High-resolution multi-omics profiling has demonstrated that recurrent AML driver mutations behave not only as static genomic lesions, but also serve as active modulators of the epigenetic landscape. However, many of these findings are derived from model systems or selected patient cohorts, and their generalizability across genetically diverse AML populations remains incompletely established. For example, chronic FLT3-ITD signaling reconfigures chromatin states by selectively enriching activating histone marks at the proliferation-associated genes while eliminating repressive marks at the cytokine-responsive loci. This epigenetic reprogramming effectively maintains active signaling from transcription factors and survival pathways, identifying chromatin modulation as a downstream hallmark of cancer-promoting, enduring FLT3 signaling [7].

The DNMT3A^R882H hotspot mutation is associated with focal DNA hypomethylation, particularly in retrotransposon-rich regions of the genome. This conclusion is supported by integrative epigenomic and transcriptomic analyses performed in DNMT3A-mutant clonal hematopoiesis and AML models, with functional validation using hypomethylating-agent exposure. However, the viral-mimicry-like interferon response should be interpreted as a context-dependent mechanism observed in specific experimental settings rather than a universal feature of all DNMT3A-mutant AML cases. In these models, azacitidine further enhanced retrotransposon derepression, interferon-stimulated gene expression, translation suppression, and apoptosis. The DNMT3A^R882H hotspot mutation is associated with focal DNA hypomethylation, especially at retrotransposon-rich genomic regions [8]. However, rather than being a passive byproduct of enzymatic loss, DNA hypomethylation has been associated with activation of a viral-mimicry–like interferon (IFN) response via exposure to double-stranded RNA in certain experimental contexts. Integrative genomic and transcriptomic analyses, together with functional studies, suggest that DNMT3A-mutant AML cells may exhibit increased sensitivity to hypomethylating agents such as azacitidine, which can further derepress retrotransposons, enhance IFN signaling, and inhibit protein translation, ultimately contributing to apoptosis [9].

In mixed lineage leukemia (MLL)-rearranged AML, SETD1B was identified as a genotype-specific dependency through a CRISPR-tiling screen targeting known H3K4 methylation modifiers in an MLL-rearranged AML model. The study combined functional genetic screening with chromatin profiling and transcriptional readouts, showing that disruption of the SETD1B catalytic SET domain reduced broad H3K4me3 domains, decreased MYC expression, and impaired leukemic proliferation. Because the primary evidence was generated in model systems, broader validation across genetically diverse primary AML cohorts remains necessary. SETD1B maintains broad H3K4me3 domains at oncogenic regions of the genome with high transcriptional output, most notably at MYC. When SETD1B was inhibited, the domains collapsed, MYC was downregulated, the cell cycle was halted, and the fraction of leukemic proliferation was reduced [10]. Therefore, recurrent AML mutations are not mere oncogenic activators at the level of DNA but reshape higher-order chromatin architecture, transcription factor occupancy, and nucleosomal positioning in complex ways that can create vulnerabilities and therapeutic opportunities.

In AML epigenome-wide association studies, DNA methylation patterns can have important mechanistic and prognostic implications; however, while several methylation-based biomarkers have shown prognostic value in individual cohorts, only a subset have been consistently validated across independent patient populations. In pediatric AML, methylation status at several CpG sites, mainly in CD34, HOXA7, and CD96 regions, strongly predicts survival [11]. Each of these genes provides important connections to stem cell identity, developmental transcriptional programs, and functional interactions with immune cells, thereby linking aberrant methylation to both leukemic stemness and immune evasion.

The extent of overexpression of B7-H3 (CD276), a likely product of promoter hypomethylation, was found to be associated with recent TP53 mutations and enrichment of immune-suppressive populations in the bone marrow in AML. B7-H3 overexpression was also associated with enrichment of epithelial–mesenchymal transition–like signatures; however, its role as a clinically actionable biomarker remains exploratory and requires validation in independent cohorts [12]. One of the prominent cytogenetically normal (CN)-AML overexpressed genes for example is CLIC4, also driven by hypomethylation, and associated with inflammatory signaling and the tumor microenvironment [13]. Another proposed prognostic factor, ANP32A, is supported by integrated multi-omics associations; however, its prognostic utility has not yet been consistently validated across large independent AML cohorts [14].

Research into histone modifications highlights the bifunctional nature of epigenetic enzymes related to leukemia. A prominent example is EZH2, the catalytic component of the PRC2 complex. The evidence for EZH2 gain- and loss-of-function effects derives largely from integrative multi-omics analyses combining chromatin, transcriptomic, proteomic, and metabolomic profiling of oncogenic EZH2 mutant models. These studies provide mechanistic insight into differential H3K27me3 deposition and downstream transcriptional consequences, but their direct clinical applicability requires validation in larger AML patient cohorts with defined EZH2 mutation status. EZH2 gain-of-function mutations act to promote H3K27me3 at loci promoting differentiation, thus leading to an epigenetic blockade of differentiation, although these mutations may already subject cells to a stem-like state. In contrast, EZH2 loss-of-function mutations will actively deregulate proto-oncogenes and promote cellular proliferation [15]. These completely opposite results suggest that therapy targeting EZH2 must be tailored to the context of the mutations. Similarly, changes in the recruitment of histone acetylation/deacetylation enzymes may also alter the enhancer landscape, working with transcriptional factor mutations, and stabilize oncogenic circuits.

Epigenetic dysregulation often co-occurs with metabolic reprogramming. For example, metformin rescues teratogenic patterns of DNA methylation and even histone modifications in DNMT3A-mutant hematopoietic stem cells, while simultaneously reducing their oxidative phosphorylation and depleting the fitness of the mutant clone [16,17,18]. In immune cells, TP53 mutations limit cytotoxicity and facilitate immune escape, although p53 reactivation pharmacologically restores effector function and improves survival in AML patient-derived xenografts [19]. Functional screens using CRISPR continue to identify mutations that confer specific dependencies on chromatin-modifying complexes, revealing context-dependent vulnerabilities (Figure 1).

Genomic and epigenomic studies in AML demonstrate that recurrent driver mutations extend beyond static DNA alterations to actively reprogram chromatin architecture, DNA methylation, and transcriptional regulation. These changes create context-dependent vulnerabilities, including altered differentiation states, immune evasion mechanisms, and metabolic dependencies. While large-scale cohort studies and functional genomics approaches have identified clinically relevant biomarkers and therapeutic targets, many findings remain context-specific and require validation across diverse patient populations to ensure generalizability and clinical utility.

Transcriptomic profiling on a massive scale has further defined subtypes of AML, enabling them to be recognized and distinguished more readily than with previous genetic or cytogenetic determinations. One important transcriptomic analysis established a large multicenter AML cohort of 655 patients in China, with RNA sequencing performed in all cases and targeted or whole-exome sequencing performed in 619 patients. Using enhanced consensus clustering, the study identified eight gene-expression subgroups (G1-G8), each associated with distinct lineage programs, transcription factor dependencies, mutational patterns, and prognostic features. Although the cohort size and integrated genomic annotation support the robustness of this classification, prospective validation and harmonization with international AML classification systems are still required before routine clinical implementation [21]. The G1 and G2 subgroups had transcriptional signatures of stem cell–like cells, with high HOX expression and transcription factor activity, including MEIS1 and PBX3, which were associated with poor outcomes. G3 and G4 contained cells with monocytic bias, and were enriched for genes targeting CEBPA and inflammatory cytokines. G5 and G6 contained transcripts associated with erythroid/megakaryoblastic programs, with the gene signatures reflecting features of acute megakaryoblastic leukemia. The G7 and G8 gene sets were related to proliferation and were characterized by common cytogenetic abnormalities associated with adverse risk. Characterization of patients using this transcriptional subtyping framework with prospective mutational data can improve prognostic models and identify patient subgroups who can be sensitized to targeted options. Additionally, single-cell RNA-seq can more fully define any subgroups and explore intra-group variations/heterogeneity, including characterization of potential subclonal states with different sensitivities to standard and novel therapies [21].

Transcriptomic data indicate mechanistic insights when combined with functional genomics. Oxidative stress-related expression scores have been proposed to stratify AML patients into risk categories; however, these models are primarily derived from retrospective datasets and require validation in independent prospective cohorts. Patients with high scores demonstrate enrichment of ROS defense mechanism pathways and chemotherapy resistance, while predicted sensitivity to dasatinib and cytarabine [22,23]. Mutations can implement differential transcriptional rewiring. The IKZF1^N159S mutation was characterized in a cohort of 475 newly diagnosed non-M3 AML patients, among whom 23 carried IKZF1 small sequence variants. RNA-sequencing–based subclassification identified three IKZF1-related AML classes, including nine patients with IKZF1^N159S mutations. These cases showed higher HOXA/B expression and native B-cell immune fractions, suggesting a distinct transcriptional and immune-associated AML subgroup. However, because the N159S subgroup was small, the finding should be considered hypothesis-generating and requires validation in larger independent cohorts [24]. Additionally, aberrant expression of NFATC4 has been associated with increased Treg infiltration and immune checkpoint activation; however, its utility as a predictive or therapeutic biomarker remains to be validated in independent patient cohorts [25]. The non-coding RNAs miR-155, miR-222, miR-424, and miR-503 act in concert to facilitate monocytic differentiation by targeting lineage-inhibitory factors [26]. Interestingly, retroelement expression from HERV-K9 is associated with improved prognosis, potentially through immune activation via IFN signaling [27].

Alternative splicing (AS) is an important driver of functional diversity within AML. Integrated splicing-mutational-immune profiling classified AML into four splicing regulation patterns associated with differences in immune function, tumor mutation profiles, signaling pathway activity, prognosis, and predicted treatment sensitivity. The study constructed and validated a splicing-related risk score as an independent prognostic factor using retrospective transcriptomic datasets. Therefore, while the analysis provides a useful computational framework for linking alternative splicing with AML biology, prospective validation and functional confirmation of the predicted therapeutic vulnerabilities remain necessary [28]. For example, SRSF2-mutant AML demonstrates immune-suppressive myeloid infiltration, while other AS patterns with higher spliceosomal protein levels are associated with improved outcomes.

RBM17 is expressed at high levels in leukemic stem cell–enriched AML populations and prevents inclusion of “poison exons” that trigger nonsense-mediated decay, thereby sustaining pro-leukemic proteins such as EIF4A2. This conclusion was supported by integrative multi-omics analyses combined with functional perturbation experiments, including RBM17 repression, splicing analysis, proteomic validation, differentiation assays, and clonogenic assays. Although these data provide strong mechanistic support for RBM17 as a leukemic stem cell dependency, therapeutic translation will require pharmacologic targeting strategies and validation in larger primary AML cohorts [29]. Dysregulation of AS also alters oncogenic pathways like PI3K-AKT, resulting in malignancy-specific spliced variants that could serve as neoantigens or therapeutic targets (Figure 2).

Transcriptomic and splicing analyses provide a central integrative layer in AML, capturing the combined effects of genomic mutations, epigenetic regulation, and microenvironmental signaling. Large-scale RNA-sequencing studies have refined AML classification into biologically and clinically meaningful subgroups, while single-cell and splicing analyses reveal intra-tumoral heterogeneity and regulatory mechanisms such as alternative splicing and non-coding RNA activity. Despite these advances, challenges remain in standardizing classification frameworks and achieving consistent validation across independent cohorts, limiting immediate clinical translation.

A landmark proteogenomic study analyzed bone marrow biopsies from 252 uniformly treated AML patients, integrating in-depth quantitative mass-spectrometry proteomics with cytogenetic profiling, mutation profiling, and RNA sequencing. This design enabled the identification of proteogenomic AML subtypes linked to metabolic states, DNA repair programs, and kinase signaling. Importantly, the uniformly treated cohort enhances clinical interpretability, but the technical complexity of mass-spectrometry workflows and the need for independent prospective validation remain barriers to immediate clinical implementation [30]. There were proteomic clusters contingent on mitochondrial oxidative phosphorylation and enriched for DNA repair proteins, suggesting these as possible therapeutic vulnerabilities. Phosphoproteomics revealed subtype-specific activated kinases, e.g., CDK and MAPK hyperactivity in proliferative AML. In murine AML, combined transcriptome-proteome analyses confirmed mitochondrial metabolic remodeling and pointed to TCA cycle and fatty acid oxidation enzymes as drug targets [31].

Proteomic studies have also mapped drug resistance networks in AML. For example, Fierro et al. mapped the interactome and substrates of the ubiquitin ligase WWP1 and identified the histone demethylase JARID1B (KDM5B) as a major target stabilized through K63-linked polyubiquitination [32]. JARID1B removes the H3K4me3 marks at DNA damage repair-related genes that aid in the chromatin configurations needed for repair and survival after chemotherapy. Functional perturbation studies indicated that cellular inhibition of WWP1 destabilized JARID1B and led to the loss of chromatin occupancy, globally increased H3K4me3 marks at JARID1B target promoters, reduced the recruitment of the DNA repair machinery to sites of damage, and extensively sensitized AML cells to the standard chemotherapeutics like cytarabine. The WWP1-JARID1B axis demonstrates the importance of post-translational control of chromatin modifiers as a major factor in chemoresistance and provides an example of how proteomic mapping of ubiquitination can expose druggable resistance mechanisms [32]. Despite these advances, proteomic studies in AML remain limited by relatively small cohort sizes, technical variability, and challenges in cross-platform reproducibility, which currently constrain their routine clinical application.

Proteogenomic profiling is also broadening the landscape of tumor-associated antigens for immunotherapy. A recent AML immunopeptidomics study integrated mass-spectrometry–based identification of MHC-I-presented peptides with transcriptomic and computational neoantigen-prioritization frameworks to characterize both canonical and non-canonical MHC-associated peptides. The study prioritized 13 candidate neoantigens, six of which were derived from non-canonical MHC-associated peptides, and further developed a prognostic risk model. These findings expand the antigenic landscape of AML, but clinical translation will require immunogenicity testing, HLA-diverse validation cohorts, and functional confirmation in T-cell–based assays [33]. Interestingly, a large proportion of ncMAPs showed higher predicted immunogenicity and AML specificity than traditional mutation-derived neoantigens. This is interesting as ncMAPs are a potentially more pipeable source of targets for T cell–based therapies.

Simultaneously, cellular engineering approaches are utilizing proteomic data derived from immune evasion. In one study, the inhibitory receptor NKG2A was targeted on CD33-directed CAR-NK cells. AML inhibits NKG2A–expressing NK cytotoxicity via the non-classical MHC molecule HLA-E. In this study, when comparing cytotoxicity against HLA-E negative and HLA-E positive AML lines, CRISPR-Cas9 knockout of NKG2A in CAR-NK cells effectively eliminated this checkpoint, ultimately leading to a statistically significant increase in anti-leukaemic activity in vitro and in the xenotransplanted model [34]. This is an example of how proteogenomic mapping of ligand–receptor interactions can directly inform the design of immunotherapy.

Post-translational modifications, such as phosphorylation, are dynamic regulators of signaling networks and phosphoproteomic profiling in AML has identified kinase-driven oncogenic circuits with prognostic and therapeutic implications. For example, increased PAK1 kinase activation was identified using phosphoproteomics and was associated with poor prognosis in primary AML cases. Functional validation was performed using PAK inhibitors, including PF-3758309, in AML cell lines and primary AML samples, where treatment reduced proliferation and induced apoptosis. These data support PAK1 as a candidate signaling vulnerability; however, the evidence remains preclinical, and clinical applicability will require validation in larger patient cohorts and assessment of inhibitor selectivity, toxicity, and biomarker-guided response [35].

Functional proteomics identified the IFN-γ-inducible lysosomal thiol reductase (GILT), which encodes a modulator of redox homeostasis in chemoresistant AML. Targeting GILT caused an increase in mitochondrial ROS levels, oxidative damage, and cell death. This was found to synergize with conventional chemotherapeutic agents, representing a ROS-priming strategy to prevent drug resistance [36]. These studies not only link phosphoproteomic signatures to drug-sensitivity assays and enable more precise kinase targeting, but they also match patients to inhibitors that have the highest probability to disrupt their dominant signaling dependencies (Figure 3). Proteomics and proteogenomics provide a functional bridge between genomic alterations and cellular phenotypes in AML by capturing protein abundance, post-translational modifications, and signaling network activity. These approaches reveal subtype-specific kinase activation, metabolic dependencies, and mechanisms of chemoresistance that are not evident at the transcriptomic level. However, their clinical application is currently limited by technical complexity, lack of standardization, and relatively modest cohort sizes. Integration with other omics layers and prospective validation will be critical for translating proteomic insights into actionable therapeutic strategies.

Microbiome-focused multi-omics studies have identified severe perturbations in the gut ecosystem during AML therapy, which have significant implications for patient immunity, metabolism, and overall treatment tolerability. In a longitudinal AML induction-chemotherapy cohort, investigators compiled amplicon sequencing data from 566 fecal samples obtained from 68 AML patients and linked these data to serum metabolomics from 260 samples collected from 36 patients, together with clinical metadata. The study validated technical quality through example analyses and provided a resource for microbiome–metabolome–clinical integration. However, interpretation of these data requires caution, as the microbiome and metabolome profiles in AML are strongly influenced by antibiotics, chemotherapy timing, hospitalization, diet, mucosal injury, and supportive-care practices [42]. This resource facilitates research into the relationships between changes in microbial composition, systemic metabolic changes, treatment response, and adverse events. The analyses revealed that therapy-related perturbations in the microbiome tend to occur very early in therapy, often leading to a loss of microbial diversity before patients experience clinical indicators of gastrointestinal dysfunction.

To support this, additional work has shown that intensive chemotherapy led to persistent dysbiosis, characterized by loss of members with beneficial commensal activity and an increase in opportunistic species that have the potential to translocate across a damaged intestinal barrier [43]. Alongside, metabolomic readouts of these patients also reflected the loss of metabolites of health importance, like short-chain fatty acids and aromatic amino acid derivatives, both of which are critical to maintaining epithelial integrity and regulating inflammatory responses. This microbiome-metabolome disruption was also associated with systemic inflammation and weight loss, contributing to cachexia. Such highlights support the therapeutic strategy of microbiome-sparing interventions, which could ameliorate toxicity and enhance resilience to treatment [41].

The rise of multi-omic technologies at single-cell and spatial levels has transformed the understanding of AML by providing highly detailed insights into the heterogeneity of leukemic populations and features of the bone marrow microenvironment at single-cell resolution. In a spatial proteo-transcriptomic study that profiled bone marrow from patients with leukemia, it was found that besides blasts, the number of mesenchymal stromal cells expanded considerably and that rather than leukemia presenting randomly throughout the bone marrow, discrete blast–stromal niches were formed that facilitated only leukemia survival and immune evasion. The niches represent organized tissues that influenced the penetration and response of therapy, and not groups of cells that randomly existed [44].

The application of clonal lineage tracing through the CloneTracer platform enhanced the ability to coordinate genetic identity with human cellular behavior, reconstruct differentiation hierarchies, and identify aberrant surface markers associated with drug response. Methodologically, CloneTracer was applied to samples from 19 AML patients, thereby providing clonal resolution to single-cell RNA-sequencing data. The study showed that residual healthy and preleukemic cells dominated the dormant stem-cell compartment, whereas active leukemic stem cells and downstream myeloid progenitors showed disease-defining transcriptional states. Although this approach provides high-resolution biological insight, the modest patient number and complexity of single-cell inference mean that validation in larger longitudinal cohorts is needed [45]. High-throughput capsule-based single-cell multi-omics is a fast and effective approach to molecularly profile many different molecular layers (genome, transcriptome, and epigenome) simultaneously in many thousands of individual cells, providing greater statistical power to study heterogeneity [46].

Longitudinal single-cell work in t(8;21) AML has demonstrated dynamic changes in immune checkpoint gene expression in the leukemic population, which suggests that the timing of immunotherapy is important [47]. Single-cell multi-omics studies of TP53-mutant leukemic evolution indicate that chronic inflammatory signaling can selectively favor expansion of TP53-mutant hematopoietic clones while suppressing wild-type progenitors. This evidence was generated using single-cell multi-omic profiling and experimental modeling of inflammatory states, supporting a non-genetic route of clonal selection. Nevertheless, because the mechanism is most directly supported in TP53-mutant contexts, it should not be generalized to all AML evolutionary trajectories without further validation [48]. Integration of advanced multi-omics approaches provides a more comprehensive picture of AML biology, therapeutic resistance, and immune interactions, uunderscoring the importance of spatial and temporal context when considering treatment strategies in the disease.

Machine learning approaches have emerged as tools to integrate multi-omics datasets into clinically actionable frameworks. In parallel, functional precision medicine strategies, such as prospective tumor board-guided studies integrating genomic profiling with ex vivo drug testing, have demonstrated the feasibility of translating multi-omics data into individualized therapeutic recommendations in AML, further bridging the gap between omics discovery and clinical implementation. The MDREAM ensemble model integrated mutational profiles, gene-expression signatures, and large-scale ex vivo drug-screening data from AML resources to predict patient-specific drug responses. Its validation relied on external cohorts and performance assessment against observed ex vivo responses rather than prospective treatment assignment. Therefore, although MDREAM demonstrates the potential of machine learning for functional precision medicine, prospective clinical validation is still required before such models can guide therapy selection in routine AML care [49]. Generative adversarial network-based frameworks have also been successfully applied to harmonizing drug sensitivity profiles and disease phenotype as different multi-omics data modalities [50]. Besides predictive approaches, network-based integration methods are also being used to identify hidden relationships among molecular features. Multilayer network survival models consider various omic layers to examine prognostically relevant molecular interactions [51]. The Integrative Network Fusion approach that takes advantage of similarity networks of each layer of omics data generates an integrated network that improves biomarker discoverability and patient stratification [52]. Resources such as AMLdb now aggregate multi-omics data with functional screening results, providing a platform for biomarker discovery and hypothesis generation [53].

Nevertheless, machine learning-based integration approaches face challenges, including overfitting, limited external validation, and the need for large, harmonized datasets, raising concerns about their robustness and generalizability in real-world clinical settings.

Studies using immune-mediated multi-omics have further defined the molecular mechanisms of immune evasion in AML and identified additional immunotherapy targets. In post–hematopoietic stem cell transplant relapse, integrated multi-omic profiling of paired diagnosis and relapse AML samples identified loss of HLA class II expression as a mechanism of immune escape associated with PRC2-mediated epigenetic repression. The study combined transcriptomic and epigenomic analyses with functional validation, showing that pharmacological PRC2 inhibition restored HLA-II expression and improved T-cell recognition of leukemic cells. This provides mechanistic and translational support for PRC2 targeting, although clinical benefit remains to be tested prospectively [54]. This mechanistic understanding is the basis for the addition of epigenetic therapies in post-transplant relapse.

In a phase II clinical trial of hypomethylating-agent therapy combined with nivolumab for AML relapse after allogeneic hematopoietic cell transplantation, single-cell immune monitoring identified immune signatures associated with response. Responders showed activated, less senescent CD8⁺ T-cell states and pro-inflammatory transcriptional programs in T-cell and myeloid compartments. Because this evidence derives from a phase II trial with correlative immune profiling, these signatures should be interpreted as candidate response biomarkers requiring validation in larger prospective studies [55]. Further studies have shown that FLT3-mutant AML harbors unique immune features that predict sensitivity to FLT3-targeted inhibitors [56]. From this integrated analysis, other immune targets, like IL1RAP in triple-mutant AML [57], are added as more potential antigens for antibody- or CAR-therapy. Large pan-cancer immunogenomic studies further identified FCN1 as an immune prognostic marker with both direct effects on leukemic proliferation and immune modulation, serving as another example of the complex relationship between innate immune pathways and leukemic cell viability (Table 1) [58].

Emerging multi-omics approaches, including microbiome-linked analyses, single-cell and spatial profiling, and machine learning–based integration, provide high-resolution insights into AML heterogeneity, microenvironmental interactions, and therapeutic response. These methods reveal dynamic ecosystem-level changes, including immune evolution, clonal architecture, and niche-specific signaling. However, their translation into clinical practice is constrained by small cohort sizes, technical complexity, computational demands, and limited prospective validation. Continued development of standardized pipelines and clinically interpretable models will be essential to harness their full potential in precision medicine (Table 2).

One interesting outcome with potential clinical implications is the uncommon t(8;16)(p11;p13) rearrangement/MYST3-CREBBP fusion that is exceptionally rare, aggressive, with significant disease biology and poor outcomes after standard-dosing chemotherapy [101]. Using multi-omic approaches, it is suggested that this fusion disrupts transcriptional regulation through histone acetyltransferase activity in the fusion protein and widespread epigenetic deregulation, increasing the potential for leukemogenesis. The clinical observations show that these patients tend to do better if they undergo early allogeneic hematopoietic stem cell transplantation in first complete remission, implying that early recognition via genomic screening of the rearrangement can drive treatment decisions.

Microenvironmental remodeling is another domain where multi-omics is producing clinically relevant insights. Stromal transcriptional profiling shows that bone marrow niche remodeling occurs during the earliest stages of AML development, even prior to the overt presentation of hematological abnormalities [99]. The remodeling is reflected by the altered composition of the extracellular matrix, alterations in cytokine secretion, and expression of adhesion molecules, which together create a pro-leukemic niche. The notion that microenvironment-targeted therapies may circumvent leukemogenic conditioning before the establishment of fully malignant disease is intriguing, and such associations might be leveraged to improve clinical outcomes in high-risk settings where pre-leukemic clonal hematopoiesis exists.

In pediatric AML, NUP98 rearrangements have emerged as a consistent indicator of poor prognosis. Through multi-omics, NUP98 fusions also appear to be associated with dysregulated transcriptional programs involving HOX gene clusters, chromatin remodeling factors, and stemness signaling pathways. It is possible that the prognostic power of this rearrangements warrants its inclusion in pediatric AML therapy risk stratification system to potentially choose more aggressive initial therapy, perhaps even early transplant, for patients with NUP98 fusions [100].

Molecular biomarkers are also advancing for prognostic evaluation in CN-AML. BAALC (brain and acute leukemia cytoplasmic) gene overexpression is associated with poor survival and, mechanistically, with erroneous activation of MYC and RAS signaling pathways [94]. High BAALC expression was proposed as a candidate criterion for intensifying induction or consolidation regimens for patients with CN-AML. Finally, immune epigenetics offers a novel perspective in addressing post-transplant relapse. Investigations of immune escape mechanisms have identified that PRC2-dependent repression of HLA class II genes impairs T-cell recognition of leukemic blasts [54]. Pharmacologic inhibition of PRC2 can revive HLA-II expression and T-cell-mediated cytotoxicity and ultimately represents a new approach to treat patients with recurrent disease following allogeneic transplantation. This mechanistic insight is particularly valuable at a time of personalized immunotherapy in which the unraveling of antigen presentation may enhance synergy with checkpoint blockade or adoptive T-cell approaches [118].

Despite the rapid expansion of multi-omics studies and their strong emphasis on clinical translation, several practical barriers continue to limit their integration into routine AML care. These challenges extend beyond biological discovery and relate to cost, infrastructure, reproducibility, scalability, and regulatory considerations. Many of the multi-omics studies described in this review rely on resource-intensive platforms, including high-depth sequencing, mass spectrometry–based proteomics, and single-cell or spatial technologies. For example, large-scale proteogenomic profiling studies involving hundreds of AML patients require complex sample preparation, advanced instrumentation, and substantial computational resources [30]. Similarly, single-cell and spatial multi-omics approaches, while highly informative, involve high per-sample costs and specialized workflows that limit scalability across clinical centers [44,45]. These financial and technical constraints currently restrict widespread clinical implementation.

Multi-omics approaches require integrated laboratory and computational infrastructure, including sequencing platforms, high-resolution mass spectrometers, and bioinformatics pipelines capable of handling large, high-dimensional datasets. While consortia-based efforts and specialized research centers can support such infrastructure, many clinical institutions lack the resources or expertise to do so. For instance, integrative frameworks such as proteogenomic profiling [30], microbiome–metabolome datasets [42], and machine-learning–based models like MDREAM [49] depend on coordinated multi-platform data generation and advanced computational integration, which are not yet standardized in routine clinical environments.

A major limitation highlighted across the cited studies is the lack of standardized protocols for sample collection, processing, sequencing depth, and data analysis. Variability across platforms, particularly in proteomics, metabolomics, and single-cell analyses, can lead to inconsistent results between studies. For example, proteomic studies in AML have identified important signaling and metabolic subtypes, but differences in mass spectrometry workflows and analytical pipelines limit cross-study reproducibility [30,68]. Similarly, transcriptomic and splicing-based classification systems show discrepancies across cohorts, reflecting differences in dataset composition and analytical methods [21,28].

Many multi-omics findings are derived from retrospective datasets, experimental models, or relatively small patient cohorts, with limited prospective validation. While large datasets such as TCGA and BeatAML provide foundational insights, numerous studies cited in this review rely on functional assays in cell lines, xenograft models, or computational validation approaches rather than clinical trials. For example, kinase dependencies identified through phosphoproteomics [35], metabolic vulnerabilities such as serine auxotrophy [38,119,120], and splicing factor dependencies like RBM17 [29] have strong mechanistic support but remain largely preclinical. Similarly, machine learning models such as MDREAM have demonstrated predictive performance using retrospective and ex vivo validation but have not yet been widely validated in prospective clinical settings [49].

Translating multi-omics into routine AML care requires rapid turnaround times and clinically interpretable outputs. However, current multi-omics workflows often involve prolonged data processing and complex interpretation, which may not align with the urgent treatment timelines in AML. Furthermore, scaling multi-layer data integration for large patient populations remains challenging due to computational demands and a lack of standardized reporting frameworks. Even large-scale databases such as AMLdb primarily function as research tools rather than clinical decision-support systems [53].

Clinical implementation of multi-omics approaches requires compliance with regulatory standards for diagnostic accuracy, reproducibility, and clinical utility. However, regulatory frameworks for multi-parameter, high-dimensional omics assays remain underdeveloped. Challenges include validating composite biomarkers, interpreting complex datasets, and ensuring reproducibility across laboratories. In addition, ethical considerations such as data privacy, management of incidental findings, and equitable access to advanced diagnostics must be addressed before widespread adoption [47,120].

Addressing these barriers will require coordinated efforts to standardize experimental and computational pipelines, reduce costs through technological innovation, and design large-scale prospective validation studies. Integration of multi-omics data into user-friendly clinical decision-support systems, combined with regulatory harmonization, will be essential to translate these advances into precision medicine for AML [48,56].