Digital Quantification of Tumor PD-L1 Predicts Outcome of PD-1-Based Immune Checkpoint Therapy in Metastatic Melanoma

Background PD-1-based immune checkpoint blockade (ICB) is a highly effective therapy in metastatic melanoma. However, 40-60% of patients are primarily resistant, with valid predictive biomarkers currently missing. This study investigated the digitally quantified tumor PD-L1 expression for ICB therapy outcome prediction. Patients and Methods Tumor tissues taken prior to PD-1-based ICB for unresectable metastatic disease were collected within the prospective multicenter Tissue Registry in Melanoma (TRIM). PD-L1 expression (clone 28-8; cut-off=5%) was determined by digital and physician quantification, and correlated with therapy outcome (best overall response, BOR; progression-free survival, PFS; overall survival, OS). Results Tissue samples from 156 patients were analyzed (anti-PD-1, n=115; anti-CTLA-4+anti-PD-1, n=41). Patients with PD-L1-positive tumors showed an improved response compared to patients with PD-L1-negative tumors, by digital (BOR 50.5% versus 32.2%; p=0.026) and physician (BOR 54.2% versus 36.6%; p=0.032) quantification. Tumor PD-L1 positivity was associated with a prolonged PFS and OS by either digital (PFS, 9.9 versus 4.6 months, p=0.021; OS, not reached versus 13.0 months, p=0.001) or physician (PFS, 10.6 versus 5.6 months, p=0.051; OS, not reached versus 15.6 months, p=0.011) quantification. Multivariable Cox regression revealed digital (PFS, HR=0.57, p=0.007; OS, HR=0.44, p=0.001) and physician (OS, HR=0.54, p=0.016) PD-L1 quantification as independent predictors of survival upon PD-1-based ICB. The combination of both methods identified a patient subgroup with particularly favorable therapy outcome (PFS, HR=0.53, p=0.011; OS, HR=0.47, p=0.008). Conclusion Pre-treatment tumor PD-L1 positivity predicted a favorable outcome of PD-1-based ICB in melanoma. Herein, digital quantification was not inferior to physician quantification, and should be further validated for clinical use.


INTRODUCTION
The introduction of immune checkpoint blockade (ICB) therapy led to a tremendous survival improvement in patients with advanced metastatic melanoma (1)(2)(3). PD-1-based ICB therapies can be used alone or in combination with CTLA-4 inhibitors (4). Despite improved long-term survival in responders, up to 60% of melanoma patients are primary resistant to PD-1-based ICB and have significantly inferior survival as a consequence (5). Approximately 40% of melanomas have a targetable tumor BRAF-V600 mutation with inhibition of the mitogen-activated protein kinase (MAPK) pathway as a viable alternative treatment option to ICB (6). There is a high need for valid pre-treatment biomarkers that predict the response to ICB to enable an optimal treatment choice for advanced melanoma patients. Nonspecific bloodbased biomarkers have been reported to predict ICB treatment outcome such as serum lactate dehydrogenase (LDH) activity, as well as blood counts of lymphocytes and eosinophils (7)(8)(9)(10). Tumor tissue-based biomarkers described to be associated with PD-1-based ICB therapy outcome are the density of tumorinfiltrating lymphocytes and the expression of PD-L1 (11). In particular, the quantification of PD-L1 expression in tumor tissue is widely used in routine clinical diagnostics of various cancer types such as non-small cell lung carcinoma (NSCLC), renal cell carcinoma, Hodgkin's lymphoma, and colorectal carcinoma (12).
The role of tumor PD-L1 expression as a predictive biomarker in melanoma, however, is not clear, mainly because of the difficulties in evaluating melanomas with overall low PD-L1 expression and high melanin content (2,(13)(14)(15). Moreover, PD-L1 staining can be detected not only on the cell membrane but also intracellularly and shows high spatial heterogeneity, so its evaluation is associated with a high interobserver variability (16). In addition, it has already been shown that PD-L1 expression can differ greatly depending on the melanoma subtype and that melanoma subtypes respond differently to ICB (17). Whole-slide imaging and digital pathology have shown an improvement in the evaluation of the immunohistochemical tumor tissue stainings for HER2 and KI67 in breast cancer and for the Gleason classification in prostate cancer (18)(19)(20). In addition, a recent study demonstrated that a digital pathology algorithm can be helpful to the pathologist in the evaluation of tumor PD-L1 expression in melanin-bleached melanoma tissue samples. However, a correlation between digital PD-L1 quantification and therapy outcome has not been performed up to now (21).
The aim of the present study was to investigate digital PD-L1 quantification versus physician PD-L1 quantification in pretreatment tumor tissue of melanoma patients as potential predictors of therapy outcome of a PD-1-based ICB.

Patients and Tissues
Formalin-fixed, paraffin-embedded (FFPE) tumor tissue samples from patients diagnosed with melanoma were prospectively collected within the multicenter translational study Tissue Registry in Melanoma (TRIM; CA209-578) performed within the framework of the skin cancer registry ADOREG of the German Dermatologic Cooperative Oncology Group (DeCOG). Out of this cohort, patients were selected for the present analysis according to the following criteria: Histologically confirmed diagnosis of melanoma of the skin, mucosa, or unknown primary; tumor tissue specimen obtained for analysis prior to a PD-1-based ICB for unresectable stage III or IV (AJCCv8) (18) metastatic disease; complete documentation of therapy outcome and follow-up, and availability of consecutive tissue slides stained for PD-L1 and control IgG, comparable in size and quality. Best overall response (BOR) was determined according to RECIST version 1.1 (22). Progression-free (PFS) and overall (OS) survival were defined as time from therapy start until disease progression or death, respectively; if no such event occurred, the date of the last patient contact was used as endpoint of survival assessment (censored observation). The study was approved by the ethics committee of the University Duisburg-Essen (15-6566-BO).

PD-L1 Staining
The PD-L1 expression was assessed in FFPE tumor tissue specimens with the use of a rabbit monoclonal anti-human PD-L1 antibody (clone  and an analytically validated automated immunohistochemical assay (PD-L1 IHC 28-8 pharmDx for Autostainer Link 48; Dako, Agilent Technologies, Santa Clara, CA, USA), as described previously (23). A consecutive tissue slide of the same specimen was prepared accordingly for each sample, stained with non-specific IgG and used as negative control. Hematoxylin was used as nuclear staining. For detailed visualization of morphological structures, an additional control tissue slide was stained with hematoxylin and eosin (H&E).
Quantification of PD-L1 Expression by the Physician PD-L1 expression in tumor tissue was quantified as the percentage of live tumor cells that exhibited specific cell surface staining of any intensity in a section containing at least 100 evaluable tumor cells, with ≥5% defined as positive staining, as previously described (23). The cutoff >5% as the definition of PD-L1 positivity is recommended by the manufacturer of the assay and is the established standard in our department. This type of quantification of PD-L1 expression was performed by either pathologists or histopathologically experienced dermatologists or both using conventional bright field microscopy, and is referred to as "physician's quantification".

Quantification of PD-L1 Expression by a Digital Algorithm
The anti-PD-L1 stained slides and the negative control slides were digitized with the whole-slide scanner Aperio AT2 (Leica, Wetzlar, Germany) using a resolution of 20x. These digitized whole-slide images were used for the quantification of PD-L1 expression by a newly defined method based on a digital algorithm. This newly developed Java-based algorithm removes artifacts present on the tumor regions and quantifies the number of PD-L1 expressing cells. Corresponding tumor regions were manually selected as regions of interest on the anti-PD-L1-stained slide and the negative control slide; Figures 1A, B. Binary masks were generated by applying an intensity threshold for PD-L1 and melanin (brown) as well as hematoxylin (blue), Figures 1C, D. The binary masks were obtained by using various thresholding methods implemented by Fiji (https://imagej.net/Fiji), each best adapted to the type of signal. To generate the tissue masks (tumor and biopsy), the "Triangle" thresholding method was used. For the Melanin and PDL1 stainings,"MaxEntropy", and "Moments" for the cellular nucleus. These methods are part of the Auto-threshold algorithms of Fiji and were chosen based on their accuracy (24). The digitized image of the respective negative control staining was used to deduct the background signal (melanin). The binary information for cellular and nuclear signals was co-registered and overlapping mask regions were used to extract the number of cells stained positive for PD-L1 (hematoxylin+/PD-L1+) or melanin (hematoxylin+/melanin+). In addition, the total number of cells (hematoxylin+) was determined to calculate the percentage of PD-L1+ cells relative to the total number of cells in the tumor area, with a percentage ≥5% defined as positive, similar to the physician evaluation. This method of PD-L1 quantification is referred to as "digital quantification". It should be noted here that the digital algorithm quantifies not only PD-L1 expression of tumor cells, but of all cells in the ROI, such as macrophages or lymphocytes as well.

Statistical Analysis
The survival endpoints (PFS and OS) were calculated using the Kaplan-Meier method for censored failure time data. The twosided log-rank test was used to compare survival rates between groups. Multivariable analyses were performed using the Cox proportional hazards model. Known prognostic and predictive parameters of metastatic melanoma were included as covariates: age (≤65 versus >65 years), sex (male versus female), disease stage (III versus IV), location of primary (skin versus others), M category of metastasis (M1a/b versus M1c/d), LDH serum activity (elevated versus normal), therapy type (anti-PD-1 monotherapy versus anti-PD-1 plus anti-CTLA-4), BRAF mutation status (mutation versus wild type), and PD-L1 expression (positive versus negative). The correlation analysis was performed using the Pearson's correlation coefficient. Best overall response (BOR) was calculated by chi-square test. P<0.05 was considered statistically significant. Survival analysis was performed with SPSS (Version 25, IBM, Armonk, NY, USA) and Graphpad Prism (Version 9, GraphPad Software, CA, USA).

Patient Characteristics and Study Flow
Of the patients participated in/were registered in the TRIM project, 388 patients started an anti-PD-1-based ICB therapy between February 2014 and July 2019 and 156 patients met the above mentioned selection criteria for the present study; Figure 2. The tumor tissue specimens examined for PD-L1   Figure 3A. The PD-L1 quantification by the physician and the digital algorithm showed a significant correlation (Pearson's correlation; r = 0.39; p <0.001; Figure 3B). In summary, 60.3% of patients showed the same assessment regarding PD-L1 positivity by physician and digital algorithm. In the cases that were classified differently by both measurement methods, the digital algorithm showed PD-L1 positive findings more frequently.

Tumor PD-L1 Positivity by Physician's Quantification Predicts Favorable
Outcome of Anti-PD-1-Based ICB Therapy We next investigated the feasibility of physician PD-L1 analysis, traditionally used in the clinic, to predict patient response to therapy and survival. Melanoma patients with PD-L1 positive tumors by physician's quantification (n=60; 38.5%) showed an improved therapy response upon anti-PD-1-based ICB (BOR=54.2%) as compared to patients with PD-L1 negative tumors (n=96; BOR=36.6%; p=0.032). The median PFS after start of anti-PD-1-based ICB in patients with PD-L1 positive tumors by physician's quantification was 10

Tumor PD-L1 Positivity by Digital Quantification Predicts Favorable
Outcome of Anti-PD-1-Based ICB Therapy The next step was to investigate whether the newly programmed digital algorithm was also suitable for PD-L1 analysis. Melanoma patients with PD-L1 positive tumors by digital quantification (n=95; 60.9%) showed an improved therapy response upon anti-PD-1 based ICB (BOR=50.5%) as compared to patients with PD-L1 negative tumors (n=61; BOR=32.2%; p=0.026). In patients with PD-L1 positive tumors by digital quantification the median PFS was 9.9 months (95% CI=5.   Table 2. Patients whose tumors were tested positive only by the physician or by the digital algorithm showed no relevant differences with regard to PFS and OS as compared to each other. The addition of PD-L1 expression analysis by the digital algorithm to conventional physician analysis has greatly improved the predictive power of PD-L1 analysis in terms of response to therapy and patient survival.

DISCUSSION
Immunohistochemical PD-L1 expression analysis has been shown to be a predictive biomarker for ICB treatment outcomes in numerous tumor entities and, in this regard, is already routinely being considered for treatment decisions in entities such as NSCLC and urogenitary cancers (12). However, in melanoma the role of PD-L1 expression as a predictive biomarker for therapy outcome is currently under debate and has still not entered the clinical routine for treatment decision making. Herein, one major reason is the high inter-observer variability of PD-L1 quantification reported for melanoma, mainly due to melanin pigmentation hampering the evaluation process. To overcome these hurdles, we here investigated the association between pre-treatment tumor PD-L1 expression and ICB therapy outcome using two independent methods of PD-L1 expression quantification, one by trained physicians and the other by a newly proposed digital algorithm. The digital quantification method harbors the advantage to be applicable regardless of the presence of melanin pigmentation. Our results showed a prolonged PFS and OS in melanoma patients whose tumors were classified as PD-L1 positive by both methods of PD-L1 quantification, with the digital quantification not being inferior to the physician quantification. The positive correlation of pre-treatment tumor PD-L1 expression with PFS and OS observed in this study is consistent with previously published data from clinical trials in metastatic melanoma (2,25,26). Interestingly, the combination of PD-L1 quantification methods, the physician's and the digital algorithm method, showed the longest survival for patients with PD-L1 positive tumors with a median PFS of 11.4 months and a median OS not reached, and proved to be independently predictive by multivariable testing. In the existing literature, tumor PD-L1 expression in melanoma is considered to play the role of a prognostic marker but to have little pre-therapeutic predictive value (27). In contrast, our analysis of pre-treatment tumor tissue samples shows that melanoma patients whose tumors were evaluated as PD-L1 positive by both the physicians and the digital algorithm had a BOR on ICB of 60.4%, indicating that tissue PD-L1 expression has predictive value.
Currently, targeted therapy with BRAF/MEK inhibition is available for melanoma patients with BRAF mutation as an alternative or an addition to ICB (28)(29)(30). In these patients, whose tumors harbor a targetable BRAF mutation, there is a lack of predictive biomarkers that help to choose the optimal individualized therapy. Here, tumor PD-L1 expression quantification could be a helpful tool, assuming that patients showing PD-L1 positivity are more likely to benefit from ICB, and patients with PD-L1-negative tumors may be more likely to benefit from targeted therapy.
Obviously, the use of PD-L1 as a biomarker is difficult, as different cut-offs and a high intratumoral heterogeneity with dynamic changes exist as well as different scoring systems are available. A total of four different scoring systems are established and in clinical use for tumor PD-L1 quantification. For melanoma, the most commonly used scoring system is the Tumor Proportion Score (TPS), which restrictively quantifies only tumor cells that exhibit linear staining of the membrane.
Other cancer entities for which the TPS is used include NSCLC and carcinomas of the head and neck. The combined positive score (CPS), which quantifies tumor cells and immune mononuclear cells, is used for urothelial and gastric carcinomas. In urothelial carcinoma, the immune cell score (IC), which quantifies all immune cells stained for PD-L1, is in use. In the melanoma score (MEL score), PDL-1-positive mononuclear immune cells and tumor cells are quantified, similar to the CPS (16,31,32). In the present study, the physicians used the restrictive TPS, whereas the digital algorithm used the CPS as classification system that includes PD-L1 expression on associated immune cells. Indeed, these scoring differences explain why the digital quantification showed higher frequencies of PD-L1 positivity compared to the physician's quantification. Interestingly, 63.5% of cases still showed the same result in terms of positivity or negativity.
Since the present algorithm is a pixel-based image analysis algorithm, its application is not limited to PD-L1 analysis and can be readily used for other immunohistochemical staining for quantification in clinical and research settings. In addition, the digital algorithm can also be used for quantification for multiplex immunofluorescence imaging in translational and basic research. The present study and the digital PD-L1 quantification method also unraveled some limitations. The digital algorithm currently is only a semi-quantitative measuring tool, as the physician still has to select the target tumor areas to be analyzed. In contrast to physician quantification with the currently recommended method TPS, where only tumor cells with PD-L1 membrane staining are counted, the digital algorithm quantifies all cells of the tumor microenvironment and does not distinguish between cytosol staining, nuclear staining or membrane staining similar to CPS. Thus, in the present study, two different investigators (physician vs. digital algorithm) were compared which used different scoring systems for quantification (TPS vs. CPS). Further modifications, e.g. the addition of artificial intelligence technologies, are required to transform the actual digital algorithm into a measuring instrument that is completely independent from the physician. Another limitation of this study is that we did not analyze melanoma samples from patients who were treated with anti-CTLA-4 monotherapy and therefore cannot conclude to what extent PD-L1 expression plays a role in these patients. However, it must be noted that anti-CTLA-4 monotherapy alone plays almost no role in melanoma therapy any longer. Notably, a large proportion of patients were pre-treated, including BRAFmutated patients with BRAF inhibitors, which may have influenced the results on treatment efficacy and identification and validation of PD-L1 as a predictive biomarker. In conclusion, our results demonstrate that pre-treatment tumor PD-L1 quantification by a digital algorithm is not inferior to the quantification by physicians as predictors of ICB therapy outcome. Moreover, the combination of both quantification methods significantly improved the predictive value. Accordingly, a digital quantification of tumor PD-L1 expression could facilitate diagnostic procedures, and improve the prediction of treatment outcomes at treatment decision making in patients with metastatic melanoma. Further studies are planned to investigate tumor PD-L1 expression by different methods in melanoma patients treated in the adjuvant setting.

DATA AVAILABILITY STATEMENT
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by Ethics committee of the University Duisburg-Essen. The patients/participants provided their written informed consent to participate in this study.

FUNDING
This work was in part supported by Bristol Myers Squibb for the multicenter translational study "Tissue Registry in Melanoma" (TRIM) within the framework of the skin cancer registry ADOREG of the German Dermatologic Cooperative Oncology Group (DeCOG).

ACKNOWLEDGMENTS
We thank the patients who participated in this study, their families, and the staff members at the study sites who cared for them.