Single-Cell RNA-Sequencing Reveals Tumor Microenvironment Composition and Prior Therapy Modulate Response to Axl Inhibition Authors: Anisha Datta, Kate Bridges, Remziye E. Wessel, Hratch M. Baghdassarian, Laura C. Bahlmann, Diana N. Gong, Erin N. Tevonian, Brian A. Joughin, Douglas A. Lauffenburger Published: 2026 ABSTRACT PurposeAxl is a therapeutic target under active clinical investigation for a variety of cancer indications. While pre-clinical mouse studies have been promising, clinical trials targeting Axl in combination with standard-of-care therapies have yielded mixed results. We hypothesize that this clinical translation gap may in part be due to an incomplete understanding of the effects of Axl inhibition specifically on the myeloid compartment of the human tumor microenvironment, as Axl mediates immunoregulatory signaling in myeloid cells. Utilizing human in vitro melanoma tumor microenvironment model systems, herein we generated and analyzed single-cell RNA-sequencing data to understand the impacts of tumor microenvironment composition and tumor cell prior therapy on myeloid cell response to Axl inhibition.MethodsWe used our previously established in vitro model system to generate single-cell RNA-sequencing samples of tumor cell-macrophage co-cultures as well as tumor cell-macrophage-dendritic cell tri-cultures from four healthy buffy coat donors. We first applied LIgand-receptor ANalysis frAmework (LIANA) to our transcriptional data to infer potential intercellular crosstalk, complemented by Tensor-cell2cell to identify prevalent patterns of potential communication across the various experimental conditions. We next probed the impact of Axl inhibition on intercellular communication with orthogonalized partial least squares discriminant modeling. Finally, we conducted gene set enrichment analysis to understand how each cell type’s state and signaling activity was impacted by each axis of experimental variation in our complex experimental design.ResultsTensor-cell2cell yielded eight patterns of communication in the inferred intercellular interaction data, and these eight patterns were paired based on their communicating cell types as well as their shared top ligand-receptor interactions. These pairings demonstrated the impact of including dendritic cells in the model system and highlighted coordination among the myeloid cells. The specific effects of Axl inhibition on potential intercellular crosstalk were clarified with supervised modeling, demonstrating impaired tumor cell invasion and migration programs, as expected. Intriguingly, gene set enrichment analysis revealed that tumor cell prior injury with standard-of-care therapy impacted myeloid cell interferon signaling in response to Axl inhibition.ConclusionsIn this work, we demonstrated that tumor microenvironment cellular composition as well as prior treatment with standard-of-care therapy modulate myeloid cell response to Axl inhibition in human in vitro melanoma model systems. While the inferred effects of Axl inhibition on intercellular crosstalk, cell state, and cell activity were largely consistent with prior literature, the impact of tumor cell prior injury on the myeloid interferon signaling response to Axl inhibition underscores the importance of intentionally incorporating treatment history as a design parameter when building model systems. Ultimately, complex human in vitro model systems that consider the nuanced effects of prior therapy on the tumor microenvironment can complement insights gained from pre-clinical mouse studies to better evaluate and iteratively design candidate therapies.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s12195-026-00930-0. Introduction Axl, a member of the receptor tyrosine kinase family composed of Tyro3, Axl, and MerTK, is a therapeutic target actively under clinical investigation for a variety of cancer indications [1]. Axl is known to promote tumor cell proliferation, survival, migration, and invasion [2–8], and it has also been implicated as a contributor to immune evasion [9–12] and drug resistance [13–18]. For these reasons, numerous studies have investigated the effects of Axl inhibition on tumor cells. Their findings—namely, inhibitor-induced downregulation of the hallmarks of cancer progression, such as tumor cell invasion and metastasis, resistance to cell death, and drug resistance—support the clinical use of Axl-targeting therapies [1, 3, 5, 8] (Fig. 1a). One such therapy includes the first-in-class Axl-specific small molecule inhibitor bemcentinib. Unfortunately, clinical trials have yielded mixed results on the therapeutic benefits of administering bemcentinib in combination with standard-of-care treatment. For example, in the context of non-small cell lung cancer (NCT02424617), a subset of patients benefitted from combination therapy with bemcentinib with reported tumor shrinkage [19]. However, in the context of metastatic melanoma (NCT02872259), combination therapy with bemcentinib did not yield increases in patient response or survival [20].Fig. 1Overview of motivation for investigating Axl as a dual target in cancer therapy and of analyses conducted in this study. a Axl, along with MerTK and Tyro3, are receptor tyrosine kinases that facilitate macrophage and dendritic cell efferocytosis of apoptotic cells and cellular debris. In tumor cells, Axl inhibition is known to suppress the hallmarks of cancer. In immune cells, Axl inhibition is hypothesized to suppress immunosuppression and promote anti-tumor immunity. Figure replicated in part from Datta et al. [32]. b We previously developed a human in vitro model of the melanoma tumor microenvironment and reported that “Untreated” (treatment-naïve) and “Injured” (targeted-therapy treated) tumor cells were the primary driver of macrophage state while Axl inhibition with bemcentinib was secondary to this distinction. We also showed that the addition of dendritic cells to the system resulted in a dampened macrophage response. DMSO = dimethyl sulfoxide, Bem = bemcentinib. Figure adapted from Datta et al. [32]. c In this current work, we generated scRNA-seq data to elucidate differences in ligand-receptor cell-cell communication as well as differences in cell signaling pathway activation that may underlie our previously reported observations. Cell schematics and ligand-receptor binding schematic created with BioRender.com Overview of motivation for investigating Axl as a dual target in cancer therapy and of analyses conducted in this study. a Axl, along with MerTK and Tyro3, are receptor tyrosine kinases that facilitate macrophage and dendritic cell efferocytosis of apoptotic cells and cellular debris. In tumor cells, Axl inhibition is known to suppress the hallmarks of cancer. In immune cells, Axl inhibition is hypothesized to suppress immunosuppression and promote anti-tumor immunity. Figure replicated in part from Datta et al. [32]. b We previously developed a human in vitro model of the melanoma tumor microenvironment and reported that “Untreated” (treatment-naïve) and “Injured” (targeted-therapy treated) tumor cells were the primary driver of macrophage state while Axl inhibition with bemcentinib was secondary to this distinction. We also showed that the addition of dendritic cells to the system resulted in a dampened macrophage response. DMSO = dimethyl sulfoxide, Bem = bemcentinib. Figure adapted from Datta et al. [32]. c In this current work, we generated scRNA-seq data to elucidate differences in ligand-receptor cell-cell communication as well as differences in cell signaling pathway activation that may underlie our previously reported observations. Cell schematics and ligand-receptor binding schematic created with BioRender.com We hypothesize that these varied clinical results may be explained in part by the effects of Axl inhibition on other cell types, including those in the myeloid compartment of the tumor microenvironment. Indeed, Axl is recognized as a dual target in cancer therapy due to its role as an innate immune checkpoint [21, 22] (Fig. 1a). Specifically, Axl signaling—along with MerTK and Tyro3 signaling—mediates the nonimmunogenic clearance of apoptotic cells, a biological process known as efferocytosis [23, 24]. Following Axl signal transduction, macrophages and dendritic cells dampen their inflammatory signaling and adopt an anti-inflammatory, pro-resolution state that is conducive to maintaining homeostasis [25–27]. We therefore hypothesize that Axl inhibition induces macrophage and dendritic cell activation. Given that the activation of these myeloid cells is a prerequisite to the T cell activation and recruitment required for durable anti-tumor immunity, we anticipate that Axl inhibition will promote an anti-tumor immune response and, ultimately, tumor cell death [28]. In vivo mouse studies support this hypothesis and have demonstrated increased intra-tumoral T cell infiltration and decreased tumor burden following Axl inhibition [10, 29–31]. Although these pre-clinical studies highlight the potential of Axl-targeted therapies to be both immunostimulatory and tumoricidal, there is a clear gap in clinical translation, as demonstrated by the mixed therapeutic efficacy of bemcentinib in the clinic [19, 20]. Thus, in this study we seek to explore the effects of Axl inhibition on myeloid cell signaling and state under physiologically relevant conditions, such as those presented by a human in vitro model system, to complement the insights gained from in vivo studies. We previously established a human in vitro model of the melanoma tumor microenvironment to study the effects of Axl inhibition on myeloid cells [32] (Fig. 1b). We used this model to compare the tumor microenvironments of treatment-naïve (“Untreated”) and targeted therapy-treated (“Injured”) tumors. We found that these tumor cell states conferred distinct immune cell profiles, ultimately affecting myeloid cell response to Axl inhibition with bemcentinib. We also demonstrated that the addition of dendritic cells to the system dampened macrophage response, suggesting that the cellular composition of the system also modulates responses to Axl inhibition. Although our previous work characterizes our in vitro system as a robust model to study myeloid responses to Axl inhibition in the melanoma tumor microenvironment, an investigation beyond this characterization is warranted to understand how Axl inhibition might impact cell-cell communication and signaling pathway activation. Herein, we generated single-cell RNA-sequencing (scRNA-seq) samples from our human in vitro model system to elucidate the differences in cell signaling that might underlie our previously reported changes in immune cell state (Fig. 1c). Specifically, to understand how ligand-receptor cell-cell communication may be impacted by tumor cell prior injury, the incorporation of dendritic cells into the model system, and Axl inhibition, we first employed LIgand-receptor ANalysis frAmework (LIANA) to calculate inferred communication scores between cells [33]. We then used these communication scores as input to Tensor-cell2cell, an analysis method that identifies biologically interpretable patterns of inferred ligand-receptor cell-cell communication that are prominent in the dataset [34]. Tensor-cell2cell is particularly well-suited for our dataset because it is amenable to complex experimental designs—an advantageous attribute given that our dataset was not solely a pairwise comparison between one treatment group and one control group. To further probe ligand-receptor cell-cell communication specifically impacted by Axl inhibition, we then performed orthogonalized partial least squares discriminant analysis (O-PLSDA) modeling on the LIANA communication scores. Finally, to understand how cell signaling activity was impacted by each axis of experimental variation, we conducted gene set enrichment analysis (GSEA) separately on the tumor cells, macrophages, and dendritic cells in our dataset. Taken together, we demonstrate that cellular composition of the in vitro model system as well as prior targeted-therapy treatment of tumor cells impact cell signaling response to Axl inhibition. Materials and Methods Cell Culture and Reagents The A375 human malignant melanoma cell line was purchased from ATCC (CRL-1619) and grown in DMEM (ThermoFisher 11965092) supplemented with 10% FBS (ThermoFisher 16000044), 1% pen/strep (ThermoFisher 15140148), and GlutaMAX (ThermoFisher 35050061). A375s were detached with 0.25% Trypsin-EDTA (ThermoFisher 25200056) and not used beyond passage 15. Healthy adult male buffy coats were obtained from Research Blood Components (Table S1). Peripheral blood mononuclear cells (PBMCs) were isolated from buffy coats via density gradient centrifugation with Lymphoprep (STEMCELL 18061). PBMCs were washed twice with phosphate-buffered saline (PBS), followed by monocyte isolation with EasySep Human CD14 Positive Selection Kit II (STEMCELL 19359). Monocytes were then cultured in RPMI 1640 (ThermoFisher 32404014) supplemented with 10% heat-inactivated FBS (ThermoFisher 10082147), GlutaMAX, and HEPES (ThermoFisher 15630080). Human-monocyte derived macrophages were differentiated from monocytes with 25ng/mL M-CSF (BioLegend 574804) over 6 days, with a media change on day 3. Human-monocyte derived dendritic cells were differentiated from monocytes with 20 ng/mL IL-4 (Peprotech 200-04) and 40 ng/mL GM-CSF (BioLegend 572903) over 8 days with a media change on days 3 and 6. Following differentiation, macrophages, which are adherent, were washed with PBS, detached with 4mM EDTA (Corning 46-034-CI), and re-seeded in fresh media for experiments. Dendritic cells, which grow in suspension, were centrifuged at 500g for 5 minutes and resuspended in fresh media prior to re-seeding for experiments. Experiments were conducted in RPMI 1640 supplemented with heat-inactivated FBS, GlutaMAX, and HEPES. All cells were maintained in tissue culture-treated plastic and incubated at 37°C with 5% CO2. Trametinib (S2673), dabrafenib (S2807), and bemcentinib (S2841) were purchased from SelleckChem. They were used at doses of 15 nM, 150 nM, and 1 μM, respectively. Dimethyl sulfoxide (DMSO, Sigma-Aldrich D2650) was used as a dose-matched vehicle control. Co-Culture and Tri-Culture Experiments As previously described [32], macrophages were washed with PBS and treated with 1 μM bemcentinib or dose-matched DMSO for 4 hours after which they were washed again with PBS. Untreated or injured A375s were generated and added to the macrophages in their respective conditioned media. Briefly, untreated A375s were treated with DMSO for 24 hours followed by 24 hours of rest. In contrast, injured A375s were treated with trametinib + dabrafenib for 24 hours followed by 24 hours of rest. It should be noted that the term “injured” reflects treatment with trametinib + dabrafenib and should not be misconstrued as prescribing anthropomorphic features to the A375s. Bemcentinib or DMSO was replenished following the addition of the untreated or injured A375s to the macrophages. After 24 hours of co-culture, the experiments were either taken down or donor-matched dendritic cells were added to each well to begin tri-culture (Fig. 2a). Tri-culture experiments were taken down 24 hours after the addition of dendritic cells. Cells were collected and immediately processed for single-cell RNA-sequencing, as outlined below, ultimately resulting in eight distinct experimental conditions (Fig. 2b). These experiments were conducted in 12-well plates with 150,000 macrophages, 50,000 A375 cells, and 75,000 dendritic cells.Fig. 2Overview of the experimental design and conditions in the single-cell RNA-sequencing dataset. a Macrophages were pre-treated for 4 hours with the Axl-specific small molecule inhibitor bemcentinib or DMSO vehicle control. Untreated and injured A375 melanoma cells were harvested and added to the macrophages, and the inhibitors were replenished. Cells were collected and prepared for single-cell RNA-sequencing following 24 hours of co-culture (top). Donor-matched dendritic cells were added to the macrophage-tumor cell co-cultures, and cells were collected and prepared for single-cell RNA-sequencing following 24 hours of tri-culture (bottom).Adapted from Datta et al. [32]. b This experimental design resulted in a total of eight experimental conditions: 2 culture types (Co = co-culture or Tri = tri-culture) by 2 A375 states (untreated or injured) by 2 inhibitor treatments (Bem = bemcentinib or DMSO vehicle). A total of 32 samples were generated: 8 experimental conditions across 4 independent donors (A, B, C, D). Created with BioRender.com Overview of the experimental design and conditions in the single-cell RNA-sequencing dataset. a Macrophages were pre-treated for 4 hours with the Axl-specific small molecule inhibitor bemcentinib or DMSO vehicle control. Untreated and injured A375 melanoma cells were harvested and added to the macrophages, and the inhibitors were replenished. Cells were collected and prepared for single-cell RNA-sequencing following 24 hours of co-culture (top). Donor-matched dendritic cells were added to the macrophage-tumor cell co-cultures, and cells were collected and prepared for single-cell RNA-sequencing following 24 hours of tri-culture (bottom). Adapted from Datta et al. [32]. b This experimental design resulted in a total of eight experimental conditions: 2 culture types (Co = co-culture or Tri = tri-culture) by 2 A375 states (untreated or injured) by 2 inhibitor treatments (Bem = bemcentinib or DMSO vehicle). A total of 32 samples were generated: 8 experimental conditions across 4 independent donors (A, B, C, D). Created with BioRender.com Single-Cell RNA-Sequencing Sample Preparation and Data Acquisition Co-culture samples for each donor were generated by pooling 6 wells across two 12-well plates (3 wells per plate) together, and tri-culture samples for each donor were generated by pooling 3 wells across one 12-well plate. After pooling the cells within each condition, samples were prepared as instructed in the Single Cell Gene Expression Flex Fixed RNA Profiling demonstrated protocols (10x Genomics Rev D) with the Chromium Next GEM Single Cell Fixed RNA Sample Preparation Kit (10x Genomics 1000414). Briefly, after cells were harvested and pooled, they were washed with PBS + 0.04% UltraPure BSA (ThermoFisher AM2616). The cell pellets were then fixed with Fixation Buffer overnight at 4°C. The next day, following centrifugation, the cell pellets were resuspended in Quenching Buffer and counted. Enhancer and glycerol (Sigma-Aldrich G5516) were then added to each sample and the samples were stored at − 80°C prior to submission to the MIT BioMicro Center for sequencing with the Element AVITI. Data Preprocessing and Integration Gene expression reads were aligned to the reference genome GRCh38-3.0.0 using the cellranger multi pipeline from the CellRanger software v8.0.1 (10x Genomics) to generate cell-by-gene matrices. Downstream analyses were performed using the Scanpy package (v1.10.0) in Python [35] and the Seurat package (v5.1.0) in R [36]. Single cells were filtered to remove doublets, dead cells, and otherwise low-quality cells. Cells with total transcript counts less than 2,000 or greater than a calculated upper bound, with total detected genes less than 300, or with mitochondrial gene counts exceeding 10% were removed from the dataset for downstream analyses. For data from each experimental condition and donor, the calculated upper bound on transcript counts was defined as the 75th quartile of total counts per cell across single cells + 1.5 * interquartile range, and ranged from 10,000 to 40,000 total transcript counts (median ~ 17,000). Following quality control filtration, the data were limited to the top 5,000 highly variable genes (HVGs) for all subsequent analyses. Data integration based on donor followed by dimensionality reduction and visualization of the data were then completed with the Scanpy implementations of Harmony [37] and Uniform Manifold Approximation and Projection (UMAP) [38], respectively. It is important to note that Harmony does not alter the data; rather, the data integration is done on principal components so that donor effects do not dominate the dimensionality reductions [37]. Data Clustering and Annotation Data were clustered using Seurat’s FindClusters function with a resolution parameter of 0.22. This clustering resolution parameter was selected based on the stability of the clusters that resulted, as assessed by generating a clustering tree (Fig. S1) with the clustree package [39] (v0.5.1) in R. To assign annotations to these 8 clusters, we used Seurat’s FindAllMarkers function to identify the top 10 differentially expressed genes in each cluster (padj < 0.05 and average log2 fold change > 1) (Fig. S2). We used Seurat’s AddModuleScore function to assess expression of M1 and M2 macrophage gene signatures, according to gene signatures defined in Oshi et al. [40] (Fig. S3). Cell-cell Communication Inference Analysis Once annotated, samples were filtered to only include single cells belonging to clusters that were expected to be in the sample based on experimental condition, e.g., co-culture injured samples were filtered to only include cells belonging to the Macrophage 1 and Injured Tumor Cell clusters (see Results for cluster annotation definitions). Following this filtration, cells were relabeled as “Macrophage”, “Tumor Cell”, or “Dendritic Cell” for subsequent ligand-receptor cell-cell communication analysis. It should be noted that this filtering and relabeling was only employed for cell-cell communication inference analysis, i.e., this filtering and relabeling was not applied to the data prior to differential gene expression analysis or gene set enrichment analysis. We used LIgand-receptor ANalysis frAmework (LIANA) [33] (v0.1.14), a method which was specifically designed to generate consensus cell-cell communication inference across commonly existing methods, to generate communication scores. Of the 4,701 ligand-receptor interactions in LIANA’s “Consensus” resource, 399 interactions were present in our dataset. We then used these scores from LIANA as input to Tensor-cell2cell [34] (v0.7.4), which applies tensor decomposition to identify patterns of communication prevalent in cell-cell communication inference results. Importantly, Tensor-cell2cell encodes these results in the form of a 4-dimensional tensor to simultaneously consider contexts (experimental conditions), ligand-receptor interactions, sender cells, and receiver cells in the identification of patterns of communication, ultimately facilitating analysis that is not limited to solely comparing two experimental conditions. When building the tensor, we subtracted the magnitude_rank values in the LIANA results from 1, i.e., 1 - magnitude_rank, and used these values as the communication scores. Any missing cells or ligand-receptor interactions were set to 0 during tensor construction, i.e., the parameters cell_fill and lr_fill were set to 0. This resulted in a 4-dimensional tensor with 32 samples x 399 interactions x 3 sender cells x 3 receiver cells. When performing tensor decomposition, we used elbow analysis (Fig. S5A) to inform the number of factors to use in the decomposition with robust optimization, as recommended in Baghdassarian et al. [41]. It should be noted that Tensor-cell2cell utilizes non-negative PARAFAC decomposition, meaning the factors that result are constrained to non-negative values [34]. To specifically identify ligand-receptor cell-cell communication that was impacted by Axl inhibition with bemcentinib, multivariate modeling was performed in MATLAB (R2024a) using a custom pipeline. Orthogonalized Partial Least Squares Discriminant Analysis (O-PLSDA) modeling was used to discriminate between bemcentinib and DMSO based on intercellular interaction scores inferred by LIANA, similar to the approach described in [42]. The inferred interaction scores (1 - magnitude_rank) were filtered to retain only interactions with sufficient ligand and receptor expression detected across all 32 samples (n = 157 interactions). All data were log10-transformed and z-scored prior to analysis. Feature selection was performed by 100 independent iterations of least absolute shrinkage and selection operator (LASSO) regularization using the entire dataset. Features that were selected in > 90% of iterations were retained and used to train the full O-PLSDA model. To determine model accuracy, a 5-fold cross validation framework was implemented over 100 independent iterations. Within each iteration and for each cross validation fold, LASSO feature selection was performed 100 times using an 80% partition of the data. An O-PLSDA model was trained on the same data partition using features selected in > 90% of LASSO iterations. Model accuracy was determined by using the resulting O-PLSDA model to predict the remaining 20% of the data. This distribution of accuracy scores was compared against the classification accuracy of null models generated using the same 80% data partition by 10 repetitions of (i) selecting an equivalent number of random features and (ii) randomly shuffling the class labels using the same LASSO-selected features. Importantly, LASSO was unable to detect stable features at a 90% stability threshold following class label permutations. Model significance was determined by pairwise two-sample, two-sided t-tests between the accuracy distributions for the true model and each null model. Pairwise Pearson correlation tests between LASSO-selected features and all remaining features were performed in MATLAB and visualized as a correlation network using Cytoscape software (version 3.10.4). P-value corrections were performed using the Bonferroni method. Only correlations with padj < 0.01 were included in the visualized networks. Univariate statistical tests were performed by Wilcoxon rank sum test with Benjamini-Hochberg correction (FDR q-value < 0.05). Differential Gene Expression Analysis The cells were pseudo-bulked by culture (co-culture or tri-culture), A375 state (untreated or injured), inhibitor treatment (bemcentinib or DMSO vehicle), donor (A, B, C, or D), and cluster annotation with the AggregateExpression function in Seurat. Pseudo-bulked samples were only used for differential gene expression analysis if they were made up of at least 30 single cells. Differential gene expression analysis was then conducted on these pseudo-bulked samples with DESeq2 [43] (v1.44.0) using a design formula of “ ~ experimental variable of interest + donor” and the contrast specified as “experimental variable of interest, group 1, group 2”. A full list of the pairwise comparisons can be found in Table S2. Gene Set Enrichment Analysis Gene set enrichment analysis (GSEA) was conducted with the fgsea package [44] (v1.30.0) on the full gene list for each pairwise comparison of interest (Table S2) using the Wald statistic calculated by DESeq2 to rank the genes. The Hallmark gene sets [45] from the msigdbr package [46] (v7.5.1) were used as a reference. Gene sets were considered to be significantly enriched with a Benjamini-Hochberg FDR-adjusted q-value < 0.05, as outputted by the fgsea package. Results Macrophage Transcriptional Profile is Driven by Both Prior Targeted Therapy Treatment of Tumor Cells and the Presence of Dendritic Cells Given our primary goal to identify differences in cell communication and signaling activity that underlie our previously reported characterization of our in vitro human melanoma model system, we generated samples for scRNA-seq following our established experimental protocol [32]. Namely, we pre-treated primary macrophages with bemcentinib or DMSO vehicle control for 4 hours. We also generated “untreated” and “injured” A375 melanoma cells by incubating cells with DMSO or trametinib + dabrafenib, respectively, for 24 hours followed by 24 hours rest. As shown in our previous work, these two tumor cell populations are indeed distinct with respect to their viability and secreted molecular cues [32]. We then harvested these A375 cells and added them—in their respective conditioned media—to the macrophages, after which bemcentinib or DMSO was replenished. These co-cultures were then incubated for 24 hours prior to sample collection for scRNA-seq (Fig. 2a, top). Tri-culture samples were generated by adding donor-matched dendritic cells to the macrophage-A375 co-cultures. Similarly, samples were collected for scRNA-seq following 24 hours of tri-culture (Fig. 2a, bottom). Thus, there were a total of 8 experimental conditions, across 4 healthy buffy coat donors, for a total of 32 samples analyzed in this work (Fig. 2b). Following data preprocessing and integration, we visualized the data in UMAP space (Fig. 3) and identified eight stable clusters (Fig. S1). To assign cell identities to these clusters, we identified the top ten differentially expressed genes for each cluster (Fig. S2). We found that three of these clusters expressed genes associated with macrophages, one cluster expressed genes associated with dendritic cells, and two clusters expressed genes consistent with tumor cell gene expression signatures (Fig. 3a). In stark contrast to in vivo scRNA-seq studies, i.e., datasets derived from tissues, the clusters we identified did not strictly correspond to distinct cell lineages or cell subsets within a lineage; rather, they corresponded to experimental conditions. Such experimental condition-driven clustering has been previously reported in the literature [47] and is an unsurprising result given the limited sources of variability inherent to in vitro systems compared to tissues. In fact, this experimental condition-driven clustering enabled further refinement of cluster annotations. For example, while the two tumor cell clusters had similar gene expression signatures (Fig. S2), coloring the UMAP visualization by the state of the A375 tumor cells in the sample clarified that the two tumor cell clusters coincided with A375 injury state (Fig. 3b) and were annotated accordingly (Fig. 3e). Similarly, by also coloring the UMAP visualization based on the other two experimental variables—inhibitor treatment (Fig. 3c) and culture type (Fig. 3d)—we observed that the macrophage clusters coincided with the A375 injury state and the culture type of the sample but not by whether the sample received bemcentinib treatment. That is, the clusters annotated as Macrophage 1 and Macrophage 3 (Fig. 3e) consisted of macrophages that were in co-culture with injured and untreated A375s, respectively (Fig. 3b), while the Macrophage 2 cluster consisted of macrophages that were in tri-culture (Fig. 3d). It should be noted that these macrophage clusters did not solely have an “M1” or an “M2” macrophage gene expression signature (Fig. S3); rather, these cells expressed genes belonging to both phenotypes, as is common for tumor-associated macrophages [48].Fig. 3UMAP visualization of the single-cell RNA-sequencing data facilitated the identification of eight stable clusters, some of which coincided with experimental variables. a The transcriptional profiles of the single cells in the data yielded eight clusters. The top ten differentially expressed genes in each cluster aided cluster annotation, resulting in three distinct macrophage clusters, one dendritic cell cluster, and two distinct tumor cell clusters, all of which were pursued in subsequent analyses. b Labeling the single cells in the UMAP visualization according to whether the sample they belonged to contained untreated or injured A375s clarified the annotation of the two tumor cell clusters and highlighted the impact of A375 state on myeloid cell transcriptional signature. c Labeling the single cells in the UMAP visualization according to whether the sample they belonged to was treated with bemcentinib or DMSO vehicle demonstrated Axl inhibitor-driven transcriptional shifts within clusters but no Axl inhibitor-driven cluster definitions. d Labeling the single cells in the UMAP visualization according to whether the sample they belonged to also contained dendritic cells underscored the influence of dendritic cell inclusion on macrophage transcriptional signature. e The macrophage clusters were annotated as “Macrophage 1”, “Macrophage 2”, and “Macrophage 3”. The dendritic cell cluster was annotated as “Dendritic Cell”, and the tumor cell clusters were annotated as “Untreated Tumor Cell” and “Injured Tumor Cell”. The two remaining clusters were termed “LAMP3 Immune Cells” and “Incomplete Efferocytosis”, neither of which were further interrogated in this work UMAP visualization of the single-cell RNA-sequencing data facilitated the identification of eight stable clusters, some of which coincided with experimental variables. a The transcriptional profiles of the single cells in the data yielded eight clusters. The top ten differentially expressed genes in each cluster aided cluster annotation, resulting in three distinct macrophage clusters, one dendritic cell cluster, and two distinct tumor cell clusters, all of which were pursued in subsequent analyses. b Labeling the single cells in the UMAP visualization according to whether the sample they belonged to contained untreated or injured A375s clarified the annotation of the two tumor cell clusters and highlighted the impact of A375 state on myeloid cell transcriptional signature. c Labeling the single cells in the UMAP visualization according to whether the sample they belonged to was treated with bemcentinib or DMSO vehicle demonstrated Axl inhibitor-driven transcriptional shifts within clusters but no Axl inhibitor-driven cluster definitions. d Labeling the single cells in the UMAP visualization according to whether the sample they belonged to also contained dendritic cells underscored the influence of dendritic cell inclusion on macrophage transcriptional signature. e The macrophage clusters were annotated as “Macrophage 1”, “Macrophage 2”, and “Macrophage 3”. The dendritic cell cluster was annotated as “Dendritic Cell”, and the tumor cell clusters were annotated as “Untreated Tumor Cell” and “Injured Tumor Cell”. The two remaining clusters were termed “LAMP3 Immune Cells” and “Incomplete Efferocytosis”, neither of which were further interrogated in this work Overall, the cell types we identified in our data were as expected, and the global transcriptional shifts were consistent with our previous results [32] (Fig. 1b): co-incubated A375 injury state drives cluster-defining transcriptional differences in macrophages, and the addition of dendritic cells to the system dampens macrophage response. Together these two perturbations had strong signals in the macrophage transcriptional data, coinciding with cluster definitions and overshadowing the smaller transcriptional shifts that corresponded to Axl inhibition with bemcentinib observed within the clusters. Dendritic Cells Coordinate the Myeloid Immune Response with Macrophages Having observed the expected global transcriptional trends in the data based on our previous characterization of the in vitro system, we were interested in investigating intercellular interactions associated with these trends. To infer potential crosstalk between dendritic cells, macrophages, and tumor cells, we employed LIANA, which generates consensus cell-cell communication inference results across existing, commonly-used methods [33]. Importantly, because the annotated clusters coincided with the A375 injury state and the culture type of the sample, we only considered cell identities that were possible within a given sample, e.g., co-culture injured samples only consisted of cells belonging to the Injured Tumor Cell cluster or the Macrophage 1 cluster (Fig. 4a). We coupled our LIANA-derived cell-cell communication networks with Tensor-cell2cell, a method that strategically encodes intercellular interaction data in the form of a 4-dimensional tensor and then decomposes this tensor into biologically interpretable patterns of communication that are prevalent in the dataset [34]. Specifically, Tensor-cell2cell accomplishes this as follows: (1) LIANA communication scores for a given ligand-receptor interaction between all 3 cell types can be summarized in the form of a matrix, (2) the matrices for all 399 interactions in the dataset can be stacked into a tensor for a given sample, (3) all 32 sample tensors can be stacked together along a fourth dimension that encodes the experimental condition of each sample, and (4) unsupervised tensor decomposition can approximate this 4-dimensional tensor as the sum of unique patterns of communication, termed “Factors” (Fig. 4b). Each factor is described by the loadings of the various experimental conditions (“Contexts”), ligand-receptor interactions (“Interactions”), sender cells (“Senders”), and receiver cells (“Receivers”). Thus, Tensor-cell2cell enables analysis of datasets with complex experimental designs, such as the dataset in this work.Fig. 4Methodology for conducting ligand-receptor cell-cell communication inference analysis with LIANA followed by identifying patterns of communication with Tensor-cell2cell. a Schematic of types of communication inferred by LIANA based on sample input data. For co-culture samples comprised of injured tumor cells, we only considered single cells annotated as Injured Tumor Cell or Macrophage 1. In contrast, we only considered single cells annotated as Untreated Tumor Cell or Macrophage 3 for co-culture samples that consisted of untreated tumor cells. In the case of tri-culture samples, single cells annotated as the appropriate tumor cell type, Macrophage 2, or Dendritic Cell were included as input to LIANA. In all cases, LIANA calculated both autocrine and paracrine communication scores between tumor cells, macrophages, and, when relevant, dendritic cells. Co = co-culture, Tri = tri-culture for culture type. Cell colors correspond to their cluster colors in Fig. 3e. Created with BioRender.com. b Tensor-cell2cell utilizes a 4-dimensional tensor to encode ligand-receptor communication scores among the cell types in the data, across all the samples in the data. Specifically, (1) for a given ligand-receptor interaction, communication scores amongst all possible sender and receiver cell type pairs are encoded into a matrix. (2) Matrices for all ligand-receptor interactions in the data are stacked into a tensor to encompass all the communication scores within a specific sample. (3) These tensors are stacked along a fourth dimension to represent all of the samples—and their specific contexts (experimental conditions)—in the data. To clarify this fourth dimension, colors corresponding to specific contexts are shown. This amounted to a 4-dimensional tensor with 32 samples x 399 interactions x 3 senders x 3 receivers. (4) Tensor-cell2cell then employs unsupervised tensor decomposition to approximate the 4-dimensional tensor as the sum of a series of the outer-product of rank-one tensors, termed factors. The loadings for each dimension in a factor enable interpretation of the unique pattern of communication captured by that factor, e.g., Factor 1 described paracrine and autocrine communication amongst dendritic cells and macrophages in tri-culture samples Methodology for conducting ligand-receptor cell-cell communication inference analysis with LIANA followed by identifying patterns of communication with Tensor-cell2cell. a Schematic of types of communication inferred by LIANA based on sample input data. For co-culture samples comprised of injured tumor cells, we only considered single cells annotated as Injured Tumor Cell or Macrophage 1. In contrast, we only considered single cells annotated as Untreated Tumor Cell or Macrophage 3 for co-culture samples that consisted of untreated tumor cells. In the case of tri-culture samples, single cells annotated as the appropriate tumor cell type, Macrophage 2, or Dendritic Cell were included as input to LIANA. In all cases, LIANA calculated both autocrine and paracrine communication scores between tumor cells, macrophages, and, when relevant, dendritic cells. Co = co-culture, Tri = tri-culture for culture type. Cell colors correspond to their cluster colors in Fig. 3e. Created with BioRender.com. b Tensor-cell2cell utilizes a 4-dimensional tensor to encode ligand-receptor communication scores among the cell types in the data, across all the samples in the data. Specifically, (1) for a given ligand-receptor interaction, communication scores amongst all possible sender and receiver cell type pairs are encoded into a matrix. (2) Matrices for all ligand-receptor interactions in the data are stacked into a tensor to encompass all the communication scores within a specific sample. (3) These tensors are stacked along a fourth dimension to represent all of the samples—and their specific contexts (experimental conditions)—in the data. To clarify this fourth dimension, colors corresponding to specific contexts are shown. This amounted to a 4-dimensional tensor with 32 samples x 399 interactions x 3 senders x 3 receivers. (4) Tensor-cell2cell then employs unsupervised tensor decomposition to approximate the 4-dimensional tensor as the sum of a series of the outer-product of rank-one tensors, termed factors. The loadings for each dimension in a factor enable interpretation of the unique pattern of communication captured by that factor, e.g., Factor 1 described paracrine and autocrine communication amongst dendritic cells and macrophages in tri-culture samples Given that macrophage cluster identities and tumor cell cluster identities were limited to specific experimental conditions, applying Tensor-cell2cell to LIANA results that retained these distinct cluster annotations yielded patterns of communication that were experimental condition-specific (Fig. S4). In order to instead identify patterns of signaling that were prevalent across the experimental conditions, we considered all macrophage clusters as “Macrophage” and all tumor cell clusters as “Tumor Cell”. This then yielded eight factors in the Tensor-cell2cell results (Fig. 5a, Fig. S5, and Supplemental Data 1–3). Factors 1, 4, and 5 corresponded to patterns of inferred cell-cell communication that were specific to tri-culture samples, while Factors 6, 7, and 8 corresponded to patterns of inferred cell-cell communication that were primarily observed in co-culture (Fig. S5b). Finally, Factors 2 and 3 corresponded to patterns of communication that were not specific to co-culture or tri-culture samples but, instead, were unique to samples that consisted of untreated tumor cells or injured tumor cells, respectively. Considering sender and receiver cells, pairs of factors emerged: (1) Factors 1 and 6 both corresponded to cell-cell communication between immune cells; (2) Factors 2 and 3 both corresponded to communication between tumor cells; (3) Factors 4 and 7 both described signaling from tumor cells to myeloid cells; and (4) Factors 5 and 8 both captured signaling from myeloid cells to tumor cells (Fig. 5a). Furthermore, hierarchical clustering (Fig. 5b) and Spearman correlations (Fig. 5c) of the ligand-receptor interactions with loadings among the top ten percent of loadings in at least one factor provided mathematical support for these factor pairings.Fig. 5Patterns of communication identified with Tensor-cell2cell demonstrated cooperation and coordination amongst macrophages and dendritic cells in mediating the myeloid immune response. a Tensor-cell2cell identified eight unique patterns of communication in the ligand-receptor cell-cell communication inference analysis communication scores from LIANA. These factors described communication specific to certain experimental conditions as well as certain interacting cell types. The factors were analyzed in pairs due to parallels in the sender-receiver cell types, e.g., both Factor 1 and Factor 6 described signaling specific to myeloid cells. Co = co-culture, Tri = tri-culture for culture type; U = untreated, I = injured for A375 state; Bem = bemcentinib, DMSO = DMSO vehicle for inhibitor treatment. Cell schematics created with BioRender.com. b Hierarchical clustering of the factors based on their loadings for the top ligand-receptor interactions in the dataset further mathematically justified the pairing of factors as shown in a. Additionally, this clustering demonstrated the general maintenance of signaling patterns present in the co-culture samples (Factors 6, 7, and 8) by their tri-culture counterparts (Factors 1, 4, and 5). Ligand-receptor interaction loadings were z-scored across the eight factors for this visualization. c Spearman correlation values of each factor’s loadings for these top ligand-receptor interactions Patterns of communication identified with Tensor-cell2cell demonstrated cooperation and coordination amongst macrophages and dendritic cells in mediating the myeloid immune response. a Tensor-cell2cell identified eight unique patterns of communication in the ligand-receptor cell-cell communication inference analysis communication scores from LIANA. These factors described communication specific to certain experimental conditions as well as certain interacting cell types. The factors were analyzed in pairs due to parallels in the sender-receiver cell types, e.g., both Factor 1 and Factor 6 described signaling specific to myeloid cells. Co = co-culture, Tri = tri-culture for culture type; U = untreated, I = injured for A375 state; Bem = bemcentinib, DMSO = DMSO vehicle for inhibitor treatment. Cell schematics created with BioRender.com. b Hierarchical clustering of the factors based on their loadings for the top ligand-receptor interactions in the dataset further mathematically justified the pairing of factors as shown in a. Additionally, this clustering demonstrated the general maintenance of signaling patterns present in the co-culture samples (Factors 6, 7, and 8) by their tri-culture counterparts (Factors 1, 4, and 5). Ligand-receptor interaction loadings were z-scored across the eight factors for this visualization. c Spearman correlation values of each factor’s loadings for these top ligand-receptor interactions We were next interested to further interpret these factor pairs. Intriguingly, Factors 1 and 6, shared many ligand-receptor signaling mechanisms, despite Factor 6 describing macrophage-specific signaling primarily in co-culture and Factor 1 describing macrophage-dendritic cell crosstalk in tri-culture (Fig. 5). This result supported our hypothesis that the addition of dendritic cells to the system dampens macrophage response because these myeloid cells cooperate and coordinate the immune response. To further pursue this hypothesis, we focused on the top ligand-receptor interactions in Factors 1 and 6. While Factor 1 and Factor 6 both had a subset of non-redundant, unique top interactions, they shared 45% their top interactions (Fig. S6 and Table S3). Such conserved signaling was also observed in the other factor pairs that involved the myeloid cells: Factors 5 and 8 shared 48% of their top interactions (Fig. S7 and Table S4) and Factors 4 and 7 shared 60% of their top interactions (Fig. S8 and Table S5). Indeed, while the shared top ligand-receptor interactions in a factor pair were comparably expressed in co-culture samples as well as tri-culture samples, e.g., LGALS9 → HAVCR2 (Fig. S6c), APOE → SCARB1 (Fig. S7c), and MIF → TNFRSF14 (Fig. S8c), the presence of non-redundant, unique interactions was impacted by the inclusion of dendritic cells in the system. For example, the expression of CCL14 → CCR1 (Fig. S6d), F13A1 → ITGA4 (Fig. S7d), and SPARC → FGFR1 (Fig. S8d) signaling was limited to tri-culture samples, meaning these ligand-receptor interactions required the presence of dendritic cells. Furthermore, in some cases, stronger expression was observed specifically when dendritic cells were the sender or receiver. Most notably though, the expression of ligand-receptor interactions unique to the factors describing patterns of signaling primarily prevalent in co-culture samples either had diminished signaling in tri-culture samples, e.g., CCL4 → CCR5 (Fig. S6e) and CXCL8 → SDC3 (Fig. S7e) or had signaling mediated by both macrophages and dendritic cells in tri-culture, e.g., SEMA7A → PLXNC1 (Fig. S8e). Taken together, these results support the claim that macrophages and dendritic cells coordinate their immune response, ultimately manifesting as an observed reduction in macrophage-specific signaling in tri-culture. The final pair of factors, Factors 2 and 3, captured signaling between tumor cells, with the key distinction between them being A375 state; Factor 2 captured signaling among untreated tumor cells while Factor 3 captured signaling among injured tumor cells (Fig. 5). In contrast to the previously discussed factor pairs, Factors 2 and 3 only shared 27% of their top ligand-receptor interactions (Fig. S9 and Table S6). This result is consistent with our prior characterization of our human in vitro culture system, in which we concluded that targeted therapy-treated tumor cells are distinct from treatment-naïve tumor cells—a consequential difference for the tumor microenvironment interactome and, ultimately, potential efficacy of therapeutic interventions[32]. As a whole, the Tensor-cell2cell results support the claim that the consistently observed dampened macrophage response upon the addition of dendritic cells to the system can be explained by dendritic cells and macrophages coordinating the myeloid immune response. Though the effects of dendritic cell incorporation on immune cell signaling were clearly evident, the effects of the other two experimental variables—A375 injury state and inhibitor treatment—on immune cell signaling were not as prominent. An argument could be made that the experimental conditions including injured tumor cells had loadings with lower magnitude as compared to those including untreated tumor cells, but such a difference in loadings between bemcentinib-treated samples and their DMSO vehicle-treated counterparts was not evident (Fig. 5a and Fig. S5b). Axl Inhibition Impairs Invasion Signaling Axes in Tumor Cells As we could not discern specific effects of Axl inhibition via bemcentinib on the inferred intercellular crosstalk in the unsupervised analysis method of Tensor-cell2cell, we turned to O-PLSDA, a supervised method that intentionally identifies features in the data that discriminate bemcentinib-treated samples from DMSO vehicle-treated samples. Feature selection with LASSO regularization resulted in eight ligand-receptor interactions, and the O-PLSDA model built from these features discriminated the samples based on inhibitor treatment with a cross-validation accuracy of 87.5% (Fig. 6a, b). Importantly, a stable feature set could not be extracted following random permutation of the class labels, emphasizing the robustness and biological relevance of the true reduced feature space. Seven out of these eight ligand-receptor interactions were univariately statistically significant (Fig. S10). The intercellular interactions that characterized the bemcentinib-treated samples were associated with inflammatory signaling in macrophages, i.e., MIF → TNFRSF14 [49], and cholesterol homeostasis signaling in tumor cells, i.e., APOE → SCARB1 [50, 51] (Fig. 6c). In contrast, the intercellular interactions that characterized the DMSO vehicle-treated samples were associated with migration and invasion signaling in tumor cells, e.g., COL6A1 → ITGA6 [52]. Furthermore, given that LASSO feature selection intentionally identifies a minimal set of features required to capture the variance in the data, we also interrogated the remaining ligand-receptor interactions in the data that correlated with these features. These networks of correlated signaling axes were also largely associated with programs of tumor cell migration and invasion (Fig. 6d), e.g., L1CAM → ERBB2 and L1CAM → ERBB3 [53]. Thus, supervised modeling with O-PLSDA on the intercellular interaction data demonstrated that Axl inhibition with bemcentinib impacts signaling axes associated with tumor progression, aligning with previous studies [1, 6, 7].Fig. 6Supervised modeling with O-PLSDA discriminated samples based on inhibitor treatment. a O-PLSDA scores plot from a model discriminating between bemcentinib and DMSO treatment conditions. The model was trained on features selected in at least 90% of 100 iterations of LASSO regularization. b Model accuracy was determined by 100 repetitions of 5-fold CV. The model outperformed null models with randomly selected features and models with randomly permuted class labels (p < 0.001, two-sided t-test). The dashed line indicates the no-information rate. c Bar plot shows the VIP scores for LASSO-selected features. Bars are oriented in the direction of their loadings on LV1. |VIP score| > 1 denotes features with above-average influence on model outcomes. d Pearson correlation network between LASSO-selected model features and other features in the data. Only correlations with padj < 0.01 after Bonferroni correction are shown. LASSO-selected features are colored according to their association with bemcentinib or DMSO experimental conditions. Correlated features are colored white. Edges are solid for positive correlation coefficients and dashed for negative correlation coefficients. Edge width is inversely proportional to the correlation padj. O-PLSDA, orthogonalized partial least squares discriminant analysis; LASSO, least absolute shrinkage and selection operator; LV, latent variable; CV, cross validation; VIP, variable importance in projection Supervised modeling with O-PLSDA discriminated samples based on inhibitor treatment. a O-PLSDA scores plot from a model discriminating between bemcentinib and DMSO treatment conditions. The model was trained on features selected in at least 90% of 100 iterations of LASSO regularization. b Model accuracy was determined by 100 repetitions of 5-fold CV. The model outperformed null models with randomly selected features and models with randomly permuted class labels (p < 0.001, two-sided t-test). The dashed line indicates the no-information rate. c Bar plot shows the VIP scores for LASSO-selected features. Bars are oriented in the direction of their loadings on LV1. |VIP score| > 1 denotes features with above-average influence on model outcomes. d Pearson correlation network between LASSO-selected model features and other features in the data. Only correlations with padj < 0.01 after Bonferroni correction are shown. LASSO-selected features are colored according to their association with bemcentinib or DMSO experimental conditions. Correlated features are colored white. Edges are solid for positive correlation coefficients and dashed for negative correlation coefficients. Edge width is inversely proportional to the correlation padj. O-PLSDA, orthogonalized partial least squares discriminant analysis; LASSO, least absolute shrinkage and selection operator; LV, latent variable; CV, cross validation; VIP, variable importance in projection Tumor Cell State Impacts Macrophage and Dendritic Cell Immunostimulatory Response to Axl Inhibition In parallel to the previously discussed cell-cell communication inference analyses we also conducted GSEA to understand how each experimental variable—inhibitor treatment, A375 injury state, and culture type—impacted cell state. We therefore pseudo-bulked the single-cell data for each cluster of interest (Fig. 3e) and performed differential expression (DE) analysis with DESeq2 followed by GSEA to elucidate the effects of each experimental variable on the enrichment of a variety of cellular programs. We performed a series of three DE analyses to investigate each experimental variable: (1) inhibitor treatment while holding the other two variables constant, e.g., co-culture injured bemcentinib Macrophage 1 compared to co-culture injured DMSO Macrophage 1; (2) A375 injury state while holding the other two variables constant, e.g., co-culture injured DMSO Macrophage 1 compared to co-culture untreated DMSO Macrophage 3; (3) culture type while holding the other two variables constant, e.g., tri-culture injured DMSO Macrophage 2 compared to co-culture injured DMSO Macrophage 1. This amounted to a total of four pairwise comparisons for dendritic cells and twelve pairwise comparisons each for macrophages and tumor cells (Table S2). To understand how a given gene set was enriched along these different axes of pairwise comparisons, we summarized the dendritic cell GSEA results as a square (Fig. 7a, left) and the macrophage and tumor cell GSEA results as a cube (Fig. 7a, right). Namely, each edge represents one pairwise comparison specified by the only experimental variable varying between the vertices of that edge, and when an edge had a statistically significant enrichment score for a given gene set, the dark red vertex along that edge corresponds to the condition in which it was enriched. To clarify the pairwise comparisons depicted in the cube visualization, we have included a schematic that shows the faces and vertices involved in each comparison (Fig. 7b). The edges that change color in this schematic correspond to the edges that represent the pairwise comparison in question.Fig. 7GSEA results demonstrated that tumor cells responded to Axl inhibition via bemcentinib treatment as expected: downregulating pathways associated with the hallmarks of cancer. a Dendritic cell GSEA results were visualized as a square (left). There were two axes of experimental variable comparisons: (1) untreated (U) versus injured (I) A375 state and (2) bemcentinib (Bem) vs DMSO vehicle inhibitor treatment. Macrophage and tumor cell GSEA results were visualized as a cube (right). There were three axes of experimental variable comparisons: (1) untreated versus injured A375 state, (2) bemcentinib versus DMSO vehicle inhibitor treatment, and (3) co-culture (Co) versus tri-culture (Tri). Each edge represents a pairwise comparison where only the experimental variable of interest changes along that edge, depicted as the only variable that changes in both name and color at the labeled vertices along that particular edge. b To clarify how each pairwise comparison is depicted in the cube visualization of the GSEA results, we have shown which faces and vertices are involved in a given comparison. The edges that change color correspond to the edges that represent the pairwise comparison. c Upon Axl inhibition with bemcentinib, tumor cells downregulated the E2F Targets gene set. Dark red = vertex at which the gene set was enriched; grey edge = the gene set was not found to be statistically significant for the pairwise comparison represented by that edge. Cell schematics created with BioRender.com. d Treatment with bemcentinib resulted in tumor cell downregulation of the G2M Checkpoint gene set, with the exception of untreated tumor cells in tri-culture. Colors as described in c. e Axl inhibition yielded a reduction in tumor cell expression of genes associated with epithelial mesenchymal transition, another hallmark of cancer. Colors as described in c. f Treatment with bemcentinib resulted in tumor cell down regulation of KRAS signaling, which promotes cell survival, invasion, and metastasis—which are hallmarks of cancer. Colors as described in c GSEA results demonstrated that tumor cells responded to Axl inhibition via bemcentinib treatment as expected: downregulating pathways associated with the hallmarks of cancer. a Dendritic cell GSEA results were visualized as a square (left). There were two axes of experimental variable comparisons: (1) untreated (U) versus injured (I) A375 state and (2) bemcentinib (Bem) vs DMSO vehicle inhibitor treatment. Macrophage and tumor cell GSEA results were visualized as a cube (right). There were three axes of experimental variable comparisons: (1) untreated versus injured A375 state, (2) bemcentinib versus DMSO vehicle inhibitor treatment, and (3) co-culture (Co) versus tri-culture (Tri). Each edge represents a pairwise comparison where only the experimental variable of interest changes along that edge, depicted as the only variable that changes in both name and color at the labeled vertices along that particular edge. b To clarify how each pairwise comparison is depicted in the cube visualization of the GSEA results, we have shown which faces and vertices are involved in a given comparison. The edges that change color correspond to the edges that represent the pairwise comparison. c Upon Axl inhibition with bemcentinib, tumor cells downregulated the E2F Targets gene set. Dark red = vertex at which the gene set was enriched; grey edge = the gene set was not found to be statistically significant for the pairwise comparison represented by that edge. Cell schematics created with BioRender.com. d Treatment with bemcentinib resulted in tumor cell downregulation of the G2M Checkpoint gene set, with the exception of untreated tumor cells in tri-culture. Colors as described in c. e Axl inhibition yielded a reduction in tumor cell expression of genes associated with epithelial mesenchymal transition, another hallmark of cancer. Colors as described in c. f Treatment with bemcentinib resulted in tumor cell down regulation of KRAS signaling, which promotes cell survival, invasion, and metastasis—which are hallmarks of cancer. Colors as described in c First, we sought to discern effects of bemcentinib on the transcriptional signature of the three cell types in our in vitro model system. Promisingly, in tumor cells, bemcentinib resulted in the downregulation of several pathways central to the hallmarks of cancer, such as the E2F targets (Fig. 7c), G2M checkpoint (Fig. 7d), epithelial mesenchymal transition (Fig. 7e), and KRAS signaling (Fig. 7f) Hallmark gene sets. That is, these gene sets were significantly enriched in tumor cells in the DMSO vehicle-treated samples; however, in the case of untreated tumor cells in tri-culture, the G2M checkpoint gene set was instead significantly enriched in bemcentinib-treated samples. We hypothesize this exception may be driven by the presence of dendritic cells. Otherwise, the effects of Axl inhibition with bemcentinib were as expected for the tumor cells because Axl is known to contribute to cancer progression by promoting cell proliferation and metastasis, among other cancer hallmarks, and numerous studies have demonstrated that its inhibition with bemcentinib impedes these processes [1, 6, 7, 54]. Intriguingly, the effects of Axl inhibition on the dendritic cells and macrophages in our in vitro model system were impacted by the state of the tumor cells in the sample. For instance, treatment with bemcentinib has previously been shown to result in increased interferon alpha signaling in dendritic cells [55], and such enrichment was also present in our data when dendritic cells were in culture with untreated tumor cells (Fig. 8a). Similarly, macrophages in culture with untreated tumor cells were also enriched for interferon alpha signaling upon bemcentinib treatment. However, both myeloid cell types downregulated this signaling pathway upon Axl inhibition when they were instead cultured with injured tumor cells. A similar reversal in the direction of gene set enrichment was also present in the interferon gamma response for the bemcentinib and DMSO vehicle comparison in the myeloid cells (Fig. 8b). Taken together, these results suggest that the injured tumor cells impede the immunostimulatory effects of Axl inhibition via bemcentinib on myeloid cell activity. Furthermore, similar tumor cell state-dependent pathway enrichment was observed in macrophage cholesterol homeostasis signaling (Fig. 8c) and macrophage MYC signaling (Fig. 8d). Both cholesterol homeostasis [56, 57] and MYC signaling [58] are associated with an immune-resolving macrophage state, which is consistent with decreased interferon signaling. Thus, these results further bolster the claim that targeted therapy-treated tumor cells and treatment-naïve tumor cells have distinct impacts on myeloid cell state, ultimately influencing myeloid response to Axl inhibition.Fig. 8GSEA results suggested that tumor cell prior treatment with molecular-targeted therapy may compromise the immunostimulatory effects of Axl inhibition with bemcentinib. a Upon Axl inhibition with bemcentinib, both dendritic cells and macrophages upregulated their interferon alpha signaling when in culture with untreated tumor cells but downregulated this signaling pathway when they were instead in culture with injured A375s. Dark red = vertex at which the gene set was enriched; grey edge = the gene set was not found to be statistically significant for the pairwise comparison represented by that edge. Cell schematics created with BioRender.com. b Bemcentinib treatment resulted in increased expression of genes associated with the interferon gamma response in dendritic cells and macrophages that were in culture with untreated A375s. When the immune cells were instead in culture with injured tumor cells, this signaling pathway was downregulated. Colors as described in a. c The cholesterol homeostasis pathway was downregulated in macrophages upon bemcentinib treatment when they were in tri-culture with untreated A375s but was upregulated in macrophages when they were instead in culture with injured tumor cells. Colors as described in a. d The MYC targets V2 gene set was downregulated in macrophages upon bemcentinib treatment when they were in culture with untreated A375s but was upregulated in macrophages when they were instead in culture with injured tumor cells. Colors as described in a GSEA results suggested that tumor cell prior treatment with molecular-targeted therapy may compromise the immunostimulatory effects of Axl inhibition with bemcentinib. a Upon Axl inhibition with bemcentinib, both dendritic cells and macrophages upregulated their interferon alpha signaling when in culture with untreated tumor cells but downregulated this signaling pathway when they were instead in culture with injured A375s. Dark red = vertex at which the gene set was enriched; grey edge = the gene set was not found to be statistically significant for the pairwise comparison represented by that edge. Cell schematics created with BioRender.com. b Bemcentinib treatment resulted in increased expression of genes associated with the interferon gamma response in dendritic cells and macrophages that were in culture with untreated A375s. When the immune cells were instead in culture with injured tumor cells, this signaling pathway was downregulated. Colors as described in a. c The cholesterol homeostasis pathway was downregulated in macrophages upon bemcentinib treatment when they were in tri-culture with untreated A375s but was upregulated in macrophages when they were instead in culture with injured tumor cells. Colors as described in a. d The MYC targets V2 gene set was downregulated in macrophages upon bemcentinib treatment when they were in culture with untreated A375s but was upregulated in macrophages when they were instead in culture with injured tumor cells. Colors as described in a Next, considering the pairwise comparison of the effects of A375 injury state on the transcriptional signature of the dendritic cells, macrophages, and tumor cells, we found that nearly all gene sets that were statistically significant were consistently downregulated in the injured condition (Fig. S11, Fig. S12, and Fig. S13). This is in agreement with our previous work reporting decreased molecular cue secretion with tumor cell injury [32] and supports our finding that inferred ligand-receptor cell-cell communication may be decreased in samples that contain injured, rather than untreated, tumor cells. Furthermore, the downregulation of many of these signaling pathways in tumor cells in particular is consistent with the expected effects of trametinib + dabrafenib on tumor cell survival and proliferation [59] (Fig. 7c-f and Fig. S13). However, there were certain deviations from the general observation that samples containing injured tumor cells had dampened pathway enrichment as compared to those including untreated tumor cells, two of which should be noted: the interferon alpha and gamma responses. Intriguingly, these two gene sets were upregulated in the injured condition in tumor cells (Fig. 8a, b). Such enrichment may be explained by the involvement of these pathways in cell death [60–62], given that the injured tumor cell population already consists of dead and dying cells [32]. Beyond this, dendritic cells and macrophages upregulated their interferon signaling when they were in culture with injured tumor cells as compared to when they were in culture with untreated tumor cells, specifically in DMSO vehicle-treated samples. Intuitively, such an enrichment in interferon signaling may be an expected response to the inflammatory cell death characteristic of the injured tumor cell population [63]. In contrast, when this same comparison was made in immune cells of bemcentinib-treated samples, the interferon alpha and gamma responses were instead upregulated when the myeloid cells were in culture with untreated tumor cells. Such a change in the direction of enrichment suggests that Axl inhibition via bemcentinib may undermine the immunostimulatory effects of tumor cell injury on macrophage and dendritic cell interferon signaling. Finally, in concurrence with our previous claims, the addition of dendritic cells to the system resulted in the downregulation of gene sets in macrophages (Fig. S12). This further strengthens the claim that adding dendritic cells to the system perturbs macrophage intracellular signaling and overall immune response. We also investigated the effects of adding dendritic cells to the system on the tumor cells, but they were too modest to draw any conclusions from (Fig. S13). As a whole, these GSEA results support our previous findings that dendritic cells dampen macrophage immune response, that tumor cell injury with prior targeted therapy treatment decreases cell signaling activity, and that tumor cell state shapes the myeloid cell response to Axl inhibition [32]. Discussion In this work, we identified ligand-receptor cell-cell communication patterns as well as changes in cell type-specific transcriptional signatures that result from the effects of three experimental variables—culture type, A375 injury state, and inhibitor treatment—on myeloid cell state and activity in our human in vitro melanoma model system. Both the Tensor-cell2cell results and the GSEA results demonstrated that the addition of dendritic cells to the system dampens macrophage immune response, and the Tensor-cell2cell results in particular support the notion that this dampening is likely a consequence of dendritic cells and macrophages working together to coordinate the myeloid immune response. Additionally, these results demonstrated that cell signaling and activity within samples comprised of injured A375 cells is decreased in magnitude as compared to their untreated A375-containing counterparts. Indeed, we previously observed that these distinct tumor cell populations propagated to differences in myeloid cell state and activity, with consequential effects on the myeloid cell response to Axl inhibition with bemcentinib [32]. Although this trickle-down effect was not pervasive in the ligand-receptor cell-cell communication inference analysis presented in this work, the GSEA results suggest that tumor cell injury may impact myeloid cell response to bemcentinib, particularly with respect to interferon signaling signatures. While the Tensor-cell2cell results yielded factors involving myeloid cells that paralleled each other, i.e., Factors 1 and 6, Factors 5 and 8, and Factors 4 and 7, each factor within a pair was comprised of its own unique top ligand-receptor interactions. That is, the shared top ligand-receptor interactions support the claims that signaling is maintained upon the addition of dendritic cells to the system and that macrophages and dendritic cells work together in conducting the myeloid immune response. However, the prevalence of unique top ligand-receptor interactions indicates that factors that paralleled each other were not entirely redundant. In other words, the inclusion of dendritic cells to create a tri-culture model is indeed a perturbation to the system that alters the signaling landscape. Though macrophages and dendritic cells have predominantly been considered to be distinct with independent roles in the immune system, the field has increasingly begun to recognize and appreciate that these myeloid cells rely on each other to coordinate the innate immune response [64]. Our results align with this view and demonstrate communication between these cells, adding to this growing body of literature [65–67]. Importantly, such communication and coordination may ultimately impact the efficacy of therapeutic interventions, and the intricacies of macrophage-dendritic cell communication in response to therapy warrant further investigation. Though Tensor-cell2cell did not identify bemcentinib-specific patterns of communication, this result should not be misconstrued as the Axl inhibitor having no impact on cell-cell communication. Rather, the lack of a bemcentinib-specific pattern can be explained by the fact that bemcentinib treatment did not have strong, cluster-defining signal in the data, and the tensor decomposition employed by Tensor-cell2cell is an unsupervised method that captures patterns associated with the most prominent signal in the data. In other words, the patterns of communication identified by Tensor-cell2cell correlate with experimental variables that provide sources of strong signal in the data. In contrast, the supervised method of O-PLSDA was able to clearly distinguish bemcentinib-treated samples from DMSO vehicle-treated samples. The ligand-receptor interactions that contributed to the discrimination between these samples primarily represented signaling programs for tumor cell migration and invasion—hallmarks of cancer which bemcentinib is known to inhibit [1, 6, 7]. Thus, the effects of Axl inhibition on inferred intercellular communication align with the literature. Finally, both the Tensor-cell2cell results and the GSEA results support the characterization of treatment-naïve tumor cells and targeted-therapy treated tumor cells as distinct populations—a distinction with potentially consequential impact on therapeutic efficacy. Indeed, though the impact of bemcentinib was as expected for most signaling pathways as assessed by GSEA, the interferon pathways in myeloid cells were influenced by the state of the A375 tumor cells in the samples. Given the pleiotropic role of interferon gamma signaling in melanoma progression [68], it is plausible that in the treatment-naïve melanoma tumor microenvironment bemcentinib may promote an anti-tumor immune response and be therapeutically beneficial while, in a tumor microenvironment compromised by prior targeted therapy treatment, Axl inhibition may not contribute to a robust anti-tumor immune response. To that end, the contribution of type I interferon signaling, i.e., interferon alpha signaling, in promoting an anti-tumor immune response is widely recognized [69, 70] and has been recently highlighted in potentiating beneficial outcomes of immune checkpoint inhibitor therapy [71]. Thus, the downregulation of the interferon alpha response upon Axl inhibition in myeloid cells that were cultured with injured tumor cells further supports the claim that injured tumor cells may compromise the development of durable anti-tumor immunity. Though our transcriptional data suggest tumor cell injury with prior molecular-targeted therapy may impede the otherwise expected immunostimulatory effects of bemcentinib, a functional experiment is required to test this hypothesis. Specifically, an experiment extending our current in vitro model to include T cells and assess their activity and function would enable the validation of this hypothesis and would be of interest in future work. Importantly, in contrast to the sudden exposure of myeloid cells to the injured A375 population that was intended to model a tumor microenvironment shaped by prior standard-of-care therapy in this in vitro system, the concentration of apoptotic and necrotic cellular debris encountered by myeloid cells likely gradually accumulates over the course of treatment in vivo. Thus, tuning the concentration of cellular debris in future studies would be of interest. Furthermore, the question of whether standard-of-care therapy, i.e., trametinib + dabrafenib, impedes the efficacy of bemcentinib in the clinical setting remains to be determined. Validating this hypothesis with clinical trial results may prove difficult because the patients received standard-of-care therapy and bemcentinib simultaneously, and there were too few evaluable patients to draw conclusions on therapeutic efficacy [20]. Beyond this, the question of whether the findings presented in this work generalize to other cancer indications warrants further investigation. Taken together, our previous characterization of our model system [32] and our present investigation of transcriptional changes underlying that characterization demonstrate that cellular composition and state, i.e., the initial conditions of the system, are not negligible variables in modeling the effects of Axl inhibition with bemcentinib on the tumor immune microenvironment. To that end, candidate therapies may not be the first-line therapy administered to a patient, meaning pre-clinical models of candidate therapies that do not emulate clinically relevant initial conditions may undermine their projected clinical success. In this regard, the potential of in vitro model systems to evaluate candidate therapies and identify otherwise unexpected complications has been underappreciated, and the potential of well-designed systems to help inform the timing of multi-arm treatments has been overlooked. There are numerous design considerations that can be tuned in the in vitro setting, such as which cells comprise the system and at what ratios, which are not easily controlled in pre-clinical in vivo mouse studies, as well as testing different dosing regimens, which is limited by financial constraints in pre-clinical in vivo mouse studies. Thus, by carefully tuning relevant design parameters, human in vitro models have the potential to complement pre-clinical mouse studies to further our understanding of why promising therapeutic candidates fall short in the clinic and in turn facilitate iterative therapeutic design to increase the likelihood of clinical success. Of course, relevant clinical data would be required to fine-tune and appraise the predictive power of complex human in vitro model systems, but it is clear from our work that these systems should be leveraged alongside pre-clinical in vivo mouse studies when evaluating candidate therapies. Supplementary Information Below is the link to the electronic supplementary material.Supplementary file1 (DOCX 5861 KB)Supplementary file2 (XLSX 20 KB)Supplementary file3 (XLSX 134 KB)Supplementary file4 (XLSX 13 KB) Supplementary file1 (DOCX 5861 KB) Supplementary file2 (XLSX 20 KB) Supplementary file3 (XLSX 134 KB) Supplementary file4 (XLSX 13 KB)