The anti-PD-L1 antibody (22C3) is used as a companion diagnostic for successful checkpoint inhibitor therapy in head and neck squamous cell carcinoma. The positive predictive value of this assay, however, is less than ideal as a high PD-L1 score does not necessarily associate with a good response, and a good response may be seen with a low or negative PD-L1 score. In principle, this could be due to lack of accuracy and/or precision of the assay, as well as biological factors, such as variable expression across tumor cells and interacting immune cells. In this study, we describe a digital image analysis algorithm to address assay-related problems and improve both accuracy and precision, as well as improve identification of PD-L1 expression in differing cell populations of the tumor microenvironment. In addition, we propose a panel of assays that in aggregate might reveal inherent heterogeneity among cases with similar PD-L1 score and thus provide vital context underlying the efficacy of checkpoint inhibitors in the clinic. Here we show that applying novel machine-learning based digital image analysis to multiplex assays together with digital scoring of PD-L1 provides an accurate and precise assessment of the real-world tumor milieu that we hypothesize will help establish the immunophenotypes that inform therapeutic efficacy of checkpoint inhibitors. We have additionally interrogated whether expression of PD-L1 in these groups are consistent with known molecular patterns assayed using novel methodologies, such as high-plex digital spatial transcriptomics through the nanoString GeoMx Digital Spatial Profiler (DSP) platform and assessed the ability of the different digital platforms to provide concordant data relating to real-world expression patterns of PD-L1 and associated biomarkers. Thus, our results point to the importance of robust methodologies, used in combination, to evaluate complex tumor immune landscapes and the advantages of digital analyses to provide accurate and precise clinical contexts for better patient outcomes. Citation Format: Clara Troccoli, Lauren Matelski, Adam Beharry, Vanessa Ly, Morgan Wambaugh, Melanie Amen, Will Paces, Geoffrey Metcalf, Roberto Gianani, Tom Turi. Digital and spacial characterization of PD-L1 expression and IVD assay performance in the immune landscape of head and neck squamous cell carcinoma: A multimodal approach [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 6634.

Background: Immune cells within the tumor microenvironment (TME) play a vital role in regulating tumor progression. Therefore, immunotherapies that elicit anti-tumor responses are of great interest for the treatment of various cancers. Macrophages are a current immune cell type of interest due to the ability to polarize into anti-tumor (M1) and pro-tumor (M2) phenotypes. The density and phenotype of macrophages within the tumor and TME have been linked to prognosis in multiple types of solid tumors. Our hypothesis is that colorectal cancer (CRC) patients with high immune infiltration and greater amounts of anti-tumor immune cells within the tumor compartment will have an increased time of survival compared to cancers with immune excluded or desert environments. Methods: One CRC tumor microarray (TMA) containing primary tumors, metastases, and normal tissue were stained via multiplex immunofluorescence (mIF) for 6 different immune markers: CD3, CD8, CD56, CD68, CD163, and PD-L1. The stained TMAs were analyzed utilizing Flagship Biosciences’ proprietary image analysis platform. Machine learning algorithms used cellular features to stratify cells as belonging to either the tumoral or stromal compartment. Core level expression data was pulled and represented on a whole-cohort basis. All staining and image analysis outputs were reviewed by a board-certified, MD pathologist. Kaplan-Meier curves were generated using survival data in relation to the low, medium, and high expression of CD68, CD163, and/or PDL1 in the whole tissue, as well as tumor and stromal compartments. Patients were also stratified into low (1 and 2) and high (3 and 4) cancer stages. Results: There is a correlation between patient survival and the presence or absence of macrophage markers CD68 and CD163 and lack of immune-suppressive marker PDL1. Specifically, lower levels of CD68+CD163- cells within the tumor compartment correlate with an increase in patient survival. Similarly, lower PDL1 expression in the stromal compartment positively correlate with prolonged patient survival in CRC patients. Conclusion: Data generated through Flagship Biosciences’ image analysis platform showed a strong relationship between CD68 and CD163 presence and localization with CRC patient survival. Altering the immune cells within the tumor to an anti-tumor immune environment could increase patient survival times in high stage CRC patient populations. Combining immune checkpoint inhibitors with current FDA approved therapies for CRC are of interest to further extend patient survival, as PDL1 expression negatively correlate with CRC patient survival. Further, utilizing Flagship Biosciences’ image analysis software to understand cancer immune microenvironments should be further utilized to aid in diagnosis and treatment decisions. Citation Format: Dannah Miller, Kate Hieber, Will Paces, Huong Nguyen, Adam Beharry, Roberto Gianani. Macrophage biomarkers CD68 and CD163 correlate with CRC patient survival [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 2530.

PD-L1 (22C3) checkpoint inhibitor therapy represents a mainstay of modern cancer immunotherapy for non-small cell lung cancer (NSCLC). In vitro diagnostic (IVD) PD-L1 antibody staining is widely used to predict clinical intervention efficacy. However, pathologist interpretation of this assay is cumbersome and variable, resulting in poor positive predictive value concerning patient therapy response. To address this, we developed a digital assay (DA) termed Tissue Insight (TI) 22C3 NSCLC, for the quantification of PD-L1 in NSCLC tissues, including digital recognition of macrophages and lymphocytes. We completed clinical validation of this digital image analysis solution in 66 NSCLC patient samples, followed by concordance studies (comparison of PD-L1 manual and digital scores) in an additional 99 patient samples. We then combined this DA with three distinct immune cell recognition algorithms for detecting tissue macrophages, alveolar macrophages, and lymphocytes to aid in sample interpretation. Our PD-L1 (22C3) DA was successfully validated and had a scoring agreement (digital to manual) higher than the inter-pathologist scoring. Furthermore, the number of algorithm-identified immune cells showed significant correlation when compared with those identified by immunohistochemistry in serial sections stained by double immunofluorescence. Here, we demonstrated that TI 22C3 NSCLC DA yields comparable results to pathologist interpretation while eliminating the intra- and inter-pathologist variability associated with manual scoring while providing characterization of the immune microenvironment, which can aid in clinical treatment decisions.

Abstract Analyzing images of tissue sections stained with biomarker-specific indicators can generate quantitative, cell-specific expression data rather than relying on pathologist's ability to count cells and recapitulate confounding scoring paradigms. Beyond biomarker expression, numerous cell features can be measured such as the cell morphology, biomarker staining patterns and intensity, and localization of the cells within defined spatial compartments or relations to one another. One of the most studied immune cells in immuno-oncology (IO) is the cytotoxic T cell, whose biological function is to identify and destroy infected or dysfunctional cells; this function is complicated by numerous factors within the tumor microenvironment (TME). Tumor cells can aberrantly express immune checkpoint molecules designed to stop CD8 T cells from performing their tumor killing function. Additionally, CD8 T cell function may be perturbed by other immune modulating factors within the TME. IO drugs modulating tumor and immune cell interactions such as PD-L1 checkpoint inhibitors have shown that an inflammatory TME, represented by high CD8 presence in the tumors (inflamed tumors), is indicative of a better therapeutic response. Investigation of CD8 T cell status in biopsied tissues typically describes each tissue as one of three main phenotypes: Immune Desert, Immune Excluded, or Inflamed. Immune Desert tissues do not express appreciable levels of CD8 throughout the tissue. Immune Excluded tissues contain CD8, but the expression is almost exclusively localized to the stroma surrounding tumor nests. Inflamed tissues show higher percentages of CD8 within the tumor nests. While this phenotypic categorization is informative, CD8 expression is often calculated as a mean of expression through the tissue and does not take in to account the heterogeneous nature of tumor biology. This may result in a tumor containing one highly inflamed tumor nest being averaged out with multiple deserted tumor nests and a tissue categorized as excluded or deserted even though inflammation is present. To better represent the heterogeneity of inflammation within tumor tissues, we present an image analysis-based algorithm which not only separates out the tumor, stroma, and tumor/stroma margin, but identifies each tumor nest within the tissue as its own discrete object. This allows for the enumeration of number and size of all tumor nests within the tissue. Each tumor nest is given its own phenotypic classification (inflamed, excluded, or deserted) and the percentage of tumor nests displaying each phenotype. We demonstrate heterogeneity of inflammation assessment alongside standard mean phenotypic evaluations of CD8 expression in non-small cell lung cancer, bladder, and melanoma tumor samples. Practical use in clinical studies can help uncover response or resistance associated phenotypes related to tumor heterogeneity. Citation Format: Charles William Caldwell, Will Paces, Sofia Reitsma, Christopher Brueckner, Si T. Lee-Hoeflich, Roberto Gianani. Characterizing heterogeneity of CD8 inflammation in biopsied tumor tissues using novel image analysis techniques [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 3137.

Dickkopf-1 (DKK1) is a secreted modulator of Wnt signaling that is frequently overexpressed in tumors and associated with poor clinical outcomes. DKN-01 is a humanized monoclonal therapeutic antibody that binds DKK1 with high affinity and has demonstrated clinical activity in gastric/gastroesophageal junction (G/GEJ) patients with elevated tumoral expression of DKK1. Here we report on the validation of a DKK1 RNAscope chromogenic in situ hybridization assay to assess DKK1 expression in G/GEJ tumor tissue. To reduce pathologist time, potential pathologist variability from manual scoring and support pathologist decision making, a digital image analysis algorithm that identifies tumor cells and quantifies the DKK1 signal was developed. Following CLIA guidelines the DKK1 RNAscope chromogenic in situ hybridization assay and digital image analysis algorithm were successfully validated for sensitivity, specificity, accuracy, and precision. The DKK1 RNAscope assay in conjunction with the digital image analysis solution is acceptable for prospective screening of G/GEJ adenocarcinoma patients. The work described here will further advance the companion diagnostic development of our DKK1 RNAscope assay and could generally be used as a guide for the validation of RNAscope assays with digital image quantification.

Abstract Dickkopf-1 (DKK1) is a secreted modulator of Wnt signaling that is frequently overexpressed in tumors and associated with a poor prognosis. In this study we demonstrate an approach for clinically validating a RNAscope chromogenic in situ hybridization (CISH) assay for determining the level of DKK1 RNA in Gastric (G) and Gastroesophageal (GEJ) tumor tissues according to CLIA guidelines. This two-step process validated first the performance of the wet assay along with the ability of a pathologist to manually score the CISH signal according to a dot-based H-score paradigm, and second the ability of image analysis (IA) software (Flagship Biosciences) to unbiasedly and reproducibly quantify DKK1 staining in the same set of samples. The DKK1 CISH assay for manual scoring passed all pre-determined criteria of sensitivity, specificity, accuracy, and precision. 100% of the 40 evaluated G/GEJ tissues demonstrated acceptable staining in target tumor cells and absence in non-target cells. 100% of the evaluated tissues passed sensitivity with a broad dynamic range of signal expression across target cells, and negligible background staining. Reproducibility was measured by blinded pathology scoring of a serial subset of 12 cases, resulting in 11/12 (92%) with concordant DKK1 H-scores. Accuracy was assessed with a DKK1 qPCR assay on a 20-sample subset and a significant correlation with the H-score data was observed. The IA algorithm also passed all pre-determined criteria of sensitivity, specificity, accuracy, and precision. 36/40 samples (90%) passed analytical specificity with the IA algorithm correctly classifying staining on true cells belonging to the tumor. Failed samples were attributed to non-specific alkaline phosphatase activity, which in practice would be disregarded by the reviewing pathologist. 100% of samples passed the sensitivity criteria of appropriate cell identification and classification. IA precision compared H-scores across 3 days of staining, with 11/12 (92%) cases having concordant DKK1 H-scores and an ICC of 0.9009 (95% CI: 0.7117-0.9601). Digital H-scores were highly correlated to the validated manual H-scores of the 40-sample set (r = 0.72, p <0.0001). Taken together, these data demonstrate a clinically validated DKK1 RNAscope CISH laboratory-derived test (LDT) for manual and IA-assisted pathologist interpretation. The DKK1 RNAscope LDT is currently being applied as part of a phase 2 clinical study of DKN-01 in combination with tislelizumab to prospectively identify previously treated G/GEJ adenocarcinoma patients with elevated DKK1 tumor expression (NCT04363801). Citation Format: Charles Caldwell, Mike Kagey, Sofia Reitsma, Will Paces, Elizabeth Bueche, Vitria Adisetiyo, Roberto Gianani. Clinical validation of a Dickkopf-1 (DKK1) chromogenic in-situ hybridization (CISH) assay for manual and image-analysis assisted pathologist interpretation [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 433.

e18031 Background: The programmed death-ligand 1 (PD-L1) (22C3) antibody is a companion diagnostic to the check-point inhibitor drug pembrolizumab for head and neck squamous cell carcinoma (HNSCC). Specifically, PD-L1 expression, as determined by a Combined Positive Score (CPS) ≥ 1, defines the clinical decision point for treatment. However, to calculate the CPS, pathologists must account for PD-L1+ tumor cells (TC) and immune cells (IC, i.e., lymphocytes, macrophages) – a cumbersome and time-consuming task that is prone to intra/inter-pathologist variability. We developed a digital image analysis (IA) based solution as a pathologist support tool that accurately and precisely scores images of PD-L1 stained HNSCC tissues. Here we compared the performance of the digital method with manual pathology. To address the literature on the differential regulation of PD-L1 by IC and TC in treatment response, we also examined how PD-L1+ IC contribute to the CPS as compared to PD-L1+ TC. Methods: 19 HNSCC samples were immunohistochemically stained for PD-L1 (22C3). Slides were assessed for CPS by manual pathology: CPS = (# PD-L1 staining cells [TC, lymphocytes, macrophages]/total # viable TC) x 100; wherein CPS < 1 (no PD-L1 expression; no treatment) and CPS ≥ 1 (PD-L1 expression; treatment). Slides were scanned at 20X and analyzed using proprietary IA solutions that identify and separate TC from IC and quantify their PD-L1 expression. Resulting IA markups of cell detection and PD-L1 expression were reviewed by an MD pathologist for accurate cell recognition and stain classification. Upon pathologist approval of the IA performance, PD-L1 staining was evaluated by digital IA for CPS, % PD-L1+ TC, and % PD-L1+ IC. Pearson’s correlation analyses were conducted to assess the concordance between manual and digital CPS along with the association between CPS and PD-L1+ IC versus PD-L1+ TC. Results: Digital and manual CPS were significantly correlated ( r = 0.76, p = 0.00016) and concordant in treatment binning for 14/19 samples; for the 5 discordant samples, manual CPS binned them for no treatment while digital CPS binned them for treatment. Digital CPS significantly correlated with the % PD-L1+ IC ( r = 0.90, p < 0.00001) along with the % PD-L1+ TC ( r = 0.98, p < 0.00001). All 6 samples binned for no treatment by digital CPS had < 1% PD-L1+ IC, whereas all but 2/13 samples binned for treatment by digital CPS had ≥ 1% PD-L1+ IC. Conclusions: The digital method for assessing PD-L1 expression in HNSCC performs similarly to manual pathology with improved accuracy for detecting and quantifying IC versus TC in the CPS calculation. Moreover, the % of PD-L1+ IC seems to contribute substantially to the CPS, to a similar degree as the % of PD-L1+ TC. Given the concordance with manual pathology, this digital method can support pathologists as a clinical diagnostic tool. As such, a clinical validation study of this digital method in a larger HNSCC cohort is underway.

Abstract The Dako PD-L1 (22C3) antibody is used in immunohistochemistry (IHC) to measure tumor programmed death-ligand 1 (PD-L1) levels, serving as a common companion diagnostic to the check point inhibitor drug pembrolizumab in non-small cell lung cancer (NSCLC). Yet for the practicing pathologist, interpretation of the PD-L1 (22C3) IHC assay is cumbersome and time-consuming resulting in poor intra- and inter-pathologist precision. We have developed a clinically validated PD-L1 (22C3) image analysis (IA) assay that was previously shown to perform equivalently to manual pathologist scoring, require less labor, and produce standardized and reproducible results with greater precision than manual scoring across repeated assessment. However, there is growing literature on the significance of macrophage PD-L1 positivity for predicting treatment response, highlighting the need to differentiate PD-L1 positivity in tumor cells versus macrophages. This need is further underscored by the tendency of manual and digital reads to confuse alveolar macrophages as tumor cells. In this study, we have applied a novel machine learning solution for tissue and alveolar macrophage detection to our PD-L1 (22C3) IA assay to improve its digital scoring accuracy in NSCLC as well as generate additional complementary data on the presence and PD-L1 positivity status of tissue and alveolar macrophages. 24 whole tissue NSCLC samples were purchased and IHC stained for PD-L1. Stained slides were scanned at 20X magnification and analyzed using Flagship's IA solutions that separate tumor and stromal compartments, identify tissue and alveolar macrophages, and quantify PD-L1 expression. Resulting image markups of cell detection and PD-L1 expression were reviewed by an MD pathologist for acceptance. PD-L1 staining was evaluated by digital IA and manual pathology in the sample's tumor compartment for Tumor Proportion Score (TPS). Applying the tissue and alveolar macrophage solution to the PD-L1 (22C3) assay successfully removed macrophages from the digital TPS assessment and generated additional data on tissue and alveolar macrophages (% present, % PD-L1 positive) without significantly impacting the integrity of the IA assay (i.e., the correlation between the digital and manual TPS remained high). This preliminary data demonstrates the robust performance of the clinically validated PD-L1 (22C3) IA assay, even when accounting for tissue and alveolar macrophages. Ongoing experiments include testing the macrophage algorithm on a larger sample of NSCLC tissues to determine the correlation and treatment cut-off concordance between digital and manual scoring obtained by three pathologists. A follow-up clinical validation study of the sensitivity, specificity, accuracy, and precision of this updated PD-L1 (22C3) IA assay is also pending. Citation Format: Roberto Gianani, Will Paces, Elliott Ergon, Vitria Adisetiyo, Charles Caldwell. A novel image analysis assay for determining programmed death-ligand 1 (22C3) in non-small cell lung cancer that accounts for tissue and alveolar macrophages [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 435.

Determination of programmed death-ligand 1 (PD-L1) level in tumor by immunohistochemistry (IHC) is widely used to predict response to check point inhibitor therapy. In particular, the Dako PD-L1 (22C3) antibody is a common companion diagnostic to the monoclonal antibody drug Keytruda® (pembrolizumab) in non-small cell lung cancer (NSCLC).1 However, for the practicing pathologist, interpretation of the PD-L1 (22C3) assay is cumbersome and time consuming. Manual pathologist scoring also suffers from poor intra- and inter-pathologist precision, particularly around the cut-off point.2 In this clinical validation study, we developed an image analysis (IA) based solution to accurately and precisely score digital images obtained from PD-L1 stained NSCLC tissues for making clinical enrollment decisions.10 NSCLC tissue samples were purchased from a qualified vendor and IHC stained for PD-L1; 4 of these samples had serial sections stained on two separate days. Stained slides were scanned at 20X magnification and analyzed using Flagship Biosciences’ IA solutions that quantify PD-L1 expression and separate tumor and stromal compartments. Resulting image markups of cell detection and PD-L1 expression were reviewed by an MD pathologist for acceptance. PD-L1 staining was evaluated by digital IA in the sample’s tumor compartment for Total Proportion Score (TPS,%). Assay specificity was defined by ≥ 90% of the tissue cohort exhibiting appropriate cell recognition (≥ 90% cells correctly recognized as determined by the pathologist), with ≤ 10% false positive rate for staining classification. Sensitivity was defined by ≥ 90% of the cohort exhibiting appropriate cell identification (≥ 90% cells correctly identified), with ≤ 10% false negative rate for staining classification. Accuracy was defined by the combination of sensitivity and specificity and precision was defined by concordance of the binned TPS (<1%, ≥ 1%, ≥ 50%) in ≥ 80% of the samples stained on multiple days.The preliminary results show that IA can yield high analytical sensitivity, specificity, accuracy, and precision in the determination of the PD-L1 score. 100% of the tissue cohort met criteria for analytical specificity, sensitivity, and accuracy and 100% of the samples stained on multiple days met the precision criteria.This data demonstrates the feasibility of an IA approach as applied to PD-L1 (22C3) scoring. Ongoing experiments include application of the developed 22C3 algorithm on a separate cohort of 20 NSCLC samples to determine the correlation of digital scoring and scoring obtained by three pathologists. Additionally, we will evaluate the precision obtained by digital scoring in relation to the intra- and inter-pathologist concordance.Incorvaia L, Fanale D, Badalamenti G, et al. Programmed death ligand 1 (PD-L1) as a predictive biomarker for pembrolizumab therapy in patients with advanced non-small-cell lung cancer (NSCLC). Adv Ther 2019;36:2600–2617.Rimm DL, Han G, Taube JM, et al. A prospective, multi-institutional, pathologist-based assessment of 4 immunohistochemistry assays for PD-L1 expression in non–small cell lung cancer. JAMA Oncol 2017;3:1051–1058.

Abstract PD-L1 positivity in tumor-associated macrophages has been related to a favorable response to anti-PD-1 and PD-L1 targeted therapies. This concept has driven interest in specifically identifying macrophages in PD-L1 stained tissues, but pathologists often have difficulty in performing this task reliably. Thus, there is a clear need for tools capable of detecting and classifying tumor-associated macrophages in the context of a standardized PDL1 assay. One such method relies on the utilization of image analysis (IA) which measures hundreds of defining cellular features. This is necessary because macrophages identification based on few features would be difficult due to their varying morphologies and similarity with tumor cells. In order to isolate macrophages using multiple aspects of their cellular characteristics, AI and machine learning algorithms were leveraged to provide robust models for their identification. In this study we compare two artificial intelligence (AI) assisted macrophage classification methods for their accuracy in specifically identifying CD68 positive macrophages. Both approaches are based on interrogating a cohort of 20 NSCLC sections with two tumor-associated macrophage recognition algorithms, the first based on a neural network (NN), and the second on a decision tree (DT) model. These classifiers were developed by training a learning model on the morphometric and chromogenic PD-L1 staining features measured on detected cells, using their fluorescent CD68 expression profile as an indicator of their macrophage lineage. These classification algorithms were trained on images developed using procedures wherein tissue sections were stained by immunofluorescence with antibody to the macrophage associated CD68 marker and subsequently scanned to obtain a digital image. After removal of the coverslip and stripping of the CD68 antibodies, the sections were then stained with the anti-PDL1 22C3 antibody and rescanned to obtain a second digital image. Coregistration of the two digital images allowed for the identification of CD68 positive cells in the context of only the 22C3 PDL1 assay. The accuracy of each classification algorithm was determined by comparing the positivity for macrophage lineage among algorithm predicted macrophage to CD68 positivity. For both the NN and the DT classifier, we demonstrated a > 90% accuracy classification of macrophages identification. The decision tree method achieved similarly accurate predictions These data indicate that both NN and DT methods allow for independent identification of macrophages using only a 22C3 IHC PD-L1 assay, making it possible to isolate the contribution of PD-L1 macrophage positivity to successful checkpoint therapy response. Citation Format: Will Paces, Roberto Gianani, Huong Nguyen, Bradley Long, Michael Argyres, Cris Luengo, Rebecca Kim, Charles Caldwell. Image analysis-based identification and quantification of macrophages in PD-L1 (22C3) stained NSCLC tissue sections without additional macrophage-specific staining [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 3879.

Understanding response to immunotherapy requires accurate and complete characterization of tumor-associated immune cells in order to fully contextualize the immuno-oncology biomarker expression. Current standard practices surrounding enumeration of biomarker-positive immune cells using image analysis necessitate a dual-labeling approach combining the biomarker of interest and immune cell identification assays. Machine Learning (ML) may be used to distinguish different tissue types in a biopsy (e.g. tumor vs non-tumor), or to identify different cell types (e.g. macrophages vs other cells). A machine learning algorithm obtains statistics for a specific tissue class or cell type based on a training set, given by “ground truth” examples. The algorithm then generalizes from the given examples to “learn” the ability to find the tissue or cell type on the rest of the digital scan of the tissue slide, or other slide scans. Here we specifically describe ML methods for macrophage identification in digital scans of cancer tissue slides. The described methods can be used independently or in combination and are both based upon the use of artificial intelligence-based computational tissue analysis (cTA®) algorithms. These technologies can recognize macrophages independently of traditional IHC or IF dual-labeling identification methods. This same methodology may also be applicable for other specific subsets of morphologically distinct immune cells. The first strategy consists of training the cTA algorithm according to pathologist identification of macrophages, without a macrophage-specific staining (e.g., CD68). In this case, the slides are stained only with a PD-L1 assay, and pathologists establish the “ground truth” to teach the ML classifier. The second strategy consists of using CD68 labeling to identify macrophages to teach the ML classifier. Upon successful training and performance testing, the developed macrophage classifier can potentially be applied to new specimens without relying on pathologist annotation or IHC or IF dual-labeling for macrophage recognition. In summary, these two novel approaches demonstrate how ML could be used for characterization of critical immune cells such as macrophages in immuno-oncology tissue evaluations, without the reliance on additional biomarkers to specifically identify the cells. In this study, we evaluated the approach in the context of standard PD-L1 IHC as is used in the Companion Diagnostic (CDx) setting. Successful ML cTA strategies may be applied to other CDx assays to gain additional information about immune cell identification and quantification while only relying on simple monoplex assays. Importantly, these ML strategies would not require a change to the existing clinical practices where higher-level multiplexing would be needed to achieve similar outcomes. Citation Format: Kelsey Weigel, Will Paces, Elliott Ergon, Jeni Caldara, Kile McFadden, Cris Luengo, Roberto Gianani, Bharathi Vennapusa. Artificial intelligence-assisted macrophage identification in tumor biopsies [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 4918.

Several studies have shown that the location and expression of infiltrating immune cells in patient tumors can better identify which patients are more likely to respond to anti- PD-1/PD-L1 therapy. In particular, immunohistochemistry-based studies have shown that the spatial location of PD-L1 expression has particular biological relevance, as PD-L1 expression in the tumor cells or immune cells in the tumor syncytium, tumor microenvionment (TME), or tumor-stroma boundary all describe differential PD-L1 biology.Here, we use Flagship's digital pathology platform (cTA®) to investigate IHC based PD-L1 and CD8 staining patterns in Non-Small Cell Lung (NSCLC) and Urothelial Carcinoma (UC) tissue biopsies. The cTA platform creates thousands of per-cell Biofeatures™ derived from the digital pathology images of the IHC stained tissue, and applies Artificial Intelligence (AI) to the data to summary score endpoints for patient and cohort classification. In this approach, each tissue's IO landscape is represented using an "IO Scorecard", which summarizes the IHC biomarker data in a summary score which captures a comprehensive analysis of the tissue sample. The AI-determined scorecard models can be used to monitor changes before and after drug treatment and/or create predictive models for patient response outcomes.In this study, NSCLC and Urothelial Carcinoma samples were sectioned and stained using either the FDA-approved Dako 22C3 or SP263 PD-L1 IHC assays. Serial sections of each tissue specimen were also stained for CD8 expression. The cTA process detected all cells, assigned them to the tumor or TME compartments, and recorded the Biofeatures™ data which characterized PD-L1 or CD8 staining in the Tumor, Tumor/TME margin, or TME compartments. The method was validated by its ability to reproduce pathologist scoring for PD-L1 and CD8. Using the accepted data, we used the AI system to create novel Scorecard based patient classifications which describe differential PD-L1 status, CD8 status, and PD-L1/CD8 status combined. The AI Scorecard approach demonstrated that certain PD-L1 staining Biofeatures™ may also predict the CD8 status of a tumor, suggesting that additional CD8 staining may not be necessary to understand important expression patterns pertaining to cytotoxic T-cells.In summary, we demonstrated how the "IO Scorecards" are able to classify patients into differential immune status cohorts using a novel AI based scoring system, which relies only on PD-L1 IHC staining, by creating a comprehensive, contextual profile of PD-L1 staining that does not require additional CD8 IHC staining to characterize the impact of cytotoxic T-cells in a tissue sample.Citation Format: Charles Caldwell, Will Paces, Jeni Caldara, Bharathi Vennapusa, Joseph S. Krueger. Using digital pathology based "IO Scorecards" to describe relationships between PD-L1 expression and CD8 positive immune cell infiltration [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 3130.

Abstract Quantification of tumor-infiltrating lymphocytes (TILs) in non-small cell lung cancers (NSCLC) is valuable for understanding patient prognosis and survival. TILs comprise a subset of tumor-infiltrating leukocytes that modulate immune evasion and response to therapy. Understanding the composition of TIL subsets, especially relative to the total tumor leukocyte population, may provide additional context for understanding NSCLC pathogenesis and patient response to treatment. However, availability of tissues and use of chromogenic assays can limit the number of TIL and leukocyte subset markers assayed in a tissue section. Therefore, this study evaluated the identification of total leukocyte component in NSCLC using morphometric parameters and routine TIL marker monoplex immunohistochemistry (IHC) assays to further identify the composition of TIL subsets. Computational Tissue Analysis (cTA™) tools were used to determine the morphometric parameters which could identify immune cells in the absence of biomarker stain. The morphometric features which characterized immune infiltrates were used to quantify the total immune cell population frequency in the tumor nests and surrounding stroma in hematoxylin-stained tissues. The leukocyte population identified with morphometric parameters was correlated with CD45+ cell frequencies identified by cTA based on biomarker staining in CD45-stained serial sections. This morphometric ruleset was then applied to CD3- and CD8-stained tissues to evaluate the frequency of CD3+ and CD8+ TILs in the context of total infiltrating leukocytes. The relative populations of CD3+ and CD8+ TILs were consistent with available literature demonstrating that the morphometric ruleset could be utilized to enable evaluation of TIL sub-types relative to total leukocyte population without the need for additional IHC stains. The approach could, therefore, provide an added dimension of analysis for tissues stained by IHC for identifying the total immune cell infiltrating component without requiring additional biomarker staining. Citation Format: Elliott Ergon, Allison S. Harney, Nathan Martin, Will Paces, Famke Aeffner, Kristin Wilson, Janet Patterson-Kane, Karen Ryall, Daniel G. Rudmann, Brooke Hirsch, Joseph Krueger. Quantifying tumor-infiltrating leukocytes in hematoxylin stained NSCLC tissue samples using morphometric features [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2017; 2017 Apr 1-5; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2017;77(13 Suppl):Abstract nr 1674. doi:10.1158/1538-7445.AM2017-1674