Résumé
Background: Acute myeloid leukemia (AML) is a complex disease marked by a multitude of genetic mutations and dysregulated gene expression profiles stemming from genetic and epigenetic alterations. The disease is marked by significant biological heterogeneity, which complicates effective treatment and contributes to the emergence of drug resistance. The molecular profiles of gene mutations, chromosomal anomalies within AML considerably impacts disease progression, therapeutic response, and survival prognosis. Several predictive and prognostic models are used but present inaccuracy at individual level. Integrating multiple layers of biological data, including transcriptomic analyses (gene expression, mutational profiling, gene fusion identification) and population-level characterization through deconvolution and multiparametric flow cytometry, offers a powerful approach to better understand the complexity of AML and improve prognostic prediction and guide treatment strategies.
Methods: We used a multi-omics data integration approach, including bulk RNAseq-based gene expression (n=45), single nucleotide variant and gene fusion data from bulk RNA sequencing (n=45), deconvolution of tumor microenvironment immune subtypes (n=45), immune subtype identification and blast characterization using multiparametric flow cytometry (MFC) (n=36), in vitro drug response (Venetoclax, n=23; 5-Azacytidine, n=23; Midostaurine, n=21, Gilteritinib, n=22), and relevant clinical metadata. This allowed us to thoroughly characterize tumor and its tumor microenvironment.
Results: Applying Multiomics Factor Analysis (MOFA) identified 9 principal factors capturing the diversity of the disease. Factor 1 was significantly associated with Venetoclax response (p<0.05). Factors 2 and 4 showed significant associations with NPM1 mutations and DNMT3A mutations, while Factors 6 and 9 were significantly associated with 5-Azacytidine response (p<0.05). Factor 9 was also significantly associated with KIT mutations (p<0.05). The correlation analysis revealed strong associations between factors and key clinical covariates including timepoint and drug response groups.
Next, we performed unsupervised clustering using the latent MOFA factors and identified four distinct clusters. Cluster 1 (C1, n=14) predominantly consisted of diagnostic samples (79%) with balanced immune infiltration. Cluster 2 (C2, n=13) showed moderate monocyte enrichment. Cluster 3 (C3, n=10) was characterized by significantly elevated regulatory T cell infiltration (p=0.017 vs C4, p=0.023 vs C1) and higher myeloid dendritic cell populations (p = 0.023 vs C2, p = 0.0051 vs C4), confirmed by both deconvolution (p=0.023 vs C2, p=0.0051 vs C4) and multiparametric flow cytometry analysis, consistent with an immunosuppressive microenvironment in this relapse-enriched cluster (60% relapsed samples). Cluster 4 was significantly enriched in monocytes (p < 0.001 vs C1 and C3).
Finally, therapeutic response profiling across molecular clusters revealed distinct patterns of drug sensitivity. C1 demonstrated favorable responses to Venetoclax, Midostaurine, and Gilteritinib. C2 exhibited poor responses to all four tested drugs, indicating a globally resistant phenotype. C3 showed resistance specifically to Midostaurine and Gilteritinib, while remaining potentially responsive to other agents. Interestingly, C4 was characterized by poor sensitivity to Venetoclax and 5-Azacytidine but responded well to Midostaurine and Gilteritinib, highlighting subtype-specific vulnerabilities that could inform personalized therapeutic strategies. Interestingly, Gilteritinib demonstrated significant activity in specific molecular subgroups (C1 and C4) independently of FLT3-ITD status, suggesting broader therapeutic potential beyond canonical FLT3-driven AML. This unexpected sensitivity, possibly linked to off-target effects such as AXL inhibition, highlights novel vulnerabilities revealed by our integrative multi-omics framework.
Conclusion: Our comprehensive multi-omics MOFA framework identified four clinically distinct AML subtypes with distinct therapeutic vulnerabilities, immune microenvironment characteristics, mutational profiles, and blast immunophenotypes. Our systematic approach demonstrates the potential of multi-omics integration for advancing personalized medicine in AML and provides a framework that will benefit from validation with larger cohorts to improve risk stratification personalized treatment approaches.