A two-axis machine learning framework for cell-of-origin inference in AML: A proof-of-concept study.

Abstract

6540 Background: Leukemogenesis mirrors hematopoietic differentiation yet proceeds through dysregulated programs. Multiparameter flow cytometry (MFC) allows leukemia immunophenotypes projection onto normal myeloid ontogeny. Here, we apply machine-learning (ML) models trained on MFC data to reconstruct a preliminary data-driven representation of inferred cell-of-origin (COO). Additionally, we explore integration of monocytic differentiation into canonical COO framework to identify dysregulated commitment programs not explained by developmental maturity alone. Methods: After IRB approval, MFC data from 261 AML patients were aligned to immunophenotypic signatures of normal myeloid ontogeny (HSC, MPP, CMP, GMP, GP, MP) as previously described (Simoes et al. , Blood Neoplasia , 2024; Vergez et al. , Blood Cancer Journal , 2022). Probabilistic COO states were inferred from core markers (CD34, HLA-DR, MPO, CD33, CD117) using ensemble Random Forest classifiers with gradient boosting [Training/Validation=70/30]. Because canonical COO algorithms omit monocytic features, we implemented a secondary Mono-Risk Layer trained on monocytic markers (CD14, CD64) and stemness/invasiveness surrogates (CD56, reflecting adhesion, migration propensity) and CD123, to generate complementary phenotypic labels. Models were trained with cross-validation in Python, and mutation enrichment was projected onto inferred COO states. Results: Mean age was 63.8 years. 114/261(44%) were male. 123/261 (47.1%) of cases produced multidimensional P-COO achieving “Ensembled (RF+ Boosting Gradient)" AUC=0.74. Confusion matrix revealed a structured pattern of misclassification predominantly between adjacent COO states, consistent with a continuous differentiation manifold (Recall for MPP, CMP, GMP, GP, MP, 50% 83%, 62%, 65%, 76%, respectively). Mono-Risk-Layer (MRL) improved discriminatory ability for MP (AUC 0.94, Recall 86%, F1 0.75). NPM1 was mapped to GMP (60%), MP (34%) and CMP (27%), p= 2.61e-09, FLT3 ITD GMP (43.3%), CMP (32.4%) and MP (12.3%), p =4.72 e-05, TP53 HSC (50%), MPP (24.3%) and GMP (16%), p =3.6e-04, RUNX1 MPP (26.1%), HSC (25%) and MP (13%), p =7.7 e-03, RAS MP (35%), MPP (19%) and HSC (13%), p =5.18 e-03. Conclusions: Machine-learning–based probabilistic cell-of-origin inference from routine flow cytometry is feasible and biologically informative in AML. Incorporation of a monocytic risk layer improves resolution of monocytic programs not captured by canonical COO and reveals distinct mutation–differentiation relationships. NPM1 and FLT3-ITD are enriched in GMP/MP-like states, TP53 in primitive HSC/MPP compartments, RUNX1 in early progenitors, and RAS in monocytic-skewed trajectories. These findings link genomic drivers to specific developmental niches and establish a scalable framework for biologically grounded AML classification and therapeutic modeling.