This study was conducted in accordance with the ethical principles stated in the Belmont Report and the US Common Rule. Tumor tissue was collected prospectively and recalled from archived tissue blocks from patients with osteosarcoma at the Johns Hopkins Hospital (Baltimore, Maryland, USA). Demographic and clinical outcomes data were obtained by retrospective chart review of the electronic medical record. This study was conducted under a Johns Hopkins Institutional Review Board approved protocol (FWA00005752) with a waiver of consent for archived tissues and patients gave written informed consent for prospective tumor collection. All samples were obtained in accordance with the Health Insurance Portability and Accountability Act.

Sixty-six formalin-fixed paraffin-embedded tissue blocks (28 untreated bone, 12 bone post-chemotherapy, 25 lung metastases, 1 soft tissue recurrence/metastasis) from 31 patients with osteosarcoma were cut into 5 µm sections and mounted onto glass slides. Each specimen was stained for H&E according to standard protocols, and for CD3, CD8, Foxp3, CD163, programmed cell death 1 (PD-1), and PD-L1 per online supplemental table 1). Additional staining for colony-stimulating factor 1 receptor (CSF1R), T-cell immunoglobulin and mucin domain-containing protein 3 (TIM-3), lymphocyte-activation gene 3 (LAG-3) and indoleamine 2,3-dioxygenase (IDO1) was performed on osteosarcoma PMs (online supplemental table 11). Whole slides were then digitally scanned to a magnification of 20x (Scanscope XT), analyzed (Halo imaging analysis software; Indica Labs, Corrales, New Mexico, USA) and densities scored as cells/mm² (CD3, CD8, Foxp3, CD163, PD-1, LAG-3) or area percentage (PD-L1, CSF1R, TIM-3, IDO1). PD-L1 positivity was scored as a percentage of the total tissue area, including both tumor and immune cells, given the expression of PD-L1 on both cell populations in our samples. In PMs, the tumor-normal lung ‘interface’ (PM interface) region (when present in the available slide) was defined as a 400 µm width region where the tumor meets the lung tissue. Slides were reviewed with a pathologist (RA and EDT) to demarcate areas of tumor versus normal tissue and to confirm our digital image analysis.

Formalin-fixed paraffin-embedded tissue blocks and H&E-stained tissue sections (10 µm, mounted using RNA precautions) of 16 PMs were used for the laser capture microdissection (Leica LMD 7000 system). For each sample, tissues were microdissected from the PM interface region as well as the tumor interior (PM interior) region (online supplemental figure S1). For comparison, a single bone tumor and a single soft tissue recurrence/metastasis were also dissected in a similar manner. RNA was isolated with High Pure Paraffin Kit (Roche) following the manufacturer’s instructions.

Transcriptome sequencing was performed using the AmpliSeq (Ion Torrent) RNA next-generation sequencing platform. This genomic analysis plugin aligns the raw sequence reads to a human reference genome that contains 20 802 RefSeq transcripts (hg19 Ampliseq Transcriptome_ERCC_V1.fasta) using the Torrent Mapping Alignment Program. Then, the number of reads mapped per gene is counted to generate raw counts files and normalized reads per gene per million mapped reads (RPM) files. On-target reads of greater than 60% were required for a library to pass quality control. Data were analyzed by unsupervised principal component analysis. Cell-type identification by estimating relative subsets of RNA transcripts (CIBERSORT) deconvolution with the leukocyte gene signature matrix LM2221 and xCell22 were used to analyze immune and other cell composition of our data. We also used gene set enrichment analysis (GSEA, Broad Institute) to analyze whether published and validated immune gene sets23–25 were significantly enriched in either the PM interface or PM interior region. A false discovery rate (FDR) of less than 25% was used as a cut-off for a gene set to be significantly enriched, and all gene sets discussed have FDR 15% or lower. xCell22 (UCSF Institute for Computational Health Sciences) is a high-resolution gene-signature-based method for cell-type enrichment for up to 64 cell types. We used xCell R package. xCell was run with the cell.types.use parameter to avoid overcompensation by the spillover correction.

cell.types.use was set to B cells, T cells, and myeloid cells to deconvolute the cellular heterogeneity within the TME from RNA sequencing data. For module visualization, we used Morpheus (Broad Institute), data were loaded as log transformed and normalized by subtracting each value from the row mean and divided by the row SD. We also specifically used Morpheus to display genes which contribute to GSEA leading edge subsets using gene expression levels to generate heatmaps. The UCSC Xena functional genomics explorer was used for generating box plots of myeloid-derived suppressor cell (MDSC) genes of interest.26

Twenty-five freshly harvested osteosarcoma tumors (15 primary bone tumors and 10 PMs) were digested using an enzymatic cocktail (0.1% DNase I and Liberase 400 µg/mL; Roche). Leukocytes were enriched and isolated by Percoll density gradient (GE Healthcare). Cells were then stored in liquid nitrogen until further analysis. MFC was performed with either an LSRII cytometer (BD Biosciences) or an Attune NxT cytometer (Thermo Fisher Scientific). TILs were stimulated in the presence of stimulation cocktail (phorbol-12-myristate-13-acetate+ionomycin; eBioscience) and GolgiStop (Monensin; BD Biosciences) according to the manufacturer’s instructions. Samples were then stained for CD45, CD3, CD4, CD8, PD-1, TIM-3, and LAG-3. Intracellular staining for cytokines and transcriptions factors was then performed for Eomesodermin (EOMES), Foxp3, T-bet, and IFNγ. Not all samples were stained for the entire MFC panel (online supplemental table 1).

IHC scoring of each histologic area (ie, primary bone tumors, PMs, and PM interface) and MFC were summarized and compared using means and scatterplots. Bone tumors and PMs were compared using the non-parametric Mann-Whitney U test, and PMs were compared with their corresponding PM interface region using the Wilcoxon matched-pairs signed-rank test. Welch’s t-test was used for comparing MDSC specific genes between PM interface and PM interior compartments. For clinical outcomes following resection of 30 PMs, time to progression was collected and defined as the number of months from the time of surgery until progression. Gene sets scores analyzed in relation to clinical outcomes were calculated based on RNA sequencing data using geometric mean. IHC and MFC measurement and gene sets scores associated with progression-free survival (PFS) were explored in patients using the univariate Cox proportional hazards model. To account for intracorrelations when multiple progressions occurred on the same patient, robust sandwich-type SE estimators were used.27 All statistical tests were two-sided, and statistical significance was set at p<0.05. Since this study was hypothesis-generating, we did not perform corrections for multiple comparisons.

After identifying and annotating areas of viable tumor tissue on pathologist review of H&E (online supplemental figure S2), we first quantified the immune cell infiltration of osteosarcoma by IHC and analyzed differences between primary bone tumors and PMs. Primary bone tumors had low infiltration of T cells (CD3+, CD8+ and Foxp3+) with relatively higher infiltration by myeloid cells (CD163+). CD163 is a scavenger receptor upregulated by alternatively activated macrophages (M2) or tumor-associated macrophages (TAM) characteristic of an anti-inflammatory TME.28 There was also very low expression of PD-1 and PD-L1 in bone tissue (figure 1A&C). There was no significant difference for these markers between untreated bone biopsies and resection of primary tumors following neoadjuvant chemotherapy (online supplemental figure S3). In contrast, our quantification demonstrated significantly higher infiltration by T cells in PMs. Furthermore, we identified two distinct compartments within the PMs: the immune cell-rich tumor-normal lung ‘interface’ region (PM interface) and the PM interior which was relatively devoid of immune cells (figure 1B). Both T cells and myeloid cells clustered at the PM interface, where PD-1 and PD-L1 expressions were predominant (figure 1B&C) (). While PD-L1 was primarily expressed by CD163+ myeloid cells at the PM interface, it was also expressed on tumor cells in the PM interior (figure 1D).

On finding increased immune cells and PD-1/PD-L1 expression in PMs which was concentrated at the PM interface, we expanded our investigation to additional targetable immunosuppressive molecules in PMs. Interestingly, we discovered that there was a higher expression of the immune checkpoint molecules TIM-3 and LAG-3, as well as the IFNγ-driven metabolic checkpoint IDO1 at the PM interface (figure 2A). Remarkably, CSF1R, which is the receptor for colony-stimulating factor 1 (a cytokine that controls the production, differentiation, and function of macrophages) was also present in PMs, but it did not exhibit the described clustering at the PM interface like the remainder of markers (figure 2A & B, online supplemental figure S4). In summary, PMs demonstrated a higher infiltration of T cells, particularly at the PM interface, accompanied by a higher expression of multiple immunoregulatory molecules. In addition, myeloid cells were present in both primary bone tumors and PMs, and in PMs they also accumulated at the PM interface.

We next questioned the immune responsiveness of T cells infiltrating PMs and analyzed them for signs of activation and exhaustion via MFC. We investigated checkpoint molecules and the corresponding intracellular cytokine expression profile of TILs in osteosarcoma specimens to further characterize functional immune properties based on tumor location (bone tumor vs PM). Our initial gating strategy is outlined in online supplemental figure S5. PMs displayed a significantly higher proportion of CD8+lymphocytes compared with primary bone tumors, and a small proportion of regulatory T cells (CD4+Foxp3+) was seen in both PMs and primary bone tumors (figure 3A). CD8+TILs in PMs expressed higher percentages of PD-1 compared with bone tumors, and these CD8+PD-1+TILs also expressed higher amounts of LAG-3 (figure 3B). Furthermore, these PD-1+TILs expressed significant amounts of the transcription factors T-bet and EOMES which are known to control the proliferation hierarchy of exhausted T cells (figure 3C). Finally, CD8+PD-1+TILs from PMs produced higher levels of the effector cytokine IFNγ indicating the cytotoxic potential of this T cell phenotype present in PM (figure 3D).

Given the remarkable regional distribution of immune cells in PMs, we speculated that the transcriptomic profiling of different identified areas would shed light on the mechanisms responsible for the exclusion of T cells from the PM interior. To explore these questions, we performed laser capture microdissection on osteosarcoma PMs and separately dissected the immune cell-rich PM interface region and the immune-excluded PM interior region (online supplemental figure S1). After isolation of RNA from the respective regions, we performed RNA sequencing studies to determine differentially expressed genes predominating in the two distinct regions. Twenty-seven libraries (12 from the PM interface regions, 11 from the PM interior regions, and one each from bone tumor interior, bone tumor-normal bone interface, soft tissue recurrence/metastasis tumor interior, and one soft tissue metastasis tumor- normal muscle interface) passed quality control. Unsupervised principal component analysis was able to segregate the global transcriptome of the PM interface regions from the PM interior regions, and the sparsely infiltrated bone tumors tended to cluster with the PM interiors (online supplemental figure S6). Overall, fewer genes were consistently overexpressed in the PM interior, while a larger number of genes were overexpressed at the PM interface (online supplemental figure S7). We determined the cell-type abundance and expression from bulk tissues via different immune deconvolution platforms. CIBERSORT analysis comparing canonical immune genes at the PM interior to the PM interface regions revealed a predominance of CD8 T cells and CD4 memory resting T cells signatures at the PM interface (figure 4A). This also fits with our IHC observations of multiple immune checkpoint and immunoregulatory molecules, including the IFNγ-regulated molecules PD-L1 and IDO1, which were indicative of increased immune activity at the PM interface. In light of these findings, we explored gene sets associated with immunoregulatory pathways25 and found that indeed both stimulatory and inhibitory molecules stood elevated at the PM interface region compared with the PM interior (figure 4B). We then further expanded our analysis to IFNγ-driven genes related to antigen presentation, chemokine expression, cytotoxic activity, and adaptive immune resistance, which have been shown to be predictive of response to immune checkpoint blockade.24 These genes were also significantly upregulated at the PM interface region (figure 4C). Next, we sought to confirm that these genes were being preferentially expressed in T cells, which would further indicate that TILs were activated but restrained by expression of checkpoint molecules. To address this question, we analyzed our data via xCell22 to generate cell-type coefficient scores. These are indicative of the specific contribution of each immune cell compartment (ie, B cells, T cells, myeloid cells) to the expression of specific genes. With this approach, we ratified that the coefficient for genes associated with markers of T cell cytotoxicity/activation and immune checkpoint molecules were predominantly upregulated at the PM interface region by the T cell compartment (figure 4D).

Tumor-infiltrating myeloid cells constitute a heterogeneous population of cells that are characterized by their diversity and plasticity.29 Many tumor-infiltrating myeloid cells originate from circulating monocytes and granulocytes, which, in turn, stem from bone marrow–derived hematopoietic stem cells. How the composition, molecular state, and spatial interactions of myeloid cells within the TME control the repertoire and the quality of the tumor-antigen-specific T cell response is not well described. Our gene expression profiling of tumor infiltrating immune cells via CIBERSORT analysis comparing the PM interior to the PM interface region revealed an increased proportion of classically activated (M1) macrophages at the PM interface region and an increased proportion of naïve M0 macrophages in the PM interior, although these trends were not statistically significant. However, a strong M2 macrophage signature was present throughout both compartments (figure 4A). This validated our IHC findings that tumor-infiltrating myeloid cells, while concentrated at the PM interface, are also present throughout the PM interior. As we know that M2 macrophages are thought to play a key role in the maintenance of an anti-inflammatory niche leading to a tumor-permissive environment, we further analyzed specific gene sets related to myeloid cells with immune regulatory capacity. It is known that the CSF1/CSF1R axis is essential for differentiation and survival of immunosuppressive myeloid cells, that is, M2 macrophages, and thus we surveyed a set of genes associated with CSF1R responsiveness.23 Notably, this gene set was also enriched at the PM interface, although these gene transcripts were still expressed in the PM interior at lower levels (figure 5A).

Based on these results, it became clear that myeloid cells are an important component of the osteosarcoma TME. Thus, we conducted a more comprehensive examination of myeloid cells in the osteosarcoma TME, including dendritic cells (DCs) and MDSCs. Genes associated with functionally different DC lineages were more highly expressed at the PM interface (figure 5B). Furthermore, we also discovered that transcription factors and apoptotic regulators, immune-regulatory genes and molecules, and cytokines involved in the development of MDSC were also elevated at the PM interface (figure 5C).

Interestingly, we noted that several of the genes that are crucial for the development, accumulation and maintenance of polymorphonuclear MDSCs (PMN-MDSCs) were specifically upregulated at the PM interface, particularly CSF3, which plays a pivotal role in growth control of granulocytic progenitors (online supplemental figure S7). Human PMN-MDSC have a gene expression profile that distinguishes them from neutrophils in patients with cancer and from healthy donors.30 Thus, our analysis uncovered that a number of canonical chemokines, cytokines and chemoattractant genes associated with PMN-MDSCs recruitment were heightened at the PM interface, alongside genes that mark DCs and endothelial cells (figure 5D). Of note, we did not observe an accumulation of normal-appearing PMN cells by morphology on this PM IF area but rather myeloid cells with more immature features. Moreover, it was in this specific PM interface region where we identified a high presence of oxidized low density lipoprotein receptor 1 (OLR1), particularly in myeloid cells, which is one of the most overexpressed transcripts in and a specific gene marker of human PMN-MDSC associated with ER stress and lipid metabolism (figure 5D and online supplemental figure S8, online supplemental table 4). Indeed, higher expression of OLR1 has been associated with inferior survival in patients with different types of cancer.30

Given the numerous immunosuppressive pathways operant at the PM interface identified by our orthogonal interrogations, we then asked how the attributes of the immune TME we had discovered might affect clinical outcomes for patients with metastatic osteosarcoma. No MFC markers were found to be significantly associated with PFS in our cohort. However, we found that by IHC, a higher expression of the immune checkpoint molecules PD-L1 and LAG-3, as well as CSF1R, at the PM interface were all associated with inferior PFS (figure 6A). In contrast, expression of gene sets associated with effective immune response, including CD8 T cells, T cell survival, and major histocompatibility complex (MHC) class 1 expression,31 as well as endothelial cells, at the PM interface were associated with improved PFS (figure 6B). We also analyzed our IHC markers from primary bone tumors and found that the infiltration of CD8 cells and the ratio of CD8:CD163 cells were both significantly associated with improved PFS from both the diagnostic biopsy and from the definitive resection, while PD-L1 was associated with significantly inferior PFS from the diagnostic biopsy (all p<0.01).