Abstract
Acute myeloid leukemia (AML) is a heterogeneous cancer, making outcomes prediction challenging. Several predictive and prognostic models are used but have considerable inaccuracy at individual level. We tried to increase prediction accuracy using a multi-omics strategy. We interrogated data from 1391 consecutive, newly diagnosed subjects with AML, integrating information on mutation topography, DNA methylation, and transcriptomics. We developed an unsupervised multi-omics classification system (UAMOCS) with these data. UAMOCS provides a multidimensional understanding of AML heterogeneity and stratifies subjects into three cohorts: (a) UAMOCS1 [high lymphocyte activating 3 (LAG3) expression, chromosome instability, myelodysplasia-related mutations]; (b) UAMOCS2 (monocytic-like profile, immune suppression and activated angiogenesis and hypoxia pathways); and (c) UAMOCS3 [CCAAT enhancer binding protein alpha (CEBPA) mutations and MYC pathway activation]. UAMOCS distinguishes overall survival rates across the cohorts (TCGA P = 0.042; GSE71014 P = 0.043; ihCAMs-AML, GSE102691 and GSE37642 all P < 0.001). The model's C-statistic is comparable to the 2022 ELN risk classification (0.87 vs 0.82; P = 0.162), but offers a more nuanced distinction between intermediate- and high-risk groups. When combined with high-throughput drug sensitivity testing, UAMOCS can accurately predict sensitivity to azacitidine (AZA) and venetoclax. The UAMOCS system is available as an R package. The UAMOCS system has the potential to redefine AML subtypes, enhance prognostic predictions, and guide treatment strategies based on patients' immune status and expected responses to therapies.