Epigenomic diagnosis and prognosis of Acute Myeloid Leukemia

Abstract

Despite the critical role of DNA methylation, clinical implementations harnessing its promise have not been described in acute myeloid leukemia. Utilizing DNA methylation from 3314 leukemia patient samples across 11 harmonized cohorts, we describe the Acute Leukemia Methylome Atlas, which includes robust models capable of accurately predicting AML subtypes. A genome-wide prognostic model as well as a targeted panel of 38 CpGs significantly predict five-year survival in our pediatric and adult test cohorts. To accelerate rapid clinical utility, we develop a specimen-to-result protocol that uses long-read nanopore sequencing and machine learning to characterize patients' whole genomes and epigenomes. Clinical validation on patient samples confirms high concordance between epigenomic signatures and genomic lesions, though uniquely rare karyotypes remained challenging due to limited available training data. These results unveil the potential for increased affordability, speed, and accuracy for patients in need of complex molecular diagnosis and prognosis.

Fig. 1. The Acute Leukemia Methylome Atlas Study.

a Overall design of the computational pipeline. b Acute Leukemia Methylome Atlas (ALMA) defined by a dimension reduction algorithm, PaCMAP, in the overall population. Each data point is a patient sample deriving from bone marrow or peripheral at diagnosis, relapse, or remission. As an unsupervised model, only 5mC values of 331,556 CpGs were used to define the clusters. c Clinical implementation of a long-read nanopore sequencing pipeline. Discovery (training) raw data analyzed in this study were obtained from Gene Expression Omnibus (GEO) under accession codes GSE190931, GSE124413, GSE133986, GSE159907, GSE152710, GSE49031, GSE147667, as well as from Genomic Data Commons (GDC) under categories GDC-TARGET-AML, GDC-TCGA-AML, GDC-TARGET-ALL. AML acute myeloid leukemia, ALL acute lymphoblastic leukemia, MDS myelodysplastic syndrome, MPAL mixed phenotype acute leukemia, APL acute promyelocytic leukemia, PaCMAP pairwise controlled manifold approximation, WHO World Health Organization, PB peripheral blood, BM bone marrow, EDTA Ethylenediaminetetraacetic acid, HMW gDNA high molecular weight genomic DNA. Panels (a and c) were created in BioRender. Marchi, F. (2025) https://BioRender.com/786kvbb.

Fig. 2. Development and benchmarking of ALMA Subtype.

a Diagram of the study design for ALMA and ALMA Subtype. b Sample count by WHO 2022 Diagnosis subtypes in train (n = 2471) and (c) test (AML02, AML08, and NOPHO AML; n = 180). Multi-class confusion matrices with normalized values for each possible prediction, along with the number of samples available, are displayed. Not confident: < 50% confidence for a subtype prediction.

Fig. 3. ALMA Subtype classification according to WHO 2022 subtypes.

Sankey diagrams displaying discovery cohort comparison between WHO 2022 diagnosis and ALMA Subtype for samples with (a) or without (d) WHO 2022 clinical annotation available. The width of the bands indicates the number of patient samples in each category. Same comparison applied to test cohorts (b, e) AML02,08 (n = 200) and (c, f) NOPHO AML (n = 142), describing samples with or without the WHO 2022 annotation available. Individual n numbers are indicated in the figures.

Fig. 4. Patient outcomes and multivariate analyses of AML Epigenomic Risk groups.

Patient outcomes by AML Epigenomic Risk groups with in MRD1 subgroups in (a) the discovery cohort and (b) the AML02,08 test cohort with EFS (top) and OS (bottom) of AML Epigenomic Risk^high (orange) and AML Epigenomic Risk^low (blue). Multivariate analysis adjusting for other confounding variables of OS in (c) discovery cohort (HR = 3.86; P < 0.0001) and (e) AML02,08 test cohort (HR = 2.83; P = 0.0032). Analysis for EFS was also performed in (d) the discovery cohort (HR = 1.78; P < 0.0001) and (f) the AML02,08 test cohort (HR = 2.65; P = 0.0009). Hazard ratios derive from Cox PH regression with two-sided hypothesis tests. Individual n numbers are indicated in the figures. MRD 1 minimal residual disease at end of first induction, FLT3 ITD FMS-like tyrosine kinase-3 internal tandem duplication, CI confidence interval, Ref. reference.

Fig. 5. Patient outcomes by Epigenomic Risk groups. within MRD1 positive negative groups.

Box plots show abundance of MRD1 positive and negative patients within AML Epigenomic Risk groups (chi-squared test; p < 0.0001 and p = 0.0330 for discovery and test, respectively. OS by MRD1 positive and MRD1 negative groups and OS within MRD1 positive and within MRD1 negative groups by AML Epigenomic Risk groups in (a) the discovery cohort and (b) AML02,08 test cohort. MRD 1, minimal residual disease at the end of the first induction. Only patients with MRD assessments were included in this analysis, which excludes patients who had events prior to the end of induction 1. Hazard ratios derive from Cox PH regression with two-sided hypothesis tests. Individual n numbers are indicated in the figures.

Fig. 6. Development and testing of 38CpG AML signature.

a Stepwise workflow for generating the 38CpG-AMLsignature model, including data preprocessing, batch correction, survival analysis, and penalized Cox PH modeling, resulting in a concise 5mC-based prognostic risk score for AML. b Manhattan plot showing risk-adjusted EWAS results with the significance of CpG probes across the genome, highlighting 200 CpGs that remained significant after a multiple comparison threshold of p < 10e-5. c Stability analysis of 200 selected CpGs, showing the frequency of non-zero coefficients across 1000 penalized model loops, which led to 38 CpGs being selected. d Distribution of 38CpG-HazardScores in the discovery and AML02,08 test cohort, dividing patients into high and low risk based on a 50% cutoff. e Kaplan-Meier survival curves for EFS and OS in the discovery cohort (n = 946) and (f) AML02,08 cohort (n = 200), comparing 38CpG-AMLsignature^high (orange) and 38CpG-AMLsignature^low (blue) groups. Hazard ratios derive from Cox PH regression with two-sided hypothesis tests. Individual n numbers are indicated in the figures.

Fig. 7

Multivariate analysis adjusting for other confounding variables for association of 38 CpG signature with OS in (a) discovery cohort (HR = 2.53; P < 0.0001) and (c) AML02,08 test cohort (HR = 2.34; P = 0.0169). Analysis for EFS was also performed in (b) the discovery cohort (HR = 1.53; P = 0.0002) and (d) AML02,08 test cohort (HR = 2.32; P = 0.0041). Hazard ratios derive from Cox PH regression with two-sided hypothesis tests. Individual n numbers are indicated in the figures. MRD1 minimal residual disease at end of first induction, FLT3 ITD FMS-like tyrosine kinase-3 internal tandem duplication, CI confidence interval, Ref. reference.

Fig. 8. AML Epigenomic Risk vs. 38-CpG AML Signature vs. standard of care.

ROC curves and AUC values for AML Epigenomic Risk model and 38-CpG AML Signature as (a) categorical and (b) continuous variables in discovery, AML02,08 and NOPHO AML cohorts. c Pearson correlation scatterplot between AML Epigenomic Risk model and 38-CpG AML Signature in discovery, AML02,08 and NOPHO AML cohorts. d Confusion matrices showing prediction rate for AML Epigenomic Risk and 38-CpG AML Signature. Individual n numbers are indicated in the figures.

Fig. 9. Genomic confirmation of ALMA Subtype predictions in selected samples.

Integrated Genomics Viewer (IGV) plots display selected genomic alterations identified in the specimen-to-result testing cohort using nanopore based sequencing. All data were generated by PromethION 2 Solo sequencing, basecalled with dorado v0.9; dna_r10.4.1_e8.2_400bps_sup@v5.0.0 and aligned to hg38 using minimap2 v2.27. Each panel highlights key pathogenic variants correlated with clinical diagnosis and ALMA subtype prediction. a AML with t(8;21)(q22;q22) (AML1-ETO); detection of RUNX1::RUNX1T1 bidirectional fusion (coverage: 6.5x, BM). b AML FAB M2 with t(6;9); DEK::NUP214 rearrangement confirmed (coverage: 1.1 x, BM). c Acute myelomonocytic leukemia NPM1 + ; frameshift insertion in NPM1 (c.860_863dupTCTG, p.Trp288fs) (coverage: 12.2 x, BM). d Relapsed refractory MLL +, FLT3 TKD + monocytic AML; identification of KMT2A::SEPTIN9 bidirectional fusion (coverage: BM 17.3 x, PB 24.6 ×). Each highlighted box indicates genomic variant regions and corresponding base-level details supporting the variant calls.