2026 ASH®
Annual Meeting & Exposition December 12-15, 2026 ▪ New Orleans, Louisiana

Paper Details

Abstracts, authors, affiliations, and disclosures.

Saturday, December 12, 2026
05:30 p.m. - 07:30 p.m.
MCCNO - Hall F
Access, Representation, Population Health, Technology and Procedures, Imaging, Molecular Testing, Omics Technologies, Pathology, Bioinformatics, Blood Banking, Computational Biology, Emerging Technologies, Profiling, Treatment Considerations, Drug Development, Supportive Care, Therapy Sequence, Biological Therapies, Chimeric Antigen Receptor, (CAR)-T Cell Therapies, Women’s Health, Workforce, Adverse Events, Artificial Intelligence (AI), Computer Vision, Deep Learning, Generative AI, Machine Learning, Natural Language Processing, Biological Processes, Apoptosis, Genomics, Hematopoiesis, Molecular Biology, Clinical Practice (Health Services and Quality), Clinical Guidelines, Quality Improvement, Quality of Care, Systems-Based Hematology, Diseases, Bleeding and Clotting, Thromboembolism, Bone Marrow Failure Syndromes, Inherited Marrow Failure Syndromes, Paroxysmal Nocturnal Hemoglobinuria, Lymphoid Malignancies, Lymphoid Leukemias, Lymphomas, Plasma Cell Disorders, Myeloid Malignancies, Acute Myeloid Malignancies, Chronic Myeloid Malignancies, Metabolic Disorders, Iron Deficiency, Measurable Residual Disease, Non-Hodgkin Lymphoma, B Cell Lymphoma, Aggressive Lymphoma, Research, Fundamental Science, AML, APL, Translational Research, MDS, MPN, Clinical Research, Clinical Trials, Real-World Evidence, Study Population, Human, Adult, Pediatric
PAPER

2583 Myeloid leukemia identity is encoded in a distributed, redundant, and low-dimensional DNA methylation space

Fabio Pires S Santos, MD, PhD, Luiz Ferreira Cortes2, Pedro V. B. Orsini, MD3, Paulo Vidal Campregher, MD, PhD4, Francisco Marchi, PhD5
1 Einstein Hospital Israelita, Sao Paulo, SP, Brazil
2 Einstein Hospital Israelita, Sao Paulo, SP, Brazil
3 Alma Genomics, Gainesville, Florida, United States
4 Einstein Hospital Israelita, Sao Paulo, SP, Brazil
5 Alma Genomics, Gainesville, Florida, United States

Background: DNA methylation classifiers can identify acute leukemia entities with high accuracy, but the architecture of this diagnostic signal remains unknown. We investigated whether leukemia identity is concentrated in a restricted set of CpGs, depends on higher-order CpG relationships, or is distributed across coordinated low-dimensional programs.

Methods: We analyzed a large, pooled dataset of array methylation in leukemias, comprising 3,314 samples, 331,556 CpGs, and 10 public cohorts. Primary analyses included 1,354 samples representing 12 WHO 2022 myeloid leukemia subtypes, each with at least 20 samples from at least 2 cohorts. The reference model was a multinomial ElasticNet classifier trained on the 10,000 most variable CpGs, selected within each training fold. Generalization was assessed across 7 leave-one-cohort-out folds. Feature selection, scaling, dimensionality reduction, module construction, and model fitting were restricted to training data. The primary metric was multiclass log-loss, with confidence intervals obtained by hierarchical cohort-level bootstrap. We evaluated matched random CpG subsets; random and directed ablation with and without model refitting; consensus co-methylation modules; principal component analysis (PCA) versus random projections; and relational models benchmarked against marginal-preserving null datasets.

Results: A linear ElasticNet model retained substantial cohort-held-out discrimination across WHO 2022 myeloid entities (log-loss, 0.72; balanced accuracy, 0.83; accuracy, 0.85), establishing a substantial linearly decodable diagnostic signal. Across panels of 100 to 1,000 CpGs, chromosome-, variance-, and CpG-context-matched random subsets did not differ significantly from supervised selections. Within the CpG ablation analysis, random removal of 90% of CpGs changed log-loss only from 0.634 to 0.659 after refitting and retained 98% of the referenced balanced accuracy, whereas log-loss increased to 2.236 when the original model was applied without refitting. Directed removal of the 90% most discriminative CpGs increased log-loss to 1.536, indicating that the signal was distributed but nonuniform. With feature count fixed at 1,000, enhancer CpGs outperformed CpG-island probes in all 7 cohorts (Δ log-loss, −0.721; 95% CI, −1.347 to −0.141). Unsupervised co-methylation analysis grouped 10,000 CpGs into 68 modules. Eighty percent showed greater-than-random internal cohesion in held-out cohorts, and their eigenscores retained 88% of the CpG-level model’s balanced accuracy. Diagnostic performance saturated substantially earlier than methylome reconstruction: 10 PCs achieved a balanced accuracy of 0.756 and 200 PCs achieved 0.837, similar to the ElasticNet reference model (0.832), although 200 PCs explained only 74% of methylome variance and approximately 1,000 PCs were required to explain 98%. Across CpG-, modifier-, and module-level relational analyses, no incremental classification gain exceeded marginal-preserving null expectations, indicating that the tested higher-order dependencies added little diagnostic information beyond marginal effects.

Conclusions: Diagnostic information distinguishing WHO 2022 myeloid leukemias is broadly distributed but nonuniform, robust to extensive random feature loss after retraining, and organized into reproducible co-methylation modules. These signals converge on a compact, linearly decodable subspace that is substantially smaller than the dimensionality required to explain near-complete methylome variance. This redundancy provides a rationale for evaluating lower-coverage, genome-distributed methylation assays as more accessible approaches to leukemia classification.

Fabio Pires S Santos, MD, PhD: GSK - Honoraria; BeOne - Honoraria; Dr Reddy's - Honoraria; AstraZeneca - Honoraria; Bristol-Myers-Squibb - Honoraria; Novartis - Honoraria; Abbvie - Honoraria; Amgen - Honoraria; Servier - Honoraria; Pint Pharma - Honoraria; Johnson & Johnson - Honoraria; Abbvie - Research Funding; AstraZeneca - Research Funding; Johnson & Johnson - Research Funding; Servier - Research Funding