Rapid Long-Read Epigenomic Diagnosis and Prognosis of Acute Myeloid Leukemia

Purpose: Despite substantial evidence underscoring the importance of the epigenome in classification and prognostication of acute leukemias, translational tools capitalizing on this knowledge have not been successfully implemented. By integrating unsupervised machine learning, epigenomics, and long-read sequencing, this study introduces a specimen-to-result workflow for accurate diagnosis and prognosis of AML patients.

Patients and Methods: Raw DNA methylation data from a total of 3,308 acute leukemia patient samples across 11 clinical trials/studies were processed and harmonized according to WHO 2022, ELN 2022, and COG clinical guidelines. A publicly available unsupervised foundation model named Acute Leukemia Methylome Atlas (ALMA) was built, which enabled the creation of two supervised, fine-tuned models: a diagnostic classifier that predicts acute leukemia epigenomic subtype and a prognostic classifier that predicts clinical outcome driven by AML epigenomic risk. Independent validation was performed using data from 201 diagnostic bone marrow specimens from two multi-center clinical trials. Finally, a rapid specimen-to-result protocol was created to integrate these predictive tools into a hematology/oncology unit setting using a customized high-molecular-weight DNA extraction protocol, needle shearing, and long-read nanopore sequencing from an input of 300uL of EDTA peripheral blood or bone marrow specimens.

Results: The diagnostic model showed overall accuracy of 0.914 and 0.931 across 25 WHO 2022 acute leukemia subtypes in discovery and validation, respectively. The prognostic model was a robust predictor of overall survival OS (HR=10.11; 95% CI=7.73, 13.22; P<0.0001; HR=3.56; 95% CI=2.12, 5.99; P<0.0001) and event-free survival EFS (HR=4.12; 95% CI=3.44, 4.94; P<0.0001; HR=2.54; 95% CI=1.65, 3.89; P<0.0001) in the discovery (COG AML trials) and validation cohorts, respectively. The prognostic model remained significantly associated with EFS and OS after adjusting for induction 1 MRD status, initial diagnostic risk-group assignment, FLT3-status, WBC-count and age in both discovery and validation. Specimen-to-result testing of four patients showed that epigenomic predictions were consistent between paired PB and BM samples despite differences in SNPs, SVs, and STRs. Additionally, samples were prepared for sequencing in 2 hours and final diagnostic and prognostic calls were made within 1 hour of sequencing, at coverages as low as 0.01x.

Conclusion: This study introduces a rapid specimen-to-result workflow that harnesses unsupervised machine learning, epigenomics, and long-read sequencing to significantly improve clinical diagnostic and prognostic accuracy for AML, particularly in the detection of ambiguous subtypes and further stratification of intermediate risk patients.