Mapping the Human Bone Marrow in Myeloproliferative Neoplasia Using Spatial Transcriptomics

Background: Myeloproliferative neoplasms (MPNs) are a group of myeloid malignancies characterised by activation of JAK-STAT signalling and increased production of haematopoietic stem and progenitor (HSPC)-derived populations. Whilst the molecular drivers of MPN are well-defined, the clinical phenotype is heterogeneous and clinical course highly variable. Whilst most patients have chronic-phase disease - most commonly essential thrombocythemia (ET) or polycythaemia vera (PV) - some develop myelofibrosis (MF) which is associated with a significantly worse prognosis. Interplay between immune, stromal and megakaryocyte populations is thought to drive bone marrow fibrosis. However, how these cellular populations underpin this process is not yet fully defined. Emerging spatial transcriptomic (ST) platforms provide in-situ profiling with up to subcellular resolution and offer novel opportunities for characterising tissue microenvironments. We use the 10x Xenium platform to map the human bone marrow across large cohort of archival diagnostic tissue, for the first time providing whole section in-situ descriptions of the bone marrow microenvironment at single cell resolution. This approach has the potential for providing novel insights into mechanisms underpinning myelofibrosis and to detect spatial and microenvironmental features characteristic the earliest stages of MPN. Integration of these data with H&E morphology provides opportunities for the development of artificial intelligence (AI)-based models with the potential for genuine translation into clinical workflows to support and enhance clinical diagnostics.

Method: We developed a workflow for generating high-quality spatially resolved data from human archival formalin-fixed paraffin-embedded bone marrow trephine (BMT) material. We perform Xenium analysis on a large (n = 30) cohort of patients with MPN (n = 11 MF, n = 3 PrePMF, n = 8 ET, n = 3 PV, and n = 5 normal BMTs). We integrate the ST output with H&E morphology, demonstrating how this supports cellular segmentation and regional annotation. We develop a graph neural network (GNN) pipeline to facilitate the application of a machine learning model to our ST data.

Results: We identify a total of 6,304,337 BM cells from BMTs from 30 patients, including lymphocyte (CD4⁺ and CD8⁺), erythroid, megakaryocyte, endothelial, antigen presenting cell (dendritic cells, monocytes and macrophages), granulocyte, and haematopoietic stem and progenitor cell (HSPC) groups. We identify distinct mesenchymal stromal cell (MSC) groups including osteogenic-MSC and osteoblasts enriched within the endosteal niche, and an adipocytic MSC group enriched within intertrabecular regions. We systematically characterise the endosteal, peritrabecular and intertrabecular microenvironments. We identify and map HSPC and progenitor populations, defining microenvironmental perturbation characteristic of MPN. We quantitively describe the cellular composition of the human marrow, demonstrating how whole-trephine cellular abundance distinguishes both MF and PrePMF patients from those with non-fibrotic MPN (ET and PV) and normal marrows. We find that erythroid and megakaryocyte groups are expanded with increased cell-cell adjacency in MF, potentially reflecting megakaryocyte bias of the shared precursor. Differential expression analysis identifies upregulation of extracellular matrix (ECM)-associated transcripts across nearly all lineages in MF and PrePMF versus the normal marrow, but not in non-fibrotic MPN. Finally, we apply a graph neural network (GNN) approach to our spatially resolved ST-BMT dataset, generating a model which can with high confidence distinguish MPN from normal marrow. This approach furthermore allows high-resolution characterisation of the cellular microenvironmental and spatial features which define MPN.

Conclusions: We present the first whole-section spatial atlas of the human marrow at single cell resolution, demonstrating how ST approaches can provide novel insights into both normal marrow homeostasis and mechanisms underpinning disease progression in MPN. We also demonstrate the opportunities afforded by spatially-resolved single cell data for generating AI-based models with potential for genuine translational applicability, supporting early and accurate diagnosis and prognostication.