Distinct Transcriptomic Cell States Differentiate Low-Risk MDS from Healthy Marrow

Introduction

Myelodysplastic syndromes (MDS) are a group of malignancies that involve ineffective hematopoiesis and often include a large pool of immature blasts. Previous studies have identified malignant ontogeny in hematopoietic stem and progenitor (HSPC) and mature myeloid cells in in vivo models, including enhanced pro-inflammatory and pro-apoptotic programs. However, how malignant cell states vary across the myeloid differentiation landscape remains elusive, particularly in low-risk MDS where there may be a smaller proportion of blasts. To address this gap, we interrogate low-risk MDS primary samples via single-cell transcriptomics and build a classifier to distinguish low-risk MDS HSPCs from normal HSPCs. We then survey how these features resolve along myeloid differentiation states for differentiation state-specific biomarker discovery.

Methods

We conducted single-cell RNA-sequencing (scRNA-seq, 10X Genomics) on bone marrow derived mononuclear cells (BMMCs) from 18 MDS samples. We also referenced an external dataset for BMMC data from 6 healthy donors (HD). Following sample quality control and batch integration, we identified major hematopoietic lineages via expert annotation and interrogated these populations for “AML-like” signatures via AUCell. We then conducted network analysis via weighted gene co-expression network analysis (WCGNA) to identify disease state-specific gene modules.

To classify MDS HSPCs from their HD counterparts, we performed LASSO regression with 10-fold cross-validation to identify a subset of informative genes from the disease-state specific gene modules. First, we split our single-cell dataset into training (70%) and testing (30%) datasets. We then curated a validation dataset containing gene expression profiles from CD34+ MDS and HD cells, and measured performance of our model by calculating a receiver operating characteristic (ROC) curve. To identify biologically relevant genes in our model gene list, we performed pathway analysis via the Enrichr tool.

To identify differentiation state-specific differences, we performed pseudotime analysis via Monocle 3. We then fitted generalized additive models (GAMs) for the myeloid lineage via tradeSeq, and calculated fold-change differences between the start and end points of the lineage via tradeSeq's association test.

Results

The majority of our MDS sample set consisted of lower risk samples (Very Low, Low, Moderate Low) according to the Molecular International Prognostic Scoring System (IPSS-M). Only two samples progressed to secondary acute myeloid leukemia whereas only three samples met the criteria for excess blasts. Comparison of the MDS and HD cell populations revealed decreased erythroid populations in the MDS samples; however, there were no major differences in the HSPC and mature myeloid populations. Further, the MDS samples were enriched for cells expressing both LSPC-like and monocyte-like AML signatures, suggesting malignant programs in not only the primitive cells but also the mature myeloid cells.

Our classification model yielded a 99-gene MDS HSPC signature. Pathway analysis of our signature revealed membrane-trafficking (PICALM, LAPTM4B) and protein translation (DPYD, EEF1G) components. Our signature also performed exceptionally well in distinguishing low-risk MDS HSPCs within the external dataset (ROC, AUC > 0.94).

Lastly, we sought to resolve differences in the MDS HSPC signature across myeloid differentiation. Following pseudotime ordering, we discovered 35 genes that were differentially expressed at the start vs. end of the lineage (log fold-change > 2, false discovery rate < 0.05). Hierarchical clustering identified early, mid, and late expression of these genes. For example, LAPTM4B expression was associated with HSPCs; however, PICALM expression was associated with mature myeloid cells. Our results suggest that distinct pathway components in the MDS HSPC signature may be expressed preferentially in differentiated myeloid cells.

Conclusions

Our analysis provides a robust model that distinguishes low-risk MDS from HD HSPCs. The results also offer potential therapeutic targets for low-risk MDS patients, including membrane-trafficking and protein translation components, which can be further resolved by differentiation state.

Bick:TenSixteen Bio: Consultancy. Ferrell:Novartis: Research Funding.