An advanced artificial intelligence (AI) prognostic tool has demonstrated remarkable accuracy in predicting and classifying acute myeloid leukemia (AML) subtypes through epigenomic signatures, confirming high concordance with genomic lesions. This breakthrough promises to radically enhance diagnostic speed and prognostic precision for AML, a severe blood cancer with diverse subtypes and varied patient outcomes. The results, validated across multiple independent cohorts and clinical trials, pave the way for quicker, cost-effective, and more accessible AML diagnostics and personalized treatment strategies.
AI Tool and Epigenomic Concordance in AML
The AI platform, known as the Acute Leukemia Methylome Atlas (ALMA), employs DNA methylation patterns—key epigenomic markers—to accurately classify 17 AML subtypes according to the latest WHO 2022 classification, as well as identify normal controls. Using rigorous clinical validation on patient samples, ALMA demonstrated a per-class concordance score range of 0.74 to 1.00 across various subtypes, with an overall accuracy nearing 90%. This highlights the tool’s ability to replicate or even surpass genomic lesion-based subclassifications using epigenomic data obtained through DNA methylation profiling.
The diagnostic submodel within ALMA, termed ALMA Subtype, leverages DNA methylation signatures to differentiate even the rare and ambiguous AML categories with high precision. The assessment of multiple independent datasets showed that the tool maintained high weighted F1 scores (above 0.9) and Cohen’s Kappa values, indicators of prediction reliability and minimal overfitting in unseen test conditions.
Prognostic Capability and Risk Stratification
Beyond diagnosis, ALMA includes a prognostic classifier named AML Epigenomic Risk. This model predicts patient survival probabilities up to five years post-diagnosis based on epigenomic risk groups defined by methylation signatures. Findings revealed that patients classified as high risk by this epigenomic signature had significantly poorer overall survival (OS), with hazard ratios indicating up to 4.4 times greater risk of death compared to low-risk groups in discovery cohorts and replicated in independent validation cohorts.
An alternative but related 38-CpG AML signature, derived via epigenome-wide association studies (EWAS), was also validated to predict 5-year time-to-death outcomes. Both epigenomic models notably outperformed current clinical risk group classifications used in AML prognosis.
The integration of epigenomic risk assessment into clinical workflow could shorten diagnostic waiting times from weeks to mere days using a simplified assay that requires only a portable sequencer. This may expand global accessibility, reduce cost barriers, and improve patient outcomes, as stated by Dr. Jatinder Lamba, co-leader of the UF Health Cancer Center Cancer Targeting and Therapeutics program.
Technological Innovations: Long-Read Sequencing and Machine Learning
A rapid specimen-to-result pipeline was developed that incorporates unsupervised machine learning, epigenomics, and long-read nanopore sequencing. This novel workflow was validated on over 3,000 acute leukemia patient samples and accurately generated diagnostic and prognostic reports within hours using minimal sample volume. The diagnostic accuracy across 25 WHO 2022 acute leukemia subtypes reached over 91%, while prognostic models robustly predicted overall and event-free survival independent of conventional clinical factors like FLT3 mutation status and minimal residual disease levels.
Machine learning methodologies behind these achievements include the use of LightGBM, a high-speed gradient boosting framework, which processes complex methylation data compressed through dimensionality reduction techniques. This computational approach uncovered numerous epigenetic biomarkers linked to AML progression and patient fate decisions, enhancing interpretability of predictive models.
Epigenomic Classification and Immune Microenvironment Insights
Independent research integrating multi-omics data from five cohorts identified two major AML epigenetic subtypes—CS1 and CS2—with substantially different prognoses. Regulatory networks characterized these subtypes based on transcription factors (such as CEBPA, RUNX1) and epigenetic modifiers (like DNMT3A, BRD4). These molecular signatures correlated with distinct immune microenvironments: high-risk patients exhibited immunosuppressive features dominated by regulatory T cells and M2 macrophages, while low-risk groups showed enhanced natural killer (NK) cell activity. Additionally, differences in mutation frequency and therapeutic vulnerabilities by subtype were observed, opening avenues for precision treatment strategies tailored by epigenomic profiling.
Clinical and Therapeutic Implications
The advanced AI prognostic tool for AML based on epigenomic signatures sets new benchmarks for speed, accuracy, and individualized risk prediction in hematologic oncology. By consolidating diagnostic and prognostic predictions into a single assay deployable with a laptop-sized sequencer, patient care is poised for transformation through quicker initiation of appropriate therapy and better stratification of treatment options.
This technological progress aligns with ongoing efforts to integrate machine learning and omics data for improved cancer management and exemplifies how epigenetics can augment the current genomic-focused paradigms. Early and precise identification of AML subtypes will empower clinicians to tailor interventions, potentially improving remission rates and survival outcomes.
AML Editor’s article was originally published in cancernetwork on August 26, 2025