Longitudinal single-cell spatial proteomics reveals predictive immune and stem cell signatures in Relapsed/Refractory AML

Abstract

Abstract Relapsed/refractory (R/R) acute myeloid leukemia (AML) remains a major clinical challenge, particularly in patients previously exposed to hypomethylating agents (HMAs). Despite the widespread use of HMAs in combination with venetoclax, the mechanisms of action and resistance remain poorly defined, and predictive biomarkers are lacking. To address this, we applied the Proximity Network Assay (PNA) (Karlsson et al. 2025. bioRxiv doi:10.1101/2025.06.19.660329), a next-generation single-cell spatial proteomics platform, to longitudinal bone marrow biopsies from 10 R/R AML patients treated with decitabine – 5 complete responders (CR) and 5 non-responders (NR). PNA constructs nanoscale protein interaction maps by combining barcoded antibodies, rolling circle amplification, and proximity-dependent ligation, enabling sequencing-based reconstruction of single-cell protein networks. Each cell yields multiplexed spatial nodes across 155 proteins, allowing simultaneous quantification of protein abundance, colocalization, and polarity with ~50 nm resolution. We employed this approach to capture the dynamic remodeling of the tumor interface and stem cell compartment during therapy. Despite similar blast count prior to treatment, patients who responded to decitabine exhibited higher proportions of monocytes and NK cells. Blasts from responding patients showed increased abundance and colocalization of immune recognition molecules (HLA, B2M), adhesion proteins (CD18, CD43, CD54), and reduced CD45RB. In contrast, CD44 abundance and clustering was increased in blasts from non-responders, with increased abundance of different adhesion proteins (CD62P, CD82, CD26), suggesting differential membrane organization and interaction with the microenvironment and immune effectors. Following decitabine, CR patients maintained lower blast counts, and persistence of higher numbers of immune effector cell frequencies were observed. Early upregulation of CD71 and CD36 on CD34+ cells with a distinct immunophenotypic “proxiome” likely indicates early erythroid recovery and metabolic activation in healthy HSC recovery. CD90 remained elevated in remission marrows, consistent with physiologic HSC expansion. In contrast, non-responding patients retained high CD34 clustering and adhesion molecule abundance, potentially representing a marker of persistent leukemic stem-like cells. Using proximity and abundance features, we trained a classifier to predict response. Prediction modeling utilizing the random forest algorithm was able to discriminate between healthy versus leukemic HSC based on PNA protein interactions (AUC 0.889), with CD36/CD82, CD71/CD71, and B2M/HLA-ABC colocalization emerging as top predictive features based on importance scores. These signatures enabled tracking of clonal evolution across timepoints. This study represents the first longitudinal spatial proteomics analysis of AML response to HMA therapy. PNA reveals dynamic remodeling of the tumor interface and identifies predictive signatures of response and resistance. Integration with methylation-aware Oxford Nanopore Technology sequencing is underway to correlate chromatin accessibility with proteomic remodeling and stem cell fate.