Different PD-L1 Assays Reveal Distinct Immunobiology and Clinical Outcomes in Urothelial Cancer Open Access

The analysis of the expression of PD-L1 by IHC is among the most extensively studied pretreatment assays to determine biomarkers of response to PD-1/PD-L1 immune checkpoint blockade (ICB) in urothelial cancer and other malignancies. The association between PD-L1 protein expression and benefit from PD-1/PD-L1 ICB in urothelial cancer has been largely inconsistent across clinical trials, a finding commonly attributed to (i) the dynamic nature of PD-L1 expression, (ii) the potential disconnect between PD-L1 expression assessed on archival primary tumors and the setting of treatment administered for metastatic disease, (iii) incomplete understanding of the relationship between PD-L1 level and outcomes with standard treatments in control arms of randomized trials, (iv) variation in the IHC assays used to detect PD-L1, and (v) heterogeneity in cellular populations scored for PD-L1 expression across different assays [i.e., cancer cells, immune cells (IC), or both] and cut points used to define “PD-L1 high” (1). The lack of harmonization of approaches to PD-L1 IHC has often been viewed as a shortcoming of the immuno-oncology field (2). However, the extent to which various approaches to PD-L1 testing may advance insights related to tumor immunology has been underexplored.

The randomized phase III IMvigor130 trial did not demonstrate improved overall survival (OS) in the intention-to-treat population when the anti–PD-L1 atezolizumab was compared with platinum-based chemotherapy in patients with treatment-naïve metastatic urothelial cancer (mUC; refs. 3, 4). However, an exploratory analysis revealed a trend toward improved survival with atezolizumab versus platinum-based chemotherapy in patients with mUC and high PD-L1 IC expression (IC2/3) assessed using the Ventana SP142 IHC assay in the overall population [HR, 0.68; 95% confidence interval (CI), 0.41–1.08] and in the subset of cisplatin-ineligible patients (HR, 0.53; 95% CI, 0.30–0.94; refs. 3, 4). Here, we sought to understand the immunological and clinical implications of PD-L1 expression using different PD-L1 assays in the IMvigor130 cohort.

Patients enrolled in the phase III IMvigor130 trial (NCT02807636) were randomized to receive atezolizumab, with or without platinum-based chemotherapy, or placebo plus platinum-based chemotherapy as first-line treatment for mUC (Supplementary Fig. S1). All patients gave written informed consent, which extended to the use of samples for this study, and IMvigor130 was carried out based on good clinical practice guidelines and the Declaration of Helsinki. Protocol approval was obtained from independent review boards (IRB) or ethics committees (EC) for each participating site. The first IRB approval was granted on May 3, 2016, for the Samsung Medical Center (Seoul, Republic of Korea) study site; IRB/EC approval was subsequently obtained across all participating study sites (a total of 221 sites). For further details on study design, see original publication (4).

Tumor tissues from 655 patients were evaluated for PD-L1 expression using the SP142 IHC assay (Ventana Medical Systems, Inc.) and the 22C3 IHC assay (Dako); no blood or control/normal tissue was used. The samples were collected prospectively as a requirement for eligibility on the IMvigor130 trial, and the PD-L1 SP142 assay was performed prospectively, as previously described (4), during the conduct of the IMvigor130 study and was used for patient stratification. The central laboratory received only one block per patient or a set of serial unstained sections from the same block. A hematoxylin and eosin stain was done on the first section. The SP142 PD-L1 assay was done on the second section. The 22C3 PD-L1 assay was done on a consecutive section from the same block or series of unstained sections.

PD-L1 status was assessed for this study using the PD-L1 IHC 22C3 pharmDx assay (Agilent Technologies) as previously described (5) at CellCarta. PD-L1 testing by IHC on IMvigor130 samples was completed retrospectively in June 2021. New slides were cut from the original formalin-fixed, paraffin-embedded (FFPE) blocks at that time. PD-L1 expression was measured using the combined positive score (CPS), defined as the number of PD-L1–staining cells [tumor cells (TC), lymphocytes, and macrophages] divided by the total number of viable TCs, multiplied by 100. All slides were evaluated by a trained pathologist blinded to the patients’ clinical data. To deconvolute the cancer cell and immune cell components of the CPS, the slides were rescored by a trained pathologist to assign separate PD-L1 staining scores to the cancer cell component and the immune cell component on each slide.

Five patient samples from the IMvigor130 trial that were stained for both SP142 (prospectively) and 22C3 (retrospectively), and with available FFPE blocks, were selected for preparing sections for restaining with the same PD-L1 IHC protocols described above (SP142 or 22C3), followed by consecutive staining for dendritic cell (DC)-LAMP3 on the same section. The PD-L1 staining was followed by staining with DC-LAMP3-CD208 (Abcam rabbit monoclonal antibody clone EPR24265-8 at a dilution of 1/200). This antibody was detected by DISCOVERY (Roche Diagnostics) anti-Rabbit HQ followed by DISCOVERY anti-HQ HRP. Visualization was done using the DISCOVERY Teal HRP Kit (SP142-DC-LAMP) or DISCOVERY Green HRP Kit (22C3-DC-LAMP) as the chromogens.

The slides were scanned using a cLIN profile on a P1000 PANNORAMIC whole-slide imager from 3DHISTECH Ltd. and analyzed by the HALO image analysis software (Indica Labs). The tumor area was analyzed for the co-localization of PD-L1 and DC-LAMP-3 by evaluating positive pixel areas with close proximity (Supplementary Fig. S2).

The specificity of SP142 or 22C3 for DC-LAMP⁺PD-L1⁺–co-expressing cells was calculated as follows: Specificity = True Negative/(False Positive + True Negative), in which true negatives were considered cells that were PD-L1⁻ DC-LAMP⁻ and false positive were considered cells that were PD-L1⁺DC-LAMP⁻.

Whole-transcriptome profiles were generated using TruSeq RNA Access technology (Illumina) as previously reported by Hamidi and colleagues (6). RNA sequencing (RNA-seq) reads were first aligned to rRNA sequences to remove ribosomal reads. The remaining reads were aligned to the human reference genome (NCBI Build 38) using GSNAP version 2013-10-10, allowing a maximum of two mismatches per 75-base sequence (parameters, “-M 2 -n 10 -B 2 -i 1 -N 1 -w 200,000 -E 1-pairmax-rna  =  200,000 –clip-overlap”). To quantify gene expression levels, the number of reads mapped to the exons of each RefSeq gene was calculated using the functionality provided by the R/Bioconductor package GenomicAlignments.

Reactome pathway enrichment analysis was completed based on the PD-L1 IHC staining pattern of the samples. Enrichment analysis was conducted using the enrichPathway function from the ReactomePA (version 1.50.0) R package, with a P value cutoff of 0.5, q value cutoff of 0.2, and Benjamini–Hochberg multiple testing correction. The effector T-cell gene signature (TGE8) was assessed as previously described by Mariathasan and colleagues (7) and included the following genes: IFNG, CXCL9, CD8A, GZMA, GZMB, CXCL10, PRF1, and TBX21.

Publicly available single-cell RNA-seq data derived from eight bladder cancer specimens along with three adjacent normal tissues and representing 52,721 single cells were downloaded as per the authors’ instructions (8). Briefly, the log₂ of counts per million and cell annotation were downloaded and processed using the Seurat package (version 3.2.2) following the standard pipeline without further log normalization. To derive a signature for the mature DCs enriched in immunoregulatory molecules (mregDC), differential gene expression was performed comparing mregDC with each of the following cell types including B, T, natural killer, fibroblast, endothelial, epithelial, and mast cells; cDC1 and cDC2; tumor-associated macrophages; and monocytes. Genes showing upregulation (fold change >1.25, adjusted P < 0.05) in mregDC across all comparisons were kept in the signature. We next deconvoluted this signature in IMvigor130 bulk RNA-seq samples. To ensure signature gene co-expression within bulk RNA-seq, we analyzed gene-to-gene correlation patterns and narrowed down the mRegDC signature to 17 genes (BIRC3, CCL17, CCL19, CD274, CSF2RA, EBI3, HMSD, LAMP3, NUB1, RAB9A, RFTN1, SEC61B, STK4, TMSB10, TRAF1, TRAFD1, and UBD). The analysis of signature association with outcome was conducted by splitting the mean log₂-transformed signature expression into tertiles.

Patients enrolled in the phase II CheckMate 275 study (NCT02387996) received nivolumab for platinum-resistant mUC. All patients gave informed consent, which extended to the use of samples for this study, and the studies were approved by their respective ethical review committees. For specific details of ethical review and study designs, see original publication (9).

Multiplex immunohistochemical consecutive staining on a single slide was performed as previously described (10) on pretreatment tumor specimens from 40 patients enrolled in the CheckMate 275 study for which there were remaining FFPE slides; no blood samples or control/normal was utilized. Briefly, 4-μm FFPE slides were prebaked at 60°C overnight, dewaxed in xylene, and loaded onto an automated staining platform (BOND RX, Leica Biosystems) with overlying covertiles (Bond Universal Covertiles, Leica Biosystems) for automatic immunostaining. Slides were incubated with a peroxide block (Bond Polymer Refine Detection Kit, DS9800; Leica Biosystems) for 15 minutes, followed by a serum-free protein block (Agilent X090930-2) for 30 minutes. The panel of primary antibodies (and dilutions) in sequential order of immunostaining were PD-L1 (1:100, E1L3N, Cell Signaling Technology), pan-cytokeratin (1:50, AE1/AE3, Dako, M3515), CD68 (1:100, KP1, Dako, M0814), DC-LAMP (1:80, 1010E1.01, Novus Biologicals, DDX0919P), and CD11b (1:250, EPR1344, Abcam, ab133357). The first primary antibody was applied (incubation time was individually optimized per antibody), followed by the biotin-linked secondary antibody [EnVision+System-HRP Labelled Polymer Anti-mouse (Agilent K400111-2) and EnVision+System-HRP Labelled Polymer Anti-Rabbit (Agilent, K400211-2)] and HRP-conjugated streptavidin binding using a polymer detection system [BOND Polymer Refine Detection Kit (DS9800), Leica Biosystems] and 3-amino-9-ethylcarbazole (AEC) substrate (ImmPACT AEC Peroxide Substrate Kit, Vector Laboratories). Slides were then counterstained using Modified Harris Hematoxylin (Sigma-Aldrich, HHS16) and incubated with an ammonia–water solution. Slides were mounted with Glycergel Mounting Medium (Agilent, C056330-2), air-dried overnight, and subsequently scanned by a high-resolution slide scanner (NanoZoomer S60, Hamamatsu) to generate whole-slide images. Coverslips were removed by incubating slides in hot water (56°C) to dissolve the mounting medium, followed by destaining in 1% hydrochloric acid and diluted ethanol solutions. Fab fragments [AffiniPure Fab Fragment Donkey Anti-Mouse (715-007-003) or Anti-Rabbit IgG (711-007-003)] were applied to block carryover staining from prior cycles. This cycle was repeated for each primary antibody.

Multiplex immunohistochemical consecutive staining on a single slide scanned whole-slide images were analyzed using QuPath version 0.3.2. Co-registration was performed with the ImageCombiner extension via affine transformation to create a concatenated overlay. Single-cell alignment was confirmed in separate regions of interest at each quadrant of the specimen, repeated individually for each immunostain image. Color deconvolution was performed to separate red, green, and blue (RGB), hematoxylin, and chromogen channels using stain and background vectors. Annotations were created on whole specimens extending to the tissue perimeter, and areas of tissue folding, degradation, or artifact were excluded from the final annotation, confirmed individually for each immunostain image. Cell segmentation was completed on whole-tissue annotations using uniform parameters. For each immunostain, thresholding was performed using multiple regions of interest located at different regions of the annotation to determine the optimal intensity for accurate stain positivity, and a positive or negative designation for each immunostain was applied to each cell detection.

TC PD-L1 status was assessed on cells positive for pan-cytokeratin, whereas IC PD-L1 status was assessed on cells positive for DC-LAMP⁺, CD11b⁺, and/or CD68⁺ (i.e., expression of PD-L1 by any of these ICs was scored for IC PD-L1 expression). The thresholds for high or low PD-L1 stratification were determined based on the median value among the cell type of interest (IC, TC, or DC-LAMP cells).

Human bladder carcinoma cell lines (5,637, HT-1376, KU-19-19, RT-112, SW 780, and RT4) were obtained in 2023 from Genentech’s common cell bank. No authentication of these cell lines was performed after they were initially purchased. Cell morphology and growth characteristics were monitored during the study to ensure their authenticity. All cell lines were validated as mycoplasma-free by PCR tests. All cell lines used in experiments were cultured for less than 10 passages. All cell lines were cultured in RPMI medium 1640 containing 10% heat-inactivated FBS (Hyclone, SH30071.03), 100 U/mL penicillin (Gibco), 100 mg/mL streptomycin sulfate (Gibco), and 1% L glutamine (Gibco, catalog number 11875093). PD-L1 expression was assessed by flow cytometry using antihuman CD274 (B7-H1 and PD-L1) antibody conjugated to fluorophore PE (BioLegend, catalog number 393608). Cells were treated with human TruStain FcX, GMP (BioLegend, catalog number 422304), which specifically blocks the FcR involved in unwanted staining without interfering with antibody-mediated specific staining of human cells, for 30 minutes at room temperature, prior to staining with anti–PD-L1. A mouse IgG1, κ isotype control (BioLegend, catalog number 981804) was used. The flow cytometry data were collected with a BD LSRFortessa (BD Biosciences) and analyzed using FlowJo software. Six bladder cell lines with PD-L1 high and low expression based on flow expression were chosen, and 1 × 10⁵ cells were plated in six-well plates and allowed to adhere overnight. The following day, the cells were dissociated using a cell detachment solution, ACCUTASE (STEMCELL Technologies, catalog number 07920), according to the manufacturer’s protocol. Following dissociation, cells were transferred to a 1.5-mL Eppendorf tube and centrifuged at 500× g for 5 minutes at 4°C to pellet the cells, media was aspirated, and the cell pellet was resuspended in 500 μL of ice-cold cryopreservation solution—50% FBS, 40% growth media, and 10% DMSO. The resuspended cells were immediately transferred to dry ice for 10 minutes and stored at −80°C until ready for sequencing.

Assay for Transposase-Accessible Chromatin using sequencing (ATAC-seq) libraries were generated using the ATAC-Seq Kit (Active Motif, catalog number 53150). Libraries were quantified using the Qubit dsDNA HS Assay Kit (Thermo Fisher Scientific) and profiled using the Bioanalyzer High Sensitivity DNA Kit (Agilent Technologies). Libraries were pooled and sequenced on HiSeq 2500 or NovaSeq 6000 (Illumina) to generate 30 million paired-end 50–base pair reads per library.

After sequencing, adapters were trimmed using Cutadapt (11). Reads were aligned to the reference genome GRCh38 with Bowtie2 (12). PCR duplicates were flagged by MarkDuplicates from Picard, and SAMTools was used to filter out duplicated, unpaired, and mitochondrial reads. BAM files were converted to paired bed files with bedtools bamtobed-bedpe, after which reads were shifted by 4 or 5 bp to account for the cutting bias introduced by Tn5 (13). MACS2 callpeak was used for peak calling with a P value threshold of 0.01 (14). Peaks were filtered for GRCh38 blacklist regions. DiffBind was used to summarize open chromatin regions across all samples and perform differential analysis. Motif enrichment within differentially accessible regions was done with Analysis of Motif Enrichment (15).

OS outcomes were analyzed by the Kaplan–Meier method with a log-rank test. Univariate Cox regressions were implemented to estimate HRs and 95% CIs.

Statistical details of experiments, number of repeats performed, and statistical tests used are stated in the figure legends or detailed in “Materials and Methods.” Unless otherwise specified, all data are presented as mean ± SE of the mean.

Qualified researchers can request access to individual patient-level clinical data through a data request platform. At the time of writing, this request platform is Vivli (https://vivli.org/ourmember/roche/). For up-to-date details on Roche’s Global Policy on the Sharing of Clinical Information and how to request access to related clinical study documents, visit https://go.roche.com/data_sharing. Anonymized records for individual patients across more than one data source external to Roche cannot, and should not, be linked due to a potential increase in risk of patient reidentification.

Qualified researchers can request access to clinical outcomes, biomarkers, and raw and processed sequencing data, deposited at the European Genome-phenome Archive under accession number EGAS50000000497 (https://ega-archive.org/studies/EGAS50000000497). Qualified researchers are those who can meet the requirements of the Data Access Agreement outlined at https://ega-archive.org/datasets/EGAD50000001100. Data related to the ATAC-seq experiments are available through the Gene Expression Omnibus (GSE286455).

Prior analyses have demonstrated that among commonly used PD-L1 IHC assays, the antibody used in the SP142 assay preferentially stains ICs versus TCs (16). We used archival tumor specimens from patients treated with atezolizumab (Arm B) or placebo plus platinum-based chemotherapy (Arm C) from the phase III IMvigor130 trial (Supplementary Fig. S1) to compare the effects of different approaches to testing the expression of PD-L1 (2, 17). PD-L1 expression using the Ventana SP142 assay was tested prospectively, as per the trial design (4). Using the available remaining specimens, we generated PD-L1 expression data using the Dako 22C3 PD-L1 IHC assay. The SP142 assay scores the percentage of tumor area occupied by PD-L1–expressing IC, and we employed the standard urothelial cancer cutoff of IC ≥5% (IC2/3) and <5% (IC0/1). The 22C3 assay scores the number of PD-L1–staining cells (TC + IC) divided by the total number of viable TCs, multiplied by 100, and we used the standard urothelial cancer cutoff of CPS ≥10 and <10 (18). The baseline characteristics of this biomarker-evaluable population (i.e., population with paired SP142 and 22C3 data; n = 655) were similar to those of the overall study population (Supplementary Table S1). Higher versus lower PD-L1 expression based on the SP142 assay (IC2/3 vs. IC0/1) was associated with improved OS among patients treated with atezolizumab (median survival, 27.5 vs. 13.08 months; HR, 0.56; 95% CI, 0.41–0.76; P = 0.0003; Fig. 1A), whereas a smaller magnitude of the effect was observed based on the 22C3 assay (median survival 15.74 vs. 14.59 months; HR, 0.76; 95% CI, 0.58–0.98; P = 0.0379; Fig. 1B).

Discordance was observed between urothelial cancer specimens identified as “PD-L1 high” using the two assays (Fig. 1C). The most prevalent phenotype was low PD-L1 expression using both assays (SP142^0/122C3^<10), whereas tumors positive for PD-L1 using only SP142 (SP142^2/322C3^<10, hereafter referred to as SP142^SP) were the least prevalent phenotype (˜4%). The longest OS among the biomarker-evaluable population treated with atezolizumab was observed in patients with “PD-L1 high” urothelial cancers according to both assays (SP142^2/322C3^≥10, hereafter referred to as PD-L1^DP; Fig. 1D), which encompassed the majority of SP142^2/3 urothelial cancers. The shortest OS was observed among patients with urothelial cancers positive for PD-L1 expression only by 22C3 (SP142^0/122C3^≥10, hereafter referred to as 22C3^SP; Fig. 1D). Similarly, poor outcomes were observed with patients harboring 22C3^SP urothelial cancers treated with platinum-based chemotherapy (Fig. 1E). Together, these findings suggest that (i) the favorable outcomes in PD-L1^DP urothelial cancers may be contributed to by the SP142 component (i.e., the vast majority SP142^2/3 urothelial cancers are captured in the PD-L1^DP group) and (ii) using the 22C3 assay alone encompasses urothelial cancers with both favorable outcomes (i.e., PD-L1^DP) as well as poor outcomes (i.e., 22C3^SP).

Bulk RNA-seq was performed using the same samples from which the PD-L1 IHC data were generated. PD-L1^DP urothelial cancers demonstrated increased expression of an effector T-cell gene signature (Fig. 2A), which has previously been correlated with response to atezolizumab in urothelial cancer (7). We next sought to understand the cellular elements, or cellular states, that might be enriched by SP142 staining (i.e., enriched in PD-L1^DP and SP142^SP urothelial cancers vs. 22C3^SP urothelial cancers). Prior studies using model systems have shown that despite being vastly outnumbered by other PD-L1–expressing cells in the tumor microenvironment (TME), PD-L1 expression on DCs is critical to regulating antitumor CD8⁺ T-cell responses (19). We performed xCell deconvolution (20) of the RNA-seq data, revealing increased inferred DC infiltration in PD-L1^DP and SP142^SP urothelial cancers versus 22C3^SP urothelial cancers (Fig. 2B). We, therefore, reasoned that the SP142 assay may be preferentially detecting PD-L1–expressing DCs and assessed co-localization of PD-L1 by SP142 or 22C3 along with the DC marker DC-LAMP. SP142 demonstrated a speckled staining pattern and was generally co-localized with DC-LAMP (Fig. 2C and D). In contrast, 22C3 displayed a more membranous staining pattern and detected PD-L1 more diffusely among cell types (Fig. 2D). We developed a protocol for completing the full SP142 or 22C3 assays along with DC-LAMP IHC on the same slide and tested the specificity of SP142 versus 22C3 staining for DC-LAMP–expressing cells using semiautomated whole-slide scoring of 5,072,809 and 6,292,450 total cells (n = 5 patient samples), respectively (Supplementary Fig. S2A–S2D). The specificity for DC-LAMP–expressing cells was 0.99 for SP142 and 0.76 for 22C3. These data revealed that PD-L1^DP urothelial cancers (representing the majority of SP142^2/3 urothelial cancers) were associated with better outcomes with atezolizumab and were characterized by increased expression of a T-cell effector gene signature and inferred DC infiltration. Although both 22C3 and SP142 stained PD-L1⁺ DCs, the SP142 assay may detect PD-L1⁺ DCs with greater specificity, potentially reconciling the better outcomes with PD-L1^DP versus 22C3^SP urothelial cancers with atezolizumab treatment.

Data from single-cell RNA-seq studies reveal a conserved transcriptional program expressed by a subset of DCs referred to as mregDCs (21). mregDCs are characterized by the high expression of genes such as CD274 (encoding PD-L1) and LAMP3 (encoding DC-LAMP), and these DCs are involved in cellular niches controlling the differentiation of progenitor-exhausted CD8⁺ T cells following ICB (21, 22). Given our finding that the SP142 assay preferentially identified cells co-expressing PD-L1 and DC-LAMP, we reasoned that mregDCs might be present in the urothelial cancer TME and confirmed this leveraging a publicly available single-cell RNA-seq urothelial cancer cohort (Fig. 2E; ref. 8). To assess the clinical relevance of the mregDC transcriptional program, we generated an mregDC gene signature (Supplementary Fig. S3A and S3B) and applied this signature to our IMvigor130 bulk RNA-seq data. Patients harboring urothelial cancers with an increased mregDC gene signature demonstrated improved OS with atezolizumab but not with chemotherapy (Fig. 2F and G).

We next turned our attention to 22C3^SP specimens given their association with poorer outcomes. The 22C3 CPS is based both on TC and IC staining. Compared with PD-L1^DP urothelial cancers, 22C3^SP urothelial cancers were enriched in TC-dominant (i.e., TC 22C3 PD-L1 scoring ≥10 and IC 22C3 PD-L1 scoring ≤10) versus IC PD-L1 expression (Fig. 3A and B). Both PD-L1^DP and 22C3^SP urothelial cancers were enriched in basal bladder cancer gene expression, whereas PD-L1 negative by both assays double-negative urothelial cancers were enriched in luminal bladder cancer gene expression (Fig. 3C). Compared with PD-L1^DP urothelial cancers, 22C3^SP urothelial cancers demonstrated decreased expression of genes (e.g., IFN, CXCL9, CXCL10, and CXCL13) and gene sets related to adaptive immunity (Fig. 3D and E).

Upregulation of PD-L1 may occur via multiple mechanisms including as a result of IFNγ in the TME (i.e., compensatory upregulation of PD-L1 on IC and TC often referred to as adaptive immune resistance) or via TC-autonomous expression downstream of oncogenic signaling programs (22). PD-L1 expression via these distinct mechanisms may likewise have different immunobiological implications and associate differently with ICB response. We surveyed a panel of six urothelial cancer cell lines to probe the TC-autonomous PD-L1 expression in urothelial cancer that might underlie our observations related to TC-dominant PD-L1 expression in 22C3^SP tumors. The heterogeneous expression of CD274 (and PD-L1 protein) was observed among urothelial cancer cell lines, and increased expression was generally observed in cell lines demonstrating basal-like urothelial cancer gene expression (Fig. 3F and G). ATAC-seq analysis revealed a significantly enriched accessible chromatin peak in the promoter region of CD274 within the basal versus luminal urothelial cancer cell lines (Fig. 3H). To identify transcription factors governing the differential chromatin accessibility, we performed motif enrichment analysis identifying significant enrichment of binding motifs for several transcription factors within accessible chromatin regions of the basal urothelial cancer cell lines (Fig. 3I) including those previously associated with PD-L1 expression in urothelial cancer and other tumor types (23, 24). Therefore, TC-autonomous PD-L1 expression may occur in a subset of urothelial cancers, particularly basal-like urothelial cancers.

To orthogonally and externally validate the association between different PD-L1 staining phenotypes and outcomes with ICB in urothelial cancer, we performed multiplex IHC (10) for a panel of markers (PD-L1 using the E1L3N antibody clone, pan-cytokeratin, DC-LAMP, CD68, and CD11b) on pretreatment tumors from a subset of patients with available tissue (n = 40) from the phase II CheckMate 275 study (Fig. 4A and B). The E1L3N antibody has previously demonstrated a high level of consistency for PD-L1 staining compared with the 22C3 antibody (25). Patients with urothelial cancers demonstrating increased PD-L1 expression on IC (with or without TC PD-L1 expression) experienced significantly longer OS (Fig. 4C). A similar trend was observed based on urothelial cancers with higher versus lower DC-LAMP⁺PD-L1⁺ cells (Fig. 4D). Alternatively, patients with urothelial cancers exhibiting PD-L1 expression predominantly on TC, but not IC, experienced significantly shorter OS (Fig. 4E). Therefore, in this external cohort of patients with mUC treated with ICB, using a complementary staining approach and whole-slide imaging, IC PD-L1 expression, regardless of TC PD-L1 expression, was associated with more favorable outcomes, whereas TC-dominant PD-L1 expression was associated with poor outcomes.

PD-L1 expression is among the most extensively studied predictive biomarkers for ICB benefit, but establishing the clinical utility of PD-L1 testing in several malignancies, including urothelial cancer, has been challenging. Leveraging specimens from the IMvigor130 trial, we generated insights derived from staining with two PD-L1 assays, which are summarized in the model in Supplementary Fig. S4: (i) PD-L1^DP urothelial cancers (representing the majority of SP142^2/3 urothelial cancer) were associated with better outcomes with atezolizumab; (ii) the SP142 assay detected DC-LAMP–expressing cells with greater specificity, potentially reflecting mregDCs; and (iii) 22C3^SP urothelial cancers were associated with poor outcomes, enriched in TC-dominant PD-L1 expression and basal-like urothelial cancer genes, and lacked transcriptional programs associated with restrained adaptive immunity. Together, our findings may not only advance knowledge about the immunobiology of urothelial cancer but also may, at least in part, explain the inconsistent performance of PD-L1 assays in urothelial cancer clinical trials to date.

Although PD-1/PD-L1 ICB has transformed the management of several malignancies, the mechanism underlying the activity of these therapies remains incompletely elucidated. A key conceptual advance related to the mechanism of ICB has been related to the potential role of ICB in preventing, rather than reversing, T-cell exhaustion within the TME and the associated role of multicellular hubs orchestrated by mregDCs in this activity (26). Using model systems, Oh and colleagues (19) demonstrated that deletion of PD-L1⁺ DCs, but not PD-L1⁺ macrophages, restricted tumor growth and enhanced CD8⁺ T-cell responses. Using single-cell RNA-seq of mouse and human tumors, Maier and colleagues (21) identified PD-L1–expressing DCs characterized by a specific transcriptional program expressed upon uptake of tumor antigens and coined mregDCs. Subsequent work by the same group revealed that progenitor CD8⁺ T cells and CXCL13⁺ T_H, which expanded after treatment in hepatocellular cancers in responders versus nonresponders to ICB, were present in pretreatment specimens in cellular hubs around mregDCs (27). Our data suggest that the SP142 assay may preferentially detect PD-L1–expressing DCs and an mregDC gene signature was associated with improved outcomes with atezolizumab in the IMvigor130 cohort.

The mechanisms underlying the preferential detection of DC-LAMP–expressing cells by SP142 are beyond the scope of this research although the speckled intracellular staining pattern is reminiscent of the endosomal localization of DC-LAMP (16). The unique staining pattern is likely related to the Ventana SP142 assay, rather than the antibody itself, as antibodies that bind a similar epitope (e.g., SP263) do not demonstrate a similar staining pattern (16). Intriguingly, CMTM6, which is highly expressed by mregDCs, has been shown to maintain PD-L1 expression and co-localizes with PD-L1 in recycling endosomes to prevent PD-L1 from lysosomal degradation (28).

Urothelial cancers that were 22C3^SP were associated with particularly poor outcomes and were enriched in TC-dominant PD-L1 expression. Using an orthogonal approach employing multiplex IHC in an independent cohort, we further demonstrated that TC-dominant PD-L1 expression was associated with poor outcomes. Both PD-L1^DP and 22C3^SP urothelial cancers were enriched in basal-like urothelial cancer genes, suggesting that these PD-L1 staining patterns may identify subsets of basal-like urothelial cancer associated with distinct immunobiology and prognosis. Indeed, using urothelial cancer cell lines, we demonstrated that PD-L1 was highly expressed in a subset of basal-like urothelial cancers and such cell lines demonstrated an open chromatin region of the PD-L1 promoter. We identified the enrichment of binding motifs for several transcription factors in these basal-like urothelial cancer cell lines, including TEADS, that have previously been linked to the cancer cell–autonomous expression of PD-L1 in other tumor types. Consistent with prior observations, we observed the enrichment of binding motifs for FOXA1 in the luminal-like, but not basal-like, urothelial cancer cell lines. Warrick and colleagues previously demonstrated that in tumors with mixed regions of pure urothelial cancer and squamous differentiation, the squamous regions were associated with increased basal-like genes and TC-specific PD-L1 expression, both of which were attributed to loss of FOXA1 expression (24). The group showed that FOXA1 knockout increased the expression of PD-L1 in urothelial cancer cell lines and that mixed histology urothelial cancer with regions of urothelial and squamous differentiation was associated with worse ICB outcomes compared with pure urothelial cancer (24). Further work is needed to define the oncogenic programs associated with TC-dominant PD-L1 expression and poor outcomes in a subset of basal-like urothelial cancers.

Our study is associated with potential strengths and limitations. Although other studies have assessed clinical outcomes with ICB related to different scoring algorithms for PD-L1 expression, many such studies have applied different scoring algorithms to staining generated using a single PD-L1 antibody clone rather than applying distinct PD-L1 assays to the same specimens. PD-L1 testing using the SP142 assay in our analysis was performed prospectively for patient stratification in the IMvigor130 study, whereas PD-L1 testing using the 22C3 assay was performed retrospectively. Our analysis comprised two clinical trial datasets and orthogonal approaches. Our study may help reconcile the inconsistent performance of PD-L1 assays in urothelial cancer trials to date. For example, different trials using the 22C3 assay may comprise different proportions of patients harboring PD-L1^DP and 22C3^SP tumors. Other observations related to PD-L1 testing remain incompletely resolved. Assays scoring only TC for PD-L1 expression have been associated with ICB outcomes in urothelial cancer. Although such assays only score TC, IC is also stained, and “PD-L1 high” urothelial cancers using such assays often comprise specimens with both TC and IC staining (rather than TC-dominant staining), particularly when using low cut points to define “PD-L1 high” specimens (29). Previous studies exploring atezolizumab in various clinical disease states of urothelial cancer have failed to establish the clinical utility of SP142 PD-L1 testing. These findings may be related to the heterogeneous and often poorly understood impact of control arm treatments on tumors demonstrating increased PD-L1 expression as well as the impact of intervening treatments on the TME (30). For example, we recently demonstrated that increased SP142 PD-L1 expression was associated with more favorable outcomes with cisplatin-based chemotherapy, potentially related to the immunomodulatory effects of such treatment (30). Finally, the extent to which various PD-L1 assays, identifying different PD-L1–staining cellular populations, associate with response to contemporary and future ICB-based combinations regimens in urothelial cancer is unknown as some combinations may theoretically overcome the putative resistance identified with TC-dominant PD-L1 expression.

Our study may foster a more nuanced understanding of PD-L1 expression in urothelial cancer and underscores the importance of dissecting and understanding specific cellular patterns of PD-L1 expression that might arise related to different underlying tumor biology (Supplementary Fig. S4). Our findings may facilitate the development of novel biomarkers for ICB in urothelial cancer and help reinforce the role of DCs in ICB response.

M.D. Galsky reports grants and personal fees from Genentech during the conduct of the study and grants and personal fees from Bristol Myers Squibb, Merck, and AstraZeneca and personal fees from Astellas, Pfizer, EMD Serono, Seagen, Janssen, Gilead, and AbbVie outside the submitted work. M. Kockx reports other support from CellCarta during the conduct of the study, and he is a shareholder and member of the board of CellCarta. J. Roels reports that she is a shareholder at Roche and employed by Genentech. R. Van Elzen reports other support from CellCarta during the conduct of the study. X. Guan reports other support from Genentech during the conduct of the study and other support from Genentech outside the submitted work. K. Yuen reports other support from Genentech during the conduct of the study. S. Gnjatic reports grants from Boehringer Ingelheim, Bristol Myers Squibb, Celgene, Genentech, Regeneron, and Takeda and personal fees from Taiho Pharmaceuticals outside the submitted work. S. Izadmehr reports grants from the NCI NIH during the conduct of the study. S. Sanjabi reports nonfinancial support and other support from Genentech/Roche during the conduct of the study. R.J. Johnston reports that he is employed by and receives salary and other compensation from Genentech/Roche. H. Koeppen reports employment with Roche/Genentech and ownership of Roche shares. S. Gupta reports other support from Bristol Myers Squibb during the conduct of the study, employment with Bristol Myers Squibb, and ownership of Bristol Myers Squibb stocks. A. Bamias reports personal fees from Roche, MSD, and Merck and grants and personal fees from Bristol Myers Squibb and AstraZeneca outside the submitted work. J.A. Arranz reports grants from Hospital General Universitario Gregorio Marañon during the conduct of the study and personal fees from Bristol Myers Squibb, MSD, Merck, Astellas, Bayer, Recordati, AstraZeneca, and Ipsen outside the submitted work. E. Kikuchi reports grants from Janssen, Chugai, Astellas, AstraZeneca, MSD, Kyorin, Otsuka, Taiho, Kyowa Kirin, Nippon Kayaku, Daiichi Sankyo, and Bristol during the conduct of the study and other support from Astellas, Bristol, Janssen, Merck Biopharma, MSD, Kissei, Chugai, Nippon Kayaku, and Taiho outside the submitted work. M. De Santis reports personal fees from Roche, Merck/MSD, Serono/Merck, and Bristol Myers Squibb during the conduct of the study. I.D. Davis reports other support from AstraZeneca, Bayer, Roche, Xennials, Eisai, and ANZUP Cancer Trials Group during the conduct of the study and other support from other sources outside the submitted work (see below); and Eastern Health, Full or part-time Employment, Personal, Professor of Medicine; Head, Eastern Health Clinical School; medical oncologist Monash University, Full or part-time Employment, Personal, Professor of Medicine and Head of Eastern Health Clinical School Astellas; local principal investigator, institutional, financial interest, and institutional support for clinical research from AstraZeneca, Bayer, Bristol Myers Squibb, Eisai, ESSA Pharma, Ipsen, Janssen, Merck/MSD, Movember Foundation, Pfizer, Roche/Genentech, and Seagen. P. Williams reports personal fees from Genentech during the conduct of the study. S. Bernhard reports employment with Roche outside the submitted work. I. Mellman reports other support from Genentech during the conduct of the study. E. Grande reports grants and personal fees from Roche and Merck during the conduct of the study. S. Mariathasan reports employment with Roche/Genentech and ownership of Roche stocks. No disclosures were reported by the other authors.

M.D. Galsky: Conceptualization, investigation, writing–original draft, writing–review and editing. M. Kockx: Conceptualization, methodology, writing–original draft, writing–review and editing. J. Roels: Conceptualization, writing–review and editing. R. Van Elzen: Conceptualization, writing–review and editing. X. Guan: Formal analysis, writing–review and editing. K. Yuen: Writing–review and editing, data analysis. D. Rishipathak: Writing–review and editing, data analysis. J.F. Anker: Investigation, writing–review and editing. S. Gnjatic: Investigation, writing–review and editing. S. Izadmehr: Investigation, writing–review and editing. S. Sanjabi: Writing–review and editing, data analysis. R.J. Johnston: Writing–review and editing, data analysis. M. Peterson: Writing–review and editing, data analysis. H. Koeppen: Writing–review and editing, data analysis. J.M. David: Writing–review and editing, data analysis. S. Gupta: Writing–review and editing, data analysis. A. Bamias: Investigation, writing–review and editing. J.A. Arranz: Investigation, writing–review and editing. E. Kikuchi: Investigation, writing–review and editing. M. De Santis: Investigation, writing–review and editing. I.D. Davis: Investigation, writing–review and editing. P. Williams: Writing–review and editing, data analysis. S. Bernhard: Writing–review and editing, data analysis. I. Mellman: Conceptualization, writing–review and editing. E. Grande: Investigation, writing–review and editing. R. Banchereau: Conceptualization, writing–review and editing. S. Mariathasan: Conceptualization, writing–original draft, writing–review and editing.

We thank the patients who participated in the study and their families. The study was sponsored by F. Hoffmann-La Roche Ltd. Editorial assistance was provided by Ashley J. Pratt, PhD, CMPP, of Nucleus Global, an Inizio Company, and funded by F. Hoffmann-La Roche Ltd. I.D. Davis is supported in part by an Australian NHMRC Investigator Grant (2016274). S. Izadmehr is supported by the NIH NCI Training Program in Cancer Biology (T32-CA78207).