Anti–PD-1 Efficacy in Patients with Metastatic Urothelial Cancer Associates with Intratumoral Juxtaposition of T Helper-Type 1 and CD8⁺ T cells

Treatment with antibodies directed against the immune checkpoint programmed cell death protein (PD)-1 or its ligand (PD-L1) leads to robust and durable clinical responses in patients with various tumor types, including melanoma and non–small cell lung cancer (1, 2). In patients with advanced and metastatic urothelial cancer (mUC), first-line treatment with pembrolizumab (i.e., anti–PD-1) and atezolizumab (i.e., anti–PD-L1) has shown promising results in single-arm trials (3, 4). Recently, a randomized trial showed that patients with mUC who received first-line treatment with chemotherapy plus atezolizumab had an improved progression-free survival as compared with patients treated with atezolizumab monotherapy or platinum-based chemotherapy plus placebo (5). Yet, no survival benefit was observed in a phase III trial comparing first-line durvalumab (i.e., anti–PD-L1) and durvalumab plus tremelimumab (i.e., anti–CTLA-4) to chemotherapy in patients with mUC (6). Furthermore, in patients with mUC with progressive disease after platinum-based chemotherapy, second-line treatment with pembrolizumab showed superior efficacy as compared with chemotherapy (objective response rates 21% vs. 11%; refs. 7, 8). Despite the limited objective response rate for second-line pembrolizumab, a small proportion of patients with mUC do obtain a durable tumor response (2-year progression-free survival of 12%; ref. 9).

PD-L1 expression in tumor tissue has been investigated extensively as a predictive marker to select patients with mUC for anti–PD-1 treatment (10), but results were conflicting in the second-line setting (11, 12). Currently, the use of PD-L1 expression to select patients for immune checkpoint inhibitors (ICI) is restricted to the first-line setting for cisplatin-ineligible patients with mUC (13, 14). As treatment with ICIs can be accompanied with severe adverse events and is associated with high costs, new markers are needed to identify patients who will benefit from treatment with ICIs. Furthermore, a better understanding of immune mechanisms that underly response and resistance to ICIs would facilitate the development of improved treatments for patients with mUC.

In a pan-cancer setting, tumor response to ICIs has generally been associated with high mutational burden and expression of signature genes downstream of interferon gamma (IFNγ), a major product of activated T cells (15, 16). In patients with mUC, tumor response to ICIs has also been correlated with markers of T-cell response, such as enrichment for T-cell receptor clones (17–20). More recently, the number and spatial organization of intratumoral T cells were observed to associate with survival and tumor response to ICIs in multiple tumor types (21–24), including UC (17, 25). Collectively, these translational studies show that recruitment of T cells and intratumoral interactions between T cells and antigen-presenting cells represent prerequisites for optimal antitumor immune responses (26–28).

In the current study, we investigated whether numerical and phenotypical differences in CD4⁺ and CD8⁺ T-cell subsets were associated with anti–PD-1 efficacy in patients with mUC. To this end, peripheral blood, plasma, and tumor biopsies were prospectively collected from patients treated with pembrolizumab in a clinical trial. Samples were analyzed using multimodality flow cytometry, ELISA and in situ stainings. We observed predominant differences between responders and nonresponders regarding CD4⁺ T-cell subsets. Responders demonstrated enhanced fractions of CD4⁺ T-cell subsets expressing cosignaling receptors in blood, and fractions of these subsets changed during treatment. Furthermore, in metastatic lesions of responders, anti–PD-1 induced the formation of immune cell niches that are rich in CD4⁺ T helper-type 1 (Th1) cells, CD8⁺ T cells, and CD11b⁺ myeloid cells.

Patients with advanced or metastatic UC were included in a phase II prospective biomarker discovery study (RESPONDER trial, NCT03263039). The study protocol was approved by the medical ethics review board of the Foundation BEBO (Evaluation of Ethics in Biomedical Research). The study was conducted in accordance with the Declaration of Helsinki principles for medical research involving human subjects. Written informed consent was obtained from all subjects before any study procedure. Between September 1, 2017 and December 31, 2019, 56 patients with mUC initiated first- (n = 5) or second-line (n = 51) treatment with pembrolizumab (200 mg i.v. every 3 weeks). Prior to the start of therapy and every 12 weeks thereafter, tumor response was evaluated using CT. Patients were classified as responders (complete or partial response according to RECIST v1.1, n = 22) and nonresponders (progressive disease according to RECIST v1.1, n = 34) according to their best overall response during treatment. For the current analyses, data cutoff was set at July 1, 2020. Patients with stable disease showed diverse tumor responses varying between limited progressive disease (<20% increase in diameter of target lesions), limited tumor response (<30% decrease in diameter of target lesions), and mixed response. For this patient group, findings are presented per type of biomaterial in Supplementary Fig. S8 and Supplementary Table S1.

PD-L1 expression was determined on baseline metastatic tumor biopsies using the companion diagnostic assay of pembrolizumab (PD-L1 IHC 22C3 pharmDx, Agilent Technologies), according to the manufacturer's guidelines (DAKO platform, Maastricht University Medical Center, Maastricht, the Netherlands). All tissues were assessed for the PD-L1 CPS by an expert genitourinary pathologist (G.J.L.H. van Leenders; refs. 3, 7). Nine of 14 patients (64%) with a CPS ≥10 obtained a tumor response, compared with 8 of 27 patients (30%) with a CPS <10 (Table 1). Patient characteristics as well as patient numbers per type of biomaterial are provided in Table 1.

Peripheral blood was collected in EDTA tubes at three time points: pretreatment [baseline (Bl)] and prior to the third and fifth administration of pembrolizumab (weeks 6 and 12; see Supplementary Fig. S1 for study design and patient numbers). Whole blood was used to enumerate immune cell populations, and PBMCs and plasma were isolated using Ficoll gradient centrifugation. Samples were stored according to standardized protocols and thawed at later time points to assess fractions of T-cell subsets (PBMCs) and levels of chemo-attractants (plasma). Tumor biopsies were collected from a safely accessible metastatic lesion at baseline, and from the same lesion after 6 weeks of treatment (see Table 1 for biopsy sites). Tissue specimens for multiplex immunofluorescence stainings were formalin-fixed and paraffin-embedded (FFPE). At baseline, an additional biopsy from the same lesion was snap frozen for RNA sequencing (RNA-seq) according to a standardized procedure (29).

Whole blood was stained, lysed to remove red blood cells, and analyzed by multiplex flow cytometry using a BD 3-laser Celesta flow cytometer and FACSDIVA 8.x software. Cell counts of major T-cell populations were determined using Flow-Count Fluorospheres (Beckman Coulter). PBMC samples were stained with a mix of antibodies directed against T-cell markers of maturation, coinhibition, costimulation, or chemotaxis as described previously (30). All antibody panels were optimized and compensated using fluorescence minus one (FMO) controls and measurements were corrected for background fluorescence. Data were gated and analyzed using FlowJo software.

CD4⁺ T-cell populations were analyzed in Python (Python Software Foundation, v3.6.8) using in-house written scripts. First, data and corresponding background from nonstained samples were normalized for spillover, and logicle transformed (31), after which the median of background was subtracted from each sample for each marker. These data were further processed using a Python adaptation of the self-organizing map algorithm FlowSOM (32). To this end, data were centered and scaled resulting in mean marker intensities of 0 and standard deviations of 1 across all samples. Second, a self-organizing map was trained using 100 nodes, and a minimal spanning tree was built to form 20 meta-clusters. These meta-clusters corresponded to the 20 most abundant CD4⁺ T-cell subsets according to the set of markers used. A meta-cluster was considered positive or negative for a certain marker if the mean intensity of that marker in the cluster was more than its standard deviation above or below the mean intensity of that marker across all data, respectively. Clusters with equal marker positivity were combined, resulting in 17 to 18 unique clusters per set of markers. Abundances of these T-cell clusters were assessed per patient, and uniform manifold approximation and projections (UMAP; ref. 33) were generated using 200,000 random data points to visualize marker intensities across clusters. Finally, differential abundance of clusters was analyzed between time points (baseline, week 6, week 12) and patient groups (responders and nonresponders).

Levels of chemo-attractants in plasma were determined with the following ELISA kits according to the manufacturer's instructions: CCL3 (Invitrogen), CCL5 (BioLegend ELISA MAX), CXCL9 (Thermo Fisher Scientific), CXCL10 (BioLegend ELISA MAX), and XCL1 (Invitrogen). Different ELISA assays were performed simultaneously to minimize biological variation, and to avoid repeated freeze and thaw cycles of samples.

Multiplex immunofluorescence was performed using OPAL reagents (Akoya Biosciences) on 4-μm sections of FFPE tumor biopsies. In brief, stainings were performed in multiple cycles of the following: antigen retrieval (15-minute boiling in antigen retrieval buffer, pH 6 or pH 9 depending on primary antibodies) followed by cooling, blocking, and consecutive staining with primary antibodies, HRP-polymer, and Opal fluorophores. Cycles were repeated until all markers were stained. Finally, nuclei were stained with DAPI. The sequence of antibody stainings was as follows: (i) CD4 (FP1600/EP204, Akoya Biosciences, 1:100) – OPAL540; (ii) CD8 (FP1601/144B, Akoya Biosciences, 1:200) – OPAL690; (iii) CXCR3 (HPA045942, Sigma, 1:100) – OPAL620; (iv) T-bet (4B10, eBioscience, 1:50) – OPAL570; (v) PD-1 (NAT105, ImmunoLogic, 1:50) – OPAL520; (vi) CD11b (EP1345Y, Abcam, 1:200) – OPAL650; (vii) Cytokeratin-Pan (AE1/AE3, Invitrogen, 1:200) – Coumarin (tyramide signal amplification coumarin system, Akoya Biosciences); and (viii) DAPI.

To assess immune cell clusters for the presence of B cells, we additionally stained consecutive tissue sections of responders with the following antibodies: (i) CD3 (SP7, Sigma Aldrich, 1:250) – OPAL520; (ii) CD20 (L26, Cell Marque, 1:1,000) – OPAL620; (iii) CD8 (144B, Cell Marque, 1:400) – OPAL570; (iv) Cytokeratin-Pan (AE1/AE3, Invitrogen, 1:200) – OPAL690; and (v) DAPI.

Whole slides were scanned and images were obtained using VECTRA 3.0 (Akoya Biosciences), after which at least four stamps (regions of interest; stamp size: 671 × 500 μm²; resolution: 2 pixels/μm; pixel size: 0.5 × 0.5 μm²) were set in nonnecrotic areas to cover >90% of tissue area. Images were spectrally unmixed using inForm software (Akoya Biosciences) to visualize markers of interest as well as autofluorescence. Subsequently, images were manually analyzed in Python using the following five steps (Tumor Microenvironment Analyzer, H.E. Balcioglu, manuscript in preparation): (i) Foreground selection: images were thresholded using all channels to generate a foreground area that covers regions positive for all signals. (ii) Tissue segmentation: cytokeratin-positive and -negative regions of the foreground were identified as tumor and stroma regions, respectively. For steps i and ii, regions that were deemed too small were excluded. (iii) Nucleus detection: background signal was subtracted from the DAPI signal through a rolling ball filter algorithm, and the nuclear mask was obtained by thresholding. (iv) Nucleus and cell segmentation: clusters of nuclei were segmented by applying a watershed segmentation algorithm to the nuclear mask, where regions that were deemed too small were excluded. Once identified, nuclei were used for cell detection through a Voronoi segmentation algorithm. (v) Phenotyping: fluorescent intensities for the channels that correspond to each marker were analyzed per cell and nucleus (in case of T-bet). Cells were assigned to either tumor or stroma regions according to the location of the center of their nuclei. Spatial orientations of individual cells were assigned according to the center of their area. Cellular densities were calculated by dividing the number of cells with a certain phenotype by the total area of that region (i.e., tumor, stroma, or both), and were averaged per patient across all stamps. Distances from a cell with a certain phenotype to the nearest cell with another phenotype were calculated by nearest neighbor analysis; this was done in the region of interest (i.e., tumor, stroma, or both), and was averaged across all cells, and per patient across all stamps.

Total RNA was extracted from freshly obtained metastatic tumor biopsies using the QIAsymphony Kit (Qiagen). Samples with a minimum of 100 ng total RNA were sequenced by paired-end sequencing on the Illumina NextSeq 550 (2 × 75 bp) and NovaSeq 6000 (2 × 150 bp) platforms. Samples were visually inspected with FastQC (v0.11.5). Sequence adapters (Illumina TruSeq) were trimmed using Trimmomatic v0.39 (34), and paired-end reads were aligned to the human reference genome (GRCh37) using STAR v2.7.1a (35) with genomic annotations from GENCODE hg19 release 30 (36). FeatureCounts v1.6.3 (37) was applied to count the number of overlapping reads per gene; only uniquely mapped reads were counted per exon and summarized per gene. RSEM v1.3.1 (38) was applied to quantify RNA expression into transcripts per million (TPM) values.

To delineate the relative presence of immune cells, RSEM read counts of all transcripts were analyzed with the R package immunedeconv (39) using the quanTIseq method (40). To capture different CD4⁺ and CD8⁺ T-cell subsets, we used previously published gene expression signatures (37, 41, 42). In addition, gene expression levels of the chemo-attractants CCL3, CCL5, CXCL9, CXCL10, and XCL1 were assessed. Gene expression values were median centered, and the mean expression of each individual gene or set of genes was displayed as gene or signature scores, respectively.

Statistical analysis was performed using R version 3.5.1. Mann–Whitney U test was used to compare continuous variables of two unpaired groups (responders vs. nonresponders). Wilcoxon signed rank test was used to compare continuous variables of two paired groups (time points). Spearman correlation was used to assess linear relationships between continuous variables. Differences were considered significant when P < 0.05. Because of the exploratory nature of this study, no correction for multiple testing was applied.

Fresh peripheral blood samples were obtained at baseline (n = 56), and after 6 (n = 43) and 12 weeks (n = 31) of treatment with pembrolizumab (Supplementary Fig. S1, see Materials and Methods for details). The numbers of T cells and their major subsets were measured in blood using Flow-Count Fluorospheres. The numbers of αβ T cells, γδ T cells, CD4⁺ T cells, and CD8⁺ T cells (per μL of blood) were not significantly different between responders and nonresponders at baseline, and remained unchanged after 6 and 12 weeks of treatment (Supplementary Fig. S2A). T cells were further classified according to the expression of 20 markers regarding cosignaling, maturation, and chemotaxis using multiplex flow cytometry. In particular, fractions of CD4⁺ T-cell subsets (Figs. 1–3; Supplementary Figs. S2 and S3, as described below), and to a lesser extent CD8⁺ T-cell subsets (Supplementary Table S2), were significantly different between responders and nonresponders.

To study CD4⁺ T cells in detail, dimensionality reduction analysis was performed. Analysis of coinhibitory receptor expression revealed 18 different clusters of CD4⁺ T cells, each with varying expression levels of PD-1 (CD279), TIM-3 (CD366), LAG-3 (CD223), and BTLA (CD272; Fig. 1A; Supplementary Fig. S3A). The abundance of 11 of 18 clusters was significantly different between time points or between responders and nonresponders. Most clusters (i.e., clusters 2, 11, 12, 13, 17, 18) contained CD4⁺ T cells with high expression of PD-1, and abundances of these clusters significantly decreased from baseline to week 6 and week 12 in responders (Fig. 1B; Supplementary Fig. S3A). Using standard flow cytometry analysis, we observed that at baseline, responders harbored higher fractions of PD-1⁺ BTLA⁺ CD4⁺ T cells as well as LAG-3⁺ BTLA⁺ CD4⁺ T cells (Fig. 1D). In addition, the fraction of PD-1⁺ CD4⁺ T cells decreased significantly during treatment in responders (Fig. 1C; median fractions of 16%, 8%, and 8% at baseline, week 6, and week 12, respectively), whereas this decrease was only observed at week 6 in nonresponders. In fact, when analyzing CD4⁺ T cells that coexpress PD-1 and BTLA, PD-1 and LAG-3, or PD-1 and TIM-3, this decrease was only observed in responders (Fig. 1D).

In addition to coinhibitory receptors, CD4⁺ T cells were investigated for the expression of the costimulatory receptors CD28, ICOS (CD278), CD40 L (CD154), 4-1BB (CD137), and/or OX40 (CD134), as well as the maturation/activation markers CD45RA, CCR7 (CD197), CD27, CD69, CD95, and/or CD103 (Fig. 2; Supplementary Figs. S2B and S2C and S3B and S3C). At baseline, the fractions of 4-1BB⁺ CD4⁺ T cells and CD28⁺ 4-1BB⁺ CD4⁺ T cells were higher in responders compared with nonresponders (Fig. 2A and B). The fractions of 4-1BB⁺ CD28⁺ CD4⁺ T cells and 4-1BB⁺ ICOS⁺ CD4⁺ T cells significantly increased from baseline to week 12 in responders (Fig. 2B), whereas an increase in the fraction of 4-1BB⁺ CD4⁺ T cells was only observed at week 6 in nonresponders (Fig. 2A). In addition, at baseline and week 6, the fraction of effector memory (CD45RA⁻ CCR7⁻) CD4⁺ T cells was higher in responders compared with nonresponders. Finally, the baseline fraction of CD4⁺ T cells with an activation phenotype (CD95⁺ CD103⁺ CD4⁺ T cells) was higher in responders compared with nonresponders (Supplementary Fig. S2C).

To address T-cell recruitment, CD4⁺ T cells were assessed for the expression of the chemo-attractant receptors CCR1 (CD191), CCR4 (CD194), CCR5 (CD195), CXCR3 (CD183), and/or CXCR4 (CD184). Dimensionality reduction and (co-)expression analysis revealed no differences between responders and nonresponders at baseline. At week 6, the fraction of CCR1⁺ CD4⁺ T cells was higher in nonresponders compared with responders. Notably, treatment-induced changes were only observed in nonresponders, where the fraction of CXCR3⁺ CD4⁺ T cells decreased at week 12 (Fig. 3A and B; Supplementary Fig. S3D). To complement flow cytometry findings, the levels of CCL3 and CCL5 (ligands for CCR1 and CCR5), CXCL9 and CXCL10 (ligands for CXCR3), and XCL1 (ligand for XCR1) were measured in plasma using ELISA. At baseline, plasma levels of CCL5 were lower, whereas levels of CXCL9 were higher in responders compared with nonresponders. At weeks 6 and 12, the levels of all chemo-attractants increased in responders, except for CCL5. In nonresponders, an increase in the levels of CXCL9 and CXCL10 was also observed during treatment, although to a lesser extent (Fig. 3C).

To obtain a better understanding of differences observed in blood, the presence of CD4⁺ T cells, CD8⁺ T cells, and CD11b⁺ myeloid cells was assessed in metastatic lesions using multiplex immunofluorescence (Fig. 4A; see Materials and Methods for details). Tumor biopsies were available for 45 of 56 patients (Supplementary Fig. S1; n = 18 responders, n = 27 nonresponders), and were obtained from different metastatic sites, including lymph node, liver, and lung (Table 1). Baseline and on-treatment biopsies were obtained from the same metastatic lesion. CD4⁺ and CD8⁺ T cells were further assessed for expression of PD-1 and CXCR3, whereas the expression of T-bet, a transcription factor typical for IFNγ-producing Th1 cells, was assessed specifically for CD4⁺ T cells. The vast majority of all immune cells was present in the stroma (i.e., cytokeratin-negative area), therefore tumor (i.e., cytokeratin-positive area) and stroma areas were not distinguished in further image analysis (Supplementary Fig. S4A–S4C). The intratumoral densities (calculated as the number of cells per mm²) of CD4⁺ T cells, CD8⁺ T cells, and CD11b⁺ myeloid cells were similar among responders and nonresponders at baseline, and remained unchanged during therapy (Fig. 4B, left). Principal component analysis demonstrated that biopsy site did not affect intratumoral immune cell densities (Supplementary Fig. S4D). In line with the observations in peripheral blood, most differences between responders and nonresponders were observed in CD4⁺ T-cell subsets (Fig. 4B, middle) rather than CD8⁺ T-cell or CD11b⁺ myeloid cell subsets (Fig. 4B, right). At baseline, the density of T-bet⁺ CD4⁺ T cells (i.e., Th1 cells) was significantly higher in responders compared with nonresponders (Fig. 4A and B). When focusing on Th1 subsets, responders had significantly higher baseline densities of PD-1⁺ Th1, and PD-1⁺ CXCR3⁺ Th1 cells (Fig. 4B, middle). Interestingly, the intratumoral density of Th1 cells correlated positively with the intratumoral density of CD8⁺ T cells (Supplementary Fig. S5). Upon treatment, no changes in the density of CD4⁺ T-cell subsets were observed.

To further assess the contribution of CD4⁺ T-cell subsets, we interrogated RNA-seq data of the same lesions for the relative presence of immune cells and for different CD4⁺ and CD8⁺ T-cell subsets using reported signatures (n = 18 patients; see Materials and Methods section and Table 1 for details). We observed a higher total immune cell fraction in responders compared with nonresponders. In addition, the scores for CD4⁺ T-cell subset signatures, specifically those capturing activated, cytotoxic, Th1, or T follicular helper cells, were higher in responders. These comparisons did not reach statistical significance due to the limited number of patients. Intratumoral gene expression of most chemo-attractants, particularly CXCL10, was also higher in responders, and it is noteworthy that high chemo-attractant expression was associated with high T-cell signature scores (Supplementary Fig. S6).

Intratumoral networks between CD4⁺, CD8⁺, and CD11b⁺ cells were assessed by measuring distances between these cells using nearest neighbor analyses (see Materials and Methods for details). At baseline, the distance between tumor cells (cytokeratin⁺) and myeloid cells, and tumor cells and Th1 cells were smaller in responders compared with nonresponders (Fig. 5A and B), and the latter correlated positively with plasma levels of CCL5 (Supplementary Fig. S5). Distances from Th1 cells to either CD8⁺ T cells or CD11b⁺ myeloid cells did not differ between responders and nonresponders at baseline. Upon treatment, distances between immune cell subsets generally became smaller and showed less variance in responders, but not in nonresponders. Particularly, the distances between Th1 and CD8⁺ T cells, Th1 and CD11b⁺ cells, and Th1 and tumor cells decreased upon treatment in responders, and were smaller compared with nonresponders at week 6 (Fig. 5B–D). Furthermore, distances between CD11b⁺ cells and CD4⁺ and CD8⁺ cells expressing CXCR3 decreased in responders upon treatment (Fig. 5C). To study whether the clusters of immune cells are related to tertiary lymphoid structures (TLS), we performed CD20 stainings on consecutive tissue sections. We observed co-occurrence of B cells in immune cell clusters in the majority of responders (Supplementary Fig. S7A), as well as a decrease in distance between CD20⁺ cells and CD8⁺ cells (Supplementary Fig. S7B). Collectively, the smaller distances between Th1 cells, CD8⁺ T cells, myeloid cells, and B cells upon anti–PD-1 treatment in responders reflect the formation of intratumoral immune cell niches (Fig. 5A, left).

In this study, prospectively collected paired sets of blood, plasma, and tumor samples from patients with mUC treated with pembrolizumab were analyzed for peripheral blood fractions and tissue localizations of T-cell subsets. In particular, pretreatment fractions and treatment-induced changes in circulating CD4⁺ T-cell subsets were different between responders and nonresponders. Responders displayed higher fractions of circulating CD4⁺ T cells expressing cosignaling receptors (i.e., PD-1⁺ BTLA⁺ and 4-1BB⁺ CD28⁺) at baseline, and showed durable changes in these fractions during treatment. Prior to treatment, responders also had high intratumoral densities of Th1 cells, which became part of niches containing immune effector cells upon treatment. These findings are summarized in Supplementary Fig. S9 and suggest that the mechanism of action of anti–PD-1 therapy in patients with mUC relies heavily on activation of CD4⁺ T cells, and intratumoral clustering of Th1 cells.

Most studies investigating ICI-induced antitumor immune responses have focused on CD8⁺ T cells (43, 44). For example, distinctive frequencies and localizations of CD8⁺ T cells were shown in patients with melanoma (23) and non–small cell lung cancer who obtained a tumor response to anti–PD-1 (45, 46). We report that patients with mUC with a tumor response to anti–PD-1 were predominantly characterized by presence and phenotype of CD4⁺ T-cell subsets, suggesting that the contribution of CD4⁺ and CD8⁺ T-cell subsets to anti–PD-1 efficacy may vary per tumor type. In line with our findings, preclinical studies in mouse models of UC and non–muscle invasive bladder cancer showed that the efficacy of anti–PD-L1 treatment depends on the presence of CD4⁺ T cells (47) and that efficacy of anti–PD-1 treatment is lost after CD4⁺ T-cell depletion (48). Furthermore, Oh and colleagues recently demonstrated that primary bladder tumors from treatment-naïve patients contained different CD4⁺ T-cell subsets, including a cytotoxic CD4⁺ T-cell subset that was able to recognize and kill tumor cells in a major histocompatibility complex (MHC) class II–dependent manner in vitro. A gene signature related to these cytotoxic CD4⁺ T cells correlated with response to anti–PD-L1 in patients with mUC (42), and higher scores of this signature were also observed in responders in our cohort. It may be that the CD4⁺ T-cell dominance that we observed is a consequence of downregulated MHC class I expression by urothelial cancer cells (49), which is a well-known immune escape mechanism resulting in the inability of CD8⁺ T cells to recognize tumor antigens.

Responders were characterized by a higher baseline fraction of circulating CD4⁺ T cells that express cosignaling receptors, and that had a more differentiated phenotype. These findings suggest that peripheral CD4⁺ T cells have an antigen-experienced phenotype in responders prior to anti–PD-1 treatment, which may reflect an earlier antitumor immune response that has become prematurely halted. Up until 12 weeks of anti–PD-1 treatment, the fraction of circulating PD-1⁺ CD4⁺ T cells decreased and the fraction of 4-1BB⁺ CD4⁺ T cells increased in responders, most likely as a result of renewed activation of CD4⁺ T cells in tumor lesions. Along these lines, the fraction of activated CD4⁺ T cells (CD95⁺, CD103⁺) correlated positively with the fraction of 4-1BB⁺ CD4⁺ T cells (Supplementary Fig. S5). The changes in fractions of CD4⁺ T cells expressing cosignaling receptors in nonresponders were not long-lasting, which most likely mirrors an insufficient antitumor immune response in these patients. Our findings extend earlier observations that T-cell costimulation, which can be provided by 4-1BB, is required to rescue T cells via anti–PD-1 treatment (50). Therefore, nonresponders may benefit from combination therapy with agonistic antibodies targeting costimulatory receptors, to compensate for the lack of cosignaling during treatment.

When assessing T-cell recruitment to tumor tissues, we observed that treatment-induced changes in circulating CD4⁺ T cells expressing chemo-attractant receptors occurred exclusively in nonresponders. In fact, there was a sharp decrease in the fraction of CXCR3⁺ CD4⁺ T cells in nonresponders, which may be in line with recent studies in which repression of CXCR3 or loss of chemo-attractant expression led to abrogation of the efficacy of anti-PD(L)1 treatment (51–53). In extension to flow cytometry, we assessed levels of chemo-attractants in tumors and plasma. At baseline, intratumoral gene expression levels of CCL5, CXCL9, CXCL10, and XCL1 were higher in responders compared with nonresponders, and this was generally associated with higher scores of T-cell signatures. Interestingly, baseline plasma levels of CCL5 were lower in responders, and correlated with lower intratumoral distances between Th1 cells and tumor cells. This may suggest that local production of CCL5 supports interactions between Th1 cells and tumor cells. In responders, plasma and intratumoral CXCL9 levels were higher at baseline. These findings extend earlier reports in which CXCL9 was related to effective antitumor immune responses and survival in several tumor types (54, 55), and favorable response to anti-PD(L)1 therapy in patients with mUC (17, 19). During treatment, plasma levels of both CXCL9 and CXCL10 significantly increased in responders and to a lesser extent in nonresponders. This may reflect a positive feedback loop to recruit more T cells to the tumor site, yet we cannot exclude this to reflect a general systemic treatment effect as we did not observe a correlation between intratumoral and plasma chemo-attractant levels.

In metastatic lesions, responders displayed a 2-fold higher density of Th1 cells at baseline. Immune cell densities remained unchanged upon treatment; however, distances between Th1 cells, CD8⁺ T cells, and CD11b⁺ myeloid cells significantly reduced in responders, suggestive of active clustering of these cells into immune cell niches. Notably, when studying the treatment response of the biopsied lesions, the baseline density of Th1 cells was higher, and distances between Th1, CD8⁺, and CD11b⁺ cells were lower in lesions with a decrease in diameter (≥30%, according to RECIST v1.1 criteria) compared with lesions with an increase in diameter (≥20%) after 12 weeks of treatment (P = 0.078 and P < 0.05, respectively). Mechanistically, pembrolizumab most likely enhances IFNγ production by Th1 cells, resulting in activation of nearby antigen-presenting CD11b⁺ cells, such as dendritic cells, which subsequently initiate the formation of immune cell niches. Interestingly, dendritic cells express XCR1 and are a major target for XCL1 (56), a chemo-attractant that increased in plasma of responders during treatment. The niches we detected also contained B cells, and resemble TLS, which generally contain CD4⁺ T cells, B cells, and other antigen-presenting cells, and to a lesser extent CD8⁺ T cells. Intratumoral presence of TLS has been observed in several cancer types, and has previously been associated with response to ICI therapy (57–60).

It is noteworthy that the immune parameters that distinguish responders from nonresponders (Supplementary Fig. S9) are partly shared by patients with stable disease and are partly unique to responders (Supplementary Fig. S8 and Supplementary Table S2). Particularly, responders were similar to patients with stable disease regarding fractions of PD-1⁺, and PD-1⁺ BTLA⁺ CD4⁺ T cells in blood, as well as plasma levels of CCL5 and CXCL9. In contrast, responders deviated from patients with either stable disease or nonresponse regarding fractions of 4-1BB⁺, CD28⁺ 4-1BB⁺, and CXCR3⁺ CD4⁺ T cells in blood. Importantly, the increased density of T-bet⁺ CD4⁺ T cells at baseline, and shortened distances between immune cells at 6 weeks, also represented unique immune characteristics of responders.

In this study, the majority of patients received pembrolizumab as second-line treatment. PD-L1 expression, mostly assessed on archival tumor tissue, did not have predictive value for treatment response in the registration trial (7). In our study, the predictive value of PD-L1 expression was also limited, as high PD-L1 expression was observed in half of the responders, and a third of the nonresponders. The intratumoral presence of CD4⁺ T-cell subsets at baseline, and the dynamic changes in cosignaling receptor-expressing CD4⁺ T cells in blood, may provide new and early predictive markers for response to anti–PD-1. Particularly, the observations in blood can be measured in routine diagnostic laboratories at relatively low cost. Before clinical implementation, our results require validation in larger cohorts of anti–PD-1–treated patients with mUC. Furthermore, the long-term evolution of CD4⁺ and CD8⁺ T-cell responses in patients with durable responses to anti–PD-1 should be assessed in follow-up studies. With respect to limitations of the study, tumor biopsies from a single metastatic lesion have precluded the assessment of intratumoral and intertumoral heterogeneity, both of which are recognized for their impacts on ICI response (61, 62). Finally, as our study is limited by (in part) correlative analyses, studies in preclinical models of mUC are recommended towards a more defined understanding of the mechanism of action of anti–PD-1 treatment in patients with mUC.

In conclusion, we showed that efficacy of anti–PD-1 treatment in patients with mUC is accompanied by dynamic changes in the fractions of peripheral CD4⁺ T cells expressing cosignaling receptors, and juxtaposition of Th1 and immune effector cells at metastatic lesions.

M.J.B. Aarts reports grants from Pfizer outside the submitted work. J. Voortman reports personal fees from F. Hoffmann-La Roche, MSD Oncology, Merck, Astellas Pharma, TEVA, Sanofi, Ipsen, and Pfizer outside the submitted work. H.M. Westgeest reports personal fees and other support from Astellas; other support from Ipsen; and personal fees from Roche outside the submitted work. J.L. Boormans reports personal fees from BMS, MSD, Ambu, Eight Medical, and Janssen outside the submitted work. R. de Wit reports grants from Merck during the conduct of the study. R. de Wit also reports personal fees from Merck; grants and personal fees from Sanofi and Bayer; and personal fees from Astellas, Janssen, and Orion outside the submitted work. M.P. Lolkema reports grants and personal fees from MSD during the conduct of the study; M.P. Lolkema also reports grants and personal fees from JNJ, Sanofi, and Astellas, as well as personal fees from Amgen, Roche, Novartis, Incyte, Pfizer, Julius Clinical, and Servier outside the submitted work. A.A.M van der Veldt reports other support from BMS, MSD, Merck, Roche, Novartis, Eisai, Ipsen, Pierre Fabre, Pfizer, Sanofi, and Bayer outside the submitted work. R. Debets reports grants from MSD during the conduct of the study; R. Debets also reports other support from Pan Cancer T outside the submitted work, as well as a patent for European patent application no. 21184727.2 pending to Erasmus MC. No disclosures were reported by the other authors.

M. Rijnders: Conceptualization, resources, data curation, formal analysis, investigation, visualization, methodology, writing–original draft, project administration. H.E. Balcioglu: Software, formal analysis, visualization, writing–review and editing. D.G.J. Robbrecht: Resources, data curation, writing–review and editing. A.A.M. Oostvogels: Investigation. R. Wijers: Investigation. M.J.B. Aarts: Resources, writing–review and editing. P. Hamberg: Resources, writing–review and editing. G.J.L.H. van Leenders: Investigation, writing–review and editing. J.A. Nakauma-González: Software, formal analysis, writing–review and editing. J. Voortman: Resources, writing–review and editing. H.M. Westgeest: Resources, writing–review and editing. J.L. Boormans: Conceptualization, resources, writing–original draft, writing–review and editing. R. de Wit: Conceptualization, resources, funding acquisition, writing–original draft, writing–review and editing. M.P. Lolkema: Conceptualization, resources, funding acquisition, writing–original draft, writing–review and editing. A.A.M. van der Veldt: Conceptualization, resources, writing–original draft, project administration, writing–review and editing. R. Debets: Conceptualization, resources, supervision, funding acquisition, methodology, writing–original draft, project administration, writing–review and editing.

This research was funded by a grant from Merck Sharpe & Dome, Kenilworth, N.J., U.S.A., through M.P.J.L. We thank all local principal investigators and the nurses of all contributing centers for their help with patient recruitment. We are particularly grateful to all participating patients and their families.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.