Association of Circulating T Cell and Tumor Microenvironment Profiles with Immune Checkpoint Blockade Outcome in Sarcoma

Abstract

Purpose:

Immune checkpoint blockade (ICB) benefits only a subset of sarcoma patients. Biomarkers of response and resistance are needed to help guide patient selection.

Patients and Methods:

We analyzed peripheral blood and tumor samples from sarcoma patients treated on five ICB-based clinical trials. Baseline peripheral blood mononuclear cells (PBMCs) underwent 11-color flow cytometry to define T cell immunotypes. Baseline tumor tissue underwent RNA sequencing to classify tumors into four tumor microenvironment (TME) subtypes using consensus clustering of 29 functional gene expression signatures. Associations between immune features and clinical outcomes were assessed. A deep-learning model was applied to baseline hematoxylin and eosin (H&E) slides to detect and quantify lymphoid aggregates in patients with available RNA sequencing.

Results:

Among 178 patients with PBMCs available for analysis, a proliferative (PRO) circulating T cell immunotype was associated with inferior overall survival (OS) compared with LAG⁻ or LAG⁺ immunotypes. RNA sequencing from 67 tumors identified an immune-enriched/non-fibrotic TME subtype associated with higher response rate, longer progression-free survival, and longer OS compared to immune-enriched/fibrotic, immune-depleted, and fibrotic subtypes. Automated analysis of 48 baseline H&E slides identified lymphoid aggregates in five tumors; four were classified as immune-enriched and two responded to ICB.

Conclusions:

Sarcoma patients with a PRO circulating T cell immunotype had inferior outcomes to ICB, while those with an immune-enriched/non-fibrotic TME had superior outcomes. Automated analysis of H&E slides showed promise in identifying patients with an immune-enriched TME. These findings support utilization of a multimodal approach toward identifying predictors of response to immunotherapy in sarcoma.

INTRODUCTION

Sarcomas are a heterogenous group of mesenchymal malignancies comprised of more than 70 distinct histologic and molecular subtypes (1). High-risk patients with sarcoma have a 5-year disease-specific survival of approximately 50% (2). The standard of care for patients with metastatic sarcoma remains chemotherapy with either doxorubicin or gemcitabine and docetaxel, leading to an overall response rate of approximately 20%, a median PFS of approximately 6 months, and median OS of less than 2 years (3). Alternative options, such as pazopanib (4), trabectedin (5), or eribulin (6) are limited to specific histologies and provide modest clinical benefit in the advanced setting. Novel treatment options are needed for these rare heterogenous diseases.

Prospective trials of PD-1 inhibition in sarcoma have demonstrated signals of efficacy within specific histologic subtypes. Reported overall response rates include 37% in alveolar soft part sarcoma (7), 8–20% in undifferentiated pleomorphic sarcoma, and 7–9% in dedifferentiated liposarcoma (DDLPS) (8–14). Because most sarcoma subtypes fail to benefit from immune checkpoint blockade (ICB), biomarkers of response and resistance are needed to maximize benefit while minimizing unnecessary exposure to these drugs.

Tumor-agnostic biomarkers predictive of response to ICB across cancers, such as a high tumor mutational burden or microsatellite instability, are rarely observed in sarcoma (15). Tumor expression of PD-L1 by immunohistochemistry (IHC) is seen in a minority of sarcomas and, similar to diseases like non-small cell lung cancer, its absence does not negate the possibility of response to ICB (16–18). One promising approach used whole transcriptome sequencing of baseline tumor samples to identify tumors with high immune infiltration and tertiary lymphoid structures (TLS), which have a high likelihood of response to the PD-1 inhibitor pembrolizumab (19). To that end, the PEMBROSARC trial investigated pembrolizumab in TLS-positive sarcomas and found a non-progression rate of 43% at 6 months, an overall response rate of 30%, and a median duration of response of 11 months. Notably, responses were seen in patients with DDLPS, epithelioid sarcoma, and leiomyosarcoma (20).

Identifying biomarkers in the peripheral blood is a less invasive alternative to sampling tumor tissue. In melanoma and urothelial cell carcinoma, a lymphocyte-activation gene 3 (LAG-3)⁺ peripheral blood T cell ‘immunotype’ was found to be associated with poor outcomes on ICB (21). Others have described five principal peripheral blood immunotypes across cancers that were associated with chemotherapy or immunotherapy outcomes (22). To our knowledge, no studies have explored the possibility that peripheral blood immune cells in sarcoma may associate with outcomes on ICB.

To identify blood and tumor-based biomarkers of ICB benefit in sarcoma, we profiled baseline peripheral blood T cell populations and performed multimodal analyses of the tumor microenvironment (TME) in a cohort of sarcoma patients treated on prospective clinical trials of ICB-based therapy at our institution. Herein, we describe how a subset of patients with a ‘proliferative’ T cell peripheral blood immunotype have poorer outcomes on ICB-based therapy, while patients with an immune-enriched TME have superior outcomes. Our analyses highlight the importance of developing a multimodal approach toward predicting which sarcoma patients are most likely to benefit from ICB-based therapies.

MATERIALS AND METHODS

Patient and sample collection

The primary data analysis set was composed of patients with unresectable or metastatic sarcoma treated at Memorial Sloan Kettering Cancer Center on one of five phase II clinical trials of ICB-based therapy. The trials included: nivolumab plus bempegaldesleukin (IL2 agonist), pembrolizumab plus epacadostat (indoleamine 2,3-dioxygenase 1 inhibitor), pembrolizumab plus talimogene laherparepvec (modified oncolytic herpes simplex virus-1), nivolumab with or without ipilimumab, and vimseltinib (CSF1R antagonist) plus avelumab (NCT03282344, NCT03414229, NCT03069378, NCT02500797, and NCT04242238, respectively). Clinical and histologic data were obtained from institutional clinical trial databases or shared organizational datasets, where noted, limiting the specificity of histologic classifications. Liposarcomas without further subtyping were classified as liposarcoma, NOS. All patients were required to undergo baseline tumor biopsies and blood collection for correlative analysis, where feasible. All included patients consented to broad research use of their biospecimens and data. This study was approved by the MSK institutional review board and conducted in accordance with the Declaration of Helsinki.

Two independent datasets were used for validation of the TME clustering technique described herein. The first set was from sarcoma patients treated with pembrolizumab monotherapy on the SARC028 trial (19,23). The second was from sarcoma patients treated with ICB at Stanford (24).

Flow cytometry of the peripheral blood

Human PBMC samples were thawed, washed, counted, and stained with a fixable Aqua viability dye (Invitrogen) and a cocktail of antibodies to the following surface markers: CD8-Qdot605 (Invitrogen, 3B5, RRID: AB_2556437), CD4-Qdot 655 (Invitrogen, S3.5, RRID: AB_11180600), PD-1-PE (BD, MIH4, RRID: AB_2033990), LAG-3-FITC (Enzo, 17B4, RRID: AB_10997322), ICOS-PE-Cy7 (eBioscience, ISA-3, RRID: AB_1518754), TIM-3-APC (R&D Systems, 344823, RRID: AB_1964725). Cells were next fixed and permeabilized with the FoxP3/Ki-67 Fixation/Permeabilization Concentrate and Diluent (eBioscience). Cells were subsequently stained intracellularly with CD3-BV570 (Biolegend, UCHT1, RRID: AB_2562124), Ki-67-AlexaFluor700 (BD, B56, RRID: AB_1964725), FoxP3-eFluor450 (eBioscience, PCH101, RRID: AB_1834364), and CTLA-4-PerCP-eFluor710 (eBioscience, 14D3, RRID: AB_2573718). Stained cells were detected using a BD Biosciences LSRFortessa, data from which was analyzed using FlowJo software (FlowJo, LLC, RRID:SCR_008520). Isotype control stains were used for establishing positivity gates for PD-1, LAG-3, ICOS, TIM-3, FoxP3, and CTLA-4.

Whole transcriptome sequencing

For the primary analysis (Memorial Sloan Kettering) cohort, RNA was extracted from frozen tissues, underwent polyA selection, and libraries were prepared using the TruSeq Stranded mRNA LT Kit (Illumina). Sequencing was performed on a HiSeq 4000 (Illumina, RRID:SCR_016386) as previously described (25). Samples were barcoded and run on a HiSeq 4000 in a paired-end (2 x 100 bp) run utilizing the HiSeq 3000/4000 SBS Kit (Illumina). An average of 41 million paired reads was generated per sample. Ribosomal reads represented 0.9–5.9% of the total reads, and the percentage of mRNA bases averaged 64%.

For the SARC028 cohort, RNA sequencing libraries were constructed and processed as previously described (26). Briefly, total RNA libraries were generated using the KAPA RNA Hyper kit and subsequently captured with the Twist Core Exome probe set. RNA libraries were then sequenced in a paired-end (2 x 75 bp) run on an Illumina HiSeq 4000 instrument, and FASTQ files were generated using Illumina’s bcl2fastq conversion software.

For the Stanford cohort, RNA sequencing libraries were prepared using SMARTer Stranded Total RNA-Seq v2-Pico Input Mammalian Kits (Takara Bio USA, Inc.) as previously reported (24). Libraries were sequenced on either an Illumina HiSeq 4000 or a NovaSeq 6000 (RRID:SCR_016387), generating 150-base pair paired-end reads.

All fastq files from internal and validation cohorts were processed using the REDISCOVERTE pipeline (27). This pipeline employs Salmon version 0.8.2 (RRID:SCR_017036) for transcript quantification, estimating relative abundance in transcripts per million (TPM) (28). Gene transcripts were then aggregated to the gene level based on Ensembl gene IDs (RRID:SCR_002344) from Gencode's human transcriptome version 26 basic annotation (GRCh38/hg38 genome assembly, available at https://www.gencodegenes.org) (29).

Automated analyses of hematoxylin and eosin-stained pathology slides

Biopsy samples (n=48) stained with hematoxylin and eosin (H&E) were scanned and digitized at high resolution (0.2516 μm per pixel), producing whole-slide H&E images for computational analysis. Our previously trained Mask R-CNN deep learning model, originally trained on breast cancer data, was used to segment and classify tumor-infiltrating lymphocytes and tumor cells (30). The Mask R-CNN model was modified for sarcoma after first normalizing H&E color intensity across all slides to eliminate differences across batches. An expert sarcoma pathologist then manually labeled tumor cells and lymphocytes in 12 slides and these annotations were used to fine-tune the model. The sarcoma-adjusted model was applied to all 48 H&E slides to segment the cells and detect lymphocytes and tumor cells, providing detailed outputs including the spatial coordinates of each cell, its area, and corresponding color intensity values.

After classifying the lymphocytes in the H&E slides, cells’ spatial coordinates were employed to detect lymphoid aggregates. Lymphoid aggregates were computationally defined and detected in multiplex immunofluorescent slides using our previous DBSCAN-based method (31), which results in more sensitive aggregate detection compared with manual identification by a pathologist (32). Lymphoid aggregates were defined as densely populated lymphocyte areas where the number of immune cells detected by the Mask R-CNN model exceeded 150 cells within a radius of 75 μm. After applying the algorithm, all slides in which lymphoid aggregates were detected were visually inspected and confirmed by our pathologist (CRA).

The advantage of using this DBSCAN-based method over previous methods was that there was no need to prespecify the number of aggregates, which aligns with our goal to detect lymphoid aggregates within H&E slides automatically, without the need for human intervention for each slide after determining the required criteria for defining an aggregate. After detecting the aggregates, a convex hull was applied, which provided more details about each aggregate, such as area, roundness, radius, skewness, and density, which can be used for further downstream analysis of the tumor microenvironment and cells’ and aggregates’ spatial distribution.

Statistical analyses

Characteristics of all patients in the primary cohort were summarized using descriptive statistics. Response rates overall, by histology, and by treatment were estimated using the proportion of evaluable patients who achieved a complete or partial response, and confidence intervals were obtained using the Clopper-Pearson method. Response by histology and treatment was illustrated using bar plots and compared using univariable logistic regression models. Survival outcomes (PFS and OS) were calculated from the first dose of treatment. Median survival estimates and confidence intervals overall, by histology, and by treatment were obtained using Kaplan-Meier methods. Survival outcomes by histology and treatment were illustrated using Kaplan-Meier curves and compared using univariable Cox proportional hazards models. Histologic composition by treatment and response was illustrated using bar plots.

For the primary cohort, a heatmap was used to illustrate the percentages of 78 T cell subpopulations for each available pre-treatment PBMC sample as measured by flow cytometry. Percentages were folded root-transformed, centered, and scaled by subpopulation for all analyses unless otherwise specified. Each sample was classified into one of three immunotypes (LAG+, LAG−, or PRO) using a previously validated multinomial logistic regression model (21), which takes four CD8+ T cell subpopulation percentages as input: % LAG-3+, % Ki67+, % TIM-3+, and % ICOS+ cells. The distribution of the three immunotypes in the present cohort, the original melanoma training cohort, and the original urothelial carcinoma validation cohort was illustrated in a bar plot. Clinical outcomes by immunotype were analyzed using the methods described above. Multivariable Cox proportional hazards models, illustrated by forest plots, were used to evaluate associations between immunotype and survival outcomes while adjusting for histology and treatment. Associations between each of the 78 T cell subpopulations and each outcome were evaluated using multivariable models adjusting for histology and treatment (logistic regression and Cox proportional hazards models). The false discovery rate (FDR) correction was applied to the set of p-values for each outcome, and subpopulations with an FDR < 0.1 were illustrated using heatmaps. Patients were split into three groups by Ward’s hierarchical clustering of % TIM-3+/CD8+PD-1- T cells. Raw percentages in each group were illustrated using box plots, and clinical outcomes by group were analyzed using the methods described above.

For the primary cohort and two independent cohorts (SARC028 and Stanford), for each available pre-treatment RNA-seq sample, single-sample gene set enrichment analysis (ssGSEA) of transcripts per million (TPM) values was used to generate enrichment scores for 29 functional gene expression signatures described previously (33). Scores were centered and scaled within each cohort for all subsequent analysis. For the primary cohort, four unsupervised TME subtypes were obtained using K-means-based consensus clustering across the 29 scores (Supplementary Fig. S1). For the SARC028 and Stanford cohorts, each sample was assigned to one of the four subtypes based on the Euclidean distance of its scores to the centroid of each subtype’s set of scores in the primary cohort. For all three cohorts, ssGSEA scores were illustrated across subtypes using heatmaps. For the primary cohort, the distribution of the four subtypes was illustrated using a bar plot, and clinical outcomes were analyzed by subtype as described above. Histologic composition by subtype and response was illustrated using a bar plot.

Two-sided p-values lower than 0.05 were considered statistically significant. For the flow cytometry discovery analysis, FDR values lower than 0.1 were considered statistically significant. All analyses were conducted using R version 4.4.0. ssGSEA was performed using GSVA (34), and consensus clustering was performed using ConsensusClusterPlus (RRID:SCR_016954) (35).

Data availability

RNA sequencing data on the 192 MSK patients who consented to data sharing is publicly available via dbGaP (accession number: phs003284). The bulk RNA-seq data from the Stanford cohort is available at GSE213065. The SARC028 dataset is available upon request for research purposes (sarc028@sarctrials.org) through the SARC clinical trial data repository: https://sarctrials.org/clinicaltrials/sarc028/. Flow cytometry and automated H&E analysis data are provided in the data supplement.

RESULTS

Outcomes on ICB-based therapies in sarcoma vary by histology and treatment (n = 195)

Patient characteristics

A total of 195 patients on five sarcoma-specific ICB-based clinical trials with pre-treatment PBMCs and/or tumor tissue available were analyzed. Patient characteristics are outlined in Table 1. The median patient age was 61 years (interquartile range [IQR], 47–70) and the majority were male (54%). Four trials utilized a PD-1 or PD-L1 antibody backbone in combination with a second immunomodulating drug: bempegaldesleukin (33%), talimogene laherparepvec (T-VEC; 18%), vimseltinib (15%), or epacadostat (15%). The fifth trial randomized patients to receive nivolumab with (8.2%) or without (10%) ipilimumab. The median number of prior therapies was 2 (IQR, 1–4). The most common histologic subtypes were undifferentiated pleomorphic sarcoma or myxofibrosarcoma (UPS/MFS; 29%), leiomyosarcoma (18%), and other fusion-positive soft tissue sarcoma (STS; 11%). Other subtypes were categorized as other STS(10%), angiosarcoma (9.7%), liposarcoma not otherwise specified (NOS; 7.7%), bone sarcoma (7.7%), or DDLPS (6.2%). Detailed histologic diagnoses are presented in Supplementary Table S1.

Table 1.

Patient characteristics. Data are presented as median (interquartile range) or n (%).

Characteristic	n = 195
Age	61 (47, 70)
Sex
Female	89 (46%)
Male	106 (54%)
Race
White	156 (81%)
Asian	16 (8.3%)
Black or African American	13 (6.8%)
Other	4 (2.1%)
Unknown	2 (1.0%)
American Indian or Alaska Native	1 (0.5%)
Missing	3
Histology group
UPS/MFS	57 (29%)
LMS	35 (18%)
Other fusion+ soft tissue sarcoma	22 (11%)
Other soft tissue sarcoma	20 (10%)
Angiosarcoma	19 (9.7%)
Bone	15 (7.7%)
Liposarcoma, not otherwise specified	15 (7.7%)
Dedifferentiated liposarcoma	12 (6.2%)
Treatment
Nivolumab + bempegaldesleukin	64 (33%)
Pembrolizumab + T-VEC	35 (18%)
Avelumab + vimseltinib	30 (15%)
Pembrolizumab + epacadostat	30 (15%)
Nivolumab	20 (10%)
Nivolumab + ipilimumab	16 (8.2%)
Prior lines of therapy	2 (1, 4)
Missing	36
Response
Complete response	1 (0.5%)
Partial response	24 (13%)
Stable disease	66 (35%)
Progressive disease	97 (52%)
Missing	7
Response	25 (13%)
Missing	7

UPS/MFS, undifferentiated pleomorphic sarcoma/myxofibrosarcoma; T-VEC, talimogene laherparepvec.

Outcomes on ICB-based therapy

One hundred eighty-eight patients were evaluable for response. Among these patients, the overall response rate by RECIST 1.1 across all trials was 13% (95% confidence interval [CI], 8.8–19%), which included 1 complete response (CR) and 24 partial responses (PR). Best response of stable disease (SD) was achieved in 66 (35%) patients, while 97 (52%) had progressive disease (PD). Survival outcomes were available for all 195 patients; median progression-free survival (PFS) and overall survival (OS) were 2 months (95% CI, 1.8–2.8 months) and 11 months (95% CI, 9–18 months), respectively.

Histology and treatment were each significantly associated with overall response, PFS, and OS (histology: p=0.003, p<0.008, and p=0.039, respectively; treatment: p<0.001, p<0.001, and p=0.019, respectively). The most responsive histologies (overall response rate [95% CI]) were angiosarcoma (47% [23–72%]) and UPS/MFS (19% [9.3–31%]), and least responsive was DDLPS (0% [0–26%]) (Fig. 1a). PFS and OS were most favorable among patients with angiosarcoma (median PFS 13 months [95% CI, 3.5–not reached (NR)]; median OS 18 months [95% CI, 11–NR]) (Fig. 1b). Across treatments, overall response rate (95% CI) ranged from 34% (19–52%) in the pembrolizumab/TVEC group to 0% (0–12%) in the avelumab/vimseltinib group (Fig. 1c). PFS and OS were most favorable in the pembrolizumab/TVEC group (median PFS 5.1 months [95% CI, 2.9–13]; median OS 25 months [95% CI, 17–41]) (Fig. 1d). The histologic composition by treatment and response is presented in Fig. 1e.

Figure 1.

Outcome on ICB-based therapy in sarcoma across trials at a single center. (A) Response rate to ICB-based therapies by sarcoma histology (n = 188 evaluable for response). (B) PFS and OS after ICB-based therapies by sarcoma histology (n = 195). (C) Response rate to ICB-based therapies by treatment regimen (n = 188 evaluable for response). (D) PFS and OS after ICB-based therapies by treatment regimen (n = 195). (E) Number of patients with each sarcoma histology by trial and response (n = 195). NE, not evaluable.

Peripheral blood T cell phenotype is associated with survival after ICB (n = 185)

Circulating T cell immunotypes are recapitulated in sarcoma

We performed 11-color flow cytometry on 185 pre-treatment PBMC samples and quantified 78 unique T cell subpopulations within each sample (Supplementary Fig. S2). Using a classifier previously developed in melanoma and validated in urothelial carcinoma, samples were categorized into one of three peripheral blood T cell immunotypes: LAG⁺ (characterized by high expression of LAG-3 across T cell subpopulations with low Ki67 expression), LAG^– (low LAG-3⁺ and other co-markers), or proliferative (PRO; a high proportion of LAG-3+ T cells with concurrently high numbers of proliferating Ki67⁺ CD8⁺ T cells and T cells expressing TIM-3 and ICOS) (21). Among the 185 samples, 20 (11%) were classified as LAG⁺, 135 (73%) as LAG^–, and 30 (16%) as PRO. The distribution of the three immunotypes was similar to the distributions previously observed in melanoma and urothelial carcinoma (Supplementary Fig. S3).

A proliferative immunotype is associated with inferior overall survival

In univariable analysis, immunotype (classification as LAG⁺, LAG^–, or PRO) was not significantly associated with overall response (PRO vs. LAG^–: OR, 1.29 [95% CI, 0.34–3.93]; LAG⁺ vs. LAG^–: OR, 2.76 [95% CI, 0.80–8.43]; p = 0.26; n = 178 evaluable) nor with PFS (PRO vs. LAG^–: HR, 1.53 [95% CI, 1.02–2.29]; LAG⁺ vs. LAG^–: HR, 0.98 [95% CI, 0.60–1.61]; p = 0.14) (Fig. 2a,b). Median PFS was 2.1 (95% CI 1.8–3.32), 2.0 (1.6–8.9), and 1.7 months (1.3–3.4) in the LAG−, LAG+, and PRO groups, respectively. While the overall association with PFS did not reach statistical significance, the pairwise significance between the PRO and LAG^– groups (indicated by the HR 95% CI, 1.02–2.29) suggested that the PRO immunotype may be associated with poorer PFS. In contrast, immunotype was significantly associated with OS in univariable analysis, with the PRO group experiencing worse OS compared to the LAG^– group (PRO vs. LAG^–: HR, 1.90 [95% CI, 1.25–2.90]; LAG⁺ vs. LAG^–: HR, 1.40 [95% CI, 0.84–2.34]; p = 0.013) (Fig. 2b). Median OS was 13 (95% CI 9.7–21), 7.6 (3.9–27), and 6.1 months (3.8–16) in the LAG−, LAG+, and PRO groups, respectively. This association remained significant in multivariable analysis accounting for histologic subtype and treatment (PRO vs. LAG^–: HR, 2.20 [95% CI, 1.39–3.48]; LAG⁺ vs. LAG^–: HR, 1.60 [95% CI, 0.92–2.80]; p = 0.003) (Fig. 2c). The histologic distribution of responders according to immunotype is presented in Supplementary Fig. S4.

Figure 2.

Peripheral blood immunotype and its association with outcome to ICB-based therapy. (A) Response rate by peripheral blood immunotype (n = 178 evaluable for response). (B) PFS and OS after ICB-based therapies by peripheral blood immunotype (n = 185). (C) Multivariable analyses of PFS and OS (n = 185). Forest plots depict the results of Cox proportional hazards models with peripheral blood immunotype, histology, and treatment regimen as covariates. (D) Heatmaps of T cell subsets significantly associated with PFS and OS while accounting for both histologic subtype and treatment (FDR < 0.1) (n = 185). Peripheral blood immunotype classification, sarcoma histology, treatment regimen, and objective response are denoted in horizontal tracks above each heatmap. Heatmap values are centered and scaled.

Next, to comprehensively assess the flow cytometry data, we evaluated associations between each of the 78 T cell subpopulations and overall response, PFS, and OS while accounting for both histologic subtype and treatment. No subsets were significantly associated with response after correction for multiple comparisons. However, 29 subsets were significantly associated with shorter PFS and 20 with shorter OS after accounting for histologic subtype and treatment (false discovery rate [FDR] < 0.1) (Supplementary Table S2). These subpopulations were heavily enriched for Ki67+, TIM-3+, and ICOS+ cells, suggesting a potential correspondence between these T cell subsets and the PRO immunotype (Fig. 2d).

A subset of TIM-3+ T cells is associated with poor outcomes independently of immunotype

To identify T cell subsets associated with clinical outcomes independently of immunotype, multivariable analyses adjusting for immunotype, histology, and treatment were performed. While no T cell subsets remained significantly associated with PFS after correction for multiple comparisons, two subsets remained significantly associated with OS (FDR < 0.1):

the percentage of (%) TIM-3⁺/CD8⁺PD-1^– T cells and % PD-1^–TIM-3⁺/CD8⁺ T cells. These two populations are biologically identical but gated differently to ensure consistency in measurement. For simplicity, we refer to these populations in the text as % TIM-3⁺/CD8⁺PD-1^– T cells, but all analyses of % PD-1^–TIM-3⁺/CD8⁺ T cells yielded similar results (data not shown).

To further characterize the TIM-3⁺/CD8⁺PD-1^– T cell population and its associations with outcomes, we trichotomized patients as having high, intermediate, or low % TIM-3⁺/CD8⁺PD-1^– T cells using unsupervised clustering (Fig. 3a,b). Response, PFS, and OS significantly differed across these three groups in univariable analysis (p=0.011, p<0.001, p<0.001, respectively). Patients with the lowest % TIM-3⁺/CD8⁺PD-1^– T cells had a response rate (95% CI) of 25% (15–39%), compared to 7.8% (3.2–15%) and 9.1% (1.9–24%) in the intermediate and high groups, respectively (Fig. 3c). PFS and OS were similarly more favorable in the low % TIM-3⁺/CD8⁺PD-1^– T cell group. Median PFS (95% CI) in the low group was 5.3 months (2.8–13), compared to 1.8 months (1.6–2.6) and 1.7 months (1.4–2.7) in the intermediate and high groups, respectively (Fig. 3d). Median OS (95 % CI) in the low group was 22 months (18–33), compared to 8.9 months (6.9–17) and 5.9 months (3.4–8.8) in the intermediate and high groups, respectively (Fig. 3d). The associations with PFS and OS remained significant when simultaneously adjusting for histology, treatment, and immunotype (p = 0.003 and p = 0.006, respectively).

Figure 3.

Percentage of (%) TIM-3+/CD8+PD-1- T cells in the peripheral blood is associated with outcome to ICB-based therapies in sarcoma. (A) Heatmap of % TIM-3⁺/CD8⁺PD-1^– T cells (N = 185). Patients were trichotomized as having low, intermediate, or high % TIM-3⁺/CD8⁺PD-1^– T cells using hierarchical clustering. Peripheral blood immunotype classification, sarcoma histology, treatment regimen, and objective response are denoted in horizontal tracks above the heatmap. Heatmap values are centered and scaled. (B) Raw percentages of TIM-3⁺/CD8⁺PD-1^– T cells in low, intermediate, and high % TIM-3⁺/CD8⁺PD-1^– T cell groups (n = 185). (C) Response rate to ICB-based therapies by % TIM-3⁺/CD8⁺PD-1^– T cell group (n = 178 evaluable for response). (D) PFS and OS after ICB-based therapies by % TIM-3⁺/CD8⁺PD-1^– group (n = 185).

These findings suggest that a low proportion of TIM-3⁺/CD8⁺PD-1^– T cells in the blood at baseline may be associated with improved survival following ICB-based therapy, independent of peripheral blood immunotype.

An immune-enriched tumor microenvironment associates with response to ICB

Gene expression analysis identifies four TME subtypes across sarcomas (n = 67)

After investigating T cell subsets in the blood, we next evaluated TME-related gene expression signatures in baseline tumor biopsy samples and evaluated their associations with clinical outcomes. RNA sequencing data were available for 67 patients. Tumors were classified into one of four TME subtypes using consensus clustering of 29 functional gene expression signatures derived from single-sample gene set enrichment analysis (ssGSEA) (Fig. 4a) (33). These signatures encompass key genes and pathways involved in angiogenesis, fibrosis, and metastasis, as well as those expressed by immune cells with anti- and pro-tumoral activity. The most common TME subtype was immune-depleted/fibrotic (“fibrotic”) (n = 25, 37%), followed by immune-enriched/fibrotic (n = 23, 34%), immune-depleted/non-fibrotic (“depleted”) (n = 10, 15%), and immune-enriched/non-fibrotic (n = 9, 13%) (Fig. 4b).

Figure 4.

TME subtype is associated with outcome on ICB-based therapy in sarcoma. (A) Heatmap of enrichment scores of 29 key gene signatures across sarcoma samples, which yield four unique TME subgroups (n = 67). Heatmap values are centered and scaled. CAF, cancer-associated fibroblast; TAM, tumor-associated macrophage. (B) Distribution of the four TME subtypes (N = 67). (C) Response rate to ICB-based therapies by TME subtype (n = 66 evaluable for response). (D) PFS and OS after ICB-based therapy by TME subtype (n = 67). (E) Number of patients with each sarcoma histology by TME subtype and response (n = 68). NE, not evaluable.

Response to ICB-based therapy was significantly associated with TME subtype, with the highest response rate (95% CI) in the immune-enriched/non-fibrotic group at 67% (30–93%) compared to 10% (0.25–45%), 9% (1–29%), and 4% (0.1–20%) in the depleted, immune-enriched/fibrotic, and fibrotic groups, respectively (p < 0.001) (Fig. 4c). PFS and OS were also significantly associated with TME subtype, with the most favorable survival in the immune-enriched/non-fibrotic cohort (median PFS 16 months [95% CI, 5.5–NR]; median OS NR) (PFS: p = 0.006; OS p = 0.006) (Fig. 4d).

The composition of histologic subtypes by TME subtype and response are presented in Fig. 4e, and the detailed histologic diagnoses are provided in Supplementary Table S3. The immune-enriched/non-fibrotic cluster included three undifferentiated pleomorphic sarcomas, and one case each of the following: myxofibrosarcoma, angiosarcoma, epithelioid sarcoma, high-grade spindle cell sarcoma, high-grade undifferentiated spindle and pleomorphic sarcoma, and undifferentiated sarcoma with a myxoid spindle and epithelioid phenotype.

TME subtyping is recapitulated in two independent datasets of sarcoma patients (n = 30; n = 38)

To determine whether the four TME subtypes identified in our primary sarcoma data were conserved in an independent cohort, the 29 functional gene expression signatures were quantified in each of the two external datasets: SARC028 cohort (n = 30) and the Stanford cohort (n = 38). Samples were then assigned to a TME subtype based on the similarity of their expression patterns to those in the original sarcoma dataset (Supplementary Fig. S5).

The distribution of TME subtypes was broadly consistent across cohorts. However, the small numbers of responders in these datasets limited the ability to robustly assess associations between TME subtype and clinical outcomes (3/30 responders in SARC028, and 4/38 in the Stanford cohort). In SARC028, a response was observed in one of four evaluable patients with an immune-enriched/non-fibrotic TME, compared to 1/5 in the depleted group, 1/10 in immune-enriched/fibrotic, and 0/10 in fibrotic (Supplementary Fig. S5a). Median PFS (95% CI) in the immune-enriched/non-fibrotic group was 3.1 months (1.5–NR), compared to 1.6 months (1.1–NR), 1.8 months (1.6–NR), and 1.7 months (1.6–NR) in the depleted, immune-enriched/fibrotic, and fibrotic groups, respectively. Similarly, in the Stanford cohort, responses were observed in two of five patients with an immune-enriched/non-fibrotic TME, compared to 0/8 in the depleted group, 1/13 in immune-enriched/fibrotic, and 1/11 evaluable in fibrotic (Supplementary Fig. S5b). Median PFS (95% CI) in the immune-enriched/non-fibrotic group was 4.7 months (2.7–NR), compared to 2.1 months (1.8–NR), 2.8 months (2.3–NR), and 2.8 months (2.1–NR) in the depleted, immune-enriched/fibrotic, and fibrotic groups, respectively. No comparisons reached statistical significance. Follow-up was not sufficient to calculate median OS across the four subtypes in in SARC028, and OS was not available for the Stanford cohort.

Automated analysis of digitized H&E slides detects lymphoid aggregates (n = 48)

Tertiary lymphoid structures (TLS) are a recognized biomarker of response to ICB-based therapy in sarcoma (19). To explore this association further and correlate our findings with TME subtypes, we aimed to detect lymphoid aggregates among the 48 patients who had both RNA sequencing and digitized H&E slides available for analysis. Deep-learning methods were used to analyze the H&E sections of core needle biopsies to automatically detect lymphoid aggregates (Fig. 5a). Aggregates were subsequently examined by an expert bone and soft tissue pathologist (CRA) to verify the model’s accuracy. Aggregates were detected in 5 patient biopsies: 1 epithelioid sarcoma, 1 undifferentiated epithelioid and spindle cell neoplasm, 1 DDLPS, 1 dedifferentiated chondrosarcoma, and 1 UPS (Fig. 5b). Four of these tumors had an immune-enriched TME subtype (2 were non-fibrotic), and two responded to ICB-based therapy.

Figure 5.

Automated detection of lymphoid aggregates on H&E slides. (A) Pseudoplot of lymphoctyes and tumor cells identified from a conventional H&E image using deep learning (left). Original low-power H&E image of core biopsy with three detected lymphoid aggregates outlined in blue boxes (labeled as 0, 1, and 2); histology and RNA-seq-derived TME subtype are indicated (center). Image of a detected lymphoid aggregate (right). (B) Original low-power H&E images of core biopsies with detected lymphoid aggregates outlined in blue boxes; histology and RNA-seq-derived TME subtype are indicated (left). Images of detected lymphoid aggregates (right).

DISCUSSION

It is well established that antitumor T cell activation, in what has been called the ‘cancer-immunity cycle,’ is a complex process. Multiple factors such as TME composition, host immune system, and the microbiome cooperate to establish antitumor immunity or lack thereof (36). It is unlikely that a single biomarker can encompass the complexity of the cancer-immunity cycle. We therefore sought to analyze both the tumor and the peripheral blood immune milieu of patients with sarcoma who were treated with ICB. Our analysis of pre-treatment peripheral blood found that patients with a proliferative T cell immunotype, characterized by high expression of Ki-67 and markers of activation (ICOS) and exhaustion (TIM-3 and LAG-3) (21), have poorer survival after ICB treatment. In parallel, our examination of key tumor and immune-related genes in the TME found that patients with an immune-enriched/non-fibrotic TME signature had a very high likelihood of response to ICB. Taken together, our findings underscore the potential value of using a multimodal approach to assess the likelihood of ICB response.

To our knowledge, this study represents the largest effort to date to profile peripheral blood immune cell subsets in sarcoma patients and assess their correlation with outcome on ICB. Peripheral blood biomarkers hold promise because they are easy to collect, involve minimal risk to patients, and are amenable to repeat measurement over time. Our results use a previously validated peripheral blood immunotyping scheme (21), which we applied to sarcoma samples for the first time. We found that LAG3^– patients have the best survival of the three dominant peripheral blood immunotypes, as seen in melanoma and urothelial carcinoma. However, whereas in prior analyses the LAG3⁺ immunotype had the worst outcome, in sarcoma the proliferative immunotype was associated with the worst survival. Perhaps this difference can be attributed to an inherent difference of the host response to sarcoma compared to melanoma or urothelial carcinoma. Additionally, these results raise the possibility that blood-based immune signatures, such as a PRO immunotype, capture prognostic biology rather than ICB-specific predictive information. A comparable phenomenon has been described in melanoma, where high LAG-3 expression in the tumor has been associated with improved outcomes even outside combination LAG-3/PD-1 blockade, suggesting that some checkpoint-related markers may be more prognostic than predictive (37).

We identified a subset of TIM-3⁺ T cells associated with poor response to ICB after accounting for histology, treatment regimen, and peripheral blood immunotype. TIM-3 is a transmembrane protein found on the surface of T cells, myeloid cells, and dendritic cells that binds to numerous ligands, including galectin-9, phosphatidylserine, high-mobility group protein B1, and carcinoembryonic antigen cell adhesion molecule 1. It is both a marker of T cell exhaustion and a negative regulator of effector T cell function (38). Multiple studies have found that TIM-3 expression correlates with cancer progression, poor clinical outcome (39,40), and resistance to PD-1 inhibition (41). Preclinical models show that blocking TIM-3 in conjunction with PD-1 can overcome this resistance (41).

TIM-3 is expressed on immune cells in many sarcomas (42,43) and its expression either in the blood or tumor is associated with a poor prognosis (7,44). To our knowledge, this is the first study to demonstrate a correlation of TIM-3 expression with outcome on ICB in sarcoma and our findings are consistent with its known mechanism of action. It is notable that the specific TIM-3⁺ T cell subset associated with outcome was PD-1-negative. In one study of non-small cell lung cancer tumor specimens, approximately 17% of TIM-3⁺ tumor-infiltrating lymphocytes (TILs) did not express PD-1. These PD-1^–TIM-3⁺ TILs had a lower cytotoxic potential than LAG-3⁺ TILs and fewer functional markers than TILs that co-expressed two or more immune-inhibitory receptors (45). It is reasonable to hypothesize that peripheral blood TIM-3⁺/CD8⁺PD1^– T cells in sarcoma have a reduced capacity for immune effector function compared to other T cell subsets, although additional studies are needed to further characterize the functional differences between individual T cell subsets.

For the tumor analyses, we used a previously validated framework for categorizing the TME into one of four subtypes that appear to be conserved pan-cancer (33). This framework is notably different from previous transcriptomic profiling efforts of the sarcoma TME (19) not only by its increased breadth (29 versus 9 gene expression signatures), but also by its inclusion of stromal components, such as cancer-associated fibroblast (CAF) and matrix signatures, and the inclusion of gene expression signatures that capture epithelial to mesenchymal transition (EMT) and tumor proliferation. Our data implies that immune enrichment of tumors alone may not be sufficient to elicit response to ICB. Rather, an effective antitumor immune response requires both an influx of immune cells and minimal stromal barriers to their activity.

The importance of the stromal barrier in precluding responses to ICB has been demonstrated in other solid tumors such as urothelial carcinoma (46–48). Recent investigations in sarcoma have highlighted the importance of glycolytic CAFs that exclude T cells from the TME by overexpressing CXCL16 (49). We posit that in mesenchymal malignancies, CAFs serve an equally crucial role by either excluding or suppressing key immune cells from trafficking and functioning within the TME. Studying the precise mechanism of CAF-mediated immunosuppression may yield alternative drug development targets for ICB synergism.

Finally, we report our results using a custom machine learning approach to detect lymphoid aggregates on digitized H&E images. Tertiary lymphoid structures (TLS) are lymphocyte-rich zones composed of T cells, B cells, plasma cells, and dendritic cells that form within the tumor stroma due to chronic antigenic stimulation (50). They are associated with response to ICB in carcinomas, melanoma, and sarcoma (19,51–54). To date, there is no universally recognized method to measure TLS, either by immunohistochemistry or by gene expression signature (55). Others have highlighted the potential of identifying TLS in carcinomas by measuring lymphoid aggregates on H&E images, which have the benefit of being widely available and inexpensive (56). We successfully identified lymphoid aggregates in 5 patients, 2 of whom responded to ICB. Although these are small numbers, this proportion of responders is consistent with the PEMBROSARC study, which found a 30% response rate among soft tissue sarcomas with TLS (20). We hypothesize that further refining the deep learning model to identify additional key pathologic features, such as quantity of individual B or T cell subsets or stromal components like fibroblasts, especially when combined with modalities discussed in this paper such as blood flow cytometry and transcriptome sequencing, can further enhance the utility of this technology in predicting which patients are most likely to respond to ICB.

Our findings have several important constraints limiting their generalizability and applicability. First, the patient population included many histologic subtypes of sarcoma that vary in their behavior, morphology, and molecular pathogenesis. The small numbers of many histologies in this dataset limits our ability to draw robust, subtype-specific conclusions. These findings underscore the need for future studies powered to evaluate biomarkers within individual sarcoma subtypes, where biology may diverge from pan-sarcoma patterns. Second, patients were treated with various ICB-based combination regimens that differ in their mechanism of action, thus potentially impacting the host immune response differently across trials. Third, our analyses of peripheral blood found discrepant results between the endpoints of response and survival. This may be due to a lack of power in the response analyses due to the limited number of responders overall in this heterogenous cohort. Despite these limitations, this study represents one of the largest studies of correlates of outcomes on ICB-based therapy in sarcoma. To further develop these potential biomarkers, our findings would benefit from validation in large independent sarcoma datasets.

Translational Relevance.

Immune checkpoint blockade (ICB) provides durable benefit for only a minority of patients with sarcoma, and predictive biomarkers are lacking. In this study, we identify two distinct features associated with outcomes: a proliferative circulating T cell immunotype linked to inferior survival, and an immune-enriched, non-fibrotic tumor microenvironment transcriptional signature linked to higher response rates. Importantly, this is the first demonstration that a peripheral blood–based immunotype may have prognostic utility in sarcoma, offering a less invasive and scalable alternative to tumor tissue profiling. In parallel, our integration of stromal and matrix gene signatures refines transcriptomic predictors of ICB benefit beyond immune infiltration alone. Together, our findings highlight the importance of a multi-modal approach to guide patient selection and underscore the need for large datasets to discover and validate biomarkers that are clinically actionable.