subclones with NRAS mutation along with the acquisition of PTPN11, FLT3^ITD, and IDH1 mutations were observed during treatment with azacytidine and enasidenib (an IDH2 inhibitor) ^[78**]. Polyclonal emergence of multiple independent kinase activating clones, including FLT3^ITD, FLT3^TKD, and Ras mutations have been described by single cell analysis also in samples who acquired resistance to venetoclax ^[83*]. Overall, these findings show that single-cell sequencing studies by dissecting clonal diversity and revealing evolution patterns have important clinical relevance and it is likely that their use will increase in the clinical management of AML.

Machine learning approach in the diagnosis of AML and MDS

The enormous advances in technology and genomics have generated a very large and heterogeneous volume of data from large cohorts of patients. At the same time, advances in hardware and computing have led to an increasing use of machine learning approaches in medicine. “Machine learning” (ML) is an application of artificial intelligence (AI) that defines a data analysis method that automatically learns from data and experience and make decisions without being explicitly programmed ^[84]. Since both AML and MDS are characterized by high genetic and phenotypic heterogeneity, they represent the good candidates for ML. Warnat-Herresthal et al ^[85*] used a transcriptomic-based ML approach to predict in an unbiased, entirely data-driven manner, genome-wide predictors of AML in 12,029 transcriptome samples from 105 different studies, including AML, MDS, acute lymphoblastic leukemia and healthy individuals. The results provided evidence that without ancillary data or expert input, the combination of large transcriptomic data with ML allows for the development of robust AML classifiers with>99% accuracy. AI-based image analysis has been successfully used in several studies to classify cells on bone marrow aspirates and predicts diagnosis with sensitivity and specificity > 95% ^[86,87]. MDS diagnosis strongly relies on morphological interpretation. ML approaches may overcome the limitation due to variable pathologic evaluations. Recently, Nagata et al ^[88] have used ML to identify patterns of co-occurrence among morphologic features and genomic events in 1079 MDS cases. Novel genotype/ morphology/prognosis associations include STAG2^mut and SRSF2^mut with myeloid dysplasia and ASXL1^mut with megakaryocytic dysplasia. By unsupervised consensus clustering, 5 distinct MDS morphological profiles with unique clinical characteristics were identified, separating patients with different prognoses. Moreover, additional genetic signatures were further classified and associated with specific morphological profiles by Bayesian graphical models and validated in an independent cohort ^[88]. Overall, these findings demonstrate the power of large data sets and computing in identifying phenotype/genotype associations and assisting in primary diagnosis.

CONCLUSION

The catalogue of mutations affecting pathophysiological and clinical features of myeloid malignancies has exponentially grown in the past decade and showed that pathogenesis is much more complex than that suggested by morphology examination alone. Integrated conventional and molecular approaches are required to comprehensively identify all combinations of mutations, guiding in classification, risk assessment and therapy.

KEY POINTS.

• Myeloid leukemia subtypes of prognostic significance are defined by combinatorial mutations, most of which are age dependent.

• Single-cell sequencing studies have dissected clonal heterogeneity in AML and MDS and showed distinct correlations between cell-type compositions and genetic lesions, patterns of mutational co-occurrence and exclusivity and clones associated with therapy resistance.

• Machine learning has been successfully used to identify phenotype/genotype associations and predict diagnosis of AML and MDS.

• Future classification of myeloid malignancies will likely integrate conventional morphological-based, molecular and computational approaches.