Single-cell spatial transcriptomics reveals immunotherapy-driven bone marrow niche remodeling in AML

Abstract

Given the graft-versus-leukemia effect observed with allogeneic hematopoietic stem cell transplantation in refractory or relapsed acute myeloid leukemia (AML), immunotherapies have been explored in nontransplant settings. We applied a multiomic approach to examine bone marrow interactions in patients with AML treated with pembrolizumab and decitabine. Using extensively trained nuclear and membrane segmentation models, we achieved precise transcript assignment and deep learning–based image analysis. To address read-depth limitations, we integrated single-cell RNA sequencing with single-cell spatial transcriptomics from the same sample. Quantifying cell-cell distances at the edge level enabled more accurate tumor microenvironment analysis, revealing global and local immune cell enrichment near leukemia cells postpembrolizumab treatment, potentially linked to clinical response. Furthermore, ligand-receptor analysis indicated potential alterations in specific signaling pathways between leukemia and immune cells following immunotherapy treatment. These findings provide insights into immune interactions in AML and may inform therapeutic strategies.

High-resolution mapping shows that immunotherapy reshapes bone marrow cell interactions.

INTRODUCTION

Acute myeloid leukemia (AML) is a deadly hematological malignancy with an estimated 5-year survival rate of 30% (1). Patients with refractory or relapsed AML (R-AML) have an extremely poor prognosis (2). Success in using immune checkpoint inhibitors (ICIs) of PD-1 signaling in other cancer types motivated the testing of this approach in patients with R-AML (3). We previously conducted a clinical trial to treat 10 patients with R-AML using the combination of pembrolizumab (an ICI) and the hypomethylating agent decitabine (NCT02996474) and reported transcriptomic and proteomic changes of T cells in these patients throughout the treatment course (4–6). Of the 10 patients treated in this single-institution, single-arm, open-label study, 6 had a response of stable disease, partial, or complete response. When we performed sequencing and transcriptional analysis of T cell receptor β, we found clonal expansion of CD8⁺ effector memory T cells in patients who experienced immune-related adverse events (4). These expanded T cell populations were largely PD-1 positive and expressed transcriptional profiles consistent with T cells in an activated, cytotoxic state. The T cell populations in the two patients with complete or partial responses, however, did not exhibit a distinct transcriptional profile. These prior results, which characterized the transcriptional and immunophenotypic changes in the T cell population but failed to explain the difference in responses by looking at a single-cell population, underscore the need to characterize not only individual cell populations but also their dynamic interactions and spatial distribution within the tumor microenvironment.

Studies quantifying immune cell infiltration in tumors have demonstrated its predictive power, showing that patients with a high level of immune cell infiltration in the tumor (also called a “hot” tumor) have superior responses to immunotherapy (7, 8). In addition, spatial signatures and patterns derived from multiplex immunofluorescent antibody staining can predict the response in multiple solid tumor types (9–12). To better understand the complex interactions between tumor and immune cells in patients with R-AML receiving ICIs and hypomethylating agents, comprehensive profiling of different cell types in their spatial context is required. We have used single-cell RNA sequencing (scRNA-seq) in combination with antibody-derived tags (ADTs) to obtain transcriptomic and immunophenotypic information, but this approach lacks the spatial information of the bone marrow niche because cells are dissociated from their tissue context for sequencing analysis. Although traditional hematoxylin and eosin staining, immunohistochemistry, and immunofluorescence imaging preserve the tissue’s spatial context and enable general spatial analysis of leukemia and immune cells, these methods do not provide sufficient information to identify and characterize the diverse immune subpopulations crucial for a comprehensive analysis of the tumor microenvironment (13–15).

Previous attempts to combine RNA detection and protein imaging have proven difficult, as the tissue digestion process required to expose RNA binding regions tends to damage the protein epitopes to which the antibodies bind. Recently, however, as many markers of cellular function that cannot be stained by antibodies are now easily detectable via spatial transcriptomic analysis, its combination with immunofluorescence imaging has allowed us to detect cell-cell interactions in the tumor microenvironment (16). This enables a more robust understanding of the tumor microenvironment’s structure and heterogeneity, including tumor-immune interactions, and allows us to even further characterize immune cells’ functional states (for instance, we can now identify T cells expressing genes indicative of cytotoxicity or exhaustion). The CosMx Spatial Molecular Imager (CosMx SMI; NanoString Technologies) can image a formalin-fixed paraffin-embedded slide, from which users select fields of view (FOVs) for high-resolution imaging and transcript detection. Our present work combines traditional scRNA-seq and ADTs with advanced spatial transcriptomic methods to investigate the spatiotemporal interactions between different immune cell populations and leukemia cells in bone marrow samples collected from six patients with CD34⁺ leukemia before and after immunotherapeutic treatment.

RESULTS

Single-cell transcriptomic analysis of bone marrow samples identifies cell type markers and distinct patient-specific leukemia populations

Patients were categorized into two groups according to their clinical response to therapy: responders, which included patients who achieved clinical benefit (complete/partial response or stable disease), and nonresponders, i.e., those with progressive disease (Fig. 1A and table S1). Previously collected bulk RNA and protein sequencing data from bone marrow and blood did not identify any substantial changes in gene or protein expression across study time points or patient response (fig. S1). We therefore conducted further experiments using the previously described advanced tissue characterization technologies to test our hypothesis that immunotherapy remodels bone marrow niche in AML.

Fig. 1. Study scheme and analysis outline.

(A) Schematic of the study design. Ten patients with R-AML underwent a pretreatment bone marrow biopsy (baseline) and aspirate collection. They then received the ICI pembrolizumab on day 1 of the first treatment cycle. On day 8 of cycle 1, another bone marrow biopsy was obtained (post-ICI time point), after which they started receiving the hypomethylating agent decitabine, which they continued to receive for 5 days. On day 15 of cycle 1, patients began another 5-day decitabine treatment. On day 1 of cycle 2, patients again received pembrolizumab but no other treatment for the remainder of cycle 2. On the last day of cycle 2 (day 21), a third and final bone marrow biopsy was taken from each patient (postcombination therapy time point). The treatment continued for eight total cycles. 10x scRNA-seq was performed on baseline bone marrow aspirate samples from all 10 patients. scRNA-seq results were used to create cell type–specific gene expression profiles for each patient. Of the 10 patients, biopsies from 6 patients with CD34⁺ leukemia were used for the spatial transcriptomic assay. Each patient’s three biopsies were loaded onto a single slide, and 5 to 15 FOVs were selected for analysis on the CosMx SMI. The picture of CosMx SMI is used with the permission of NanoString Technologies Inc. (B) The image processing pipeline used for analysis consisted of image enhancement, cell segmentation, assignment of transcripts to individual cells, cell type identification, and further spatial analyses. The cell type–specific gene expression profiles generated for each patient from the 10x scRNA-seq data were used to identify cell types in the SMI data. Created in BioRender. M.A.B. (2025); https://biorender.com/ua6bwo1.

We obtained baseline bone marrow aspirate samples from all 10 patients in the clinical trial for 10x Genomics 3′ scRNA-seq with 77 ADTs. Three bone marrow samples from healthy donors were used as the reference for data integration and cell type annotation (Fig. 1B). Each sample underwent individual quality control and normalization, whereupon their data were integrated, resulting in 55,635 total cells (Fig. 2A, fig. S2A, and table S3).

Fig. 2. Cell type annotation and visualization for 10x scRNA-seq with cell surface antibodies and CosMx platforms.

(A) UMAP visualization of all cells from 10 patient and 3 healthy donor samples by 10x with cell type annotation. The colors represent different cell types and are used across the figure. (B) Normalized marker gene expression across different cell types by 10x. (C) Cell typing workflow for CosMx data. RBCs were removed from further analysis, and leukemia cells were identified by CD34 staining. All remaining cells were identified by transcriptional profiling. (D) Normalized marker gene expression across different cell types by CosMx. (E) Cell type distribution across all samples for six patients at three time points of CosMx data. (F) Comparison of cell type percentages for cells larger or smaller than 20,000 pixels in area. Those annotated as megakaryocytes have relatively larger cell size. (G) A megakaryocyte identified by size. The scale bar represents 10 μm. (H) Cell typing result in the same cropped region in which megakaryocytes are shown in light pink. The scale bar represents 10 μm. Created in BioRender. M.A.B. (2025); https://biorender.com/nyx1mok.

The median number of cells for all samples was 4579 (range: 2517 to 6207), with a median number of transcripts of 3787 [interquartile range (IQR): 2335 to 5997] and a median number of ADTs of 599 (IQR: 390 to 904). Unsupervised clustering of the integrated data did not show sample-specific or response-specific patterns, mainly separating healthy donor cells from patients (fig. S2B). Differential protein expression between clusters consisting only of cells from either healthy donors or patients confirmed that the cells in the cluster dominated by patient cells had enriched stem cell–related protein signatures, such as CD34 and CD117 (fig. S2C). Meanwhile, the cluster of predominantly healthy donor cells was annotated as consisting of mature B cells, a population typically diminished in patients with AML (fig. S2D) (17). Cell types of all clusters were annotated by using the top differentially expressed ADT signatures, which were further supported by the top gene markers (Fig. 2B). These gene markers were used to identify reference cells in the downstream spatial transcriptomic analyses. A total of 5468 cells was labeled as leukemia cells, consisting of 13.8% of all patient cells (range per patient: 3 to 38%) after excluding erythrocytes.

The clinical data suggested that patients had different leukemia-associated immunophenotypes (LAIPs) and cytogenetics at initial diagnosis (4). These differences would be useful for validation of our malignant cell identification. The data from each individual patient was subset and separately underwent dimension reduction, clustering, and cell type annotation, showing that the LAIP of patient-specific malignant cell clusters matched what was reported clinically (fig. S2E) (4). In addition, 91% of the leukemia cells identified in the integrated dataset overlapped with the patient-specific malignant cells, demonstrating agreement between the two assignment methods. Most of the leukemia cells identified in individual samples but not the integrated dataset were progenitor cells: 78% were early erythroid precursors, granulocyte-monocyte progenitors, thymocytes, and T cell progenitors. Patients 3 and 6 were selected for copy number variation analysis because their cytogenetic analyses found chromosomal loss or gain in 20 metaphases. Compared to gene expression from hematopoietic stem and progenitor cells from healthy donors, the posterior estimation of copy number changes in the malignant cells from patient 3 showed deletions in chr5/chr7 and a gain in chr21, consistent with the cytogenetic report at diagnosis (fig. S2F). Although the population of leukemia cells was smaller in patient 6, the results still matched their clinical reports and showed a loss in chr7 and a gain in chr8.

Further training of the combined nucleus-cell membrane model yields accurate cell segmentation

Before the cell type classification and spatial analysis of the tumor microenvironment with CosMx data are possible, the detected RNA transcripts must be assigned to single cells, which must first be detected via cell segmentation. NanoString uses Cellpose (18, 19), an open-source, state-of-the-art segmentation algorithm for cell microscopy data, which is meant to detect all the cells and identify their boundaries in each FOV. Cellpose is among the most accurate cell segmentation algorithms available, and yet the NanoString-generated segmentation of our CosMx SMI data did not always accurately represent cell borders. Comparing their segmentation to manually annotated ground truth showed that the annotated borders were notably expanded to capture more transcripts than were truly inside the cell. In addition, cells in our dataset exhibiting unique morphology, such as endothelial cells and megakaryocytes, were often not accurately segmented.

To improve the segmentation accuracy and to fully capture the diversity of cell types in our imaging data, we trained our own nuclear and membrane segmentation models using the Cellpose 2.0 “train-your-own model” feature (18). We generated a large training dataset by manually annotating 448,439 nuclei and 318,401 cell membranes on the basis of the nuclear and cell membrane staining of the CosMx SMI data using these comprehensive datasets as the ground truth for training and eventually testing our model (fig. S3, A to F).

In addition, we created an algorithm to combine nuclear and membrane segmentation results such that each cell could be represented by a single mask with a single-cell identification. It was important to train both nuclear and membrane segmentation models because while membrane segmentation most accurately represents a cell’s true border, some cells had very weak or no visible membrane staining, which was likely due to staining artifacts. In these cases, detected nuclei were expanded to estimate the cell’s border (fig. S3G).

To evaluate the performance of our segmentation and merging process, we compared its F1 score to that of the default segmentation generated by NanoString from the CosMx data. The F1 score, which balances precision and recall, was calculated at an intersection-over-union (IOU) threshold of 0.7 (see Materials and Methods for the detailed calculation). The optimal IOU threshold to use for model evaluation depends on the application. Typical IOU thresholds range from 0.5 to 0.9, but 0.7 is a common IOU used to evaluate cell segmentation methods (20). A 0.5 IOU was too lenient for our purposes, as it only required a true-positive segmentation mask to overlap 50% with the actual cell. On the other hand, 0.9 was too stringent, calling detected objects with less than a 90% overlap between the mask and the true cell false positives. At an IOU threshold of 0.7, our segmentation result showed substantial improvement in accuracy compared to the original CosMx SMI platform output (16) in our images: The F1 score of the NanoString segmentation was 0.23, while our approach attained an F1 score of 0.57 (fig. S3H).

Our trained and merged segmentation result not only improved the F1 score but also better captured the variability in cell size present in our dataset. This was a key area for improvement, as our images contained a diversity of cell types with vastly different morphologies, ranging from small, round lymphocytes to large, multilobed megakaryocytes. NanoString’s default segmentation result detected a total of 593,279 cells across all six patients, and our trained nuclear-membrane model combination detected 629,051 cells (6% more than NanoString’s default segmentation result). Furthermore, cells segmented by our method were on average smaller than those in NanoString’s segmentation, which generated segmentations with a median size of 58.3 μm² (range: 54.2 to 70.1 μm²), while our method identified cells with a median size of 46.1 μm² (range: 41.6 to 57.0 μm²). Because we segmented nuclear and membrane image channels separately, we were able to limit our use of nuclear expansion to cells without a detectable membrane, leading to fewer instances of cell size inflation via overexpansion, which improved the accuracy of transcript-based cell characterization (fig. S4).

Cell typing in situ by protein expression, transcriptomic profile, and morphological features

Before cell typing the CosMx data, we used several criteria to assess the heterogeneity of the samples. The transcripts and CD34 protein channel were plotted by the measured local coordinates per FOV, which identified different cell densities and distributions across FOVs, showing the heterogeneity of cell distributions within the bone marrow biopsy and background tissue binding (fig. S5A). There were 141 FOVs in total (range: 21 to 25) and a median of 4146 cells per FOV (range: 1041 to 9408). The average transcripts per cell showed strong sample- and FOV-specific variations (fig. S5B), with a median of 102 considering all FOVs (IQR: 45 to 177).

Data from the CosMx SMI provide the unique opportunity to capture both the morphology and transcriptome of a given cell, enabling more robust cell type assignment and validation than would be possible from single-cell transcripts alone. Before proceeding with transcript-based cell type identification, we had to identify red blood cells (RBCs) by morphology and remove them from the downstream analysis (Fig. 2C and fig. S6). Even though RBCs lack a nucleus and are not clearly stained by the B2M/CD298 membrane marker, their autofluorescence makes them visually detectable. After segmentation, we fed cropped images of each individual cell’s B2M/CD298 and 4′,6-diamidino-2-phenylindole (DAPI) channels into EfficientNet (21), a convolutional neural network designed to classify images. As its final output, EfficientNet provides a vector of probabilities, known as an embedding, that quantifies how likely the input image is to belong to one of 1000 predefined classes. Rather than letting the classifier run to completion, we extracted the 672 features used to describe each single-cell image from the second to last (B0) layer of the convolutional neural network. We then used Phenograph (22) for unsupervised clustering to cluster cells by their extracted features, independently for samples from each patient (23). By labeling each cell in the image with its cluster number, we were able to identify which clusters were predominantly composed of RBCs. The number of RBCs varied per patient: 5315 RBCs were identified among segmented cells in FOVs from patient 1, 26,881 from patient 2, 4624 from patient 3, 1476 from patient 4, 30,850 from patient 5, and 1437 from patient 6. After the RBCs were identified by morphology, they were removed from subsequent analysis. Most of the morphologically identified RBCs had low gene expression, with 95% having a total gene count of 44 or fewer.

Because all selected patients had clinically confirmed CD34-positive leukemia cells via flow cytometry and pathology at the time of biopsy collection, we used the CD34 protein channel to identify leukemia cells. We once again used the EfficientNet classifier to extract embeddings from cropped images of every cell in the dataset, but rather than including the B2M/CD298 and DAPI channels, the crops included only the CD34 protein channel of each segmented cell. Leukemia cells were therefore identified on the basis of cell morphology and the CD34 protein expression pattern, which enabled us to exclude CD34-expressing endothelial cells. Because the clinical information and scRNA-seq data indicated that normal hematopoietic stem cells were extremely rare in these samples, by this method, a total of 40,112 leukemia cells was identified via unsupervised clustering: 5005 from patient 1, 19,833 from patient 2, 5005 from patient 3, 6036 from patient 4, 2332 from patient 5, and 1901 from patient 6 (fig. S7).

After the annotated RBC and leukemia cell populations were removed, additional low-quality FOVs and cells were removed from the analyses. To further resolve other cell types in the spatial transcriptomic data, we used cell-specific gene signatures derived from matched scRNA-seq. This cohort-specific reference was critical, as our AML samples originated from a pilot immunotherapy trial and may differ from previously published datasets. Given the targeted nature of the CosMx gene panel, using validated signatures from our own cohort enhanced annotation accuracy within these technical constraints. The genes in the spatial gene panel that had been previously identified as cell type markers in the 10x scRNA-seq data were used to select reference cells. These overlapping genes included HBB and HBA1 for erythroid progenitors; PF4 and PPBP for megakaryocytes; IL3RA, CD33, and LYZ for dendritic cells (DCs); CD3E, CD3D, and CD3G for T cells; CD19 and CD79A for B cells; MPO and ELANE for granulocyte-monocyte progenitors; CD14, CSF3R, S100A9, CD33, and LYZ for CD14⁺ monocytes; GNLY, NKG7, and FCGR3A for natural killer cells; and FCGR3A (without GNLY or NKG7) for CD16⁺ monocytes. The lymphocyte category was further subdivided on the basis of the expression of CD4, CD8A, CD8B, IL7R, CCR7, CCL5, GZMH, and GZMK for T cells and MS4A1, TCL1A, and JCHAIN for B cells. Cells expressing these overlapping genes most highly were selected as the reference cells for the cell type prediction algorithm InSituType (24), which annotated cell types for the remaining cells. By using the posterior probability of the InSituType output, cells were annotated with the expected differential marker gene expression (Fig. 2, D and E). To validate the inferred cell type annotation, we leveraged imaging data–extracted cell size information and the presence of large megakaryocytes, finding that a total of 98% of cells larger than 648 μm² (20,000 pixels) was correctly labeled as megakaryocytes (Fig. 2, F to H).

Representing leukemia cell neighborhoods with a linear mixed model reveals shifts in cell composition after ICI treatment

Common methods of studying cell-cell and neighborhood interactions within the tumor microenvironment begin by computing the distance between each cell and all other cells. Now, it is a common practice to reduce each cell to a single coordinate location by finding its centroid and then measuring the distance between each cell’s centroid and all other cell centroids (25). However, this method has several limitations: Not only are irregular cells not accurately represented by a simple centroid location, but centroid measurements are also unable to accurately quantify the number of directly touching cells (fig. S8). When examining the composition of each leukemia cell’s neighborhood, we developed a computational approach to quantify cell-to-cell distances using the distance between the closest points of each cell’s edge, as opposed to simply the distance between centroids. After applying this method to our dataset to count the number of cells in direct contact with each cell, we found that the mode of this measurement was three directly touching cells, which represent three equidistant neighbors. Had we used centroid-to-centroid distances, these three neighbors would have been ranked with respect to their proximity to the reference cell, which would be an inaccurate representation of the biological reality.

The spatial transcriptomic data enabled us to determine whether there were any changes in cell type abundance in the regions surrounding leukemia cells specifically. To study the cell type density changes around leukemia cells across patient response and study time points, we fit a series of Poisson linear mixed effect models to the count data extracted for each cell type with patients grouped as responders and nonresponders. This approach accounts for the count nature of our data while considering the repeated-measures structure and potential variability between patients. In doing so, we considered that patients who responded similarly to the ICI would display similar cell distribution patterns to surrounding leukemia cells. Ring-shaped neighborhoods at radial distances of 0 to 5, 5 to 15, 15 to 25, 25 to 35, and 35 to 45 μm were defined around the cell membrane of each leukemia cell (Fig. 3A), and the abundance of each different cell type in each of these radial neighborhoods was analyzed as the dependent variable in the linear mixed model. We used the total number of cells inside each neighborhood as the offset. This enabled us to compare the prevalence of different cell types at different proximities to a leukemia cell and to determine whether their proportions varied across study time points and/or with increasing distance from the central leukemia cell.

Fig. 3. Comparison of leukemia cell neighborhood composition by time point and response group.

(A) Schematic depicting cell type composition determination in leukemia cell neighborhoods quantified in neighborhoods 0 to 5, 5 to 15, 15 to 25, 25 to 35, and 35 to 45 μm from the edge of each leukemia cell. (B) Comparing the densities of CD14+ monocytes at increasing distance from leukemia cells between baseline and post-ICI samples among nonresponders (left) and responders (right). In neighborhoods at radii of 5 to 15, 15 to 25, 25 to 35, and 35 to 45 μm from leukemia cells, there were higher proportions of CD14+ monocytes in the post-ICI samples, but these trends were substantial only in responders. (C) Representative image of segmented cells showing CD14+ monocyte density around leukemia cells in one responder at the baseline. Concentric rings mark 5-μm intervals around each leukemia cell from 0 to 45 μm. Leukemia cells are in magenta; CD14+ monocytes range from white to dark green as they increase in distance from the nearest leukemia cell. (D) Representative image of segmented cells showing CD14+ monocyte density around leukemia cells in one responder in the post-ICI sample. (E) Density comparison of DCs among nonresponders (left) and responders (right) at increasing distance from leukemia cells between baseline and post-ICI samples. Proportions of DCs were higher in post-ICI samples at radii of 5 to 15, 15 to 25, 25 to 35, and 35 to 45 μm from leukemia cells in nonresponders. (F) Representative image of segmented cells showing DC density around leukemia cells in one nonresponder at the baseline. Leukemia cells are in magenta; DCs range from white to dark brown as they increase in distance from the nearest leukemia cell. (G) Representative image of segmented cells showing DC density around leukemia cells in one nonresponder in the post-ICI sample. Scale bars represent 20 μm in (C), (D), (F), and (G). P values in the figure were adjusted P values considering the size of the model.

From the separate models we fit for different cell types considering the interactions between patient response, distance groups, and treatment time point, we were able to investigate how ICI treatment perturbed the prevalence of different cell types within leukemia cell neighborhoods in both responders and nonresponders (table S2). We started with examining whether the proportion of cells of any cell type changes around the nearest leukemia cell when comparing responders with nonresponders at the baseline. We observed that there were fewer granzyme K⁺ CD8⁺ T effector cells and CD14⁺ monocytes in leukemia cell neighborhoods in responders than in nonresponders at the baseline (fig. S9, A and B). Previous studies have found that patients with higher levels of classic monocytes before ICI treatment tend to have lower overall survival (26), which aligns with our observation that CD14⁺ monocytes were more prevalent at the baseline among nonresponders. We also observed increased proportions of granulocyte-monocyte progenitor cells in responders than in nonresponders at the baseline (fig. S9C). Higher numbers of CD14⁺ monocytes were seen at the post-ICI time point than at the baseline in responders, suggesting that the elevated numbers of granulocyte-monocyte progenitor cells among responders at the baseline may have differentiated into CD14⁺ monocytes after treatment. Studies in mice have found that anti–PD-1 promotion can catalyze myeloid progenitors to differentiate into antitumor effector cells (27).

Among responders, the proportion of CD14⁺ monocytes in leukemia cells increased after ICI treatment across all defined radial neighborhoods [rate ratio (RR) range: 2.5 to 2.7; adjusted P values <0.001]. This trend was observed but not significant in nonresponders (Fig. 3, B to D). This showed that there was an overall increase in CD14⁺ monocytes post-ICI therapy for responders, irrespective of leukemia cell proximity. Classic monocytes express chemokine receptors and are known to migrate toward inflammation (28). Their generalized influx after ICI therapy in responders could suggest a response to inflammation around the leukemia cells. Alternatively, this observation may reflect treatment-induced maturation of leukemia cells into CD14⁺ monocytes. Further mechanistic studies will be necessary to distinguish between these possibilities.

Among nonresponders, we observed an increase in the proportion of DCs in all neighborhoods, suggesting an increase in population at the post-ICI time point compared with the baseline (RR range: 1.95 to 2.01; adjusted P values <0.001). The trend was the opposite in responders but was not statistically significant (Fig. 3, E to G). It has been established that DCs are of limited utility after activating neighboring T cells by antigen presentation, and their accumulation may yield overactivation of the immune system. Previous studies have shown that DCs can undergo attack and elimination by CD8 T cells after serving as antigen-presenting cells. It is possible that this post-ICI DC accumulation among nonresponders is related to an ineffective CD8 T cell response (29). All the findings from these models indicated that the notable differences identified were across all prespecified distance groups, implying the potential shift in cell distribution related to therapy and clinical response.

Changes in patient-specific spatial cell distribution are identified by density comparison of time points across cell types

The linear mixed model assumed that there were consistent changes across patients who responded to the therapy compared with those who did not and used a simple discretization of the cell distances. It also did not account for overlapping neighborhoods. Grouping the patients by their response showed us the high-level trend of cell enrichment around leukemia cells. In the meantime, we observed high heterogeneity across samples (fig. S3); thus, we hypothesized that there could be patient-specific signatures that were obscured by considering the patients in groups. We developed an algorithm that was able to detect patient-specific shifts in cell type density around leukemia cells between two time points of interest, which we took to be the baseline and the end of cycle 2 of treatment (Fig. 4A and fig. S10).

Fig. 4. Leukemia-T cell density shift from the baseline to the post-ICI time point (purple) compared with simulated background distribution (gray) in pixels.

(A) Schematic representing this comparison. Density shift in (B) CD4 naïve T cells showed a consistently significant trend of enrichment at the baseline around leukemia cells, and (C) CD8 naïve T cells showed an opposing trend for patients 3 and 6 versus patient 2, (D) additional significant shifts in myeloid cells for patient 3, and (E) additional shifts for patients 2 and 4 in granulocyte-monocyte progenitors. The distance was limited to 45 μm (1 pixel = 0.18 μm) with the density calculated. The plot was normalized to the median of the simulated curves at each distance.

Responder patients 2 and 6 and nonresponder patient 3 exhibited a consistently notable trend of enrichment in the naïve CD4⁺ T cell population at the baseline around leukemia cells compared with the end of the second treatment cycle (Fig. 4B). For the naïve CD8⁺ T cell population, patient 2 showed a trend of having more naïve CD8⁺ T cells at the baseline than at the end of cycle 2; however, patients 3 and 6, who had different responses, consistently displayed an opposing trend compared with patient 2, with a positive density shift for CD8-naïve T cells around leukemia cells at the end of cycle 2 from the baseline (Fig. 4C).

There were additional significant shifts in the myeloid cell population density in nonresponder patient 3. There was an enrichment of DCs, granulocyte-monocyte progenitors, CD14 monocytes, and CD16 monocytes around leukemia cells at the baseline compared with the end of cycle 2 (Fig. 4D). For other patients in terms of myeloid cells, patients 2 and 4, who were both responders, had a positive density shift in granulocyte-monocyte progenitors with treatment (Fig. 4E). The other positive shift for patient 3 comparing the two times was for megakaryocytes (fig. S11). Patient 6, who was one of the responders, had an increase in progenitor B cells and a decrease in mature B cells (fig. S11).

Ligand-receptor analysis reveals leukemia-to-myeloid TWEAK signaling post-ICI therapy

Beyond simply understanding the changes in the spatial distribution of cells in the tumor microenvironment, we wanted to identify ligand-receptor pairs involved in the interactions between leukemia cells and other cell types nearby. Given the relatively low read counts of our dataset, especially compared with traditional scRNA-seq studies, we used a pseudo-bulk approach. We aggregated transcripts captured within leukemia cells of each FOV to determine which ligand or receptor genes were more highly expressed in leukemia cells than in other cell types. Then, for each individual cell type, we aggregated transcript counts from all cells of that type located within 5 μm of the nearest leukemia cell. To determine whether the complementary ligand/receptor gene to that expressed by the leukemia cell was enriched in leukemia cell neighbors specifically and not just in all cells of that type across the FOV, we compared ligand/receptor gene expression in this group of cells within 5 μm of a leukemia cell with the expression of the same transcript in all cells of that cell type further than 30 μm from the nearest leukemia cell (Fig. 5A).

Fig. 5. Ligand-receptor analysis.

(A) Schematic illustrating how the pseudo-bulk method was used to separate cells into “close” (0 to 5 μm to the nearest leukemia cell) and “far” (30+ μm to the nearest leukemia cell) groups. Created in BioRender. M.A.B. (2025); https://biorender.com/vbhr8nf. (B) Heatmap of ligand-receptor pairs where the ligand gene is expressed by the leukemia cell (x axis) and the receptor gene is expressed by another cell type (y axis) in patient 3. The P value was calculated on the basis of difference in the expression level of the receptor gene between a given cell type’s close and far groups. The two pairs with significantly higher receptor expression in the close group (P value <0.05) were TNFRSF12A expressed by DCs/TNFSF12 expressed by leukemia cells and TNFRSF12A expressed by naïve CD4 T cells/TNFSF12 expressed by leukemia cells. Pair TNFRSF12A expressed by granulocyte-monocyte progenitor cells/TNFSF12 expressed by leukemia cells (P = 0.07). Created in BioRender. M.A.B. (2025); https://biorender.com/z3bxnsp. (C) Heatmap of ligand-receptor pairs where the receptor gene is expressed by the leukemia cell (x axis) and the ligand gene is expressed by another cell type (y axis). The P value was calculated on the basis of difference in the expression level of the ligand gene between a given cell type’s close and far groups. No significant pairs of biological interest were identified. Box plot comparing the mean expression per FOV of receptor gene TNFRSF12A of all naïve CD4 T cells (D), DCs (F), granulocyte-monocyte progenitors (GMP) (H) in the close versus far groups from patient 3 at post-ICI time point and representative images of segmented cells showing naïve CD4 T cells (E), DCs (G), and GMP (I) expressing TNFRSF12A within a 5-μm distance of a leukemia cell expressing TNFSF12. Leukemia cells are in magenta; CD4 T cells, DCs, and GMP are in blue, yellow, and green, respectively. Scale bars represent 10 μm.

Ligand-receptor pairs found to be significantly enriched were all observed at the post-ICI treatment time point (Fig. 5B). Most notably, leukemia cells in patient 3 showed elevated expression of TNFSF12, which codes for a pro-inflammatory cytokine in the tumor necrosis factor (TNF) family called TWEAK, or TNF-like Weak Inducer of Apoptosis (also known as APO3L and CD255). This cytokine can trigger apoptosis in neighboring cells. Unexpectedly, the nearby naïve CD4 T cells (Fig. 5, D and E) and DCs (Fig. 5, F and G) had significantly higher expression of TNFRSF12A, which codes for the corresponding receptor TWEAKR (also known as FN14 and CD266) than that of the same cell types at least 30 μm away from the closest leukemia cell. Granulocyte-monocyte progenitor cells also expressed elevated levels of TNFRSF12A (Fig. 5, H and I) with a P value trending toward significance (P = 0.07).

Prior studies have reported that TWEAK binding stimulates activation of the NF-κB (nuclear factor κB) pathway in neighboring TWEAKR⁺ cells, leading to TNF-α production and proapoptotic autocrine signaling (30, 31). Studies have shown that TWEAK signaling can be particularly lethal to monocytes, promoting monocyte clearance alongside TRAIL and Fas (31). In a patient with progressive disease, the expression of TNFSF12 in tumor cells indicates potential leukemia-instigated apoptosis of neighboring myeloid cells. In addition to promoting apoptosis, the TWEAK ligand can also enhance cell proliferation while inhibiting differentiation (32). This has been observed in myoblasts that were stimulated by TWEAK to proliferate but failed to differentiate (33). This TWEAK function is particularly interesting in our case, as one of the cell types expressing the receptor TWEAKR was granulocyte-monocyte progenitor cells. It is possible that the TWEAK ligand from leukemia cells could be inhibiting the immune response by dysregulating granulocyte-monocyte progenitor cells from differentiating into functional monocytes.

DISCUSSION

In this study, we combined the power of traditional scRNA-seq with cutting-edge spatial transcriptomic technology, enabling us to gain insight into changes to the tumor microenvironment of R-AML induced by treatment with an ICI and hypomethylating agent. After immunotherapy, we observed remodeling of the AML bone marrow niche: Responders became more enriched with CD14⁺ monocytes, while DCs became more prevalent in nonresponders. Although we did not find a statistically significant enrichment of CD8 T effector cells from the baseline to the post-ICI time point, it has been established that in an effective CD8 killing response, CD8 T cells undergo apoptosis after executing the killing function (34). It is therefore unlikely that CD8 T cells would have been prevalent at the time of the second bone marrow biopsy (8 days after ICI administration). The density shift algorithm identified patient-specific spatial patterns in the cell distribution around leukemia cells before and after ICI treatment. Given that spatial transcriptomic data capture both a cell’s position and its transcriptome, we examined potential ligand-receptor gene pairs being expressed in neighboring cells. Although limited by our sample size, CosMx data showed potential post-ICI TWEAK signaling between leukemia cells and adjacent DCs, naïve CD4 T cells, and granulocyte-monocyte progenitor cells.

While spatial transcriptomic technology continues to develop, for the time being, some intrinsic limitations remain. Here, we used the CosMx 1000-plex gene panel, which can detect up to 960 different transcripts—still much less than that of traditional scRNA-seq (35, 36). Hence, not all interactions between tumor and immune cells are fully captured because of suboptimal gene coverage. Moreover, spatial transcriptomic technologies capable of detecting transcripts in thicker samples will further enhance detection capabilities. In addition, protein staining of lineage and functional cell markers helps with cell type classification, tumor versus immune for instance. However, finer-grained classification requires multiple protein markers, which is very challenging with the current available spatial transcriptomic technologies. For cells not identifiable through protein staining, the limited number of transcripts captured per cell in our spatial transcriptomic data made accurate cell typing extremely challenging. We thus chose to use a reference-based method to annotate cells by using overlapping marker gene expression signatures across different cell types from 10x scRNA-seq, which offers a more robust result of cell typing compared with the use of spatial transcriptomic data alone. Other platforms for spatial transcriptomics will have different properties with respect to the area they can analyze, the number of transcripts per cell, the ability to stain for multiple proteins, and the ability to perform cell membrane staining.

Another intrinsic limitation of spatial transcriptomic methods is that for true single-cell level data, they require comprehensive cell segmentation across the entire area of interest. Imprecise delineation of cell boundaries leads to the incorrect attribution of liminal transcripts, which obscures the true cell–specific gene expression patterns and hinders the identification of rare or subtle cellular states, especially at the currently low read depth. Moreover, inaccurate segmentation can confound the computational analysis of intercellular spatial relationships. The utility of single-cell level spatial transcriptomic results is therefore dependent on accurate single-cell level segmentation. To this end, we invested a large amount of effort in optimizing the cell segmentation process, which showed substantial improvement over the vendor-supplied segmentation result. Our comprehensive training dataset and optimized segmentation approach can serve as a flexible foundation for developing and benchmarking various cell segmentation models, ultimately enabling more precise and robust spatial analysis of complex tissues. We have found that it is more effective to train specialist models on membrane protein channel images and combine their segmentation results than to rely on generalist models, at least for the time being. A true cell segmentation foundation model, i.e., a model that can accurately segment cell borders across diverse tissue types in the full range of experimental conditions, will require training on an even larger dataset.

In addition to technological limitations, spatial transcriptomic methods are also prone to substantial sampling bias. In our study, we imaged needle biopsies from the bone marrow, which at the outset provided only a limited view of the tumor microenvironment. Within each biopsy, 5 to 15 FOVs sized 0.985 mm by 0.657 mm were selected for imaging, each covering an area of about 0.65 mm². Between these same-sample replicates, we observed variability in both the transcriptomic and imaging data, providing further evidence that one or more FOVs selected for spatial transcriptomic analysis might not be representative of an entire sample. While capturing the entire tissue area would be ideal, this is usually not logistically and technically feasible. In addition to the increased time and resources necessary to acquire additional FOVs, the process of repeatedly flowing reporters into and out of the flow cell can gradually damage the tissue and cause it to float off the slide. We expect the issue of inter-FOV variability to be a substantial issue in the analysis of spatial data.

As a study designed to assess the clinical feasibility of pembrolizumab-decitabine combination therapy in patients with R-AML, this early phase clinical trial consisted of 10 patients. The criteria that patients needed to have CD34⁺ leukemia for inclusion in the spatial transcriptomic analysis limited the cohort size to six patients. Among these six patients, treatment outcomes varied substantially: One patient had a complete response, another had a partial/complete response, two finished the regimen with stable disease, and two developed progressive disease. Among these vastly different outcomes, we categorized patients into responder (stable disease/response, n = 4) and nonresponder (progressive disease, n = 2) groups. The scarcity of extreme responses, such as complete response, made analysis extremely challenging. In the future, it would be valuable to obtain a direct mutation profile of leukemia cells using spatial transcriptomic data, although this remains challenging with the current spatial technologies. Heterogeneity between patients made it difficult to distinguish treatment-dependent effects from natural patient-to-patient variability. A large-scale clinical study would be necessary to further confirm and validate our findings of the immunotherapy-driven bone marrow niche remodeling.

Despite the limitations of spatial transcriptomic technology and the limited sample size, we show that integrating complex single-cell proteogenomic analysis with spatial-temporal multiomic analysis in patients with AML is feasible and may lead to a deeper understanding of interactions between leukemia cells and immune cells in the bone marrow niche, as evidenced by our finding of the shifts in cell composition in leukemia neighborhood and leukemia cell–mediated TWEAK signaling. Applying our spatial-temporal multiomic analysis method to a larger clinical cohort should reveal the interactions between tumor and immune cells in the tumor microenvironment and identify potential targets for developing better therapeutics.

MATERIALS AND METHODS

Patient cohort

Bone marrow samples were from the pilot clinical study 17H-0026 (PDAML, NCT02996474, approved by NHLBI Institutional Review Board) with 10 adult patients with R-AML to test the feasibility of the combination of pembrolizumab and decitabine. All patients consented to the trial. The treatment and sample collection times were previously reported (4). The baseline data were collected before any drug administration. The second and third time points corresponded to 8 days after the first dose of pembrolizumab just before decitabine initiation (post-ICI treatment, C1D8) and the end of cycle 2 (postcombination therapy, EOC2), respectively. Bone marrow samples were also collected from healthy donors within the age distribution of the patients for scRNA-seq as the reference datasets (5).

scRNA-seq experiments and analysis

Cell surface staining with oligo-tagged antibodies

Cryopreserved bone marrow–derived mast cells were thawed into RPMI-1640 supplemented with 10% fetal bovine serum, washed twice with centrifugation at 300g for 5 min at 4°C in a swing-arm rotor centrifuge, and counted with trypan blue staining on an automated hemocytometer. Between 500,000 and 1 million cells were resuspended in 100 μl of phosphate-buffered saline (PBS) with 1% bovine serum albumin (BSA). Cells were incubated with 5 μl of Human TruStain FcX FC receptor blocking solution for 5 min at 4°C, after which 1 μl (0.5 μg) each of the 32 TotalSeq-A oligo-tagged antibodies (BioLegend) was added. Cells were incubated with antibodies for 30 min at 4°C. After incubation, cells were washed twice with 1 ml of PBS and 1% BSA, filtered through a 40-μm strainer, and counted with trypan blue staining on an automated hemocytometer. Cells were washed a final time with PBS and 0.04% BSA, and the cell concentration was adjusted to between 700 and 1200 cells/μl. Cells were kept on ice until acquisition.

scRNA-seq library construction

The 10x Genomics 3′v3 Single Cell Profiling platform was used for scRNA-seq. After staining and cell concentration adjustment, cells were loaded onto a Chromium Chip A with master mix, gel beads, and partitioning oil according to the manufacturer’s instructions. The desired number of captured cells was 4000. Reverse transcription of mRNA into barcoded first-strand cDNA was performed on the emulsion by incubating at 53°C for 45 min in an Eppendorf Mastercycler X50a. After reverse transcription, emulsions were broken, and cDNA was purified using Dynabeads MyOne SILANE. At this stage, cDNA was amplified by polymerase chain reaction (PCR; 98°C for 45 s; 13 cycles of 98°C for 20 s, 67°C for 30 s, and 72°C for 1 min; 72°C for 1 min). Additional primers were added to ensure adequate amplification of oligo tags to later generate ADT sequencing libraries. Amplified cDNA was mixed with 0.6X SPRIselect magnetic beads (Beckman Coulter); after magnetic separation, the supernatant was transferred for ADT library generation, and the remaining pellet was used for constructing the gene expression (GEX) library. The concentration of amplified cDNA was measured at a 1:10 dilution with a High Sensitivity (HS)-D5000 chip on an Agilent Tapestation.

For the GEX library, 50 ng of amplified cDNA was carried forward for library construction. Fragmentation, end-repair, and A-tailing were performed (32°C for 5 min; 65°C for 30 min), followed by double-sided size selection using SPRIselect beads. Adaptor ligation (20°C for 15 min), cleanup with SPRIselect beads, and PCR amplification with sample indexing primers (98°C for 45 s; 14 cycles of 90°C for 20 s, 54°C for 30 s, and 72°C for 20 s; 72°C for 1 min) were performed in that order. A final double size selection was again done using SPRIselect beads. To make ADT libraries, sample index PCR using unique indices was performed with the transferred supernatant taken immediately after initial cDNA amplification (98°C for 45 s; nine cycles of 98°C for 20 s, 54°C for 30 s, and 72°C for 20 s; 72°C for 1 min) using Kapa HiFi Hotstart Readymix (Kapa), followed by a single-sided size selection cleanup with SPRIselect beads.

scRNA-seq library quality control and sequencing

GEX and ADT libraries were run on an HS-D5000 chip using an Agilent Tapestation at a 1:10 dilution, and the average size of all libraries was calculated [GEX libraries were around 500 base pairs (bp) in size; ADT libraries were about 200 bp]. The libraries were quantitated in triplicate using quantitative PCR (Kapa) at four dilutions from 1:40,000 to 1:5,000,000 to calculate nanomolar library concentrations. The scRNA-seq libraries were then pooled in the appropriate ratio so that each library would be sequenced at the correct depth and sequenced on an Illumina HiSeq 3000 using read lengths of 150 bp from read 1, 8 bp of the i7 index, and 150 bp from read 2. Target sequencing depths were 50,000 read pairs per cell for the GEX libraries and 5000 read pairs per cell for ADT libraries.

Single-cell data preprocessing, integration, and clustering

Sequencing libraries were preprocessed using Cell Ranger version 3.1.0 (10x Genomics) to output filtered unique molecular identifier (UMI) matrices for downstream analysis (37). The number of barcode mismatches was set to be 0 to minimize the demultiplexing error. Cells with overlapping cell barcodes between GEX and ADT results were included in the multimodal analysis.

Seurat version 3 was used for sample basic quality control, normalization, and integration (38). Data were normalized on the basis of regularized negative binomial regression, and 2000 variable features were selected for each sample on the basis of a variance stabilizing transformation. Anchor cells, which are pairs of cells in matched biological states, were identified across samples to integrate and match the overlapping cell populations. Integrated datasets were scaled, and the number of principal components was chosen to represent the heterogeneity in gene expression. The cell clusters were constructed by applying a graph-based clustering approach, and the data were explored by uniform manifold approximation and projection (UMAP) for visualization. Highly expressed proteins were identified for cell type annotation. The expression of several categories of genes including exhaustion, cytotoxicity, and naïve signatures was extracted to calculate the respective gene scores for each cell type. Leukemia-associated clusters were labeled on the basis of markers from previous clinical flow cytometry records, and the labeling of cell clusters was confirmed by performing clustering analysis on proteins for each patient sample. Protein-only clustering done separately for each sample was also used to confirm other major cell types such as T cells. The InferCNV (39) pipeline was implemented to identify cells with a known copy number variation for samples from patients with reported cytogenetic abnormalities. Malignant cell populations identified using different methods were compared with each other by calculating the overlap percentages and using the joint UMAP for visualization.

Spatial imaging processing with proteomic and transcriptomic profiling

Sample preparation

Formalin-fixed paraffin-embedded slides prepared from bone marrow biopsies at time points A (pretreatment), B (postpembrolizumab), and C (postpembrolizumab and postdecitabine) from six patients were analyzed using a precommercial CosMx SMI (NanoString Technologies Inc.). All three biopsy samples from a given patient were placed on the same glass slide (Fig. 1A). To improve tissue adherence, the slides were baked overnight (60°C). Next, the Leica Bond RX system was used to expose RNA targets on the samples through a process of deparaffinization, proteinase K digestion (3 μg/ml, Thermo Fisher Scientific; incubated at 40°C for 30 min), and heat-induced epitope retrieval (using Leica buffer ER1 at 100°C for 15 min). Samples were then rinsed twice with diethyl pyrocarbonate (DEPC)–treated water and incubated for 5 min at room temperature with fiducials (Bangs Laboratory) diluted 1:1000 in a 2× SSCT solution (2× saline sodium citrate with 0.001% Tween 20). After excess fiducials were removed by rinsing the samples with 1× PBS, the samples were fixed for 5 min at room temperature in 10% neutral buffered formalin and then rinsed for 5 min each with tris-glycine buffer (0.1 M glycine and 0.1 M tris base in DEPC-treated water) and 1× PBS. Next, samples were blocked for 15 min at room temperature with 100 mM N-succinimidyl acetate (NHS-acetate; Thermo Fisher Scientific) in NHS-acetate buffer (0.1 M NaP and 0.1% Tween 20 at pH 8 in DEPC-treated water). After a 5-min 2× SSC rinse, samples were protected with an Adhesive SecureSeal Hybridization Chamber (Grace Bio-Labs) before overnight hybridization with RNA in situ hybridization (ISH) probes. The RNA ISH probe panels included 960 probes and 19 negative control probes. To prepare the ISH probe mixture, RNA ISH probes were denatured for 2 min at 95°C and then placed immediately on ice. The probe mix was then pipetted into the hybridization chamber enclosing the sample and covered with adhesive tape to prevent the hybridization solution from evaporating during the 37°C overnight hybridization. After hybridization, samples were washed twice with a 50% formamide (VWR) solution in 2× SSC, each for 25 min at 37°C. They were next rinsed twice for 2 min in 2× SSC alone at room temperature. Before the samples were encased on each slide in a custom flow cell, they were blocked for 15 min with 100 mM NHS-acetate and washed twice for 2 min in 2× SSC alone at room temperature.

Data acquisition

The SMI RNA detection and imaging procedure is described by He et al. (16) and summarized here. The prepared flow cell was loaded into the SMI, where several rounds of reporters would be cycled in, imaged, and removed to detect the presence of all transcripts in the panel. Before beginning the cycling, reporter wash buffer [1× SSPE, 0.5% Tween 20, SUPERase•In RNase Inhibitor (0.1 U/μl) at 20 U/μl, 0.1% Proclin 950, and DEPC-treated water] was flowed into the flow cell to rinse the samples and remove any air bubbles. Next, the SMI performs a preliminary scan of the whole area covered by the flow cell, at which point 5 to 10 FOVs were selected from each biopsy that would be captured by the camera in the coming reporter cycles. To begin the first RNA readout cycle, 100 μl of reporter pool 1 flowed into the flow cell and was left to incubate for 15 min before 1 ml of reporter wash buffer flowed in to remove any unbound probes. Before the instrument imaged the fluorescent probes, the wash buffer was replaced with an imaging buffer [80 mM glucose, pyranose oxidase from Coriolus sp. (0.6 U/ml), catalase from bovine liver (18 U/ml), 1:1000 Proclin 950, 500 mM tris-HCl buffer, pH 7.5, 150 mM sodium chloride, and 0.1% Tween 20 in DEPC-treated water]. The SMI captured nine z-stacked images, at a step size of 0.08 μm, from each FOV. After imaging, the fluorophores on the reporter probes were ultraviolet cleaved and washed away with strip wash buffer (0.0033× SSPE, 0.5% Tween 20, and 1:1000 Proclin 950). This reporter pool, imaging buffer, and wash buffer cycle was repeated 15 more times for a total of 16 cycles total (once per reporter pool). In total, nine rounds of this 16-cycle process were repeated. The last step was to capture sample morphology. The samples were incubated for 1 hour with a solution of fluorescently labeled antibodies, which included CD298 at 1:40 dilution (Abcam, EP1845Y), B2M at 1:40 dilution (Abcam, EP2978Y), CD34 at 1:40 dilution (Abcam, Qbend/10; conjugated to AF647), and DAPI. One additional protein marker, CD3 at 1:20 dilution (Abcam, F7.2.38), was not directly conjugated to a fluorophore, so 1 hour of CD3 incubation was followed by 1 hour of donkey anti-mouse AF594 incubation (1:80 dilution). After excess antibody was washed away with reporter wash buffer and the wash buffer was replaced with imaging buffer, the instrument captured nine z-stacks per FOV.

Image preparation: Normalization and deconvolution

Of the five channels imaged, only two were needed for cell segmentation: DAPI for nucleus segmentation and CD298/B2M for membrane segmentation. To create a crisper image for image annotation, we deconvoluted the DAPI and CD298/B2M channels using Huygens Professional version 22.04 (Scientific Volume Imaging, The Netherlands, http://svi.nl). Finding the cell boundaries easier to detect but the nuclei dimmed, we layered the deconvoluted DAPI channel with the raw DAPI channel. Next, both the deconvoluted CD298/B2M channel and hybrid raw/deconvoluted nucleus channel were clipped to a range between the 5th and 99.99th percentiles of their raw grayscale values before each protein channel was mapped to its RGB color channel (DAPI in blue and CD298/B2M in green).

Model training

We used the Cellpose (19) train-your-own model feature to train two models: one for nucleus segmentation and one for cell membrane segmentation (using the B2M/CD298 membrane marker). Using manual annotation, we first generated a total of 448,439 ground truth nucleus masks. Once enough nuclei had been annotated to train our own model, we switched to a semiautomatic annotation approach. We used our trained nucleus model to predict nucleus masks and manually corrected them. We used the same approach to generate a membrane prediction model, with 318,401 ground truth membrane masks used in the final dataset.

Implementing segmentation models

Of the 142 FOVs imaged, nucleus masks were manually corrected on 76 and membrane masks on 22. For the remaining FOVs, our trained nucleus and membrane models were used to generate predicted masks of each respective class. Wanting to have a single mask to represent each cell in (including both nucleus and membrane), we merged nucleus and membrane masks that corresponded to the same cell for all FOVs. When evaluating the output of the trained model in comparison with the ground truth, we computed an F1 score using the formula

$$\mathrm{F}1=\frac{\text{TP}}{\text{TP}+\left(\frac{\text{FP}+\text{FN}}{2}\right)}$$

where TP is the number of true positives, FP is the number of false positives, and FN is the number of false negatives. To compute this metric, we used an IOU threshold of 0.7, meaning that for a single cell label to be considered a true positive, it must overlap with a ground truth label by 70% or more. Any less overlap and that label would count as a false positive.

Cell type identification for spatial transcriptomic data

Before proceeding with assigning cell types by the gene expression profile, we identified leukemia cells using an approach based on cell morphology and CD34 protein expression. Raw CD34 channel intensities were clipped below the 5th percentile and above the 99.9th percentile and then normalized from 0 to 255 for each FOV. Individual images of every cell’s normalized CD34 protein channel were resized to 224 by 224 pixels and triplicated to create a pseudo-RGB image. For each patient, the pseudo-RGB images of all cells imaged from that patient were fed into the EfficientNet classifier (21). The vector of 672 features extracted from each image in the B0 layer was then used to cluster the cells via UMAP. Overlaying the mean CD34 protein intensity expressed in each cell on the UMAP revealed, as anticipated, an intensity-dependent clustering. Because cell shape is also encoded in the features of each image, leukemia cells clustered separately from other CD34⁺ cells with different morphology (chiefly endothelial cells of the blood vessels).

RBCs, which are morphologically distinguishable from other cells owing to their lack of nucleus and distinct shape, were identified and removed before transcript-based cell type assignment. Like the process used to identify leukemia cells (but using normalized RGB images that contained the DAPI, B2M/CD298, and CD34 channels), single-cell image features extracted from the B0 layer of the EfficientNet (21) classifier were used to cluster cells by morphology. Representative cells from each cluster on UMAP were visualized to identify which clusters contained RBCs. All cells of suspected RBC clusters were highlighted on each FOV and reviewed to verify that the RBCs had been identified. This process was completed independently for each patient.

The total transcript count was used as the quality control metric to remove low-quality cells (<20 transcripts assigned to the identification), and three FOVs with less than 300 cells remaining were removed from the analyses. InSituType (39), a likelihood-based cell typing method, was used to annotate the remaining cells. Cell types identified in 10x single-cell experiments were included in the supervised cell typing pipeline, and the marker genes were selected on the basis of prior knowledge and a differential expression list to create the reference profiles. Because the data from different samples were heterogeneous, the following cell type annotation steps were performed within each sample so that the sample variability would not affect the annotation result. The expression matrices for all cell typing marker genes were extracted, and the average expression of each cell type was calculated for each cell. Reference cells were selected by choosing those with only one cell type marker expressed, and the reference profiles were created by taking the average expression of all genes from cells labeled as reference cells for each cell type. The reference matrix was the input for the algorithm with genes in rows and cell types in columns. Average negative control values were provided to adjust the count and provide the normalized count, and the posterior probability was used to determine the cell types. The first round of the analysis determined the broad cell types, and an additional round of creating reference and calculating posterior probability was performed for immune cell subtypes to obtain more specific annotation for B cells and T cells. The average expressions of marker genes for the final cell types were compared using a heatmap to confirm the accuracy of the gene expression profile of each population. External validation was performed by using the cell size information obtained from the cell segmentation step.

Spatial distance measurement and cell distribution modeling

Distance determination

The spatial genomic data contained local coordinates of all transcripts within each FOV. Each cell, however, was represented as a polygon, and cell-to-cell distances were found by computing the minimum distance between the edges of every pair of cells in each FOV.

Generalized linear mixed effect model for spatial distribution

For each leukemia cell, every other cell was labeled with a distance group by how far it was to this leukemia cell, with 0 representing directly touching and every other group labeled by increasing distances of 5 μm. Then, we summarized the number of cells for each cell type around this leukemia cell. The input for the Poisson generalized linear mixed model was the count for different cell types across distance groups for each leukemia cell.

Because we were interested in testing whether there were significant changes in any cell types across patient response and time points, one Poisson generalized linear mixed model (40) was fit for each cell type, with the count for the cell type of interest as the response variable; the three-way interaction of patient response (responder/nonresponder), sample collection time points, and distance group as the main effects; and FOVs as a random effect. Because the total cell count of each FOV across different distance groups strongly correlated with the count of each cell type in that region, we considered this total cell count as an offset of the model.

P values were recalculated using the fitted model for different combinations of coefficients of interest to investigate the difference for responders between time points and the difference between responders and nonresponders at the baseline. Bonferroni correction was used to generate the adjusted P values considering all models fitted to determine any significant changes, with two-sided adjusted P values of 0.05 as statistically significant.

Cell type density shift

To examine the spatial distribution shift between time points, we started with defining the minimum distance of one cell to another cell type within one FOV as the shortest distance between the cell and any cells of the other cell type. For example, if the two cell types of interest were leukemia cells and T cells, suppose for one FOV, there were $M$ leukemia cells $m_1,m_2,\dots ,m_M$ and $N$ T cells $n_1,n_2,\dots ,n_N$ . The pairwise distance matrix for all leukemia cells and T cells was defined as

$$\begin{array}{c}{m}_1{n}_1,m_2{n}_1,m_3{n}_1\dots ,m_M{n}_1\\ m_1{n}_2,m_2{n}_2,m_3{n}_2\dots ,m_M{n}_2\\ \begin{array}{c}\dots \\ m_1{n}_N,m_2{n}_N,m_3{n}_N\dots ,m_M{n}_N\end{array}\end{array}$$

The minimum distance of one T cell $n_i$ to leukemia cells was defined as the minimum value of that row of distance values

$$d_i=\text{min}(m_1{n}_i,m_2{n}_i,m_3{n}_i\dots ,m_M{n}_i)$$

For all the T cell $N$values in this FOV, the distribution of minimum distance was estimated by getting the minimum value of each row in that matrix

$$D_T=\{d_1,d_2,\dots d_N\}$$

The distribution of $D_T$ was the minimum distance distribution from T cells to leukemia cells for this FOV. We considered the FOVs from one time point of one patient as sample replicates, and the distribution of one sample was the aggregated results from all FOVs

$$\begin{array}{c}\begin{array}{c}\begin{array}{c}{D}_{T,A}=\left\{d_{1,A_1},d_{2,A_1},\dots ,d_{N_{A_1},A_1},d_{1,A_2},d_{2,A_2},\dots ,d_{N_{A_2},A_2},\dots ,d_{1,A_F},d_{2,A_F},\dots ,d_{N_{A_F},A_F}\right\}\hfill \end{array}\hfill \end{array}\hfill \end{array}$$

For one patient sample at time point $A$ , with $F$ numbers of FOVs from this sample, and for FOV $f\in \{1,2,\dots ,F\}$ , there were $N_{A_f}$ number of T cells.

To create the background distribution for the density shift test, we started with constructing the background distribution of any cells to leukemia cells. Similar to what was described, for the same FOV with $M$ leukemia cells and $K$ other cells including $N$ T cells $(N<K)$ , the background distribution from a random cell group of the same number of cells to leukemia cells was obtained by 100 random permutations. For each permutation, we randomly sampled $N$ cells from $K$ cells, $n_1^{\prime },n_2^{\prime },\dots ,n_N^{\prime }$ , and obtained the following pairwise distance matrix

$$\begin{array}{c}{m}_1{n}_1^{\prime },m_2{n}_1^{\prime },m_3{n}_1^{\prime }\dots ,m_M{n}_1^{\prime }\\ m_1{n}_2^{\prime },m_2{n}_2^{\prime },m_3{n}_2^{\prime }\dots ,m_M{n}_2^{\prime }\\ \begin{array}{c}\dots \\ m_1{n}_N^{\prime },m_2{n}_N^{\prime },m_{3}n{n}_N^{\prime }\dots ,m_{M}n{n}_N^{\prime }\end{array}\end{array}$$

The same definition was used to get the background distribution $D_{T,A}^{\prime }$ for one sample.

For two time points of interest $A$ and $B$ (here, we considered the baseline as time point $A$ and the post-ICI time point as time point $B$ ), we obtained the distributions $D_{T,A}$ and $D_{T,B}$ and used the $\text{density}$ function in R to get the respective probability density functions ${\mathit{pdf}}_{T,A}$ and ${\mathit{pdf}}_{T,B}$ for $D_{T,A}$ and $D_{T,B}$ , with the same distance range $x=(x_1,x_2,\dots ,x_l,\dots ,x_L)$.

$${\mathit{pdf}}_{T,A}=(p_{x_1}^{T,A},p_{x_2}^{T,A},\dots ,p_{x_L}^{T,A}),{\mathit{pdf}}_{T,B}=(p_{x_1}^{T,B},p_{x_2}^{T,B},\dots ,p_{x_L}^{T,B})$$

The density shift distance from time $A$ to time $B$ was defined as

$$\begin{array}{c}{S}^{T,A-B}={\mathit{pdf}}_{T,A}-{\mathit{pdf}}_{T,B}\\ =(s_{x_1}^{T,A-B}=p_{x_1}^{T,A}-p_{x_1}^{T,B},s_{x_2}^{T,A-B}=p_{x_2}^{T,A}-p_{x_2}^{T,B},\dots ,s_{x_L}^{T,A-B}=p_{x_L}^{T,A}-p_{x_L}^{T,B})\end{array}$$

With the same definition, for any random permutation $j(j=1,2,\dots ,J)$ , we could obtain its density shift distance

$$S_j^{T,A-B}=(s_{j{x}_1}^{T,A-B},s_{j{x}_2}^{T,A-B},\dots ,s_{j{x}_L}^{T,A-B})$$

To determine whether the density shift was significant, we checked for each $s_{x_l}^{T,A-B}$ , where it fell with regard to the distribution of $(s_{1x_l}^{T,A-B},s_{2x_l}^{T,A-B},\dots ,s_{J{x}_l}^{T,A-B})$ . The shift was determined to be significant if the data $s_{x_l}^{T,A-B}$ were outside of 1 standard deviation of the simulated background (either direction). For visualization, the color of the shift in density was determined by the magnitude of the shift.

A simulation study was implemented by ordering the cell-to-cell distance and setting the closest or furthest cell as a T cell to test whether the algorithm could identify the pattern shift. This process was repeated for all cell type pairs of interest to identify potential patterns in spatial shift across treatment time.

Ligand-receptor analysis

The human ligand-receptor gene pairs were downloaded from CellTalkDB (41) with 3399 pairs in total and 719 pairs overlapping with the spatial gene panel. For each cell other than leukemia cells, the distance between the cell and its nearest leukemia cell was calculated. Two cell groups were extracted for the analysis, one with minimum distance less than or equal to 5 μm (close group) and the other with minimum distance greater than or equal to 30 μm (far group). Because the data for each FOV were sparse, we adopted a bulk approach to study potentially different ligand-receptor pairs. The expression data were aggregated for each FOV of each sample (patient and time point) by calculating the mean expression of genes in the ligand-receptor reference for each cell type. We analyzed the data two ways by considering genes from the ligand list on the leukemia cells or from the receptor list on the leukemia cells. For each way of considering the gene list, the median expression level of the genes was aggregated for each cell type within one FOV and genes highly expressed in leukemia cells were defined as those with median expression greater than 2 standard deviations away from the expression of other cell types. To test the other gene list on other cell types, the Mann-Whitney U test was used to compare the median expression level of genes between two distance groups (far group versus close group) within each sample for different cell types, considering FOVs as replicates. Ligand receptor results were summarized based on the two different tests on ligand and receptor gene lists, and adjusted P values are reported.

Software

R version 4.3.2 was used for the analyses and figure creation. Python 3 was used for spatial transcriptomic analyses (minor version varied by the virtual environment). Illustrations were created with BioRender.com and Adobe Illustrator.

Supplementary Materials

The PDF file includes:

Figs. S1 to S11

Legends for tables S1 to S3

Other Supplementary Material for this manuscript includes the following:

Tables S1 to S3