AUTHORS
Beatrice Zhang, Stephen Martis2, Robert Stanley, MD, PhD3, Kimon Argyropoulos, MD, PhD4, Zoe Katsamakis5, Ning Fan6, Kenton Wu7, Catherine Snopkowski8, Samuel Taylor, MD, PhD9, Jeetayu Biswas10, Matthew Zatzman11, Roni Tamari, MD12, Jonathan Peled, MD/PhD13, Gunjan L. Shah, MD, MS14, Joseph Chan, MD, PhD15, Ronan Chaligne, PhD16, Benjamin Greenbaum17, Omar Abdel-Wahab, MD18, Susan DeWolf, MD19
AFFILIATIONS
1 Memorial Sloan Kettering Cancer Center, New York, New York, United States, Memorial Sloan Kettering Cancer Center, New York, New York, United States, Weill Cornell Medicine, New York, New York, United States
2 Memorial Sloan Kettering Cancer Center, New York, New York, United States
3 Memorial Sloan Kettering Cancer Center, New York, New York, United States, Memorial Sloan Kettering Cancer Center, New York, New York, United States
4 Memorial Sloan Kettering Cancer Center, New York, United States
5 Memorial Sloan Kettering Cancer Center, New York, New York, United States
6 Memorial Sloan Kettering Cancer Center, New York, New York, United States
7 Memorial Sloan Kettering Cancer Center, New York, New York, United States
8 Memorial Sloan Kettering Cancer Center, New York, New York, United States
9 Memorial Sloan Kettering Cancer Center, New York, New York, United States
10 Memorial Sloan Kettering Cancer Center, New York, New York, United States
11 Memorial Sloan Kettering Cancer Center, New York, New York, United States
12 Memorial Sloan Kettering Cancer Center, New York, New York, United States
13 Memorial Sloan Kettering Cancer Center, New York, New York, United States
14 Memorial Sloan Kettering Cancer Center, New York, New York, United States
15 Memorial Sloan Kettering Cancer Center, New York, New York, United States, Memorial Sloan Kettering Cancer Center, New York, New York, United States
16 Memorial Sloan Kettering Cancer Center, New York, New York, United States
17 Memorial Sloan Kettering Cancer Center, New York, New York, United States, Memorial Sloan Kettering Cancer Center, New York, New York, United States
18 Memorial Sloan Kettering Cancer Center, New York, New York, United States, Memorial Sloan Kettering Cancer Center, New York, New York, United States
19 Memorial Sloan Kettering Cancer Center, New York, New York, United States, Memorial Sloan Kettering Cancer Center, New York, New York, United States
ABSTRACT
The diagnosis of MDS relies on pathological evaluation of cellular morphology combined with genetic analyses. Nevertheless, diagnosis and treatment of MDS, as well as prediction of progression to AML, remain challenging. Given the importance of both molecular characteristics and spatial organization of the bone marrow, we hypothesized that spatial transcriptomics, which integrates these complementary features, could improve diagnostic discrimination while uncovering targetable molecular interactions within the tissue microenvironment. Here we generated single cell spatial transcriptomic data of >7 million cells from bone marrow biopsies of healthy individuals and patients with MDS, AML, and myelofibrosis (MF). From these data, we developed a machine learning model to classify myeloid malignancies and identify molecular interactions that distinguish each disease state.
We performed spatial transcriptomic profiling (10x Genomics Xenium) on whole slides from 81 patient bone marrow biopsies comprising 41 untreated MDS, 8 untreated AML, 10 untreated MF, 17 age-matched healthy controls, and 5 post-allogeneic transplant samples of MDS patients in remission for >2 years. The MDS cohort contained 24 MDS-LB, 12 MDS-IB1, and 5 MDS-IB2 samples.
We first sought to determine whether spatial features of cellular organization alone could discriminate MDS from control samples without predefined clinical information. We developed a computational method to identify enrichment of local scale spatial cellular interactions within cell niches (fixed size of k=5-10 cells). This revealed that the MDS bone marrow has increased spatial cellular heterogeneity relative to healthy subjects, a finding which normalized in MDS patients in long-term remission following allogeneic transplantation. We further identified spatial features that stratified MDS samples along two principal axes: most samples—including all MDS-IB2 cases—were enriched for myeloid-dominant niches, whereas a smaller subset exhibited enrichment of erythroid-dominant niches.
These results established a foundation for the development of a spatially informed disease classifier. To construct the classifier, spatial transcriptomic data were represented as neighborhood graphs for each sample, where nodes corresponded to genes and edges were weighed by spatial co-occurrence between gene pairs. An elastic-net regularized multinomial logistic classifier was trained on sample graph representations to predict clinical categories of healthy, MDS, MF, and AML. Model performance was evaluated using stratified 5-fold cross-validation, achieving an overall accuracy of 91% and class-wise AUROCs ranging from 0.95-0.99. The primary source of model confusion was distinguishing MDS-LB from healthy control samples.
Interrogation of the learned feature weights from our model revealed co-localized gene modules distinguishing each disease state. Control marrows were characterized by interactions with B cells, demonstrated by B cell-defining transcripts co-localizing with myeloid transcripts. By contrast, MDS marrows were distinguished by co-localization of cytokine signaling transcripts (CXCL2, CXCL9, CCR7) with myeloid-lineage markers, consistent with the identification of an altered immune microenvironment in MDS characterized by large aggregates of T-, B-, and plasmacytoid dendritic cells localized to cytokine-rich regions. Further, the model-predicted probability of the AML class increased progressively from MDS-LB to IB1 to IB2, suggesting that spatial molecular features characteristic of AML are progressively acquired during MDS progression.
Our study defines spatial signatures of MDS across multiple biological scales. To address the heterogeneity of MDS, we applied an unsupervised learning framework to characterize local scale cellular niches based on cell-type composition, identifying two spatially distinct subsets of MDS samples. Complementarily, we developed a supervised machine learning classifier that identified gene-level spatial interactions distinguishing healthy controls, MDS, AML, and MF. While prior machine learning approaches in MDS have trained on H&E images, our spatial transcriptomics-based classifier captures a feature space distinct from current diagnostic practice. Together, these findings demonstrate the utility of spatial transcriptomics for both improving disease classification and uncovering molecular alterations associated with progression from MDS to AML.
DISCLOSURES
Beatrice Zhang: No relevant conflicts of interest to declare.