Background

Metastatic/unresectable urothelial carcinoma (mUC) is detected in approximately 20% of cases of invasive UC.1 Combination chemotherapy involving platinum agents has been the main component of the principal therapeutic scheme for several decades. Nonetheless, despite achieving a high initial response rate with chemotherapy, few patients have long-term response and extended survival.1 Immune checkpoint inhibitors (ICIs), such as anti-programmed death-1 (PD-1) and anti-PD-ligand-1 (PD-L1) blockades, have advanced the management of patients with mUC who are resistant to platinum-based chemotherapy.1 More recently, the CheckMate 274 trial demonstrated a clinically significant benefit of adjuvant PD-1 immunotherapy compared with placebo in improving disease-free survival (DFS) in patients with muscle-invasive UC undergoing radical surgery.2 Despite such substantial efforts, issues regarding primary and secondary/acquired resistance to ICI therapy remain.3 Thus, the clinical application of ICI combination treatment has been investigated to attain synergetic effects with improved and durable responses.4 5

In addition to PD-(L)1 and cytotoxic T lymphocyte antigen-4 (CTLA-4), next-generation coinhibitory receptors (IRs) found on T-cells, such as T-cell immunoreceptor with Ig and ITIM domains (TIGIT), T-cell immunoglobulin and mucin domain 3 (TIM-3), and lymphocyte activation gene 3 (LAG-3), have recently gained attention for their use in cancer therapy.6 7 To date, several clinical trials have proven the efficacy of these IR-targeting agents. In non-small-cell lung cancer (NSCLC), tiragolumab (anti-TIGIT antibody) plus atezolizumab therapy resulted in significantly improved objective response and progression-free survival (PFS) rates in the study group compared with the placebo plus atezolizumab group.8 The potential for leveraging TIM-3 for anticancer therapy is currently being evaluated in early studies on several solid malignant tumors.9 Relatlimab (anti-LAG-3 antibody) and nivolumab provided greater benefits in terms of PFS than nivolumab alone in patients with untreated metastatic/unresectable melanoma.10 Moreover, a strong association between higher LAG-3 expression and response to combination anti-LAG-3 and anti-PD-1 immunotherapies has been reported.11 Although the conjugative therapy of these IR-targeting agents with PD-(L)1 inhibition may be promising, the clinical significance of these IR expressions has yet to be elucidated in UC.

In this study, we initially assessed previously published bulk RNA-sequencing data to reveal how the expression of next-generation immune checkpoint receptors (LAG-3, TIGIT, and TIM-3) and their ligands influence tumor progression and immunoevasive mechanisms explaining resistance to ICIs in patients with mUC subjected to anti-PD-(L)1 treatment. Based on these outcomes, we used the multiplex immunohistochemistry (mIHC) method to clarify the effect of spatial expression and the intercellular interactions of these molecules on patient prognosis.

MethodsPatients and specimens

The study design with inclusion and exclusion criteria is summarized in online supplemental figure S1. We included 348 patients treated with atezolizumab in the IMvigor210 trial, which was a multicenter, single-arm, phase 2 study that explored the clinical efficacy and safety of atezolizumab in mUC (cohort 1: n=348; online supplemental table S1). The data were obtained from a free data resource (http://research-pub.gene.com/IMvigor210CoreBiologies).12 13 For validation, we incorporated data from 89 patients with mUC who received anti-PD-(L)1 inhibitive therapy at the University of North Carolina, obtained from the Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE176307) (cohort 2: n=89; online supplemental table S2).14 We also included 296 patients with advanced UC of The Cancer Genome Atlas bladder cancer (TCGA-BLCA 2020) according to cBioPortal (https://www.cbioportal.org/) (cohort 3: n=293; online supplemental table S3). After obtaining institutional review board (IRB) approval, we recruited 56 mUC patients treated with pembrolizumab and 91 UC patients with locally advanced UC who underwent radical cystectomy at KMU, who were ≥20 years of age, both male and female. Of all patients, we extracted cases whose formalin-fixed paraffin-embedded (FFPE) primary tumor tissue blocks were analyzed (cohort 4: n=29 and cohort 5: n=90; online supplemental tables S4 and S5). Our study followed the principles of the Declaration of Helsinki.

The best objective response was estimated based on the Response Evaluation Criteria in Solid Tumors V.1.1 (complete response: CR; partial response: PR; stable disease: SD; and progression disease: PD). Overall survival (OS) was defined as the period from ICI initiation to death caused by all conditions or the date of the last follow-up. PFS was defined as the time from the onset of ICI to disease progression or death from any cause. DFS was defined as the period from the date of radical cystectomy to the time of recurrence or death.

RNA-sequencing dataset analysis

Bulk RNA-sequencing data for IMvigor210, GSE176307, and TCGA were retrieved from each data set. The median expression values were applied to dichotomize the sample into high-RNA and low-RNA level groups, similar to previous studies.15 16 RNA expression signatures were outlined based on previous studies and summarized in online supplemental table S6. Each gene signature score was determined: the z scores of signature genes were adjusted to a normal distribution in the sample, and a single signature score for each patient was calculated.17 The volcano plot was obtained in ggVolcanoR, an R-based Shiny application (https://ggvolcanor.erc.monash.edu/). Gene set enrichment analysis (GSEA) was performed to investigate possible molecular interactions using Hallmark gene and immunologic signature gene sets.18 19

Genomic and variant evaluation

We obtained whole-exome sequencing (WES) data from https://www.cbioportal.org/, and tumor mutation burden (TMB) was defined as total nonsilent somatic mutation counts in the TCGA-BLCA 2020 cohort. To visualize WES data, maftools package (r package, V.2.2.10) was used.20 Gene alterations were outlined as gene mutation complexes and copy number variations, including nonsense, missense, frameshift, splice-site variants affecting consensus nucleotides, or deleterious homozygous deletions and amplifications.21

Multiplex immunohistochemistry

Tissue microarrays (TMA) were constructed from 2 mm cores of an FFPE tissue block with cohorts 4 and 5. An expert genitourinary pathologist (CO) prepared HE-stained slides, which had two representative locations, and two cores were sampled from each case. Subsequently, mIHC was performed with 4 μm thick sections sliced from TMA blocks using the Opal 7 Solid Tumor Immunology Kit (AKOYA Biosciences, Massachusetts, USA). The methodology for mIHC was succinctly outlined in online supplemental methods S1. Online supplemental table S7 shows the details of the primary and secondary antibodies, and online supplemental figures S2 and S3 display representative images.

In previous studies, multiplex immunofluorescence panels were rigorously validated using antibodies for CD8, CD4, Foxp3, and pan-cytokeratin (PanCK).22–24 We confirmed the equivalence between the monoplex and mIHC of LAG-3 and fibrinogen-like protein 1 (FGL1) before this study (online supplemental figure S4).

Multispectral image analysis

The stained whole slides were scanned using a PhenoImager HT Vectra Polaris (AKOYA Biosciences). The inForm V.2.6 (AKOYA Biosciences) image analysis program was used to separate and measure spectrally overlapping markers in multiplex assays at a single-cell level. Based on the 4',6-diamidino-2-phenylindole (DAPI) and PanCK signals, the tissues were divided into tumor, stromal, and other areas with no tissues. The cells were segmented using a machine learning system equipped with inForm software, where the nucleus was defined by DAPI signals. The supplementary CD4 and PanCK signals were assessed to detect the cytoplasm and membrane, respectively. Furthermore, multiplexed phenotyping was performed by evaluating an optimized intensity-based threshold calculated using the percentile rank formula in Spotfire software V.7.8 (TIBCO Software, California, USA). CD8+LAG-3+, CD4+FOXP3+, and PanCK+FGL1+ cells were considered LAG-3+ cytotoxic T-cells, CD4+ regulatory T-cells (Tregs), and FGL1+ tumor cells (TCs), respectively (online supplemental figure S5). The cell density in a respective tumor/stroma area was determined by estimating cell counts from obtained images and equilibrating their distribution (cell/mm2).25 Furthermore, intercellular distance spatial analysis was also performed (online supplemental methods S2). The mean values of the two cores of TMA were defined as representative for each case.

Statistical analysis

The χ2 test was used to analyze categorical variables. Welch’s t-test or one-way analysis of variance was used for parametrically distributed continuous variables, and the Mann-Whitney U test or Kruskal-Wallis test was used for non-parametric distributions to detect statistically significant differences between two or more groups. For multiple comparisons, the Holm-Bonferroni method was used. Correlations between two variables were evaluated using Spearman’s rank correlation test. Survival analysis was conducted using the Kaplan-Meier method, along with the log-rank test and the Cox proportional hazard model. HRs and corresponding 95% CIs were estimated from Cox analyses. Harrell’s concordance index (c-index) was used to estimate the predictive accuracy of the biomarkers for OS. Based on previous studies,23 26 the receiver operating characteristic curve was analyzed to obtain a potential cut-off for cell density and distance, aiming to discriminate PFS in cohort 4. Subsequently, the cut-off value was applied to OS and DFS in cohorts 4 and 5, respectively. All statistical analyses were performed using R statistical program V.4.3.1 (R Foundation for Statistical Computing, Vienna, Austria) and GraphPad Prism V.9.0 software (RRID:SCR_002798), and EZR V.1.54 (Saitama Medical Center, Jichi, Japan). A two-sided p<0.05 was considered statistically significant.

ResultsClinical impact of next-generation IRs and ligands on oncological outcomes

LAG-3, TIGIT, and TIM-3 mRNA expression levels were significantly intercorrelated in cohorts 1, 2, and 3 (all, r>0.558; p<0.001; figure 1A and online supplemental figure S6A). Furthermore, all IR expression levels were correlated with PD-1 and PD-L1 expression (all, r>0.424, p<0.001; online supplemental figure S6B). Patients with high LAG-3 expression levels showed a stronger association with ICI response than those with low expression levels (p=0.027, figure 1B), and no associations were found between TIGIT and TIM-3 expression and response rates. A similar tendency was observed among the IR expression levels in cohort 2 (p=0.081, online supplemental figure S7A). Regarding the predictive accuracy for CR/PR, LAG-3 expression had the highest area under the curve of 0.63 compared with that of the other two IRs (all, p<0.05; figure 1C). Patients with high LAG-3 or TIGIT expression levels exhibited significantly higher OS rates than those with low expression levels (log-rank p=0.044 and p=0.031, respectively; figure 1D). In the validation cohort, the dichotomized levels of all three IR expression levels distinguished OS rates (online supplemental figure S7B). According to the Cox proportional hazard model, LAG-3 and TIGIT had higher c-index values for OS compared with TIM-3 (figure 1E).

Figure 1

Clinical effect of immune coinhibitory receptors (LAG-3, TIGIT, and TIM-3) and ligands in patients with metastatic urothelial carcinoma (UC) who underwent therapy with immune checkpoint inhibitors (ICIs). All analyses were performed using cohort 1 (n=348). (A) 3D scatter plots with a 3D confidence ellipsoid showing mRNA expression levels of LAG-3, TIGIT, and TIM-3. (B) Comparison of mRNA expression levels of LAG-3, TIGIT, and TIM-3 between CR/PR and SD/PD groups. P values refer to those obtained using the Mann-Whitney U test. (C) The area under the receiver operating characteristic curves of mRNA expression levels of LAG-3, TIGIT, and TIM-3 representing the prediction of the CR/PR response. Each predictive accuracy was compared using the DeLong test. (D) Kaplan-Meier curves for OS, stratified by the median expression levels of LAG-3, TIGIT, and TIM-3. Log-rank test was used for statistical analysis. (E) C-indices of LAG-3, TIGIT, and TIM-3 expression for OS, as assessed by the Cox proportional hazard model. *p<0.05. (F) Kaplan-Meier curves for OS stratified by the median expression levels of LAG-3 ligands, including FGL1, MHCII, and GAL-3. Log-rank test. (G) C-indices of FGL1, MHCII, and GAL-3 expression for OS, as assessed by the Cox proportional hazard model. *p<0.05. (H) Kaplan-Meier curves for OS based on the median expression levels of TIGIT ligands, including CD155 and CD112. Log-rank test. (I) C-indices of CD155 and CD112 expression for OS, as shown in the Cox proportional hazard model. (J) Kaplan-Meier curves for OS, stratified by the median expression levels of TIM-3 ligands, including CEACAM1, HMGB1, and GAL-9. Log-rank test. (K) C-indices of CEACAM1, HMGB1, and GAL-9 expression for OS, as outlined in the Cox proportional hazard model. 3D, three dimensional; CR, complete response; OS, overall survival; PD, progression disease; PR, partial response; SD, stable disease.

We focused on LAG-3-related ligands, including FGL1,27 major histocompatibility complex II (MHC II) signature,28 29 and galectin-3 (GAL-3).30 Although no differences in OS rates were observed based on MHCII or GAL-3 expression, higher FGL1 expression was significantly correlated with lower OS rates (p=0.045; figure 1F), which was relatively consistent with the result of cohort 2 (p=0.06; online supplemental figure S7C). The c-index of FGL1 expression for OS was the highest, followed by MHCII and GAL-3 expressions (figure 1G). Regarding TIGIT-related ligands,31 CD155 and CD112 expressions did not affect OS across Cohorts 1 (both p>0.05; figure 1H) and CD155 did not affect OS for cohort 2 (p=0.892; data for CD112 could not be analyzed due to the lack of data; online supplemental figure S7D). The c-index of CD155 expression for OS was slightly higher than that of CD122 (figure 1I). Similarly, the expression levels of TIM-3-related ligands, including colonic CEA cell adhesion molecule 1 (CEACAM1),32 high mobility group box-1 (HMGB1),33 and galectin-9 (GAL-9),34 showed no correlation with OS in both cohorts (all p>0.05; figure 1J and online supplemental figure S7E). The c-indices of CEACAM1 and HMGB1 expressions for OS were higher than those of GAL-9 (figure 1K).

Together with the outcomes related to the ICI response and OS, LAG-3, and its ligand, FGL1 was the most significant clinical marker among the next-generation IRs in patients with mUC subjected to PD-(L)1 blockade.

Clinical and molecular features of LAG-3/FGL1 subtypes

Next, we divided the patients into four LAG-3/FGL1 subtype groups, namely LAG-3highFGL1high, LAG-3highFGL1low, LAG-3lowFGL1high, and LAG-3lowFGL1low, and clarified whether these subtypes were associated with ICI response and OS rate. The baseline characteristics were similar in all groups, except for the PD-1/PD-L1 subgroup, which was defined similarly to the LAG-3/FGL1 subgroup (online supplemental tables S1–S3). LAG-3highFGL1low expression was enriched in the CR/PR group vs the SD/PD group (cohort 1: 44.1% vs 20.4%, p=0.001; cohort 2: 37.5% vs 12.5%, p=0.027; figure 2A). In cohorts 1 and 2, patients with the LAG-3highFGL1low subtype had a significantly higher OS rate compared with those with the LAG-3highFGL1high subtype (p=0.031 and 0.033, respectively; figure 2B).

Figure 2

Clinical and molecular characterization of LAG-3/FGL1 subtypes in patients with metastatic UC treated with ICIs. Analyses were performed involving cohort 1 (n=348) and cCohort 2 (n=89). (A) Distribution of patients within each LAG-3/FGL1 subtype based on the best objective response. χ2 test. (B) Kaplan-Meier curves for OS, stratified by LAG-3/FGL1 subtypes. Log-rank test with Holm-Bonferroni method. (C) Relationship between LAG-3/FGL1 subtypes and main molecular pathways in cohort 1. Rows of the heat map reflect gene expression (Z scores) classified by pathway. Detailed comparisons between each subtype are shown in online supplemental figure S4. (D) Kaplan-Meier curves for OS stratified by PD-1/PD-L1 subtypes. The log-rank test. (E) Forest plots depicting the multivariate Cox hazard regression models integrating PD-1/PD-L1 and LAG-3/FGL1 axes for OS. Angio, angiogenesis; APM, antigen presentation machinery; DDR, DNA damage response and repair; EMT, epithelial-mesenchymal transition; F-TBRS, pan-fibroblast TGFβ response signature; Hi, high; Lo, low; OS, overall survival; TMB, tumor mutation burden; TNB, tumor neoantigen burden.

The relationship between the LAG-3/FGL1 subtype and transcriptomic features is revealed in cohort 1 (figure 2C and online supplemental figure S8A). The FGFR3 signature was enriched in the LAG-3-low groups compared with the LAG-3-high groups. In contrast, tumor immune microenvironment (TIME)-related signatures, such as CD8+ effector T-cell (CD8+ Teff), antigen presentation machinery (APM), immune checkpoint, were higher in the LAG-3-high group than in the LAG-3-low group. The CD8+ Teff, APM, and immune checkpoint expression levels were not significantly different between the LAG-3highFGL1low and LAG-3highFGL1high groups. Meanwhile, the expression of pan-fibroblast TGFβ response signature (F-TBRS), which is associated with TGFβ and the epithelial-to-mesenchymal transition (EMT),13 was significantly more pronounced in the LAG-3highFGL1high subtype compared with the LAG-3highFGL1low subtype.

IMvior210 data covered PD-L1 protein expression on TCs and ICs analyzed using the SP142 IHC assay.12 PD-L1 expression was high in both cell types in the LAG-3-high cohorts compared with the LAG-3-low cohorts (both p<0.001; figure 2C and online supplemental figure S8B), and this trend corresponded to the score of the immune checkpoint gene signature demonstrated in online supplemental figure S8A.

We analyzed the correlations of TMB and genetic alteration with LAG-3/FGL1 subtypes in cohorts 1 and 3. Overall, although the LAG-3highFGL1low subtype tended to have higher TMB and tumor neoantigen burden (TNB) levels than other subtypes, each parameter had no significant statistical difference between LAG-3highFGL1high and LAG-3highFGL1low (figure 2C and online supplemental figure S8C and figure S9A). Regarding the parameters of genomic alterations, the frequency of RB1 alteration, a potential target of CDK4/6 inhibition,35 was the highest in the LAG-3highFGL1low subgroup, followed by the LAG-3lowFGL1high and LAG-3lowFGL1high subgroups (p=0.012; online supplemental figure S9A). The FGFR3 variation, a target of erdafitinib,36 was associated with LAG-3/FGL1 subtypes. Interestingly, its frequency was high in the LAG-3-low group, regardless of FGL1 expression level (p<0.001; online supplemental figure S9A).

Online supplemental figure S6B shows that LAG-3 and PD-(L)1 expressions were highly correlated; however, it remains unclear whether these markers can be confounding factors in response to ICI and OS. Thus, we aimed to elucidate the independence of the LAG-3/FGL1 and PD-1/PD-L1 axes in terms of OS. The four subtypes of PD-1/PD-L1 did not show statistically distinct OS rates in cohorts 1 and 2 (p=0.294 and 0.101, respectively; figure 2D). The outcomes of multivariate Cox regression analysis showed that the independent factors in predicting OS in cohort 1 were ECOG-PS (<2 vs ≥2: HR: 2.02; p=0.039), previous platinum use (HR: 2.02; p=0.026), and LAG-3/FGL1 subtypes (LAG-3highFGL1low vs LAG-3highFGL1high: HR: 0.59; p=0.006) (figure 2E). Similarly, the LAG-3/FGL1 subtype (LAG-3highFGL1low vs LAG-3highFGL1high: HR: 0.21; p=0.016) was an independent factor for OS in cohort 2 (figure 2E). Accordingly, the PD-1/PD-L1 axis was not a confounding factor for the LAG-3/FGL1 axis. The LAG-3highFGL1low subtype was found to be an independent predictive factor for OS in patients treated with PD-(L)1 blockade.

Identification of immunoevasive molecular features related to LAG-3/FGL1 subtypes

We subsequently focused on assessing the influence of CD8+ T-cell status within the tumor microenvironment (TME) on the differential prognoses of LAG-3/FGL1 subtypes, particularly between the FGL1-low and FGL1-high subgroups within the LAG-3-high cohort. The tumor immune phenotype data (desert, excluded, and inflamed) from the CD8 IHC assay from IMvigor21013 revealed that the inflamed phenotype was highly frequent in the LAG-3-high population (p<0.001), but the difference was not significant between the LAG-3highFGL1low and LAG-3highFGL1high patient groups (figure 3A). A compatible pattern was identified in the CD8+ Teff gene signature across cohorts 1–3 (online supplemental figure S8A and figure S9B). Furthermore, in cohorts 1 and 2, characterized by high CD8+Teff signatures, patients with LAG-3highFGL1high exhibited significantly worse OS compared with those with LAG-3highFGL1low (p=0.003 and 0.015, respectively; figure 3B). Thus, high FGL1 expression was strongly associated with poor oncological outcomes following PD-(L)1 blockade therapy, even with the enrichment of CD8+T-cells.

Figure 3

Identification of immunoevasive and tumor progressive contextures associated with the FGL1 expression level in LAG-3-high cohorts. (A) Representative histological images13 and distribution of tumor immune phenotypes based on CD8 IHC, including desert, excluded, and inflamed between LAG-3/FGL1 subtypes in cohort 1 (n=348). Kruskal-Wallis test with Steel-Dwass test. (B) Kaplan-Meier curves for OS, stratified by levels of CD8+ Teff signature score and FGL1 expression in LAG-3-high in cohorts 1 (n=175) and 2 (n=32). The log-rank test using the Holm-Bonferroni method. (C) A volcano plot depicting differentially enriched pathways between the FGL1-high and FGL1-low expression groups in LAG-3-high in cohort 1 (n=175). (D) Hallmark and immunological signature gene set enrichment analysis revealing pathways associated with FGL1-high versus FGL1-low expression. (E, F) T-cell exhaustion markers and immune cell infiltration-related gene signatures, respectively. *p<0.05, **p<0.01 for Mann-Whitney U test. (G) Correlation between the expression of NRP1 and CD4+Treg gene signatures. Spearman’s rank correlation coefficient. (H) Association of cytotoxic gene expressions (IFNγ, GZMB, and PRF1) with FGL1 expression levels. (I) Pathway-related gene signatures based on FGL1 expression levels. Mann-Whitney U test. DC, dendric cell; EMT, epithelial to mesenchymal transition; IHC, immunohistochemistry; OS, overall survival; Teff, effector T-cell; Treg, regulatory T-cell.

To better understand the immunosuppressive composition associated with FGL1 expression, a comparative analysis was performed on the differences in TME features between the LAG-3highFGL1low and LAG-3highFGL1high subtypes. Differential gene expression analysis showed that TIME-related (NRP1, ECM2, STAT3, and CD163), TGF-β signaling pathway (TGFB1, TGFB3, and TGFBR1), and EMT-related genes (VIM, SNAL1, COL4A1/2, and ACTA2) were highly upregulated in FGL1-high versus FGL1-low tumors (figure 3C). Additionally, GSEA revealed the upregulation of pathways involving EMT, KRAS signaling, CD4+T-cell infiltration, angiogenesis, and CD8+T-cell exhaustion in FGL1-high versus FGL1-low tumors, with the highest-ranking pathway identified (figure 3D).

To further outline the immunosuppressive features correlating with FGL1 expression in CD8+LAG-3+-high tumors, subgroup analyses were conducted on the data from patients with CD8+Teff signature- and LAG-3-high tumors in cohorts 1 and 3. The T-cell exhaustion markers TIM-3, CD39, and NRP1 were highly upregulated in FGL1-high tumors compared with FGL1-low tumors (figure 3E and online supplemental figure S9C). Considering immune cell (IC) infiltration-related gene signatures, those of CD4+ Treg, M2-like macrophages, and dendritic cells were upregulated in FGL1-high tumors in Cohort 1, and the CD4+ Treg signature showed a significant increase in cohort 3 (figure 3F and online supplemental figure S9D). Moreover, NRP1 expression, which is essential for the maintenance of intratumoral Treg stability and function,37 38 showed a positive correlation with the CD4+Treg signature (r=0.493, p<0.001) (figure 3G). Additionally, among effector genes, only interferon-γ (IFNγ) expression was linked to FGL1 expression levels (p=0.033; figure 3H).

Finally, we investigated FGL1-mediated signaling pathways involved in ICI resistance.39 In this regard, EMT, angiogenesis, TGFβ, and F-TBRS scores were significantly higher in FGL1-high tumors than in FGL1-low tumors (all, p<0.01; figure 3I).

Quantitative analysis and spatial distribution of targeted markers in UC

Despite capturing trends in gene expression associated with the LAG-3/FGL1 axis through bulk RNA-sequencing data, as demonstrated above, there was insufficient comprehensive spatial information on marker expression. Thus, a single-cell-level mIHC assay was performed to evaluate the validity of these findings. The entire area of TMA samples across cohorts 4 (n=29) and 5 (n=90) was analyzed using an mIHC panel containing markers such as DAPI, CD4, CD8, FGL1, Foxp3, LAG-3, and PanCK (figure 4A). The detailed method of cell segmentation was defined in the Methods section. All immune markers, except for Foxp3 (nuclear staining), were quantified based on the segmentation of membrane staining.40 41 Figure 4B and online supplemental figure S10A show the median cell density for all markers, including double positivity in each tumor/stroma area. The density of tumor FGL1+ cells was significantly elevated (cohorts 4: 554.97 cells/mm², cohort 5: 615.8 cells/mm²), followed by stromal CD4+ cells (cohort 4: 328.44 cells/mm², cohort 5: 166.87 cells/mm²), stromal CD8+ cells (cohort 4: 217.99 cells/mm², cohort 5: 145.31 cells/mm²), and stromal Foxp3+ cells (cohort 4: 209.21 cells/mm², cohort 5: 134.66 cells/mm²).

Figure 4

Evaluation of tumor immune microenvironment using multiplex fluorescence immunohistochemistry using inForm analysis in UC. Analyses were performed using cohort 4 (n=29). (A) Representative low-power and high-power field images stained with seven-color mIHC panel (DAPI in blue, CD4 in turquoise, CD8 in yellow, FGL1 in green, Foxp3 in orange, LAG-3 in red, and PanCK in magenta) (left two panels), tissue segmentation map of tumor (red) and stromal (green) regions (center panel), and cell segmentation map (nucleus in green, membrane outlined in red, and cytoplasm located between nucleus and membrane) (right panel). (B) The overall distribution of spatial cell densities of immune and tumor cell populations. (C) Mean tumor/stroma density distribution and representative image for each immune or tumor cell for each patient across the cohort. The membranous stain was detected for all immune or tumor cells except Foxp3 (nuclei stain). *p<0.05, ****p<0.001. Mann-Whitney U test. (D) A Spearman matrix heatmap comparing the relationship between the cell densities of each marker. (E, F) Correlation of cell density between a tumor FGL1 cell and tumor/stroma CD8+LAG-3+ cell and tumor/stroma CD4+Foxp3+ cell. Spearman’s rank correlation coefficient. CK, cytokeratin; DAPI, 40,6-diamidino-2-phenylindole; mIHC, multiplex fluorescence immunohistochemistry; n.s., not significant; UC, urothelial carcinoma.

We then demonstrated the distribution of cell density in histograms, with each individual arranged horizontally for each marker, confirming the overall validity of distribution between tumor and stroma across both cohorts (figure 4C and online supplemental figure S10B). While CD8+ cell ratios between tumor and stroma varied significantly among the individuals, CD4+ and CD4+FoxP3+ cells consistently showed higher expression in the stroma than in the tumor. Mean bar graphs indicate that CD8+, CD4+, and CD4+FoxP3+ cells were significantly more abundant in the stroma than in the tumor (p<0.05, p<0.001, and p<0.001, respectively); FGL1+ cells, however, exhibited the opposite trend (both p<0.001).

The relationship across markers is depicted in figure 4D and online supplemental figure S10C. We focused on the relationship between the density of FGL1+ cells within the tumor area and CD8+LAG-3+ cells contained in tumor and stromal areas, revealing modest but statistically significant positive correlations in the mUC cohort (Pearson’s r: 0.524 and r: 0.260, respectively; figure 4E). Furthermore, the density of FGL1+ cells within a tumor area and CD4+Foxp3+ cells within the stroma showed a positive linear correlation (Pearson’s r: 0.407; figure 4F), corresponding to the association of the CD4+Treg signature with FGL1 mRNA expression (figure 3F).

Clinical significance of the LAG-3/FGL1 axis spatially analyzed in UC

We initially evaluated the association between cell densities of CD8+, LAG-3+, CD8+LAG-3+, and FGL1+ in the tumor/stroma and the survival rate after pembrolizumab therapy in cohort 4. To assess the prognostic accuracy of each marker for PFS and OS, we derived c-indices from univariate Cox proportional hazard models and compared tumor and stromal areas (figure 5A). The c-indices for the densities of CD8+, LAG-3+, and CD8+LAG-3+ cells in the stromal area were higher than those in the tumor area for PFS and OS, highlighting their greater clinical importance in the stromal area. Conversely, higher c-indices for survival outcomes were observed for cell density of FGL1+ in the tumorous compared with the stromal area (PFS: 0.632 vs 0.520; OS: 0.580 vs 0.566).

Figure 5

Clinical implication of spatially distinct immune and tumor cell densities related to the LAG-3/FGL1 axis. Analyses were performed using cohort 4 (n=29). (A) C-indices of stroma/tumor CD8+, LAG-3+, CD8+LAG-3+, and FGL1+ cell densities for PFS and OS, as outlined by the Cox proportional hazard model. (B) Kaplan-Meier curves for PFS and OS based on stroma CD8/LAG-3 and tumor FGL1 expressions. The log-rank test. (C) Distribution of patients with stroma CD8+LAG-3+/tumor FGL1+ subtypes based on the best objective response to pembrolizumab. (D) Kaplan-Meier curves for PFS and OS, stratified by stroma CD8+LAG-3+/tumor FGL1+ subtypes. Log-rank test and Holm’s test. (E) Kaplan-Meier curves for DFS stratified by stroma CD8+LAG-3+/tumor FGL1+ subtypes in Cohort 5. The log-rank test and *Cox proportional hazard model. CR, complete response; DFS, disease-free survival; OS, overall survival; PD, progression disease; PFS, progression-free survival; PR, partial response; SD, stable disease.

To visualize the survival effect of clinically relevant markers demonstrated above, Kaplan-Meier analyses with a binary class stratification were performed (figure 5B). As expected from the RNA-seq results, the increased densities of stromal CD8+, LAG-3+, and CD8+LAG-3+ cells correlated with significantly or marginally longer PFS and OS in patients treated with PD-1 blockade. Compared with low-density cases, a high TCs density of FGL1 was significantly associated with lower PFS (p=0.005), but not with OS.

Given the specific focus on the correlated expressions of LAG-3 on cytotoxic T-cells and FGL1 on tumors, the tumor types were categorized into four subtypes based on the density of stromal CD8+LAG-3+ cells (sCD8LAG-3)/tumor FGL1+ cells (tFGL1). Subsequently, we estimated their correlation with the ICI response and prognosis following anti-PD1 treatment. The CR population (n=2) was exclusively identified in the sCD8LAG-3hightFGL1low subtype, whereas no CR/PR population was detected in the sCD8LAG-3hightFGL1high and sCD8LAG-3lowtFGL1high subtypes (figure 5C). Compared with patients having the sCD8LAG-3hightFGL1high subtype, patients with the sCD8LAG-3hightFGL1low subtype exhibited a significantly better PFS rate, but not OS (p=0.006 and 0.33, respectively; figure 5D).

Subsequently, we evaluated the clinical significance of the LAG-3/FGL1 axis for DFS in patients who underwent radical cystectomy (cohort 5), exploring the indications for postoperative adjuvant therapy with ICIs. The c-indices for the densities of LAG-3+ and CD8+LAG-3+ cells in the stromal area, but not for CD8+ cells alone, were higher than those in the tumor area in terms of DFS. Meanwhile, the relevance of FGL1+ cells within TCs was more apparent than that in stromal cells (online supplemental figure S11A). Applying the same cut-off values of markers analyzed in figure 5B, the Kaplan-Meier curves also showed a positive correlation of stromal CD8+, LAG-3+, and CD8+LAG-3+ cell densities with significantly higher DFS. Furthermore, a higher TCs density of FGL1+ showed a marginally significant correlation with worse DFS compared with those with a low density (p=0.086; online supplemental figure S11B).

When classifying the tumor types into four subtypes in a manner compatible with cohort 4, patients with the sCD8LAG-3hightFGL1low subtype had a significantly higher DFS rate than those with the sCD8LAG-3hightFGL1high