Identifying treatment response in AML through a multi-omics classifier of malignant cell states

Abstract Background: Therapy resistance in Acute Myeloid Leukemia (AML) remains a major clinical challenge, yet its underlying cellular and molecular mechanisms are poorly understood. We developed CSCO-AML (Classification System based on Cellular and Omics profiling in AML patients), a single-cell-based classifier to define malignant progenitor cell states and identify the key characteristics that determine whether patients will respond to induction therapy. Methods We profiled 306 bone marrow samples from 198 newly diagnosed (Dx) AML patients, including 46 non-responders (NR), 48 complete remissions (CR), and 14 relapse (RE) cases. Multi-omics profiling was performed, including single-cell RNA sequencing (scRNA-seq) on 3,283,568 cells, as well as whole-genome sequencing (WGS) and whole-genome bisulfite sequencing (WGBS). Clinical data and omics features were integrated to train a machine-learning model aimed at predicting the most effective therapeutic strategies for individual patients. Results PCA revealed significant heterogeneity across leukemic and immune cell subpopulations. Notably, Dx-NR (NR at diagnosis) showed an increased abundance of stem-like malignant progenitor cells (HSC-L, MPP-L, LMPP-L) and a globally hypermethylated genome, indicating a chemo-resistant, stemness-associated phenotype compared to Dx-CCR (CR at diagnosis) patients. To further explore this, we performed unsupervised clustering of scRNA-seq data to define functional subtypes of malignant progenitor cells. Focusing on HSC-L, MPP-L, and LMPP-L cell types, we identified four distinct subgroups within each. These subgroups exhibited strong clustering behavior, and were subsequently consolidated into four major groups, designated as CSCO-G1 through G4. CSCO-G1/G3 were predominantly enriched in Dx-NR and NR samples, exhibiting immune-silent characteristics, accompanied by activation of immune evasion pathways. In contrast, CSCO-G2/G4 were enriched in Dx-CCR samples, showing immune activation and lower metabolic stress, which were consistent with more favorable therapeutic responses. The identification of these four distinct groups highlights the functional relevance of AML progenitor cell states in determining therapeutic outcomes. To further investigate these subpopulations, we conducted transcriptomic profiling of CSCO groups and identified a distinct gene expression signature elevated in CSCO-G1/G3 that was significantly associated with adverse clinical outcomes (p < 0.0005, TCGA), leading to their classification as the Unfavorable Type. This gene signature were associated with ferroptosis regulation and lipid metabolism, marked by pronounced upregulation of CD36, GPX4, and SLC27A4. In contrast, transcriptomic profiling of CSCO-G2/G4 revealed a distinct gene expression signature significantly associated with favorable prognosis (p < 0.05, TCGA), leading to its classification as the Favorable Type，and demonstrated opposite trends in ferroptosis and lipid metabolism pathways. These results highlight the ability of the CSCO-AML to delineate biologically and clinically distinct AML subtypes with unique molecular programs. Notably, CSCO-AML not only captures malignant progenitor states but also integrates features of immune dysfunction. We independently analyzed the tumor immune microenvironment (TME) characteristics of the 306 samples and stratified the TME into five subtypes. The Unfavorable Type exhibited an immunosuppressive profile with exhausted CD8⁺ T cells and enhanced immune evasion signaling (e.g., NECTIN2–TIGIT). Moreover, CSCO clustering revealed distinct cell-state programs linked to response to Venetoclax (VEN)-based induction therapy. ​​Leveraging these molecular signatures, we constructed a robust prognostic classifier demonstrating excellent predictive accuracy (AUC = 0.95, 95% CI: 0.878-0.995) in clinical outcome stratification. In VEN-CR patients, 87% of predicted regimens matched clinical treatment. For VEN-NR patients, 15% of predictions aligned with clinical choices, suggesting that 85% of NR patients may benefit from alternative therapies. Conclusion CSCO-AML enables functional stratification of malignant progenitor states, links immune-metabolic programs to resistance, and offers a predictive tool for personalized therapy in AML.