Comparative analysis of machine learning approaches to classify tumor mutation burden in lung adenocarcinoma using histopathology images.

Both histologic subtypes and tumor mutation burden (TMB) represent important biomarkers in lung cancer, with implications for patient prognosis and treatment decisions. Typically, TMB is evaluated by comprehensive genomic profiling but this requires use of finite tissue specimens and costly, time-consuming laboratory processes. Histologic subtype classification represents an established component of lung adenocarcinoma histopathology, but can be challenging and is associated with substantial inter-pathologist variability. Here we developed a deep learning system to both classify histologic patterns in lung adenocarcinoma and predict TMB status using de-identified Hematoxylin and Eosin (H&E) stained whole slide images. We first trained a convolutional neural network to map histologic features across whole slide images of lung cancer resection specimens. On evaluation using an external data source, this model achieved patch-level area under the receiver operating characteristic curve (AUC) of 0.78-0.98 across nine histologic features. We then integrated the output of this model with clinico-demographic data to develop an interpretable model for TMB classification. The resulting end-to-end system was evaluated on 172 held out cases from TCGA, achieving an AUC of 0.71 (95% CI 0.63-0.80). The benefit of using histologic features in predicting TMB is highlighted by the significant improvement this approach offers over using the clinical features alone (AUC of 0.63 [95% CI 0.53-0.72], p = 0.002). Furthermore, we found that our histologic subtype-based approach achieved performance similar to that of a weakly supervised approach (AUC of 0.72 [95% CI 0.64-0.80]). Together these results underscore that incorporating histologic patterns in biomarker prediction for lung cancer provides informative signals, and that interpretable approaches utilizing these patterns perform comparably with less interpretable, weakly supervised approaches.

Lipid accumulation impairs natural killer cell cytotoxicity and tumor control in the postoperative period.

BACKGROUND: Natural killer (NK) cell dysfunction following cancer surgery has been shown to promote metastases. Recent studies demonstrate an emerging role for lipids in the modulation of NK cell innate responses. However, the mechanisms involved in lipid modulation of NK cell postoperative anti-tumor function are unknown. This current study will determine whether the lipid accumulation via scavenger receptors on NK cells is responsible for the increase in postoperative metastasis. METHODS: Lipid content in mouse and human NK cells was evaluated by flow cytometry. NK cell scavenger receptor (SR) expression was measured by microarray analysis, validated by qRT-PCR and flow cytometry. NK cell ex vivo and in vivo tumor killing was measured by chromium-release and adoptive transfer assays, respectively. The mediating role of surgery-expanded granulocytic myeloid derived suppressor cells (gMDSC) in SR induction on NK cells was evaluated using co-culture assays. RESULTS: NK cells in surgery-treated mice demonstrated increased lipid accumulation, which occurred via up-regulation of MSR1, CD36 and CD68. NK cells with high lipid content had diminished ability to lyse tumor targets ex vivo. Adoptive transfer of lipid-laden NK cells into NK cell-deficient mice were unable to protect against a lung tumor challenge. Granulocytic MDSC from surgery-treated mice increased SR expression on NK cells. Colorectal cancer surgical patients showed increased NK cell lipid content, higher CD36 expression, decreased granzyme B and perforin production in addition to reduced cytotoxicity in the postoperative period. CONCLUSIONS: Postoperative lipid accumulation promotes the formation of metastases by impairing NK cell function in both preclinical surgical models and human surgical colorectal cancer patient samples. Understanding and targeting the mechanisms underlying lipid accumulation in innate immune NK cells can improve prognosis in cancer surgical patients.

Real-World Clinical Impact of Immune Checkpoint Inhibitors in Patients With Advanced/Metastatic Non-Small Cell Lung Cancer After Platinum Chemotherapy.

BACKGROUND: The real-world effect of anti-programmed death ligand 1 (PD-L1) therapies is unclear. We compared US patients who received second-line therapy for non-small-cell lung cancer (NSCLC) before and shortly after US Food and Drug Administration (FDA) approval of PD-L1 inhibitors. PATIENTS AND METHODS: Patients in the Flatiron Health database (≥18 years; received first-line platinum therapy for advanced/metastatic NSCLC; ≥6 months of follow-up) were assessed before ("historical": January 1, 2011 to December 31, 2013) and after ("current": January 1, 2015 to May 31, 2017) FDA approval of anti-PD-L1 therapies for NSCLC. Index was start of second-line therapy. Baseline variables, treatment patterns, and overall survival (OS) were reported. RESULTS: A greater proportion of patients in the current cohort received second-line treatment than in the historical cohort (n = 4240 [57.0%] vs. n = 2357 [37.4%]); 48.8% [n = 2071] of the current second-line patients received anti-PD-L1 therapy. Current patients were more likely to receive second-line anti-PD-L1 therapy if they had poorer Eastern Cooperative Oncology Group (ECOG) performance status (≥2), had squamous histology, or had no epidermal growth factor receptor (EGFR), anaplastic lymphoma kinase (ALK), or ROS proto-oncogene 1 mutations. Median OS from index was higher in the current cohort (9.4 [95% confidence interval (CI), 8.9-9.9] months) than the historical cohort (7.3 [95% CI, 6.9-7.8] months). Adjusted for sex, race, ECOG performance status, disease stage, and Kirsten rat sarcoma viral oncogene homolog, EGFR, and ALK status, OS was improved by 15% in the current cohort. CONCLUSION: Contemporary patients are more likely to receive second-line therapy and have longer OS than patients who received care before approval of anti-PD-L1 therapies.

Exceptional response and multisystem autoimmune-like toxicities associated with the same T cell clone in a patient with uveal melanoma treated with immune checkpoint inhibitors.

Balancing the potential for durable remissions with autoimmune-like toxicities is a key clinical challenge in the use of immune checkpoint inhibitors (ICI). Certain toxicities are associated with an increased response rate; however, the molecular underpinnings of this association are poorly understood. Here, we report a patient with wide spread uveal melanoma who had an exceptional response to treatment with ipilimumab and nivolumab, but suffered severe immune-related sequelae, including central serous retinopathy with retinal detachment, tinnitus, and vitiligo resembling Vogt-Koyanagi-Harada disease, and refractory enteritis. TCR-sequencing of the primary tumor, a hepatic metastasis, duodenal biopsy and peripheral blood mononuclear cells, identified the identical T cell clone in all four tissues. This case provides preliminary evidence for cross-reactivity as a mechanism for the association between effect and toxicity of ICIs.

A real-world, comparative study of FDA-approved diagnostic assays PD-L1 IHC 28-8 and 22C3 in lung cancer and other malignancies.

AIMS: At the time of analysis, two widely used, drug-specific, tumour-cell programmed death ligand 1 (PD-L1) assays were approved by the US Food and Drug Administration for anti-PD-1 therapies: the Dako PD-L1 immunohistochemistry (IHC) 28-8 pharmDx assay and the Dako PD-L1 IHC 22C3 pharmDx assay. Given that the majority of current PD-L1 testing in US clinical practice is performed at commercial reference laboratories, we aimed to evaluate the concordance of the 28-8 and 22C3 assays in a real-world setting. METHODS: Matched PD-L1 IHC 28-8 and 22C3 results from routine assessment were obtained from 1930 patients, including 412 confirmed to have lung cancer, submitted from hospitals in over 38 US states/territories. Biopsies were stained, reviewed and scored by trained/certified pathologists at a single cancer reference laboratory between 2015 and 2017. Rate of concordance between assay findings was assessed by Bland-Altman analysis; overall per cent agreement (OPA), positive per cent agreement and negative per cent agreement; and Cohen's kappa. RESULTS: PD-L1 IHC 28-8 and 22C3 displayed strong correlation across all samples and in samples with a confirmed lung cancer diagnosis irrespective of biopsy site. The OPA was 97%-98% for all samples, depending on the expression level defining PD-L1 positivity. In the Bland-Altman analysis, the mean difference in percentage of tumour cells positively stained for PD-L1 between the paired assay findings was -0.80% for all samples and -0.93% in samples with a confirmed lung cancer diagnosis. CONCLUSIONS: These data, in conjunction with recent findings, support the analytical concordance of the PD-L1 IHC 28-8 and 22C3 assays for assessing per cent tumour-cell membrane PD-L1 expression.

A Prospective, Multi-institutional, Pathologist-Based Assessment of 4 Immunohistochemistry Assays for PD-L1 Expression in Non-Small Cell Lung Cancer.

IMPORTANCE: Four assays registered with the US Food and Drug Administration (FDA) detect programmed cell death ligand 1 (PD-L1) to enrich for patient response to anti-programmed cell death 1 and anti-PD-L1 therapies. The tests use 4 separate PD-L1 antibodies on 2 separate staining platforms and have their own scoring systems, which raises questions about their similarity and the potential interchangeability of the tests. OBJECTIVE: To compare the performance of 4 PD-L1 platforms, including 2 FDA-cleared assays, 1 test for investigational use only, and 1 laboratory-developed test. DESIGN, SETTING, AND PARTICIPANTS: Four serial histologic sections from 90 archival non-small cell lung cancers from January 1, 2008, to December 31, 2010, were distributed to 3 sites that performed the following immunohistochemical assays: 28-8 antibody on the Dako Link 48 platform, 22c3 antibody on the Dako Link 48 platform, SP142 antibody on the Ventana Benchmark platform, and E1L3N antibody on the Leica Bond platform. The slides were scanned and scored by 13 pathologists who estimated the percentage of malignant and immune cells expressing PD-L1. Statistical analyses were performed from December 1, 2015, to August 30, 2016, to compare antibodies and pathologists' scoring of tumor and immune cells. MAIN OUTCOMES AND MEASURES: Percentages of malignant and immune cells expressing PD-L1. RESULTS: Among the 90 samples, the SP142 assay was an outlier, with a significantly lower mean score of PD-L1 expression in both tumor and immune cells (tumor cells: 22c3, 2.96; 28-8, 3.26; SP142, 1.99; E1L3N, 3.20; overall mean, 2.85; and immune cells: 22c3, 2.15; 28-8, 2.28; SP142, 1.62; E1L3N, 2.28; overall mean, 2.08). Pairwise comparisons showed that the scores from the 28-8 and E1L3N tests were not significantly different but that the 22c3 test showed a slight (mean difference, 0.24-0.30) but statistically significant reduction in labeling of PD-L1 expression in tumor cells. Evaluation of intraclass correlation coefficients (ICCs) between antibodies to quantify interassay variability for PD-L1 expression in tumor cells showed high concordance between antibodies for tumor cell scoring (0.813; 95% CI, 0.815-0.839) and lower levels of concordance for immune cell scoring (0.277; 95% CI, 0.222-0.334). When examining variability between pathologists for any single assay, the concordance between pathologists' scoring for PD-L1 expression in tumor cells ranged from ICCs of 0.832 (95% CI, 0.820-0.844) to 0.882 (95% CI, 0.873-0.891) for each assay, while the ICCs from immune cells for each assay ranged from 0.172 (95% CI, 0.156-0.189) to 0.229 (95% CI, 0.211-0.248). CONCLUSIONS AND RELEVANCE: The assay using the SP142 antibody is an outlier that detected significantly less PD-L1 expression in tumor cells and immune cells. The assay for antibody 22c3 showed slight yet statistically significantly lower staining than either 28-8 or E1L3N, but this significance was detected only when using the mean of 13 pathologists' scores. The pathologists showed excellent concordance when scoring tumor cells stained with any antibody but poor concordance for scoring immune cells stained with any antibody. Thus, for tumor cell assessment of PD-L1, 3 of the 4 tests are concordant and reproducible as read by pathologists.

PD-L1 Expression by Two Complementary Diagnostic Assays and mRNA In Situ Hybridization in Small Cell Lung Cancer.

INTRODUCTION: Therapeutic antibodies to immune checkpoints show promising results. Programmed death-ligand 1 (PD-L1), an immune checkpoint ligand, blocks the cancer immunity cycle by binding the PD-L1 receptor (programmed death 1). We investigated PD-L1 protein expression and messenger RNA (mRNA) levels in SCLC. METHODS: PD-L1 protein expression and mRNA levels were determined by immunohistochemistry (IHC) with SP142 and Dako 28-8 PD-L1 antibodies and in situ hybridization in primary tumor tissue microarrays in both tumor cells and tumor-infiltrating immune cells (TIICs) obtained from a limited-disease SCLC cohort of 98 patients. An additional cohort of 96 tumor specimens from patients with extensive-disease SCLC was assessed for PD-L1 protein expression in tumor cells with Dako 28-8 antibody only. RESULTS: The overall prevalence of PD-L1 protein expression in tumor cells was 16.5%. In the limited-disease cohort, the prevalences of PD-L1 protein expression in tumor cells with SP142 and Dako 28-8 were 14.7% and 19.4% (tumor proportion score cutoff ≥1%) and PD-L1 mRNA ISH expression was positive in 15.5% of tumor samples. Increased PD-L1 protein/mRNA expression was associated with the presence of more TIICs (p < 0.05). The extensive-disease cohort demonstrated a 14.9% positivity of PD-L1 protein expression in tumor cells with Dako 28-8 antibody. CONCLUSIONS: A subset of SCLCs is characterized by positive PD-L1 and/or mRNA expression in tumor cells. Higher PD-L1 and mRNA expression correlate with more infiltration of TIICs. The prevalence of PD-L1 in SCLC is lower than that published for NSCLC. The predictive role of PD-L1 expression in SCLC treatment remains to be established.

VEGF-Mediated Induction of PRD1-BF1/Blimp1 Expression Sensitizes Tumor Vasculature to Oncolytic Virus Infection.

Oncolytic viruses designed to attack malignant cells can in addition infect and destroy tumor vascular endothelial cells. We show here that this expanded tropism of oncolytic vaccinia virus to the endothelial compartment is a consequence of VEGF-mediated suppression of the intrinsic antiviral response. VEGF/VEGFR2 signaling through Erk1/2 and Stat3 leads to upregulation, nuclear localization, and activation of the transcription repressor PRD1-BF1/Blimp1. PRD1-BF1 does not contribute to the mitogenic effects of VEGF, but directly represses genes involved in type I interferon (IFN)-mediated antiviral signaling. In vivo suppression of VEGF signaling diminishes PRD1-BF1/Blimp1 expression in tumor vasculature and inhibits intravenously administered oncolytic vaccinia delivery to and consequent spread within the tumor.

Reciprocal cellular cross-talk within the tumor microenvironment promotes oncolytic virus activity.

Tumors are complex ecosystems composed of networks of interacting 'normal' and malignant cells. It is well recognized that cytokine-mediated cross-talk between normal stromal cells, including cancer-associated fibroblasts (CAFs), vascular endothelial cells, immune cells, and cancer cells, influences all aspects of tumor biology. Here we demonstrate that the cross-talk between CAFs and cancer cells leads to enhanced growth of oncolytic virus (OV)-based therapeutics. Transforming growth factor-ß (TGF-ß) produced by tumor cells reprogrammed CAFs, dampened their steady-state level of antiviral transcripts and rendered them sensitive to virus infection. In turn, CAFs produced high levels of fibroblast growth factor 2 (FGF2), initiating a signaling cascade in cancer cells that reduced retinoic acid-inducible gene I (RIG-I) expression and impeded the ability of malignant cells to detect and respond to virus. In xenografts derived from individuals with pancreatic cancer, the expression of FGF2 correlated with the susceptibility of the cancer cells to OV infection, and local application of FGF2 to resistant tumor samples sensitized them to virotherapy both in vitro and in vivo. An OV engineered to express FGF2 was safe in tumor-bearing mice, showed improved therapeutic efficacy compared to parental virus and merits consideration for clinical testing.

Microtubule disruption synergizes with oncolytic virotherapy by inhibiting interferon translation and potentiating bystander killing.

In this study, we show that several microtubule-destabilizing agents used for decades for treatment of cancer and other diseases also sensitize cancer cells to oncolytic rhabdoviruses and improve therapeutic outcomes in resistant murine cancer models. Drug-induced microtubule destabilization leads to superior viral spread in cancer cells by disrupting type I IFN mRNA translation, leading to decreased IFN protein expression and secretion. Furthermore, microtubule-destabilizing agents specifically promote cancer cell death following stimulation by a subset of infection-induced cytokines, thereby increasing viral bystander effects. This study reveals a previously unappreciated role for microtubule structures in the regulation of the innate cellular antiviral response and demonstrates that unexpected combinations of approved chemotherapeutics and biological agents can lead to improved therapeutic outcomes.

High-throughput titration of luciferase-expressing recombinant viruses.

Standard plaque assays to determine infectious viral titers can be time consuming, are not amenable to a high volume of samples, and cannot be done with viruses that do not form plaques. As an alternative to plaque assays, we have developed a high-throughput titration method that allows for the simultaneous titration of a high volume of samples in a single day. This approach involves infection of the samples with a Firefly luciferase tagged virus, transfer of the infected samples onto an appropriate permissive cell line, subsequent addition of luciferin, reading of plates in order to obtain luminescence readings, and finally the conversion from luminescence to viral titers. The assessment of cytotoxicity using a metabolic viability dye can be easily incorporated in the workflow in parallel and provide valuable information in the context of a drug screen. This technique provides a reliable, high-throughput method to determine viral titers as an alternative to a standard plaque assay.

Model-based rational design of an oncolytic virus with improved therapeutic potential.

Oncolytic viruses are complex biological agents that interact at multiple levels with both tumour and normal tissues. Antiviral pathways induced by interferon are known to have a critical role in determining tumour cell sensitivity and normal cell resistance to infection with oncolytic viruses. Here we pursue a synthetic biology approach to identify methods that enhance antitumour activity of oncolytic viruses through suppression of interferon signalling. On the basis of the mathematical analysis of multiple strategies, we hypothesize that a positive feedback loop, established by virus-mediated expression of a soluble interferon-binding decoy receptor, increases tumour cytotoxicity without compromising normal cells. Oncolytic rhabdoviruses engineered to express a secreted interferon antagonist have improved oncolytic potential in cellular cancer models, and display improved therapeutic potential in tumour-bearing mice. Our results demonstrate the potential of this methodology in evaluating potential caveats of viral immune-evasion strategies and improving the design oncolytic viruses.

Quantitative epistasis analysis and pathway inference from genetic interaction data.

Inferring regulatory and metabolic network models from quantitative genetic interaction data remains a major challenge in systems biology. Here, we present a novel quantitative model for interpreting epistasis within pathways responding to an external signal. The model provides the basis of an experimental method to determine the architecture of such pathways, and establishes a new set of rules to infer the order of genes within them. The method also allows the extraction of quantitative parameters enabling a new level of information to be added to genetic network models. It is applicable to any system where the impact of combinatorial loss-of-function mutations can be quantified with sufficient accuracy. We test the method by conducting a systematic analysis of a thoroughly characterized eukaryotic gene network, the galactose utilization pathway in Saccharomyces cerevisiae. For this purpose, we quantify the effects of single and double gene deletions on two phenotypic traits, fitness and reporter gene expression. We show that applying our method to fitness traits reveals the order of metabolic enzymes and the effects of accumulating metabolic intermediates. Conversely, the analysis of expression traits reveals the order of transcriptional regulatory genes, secondary regulatory signals and their relative strength. Strikingly, when the analyses of the two traits are combined, the method correctly infers ~80% of the known relationships without any false positives.

Chromosomal position effects are linked to sir2-mediated variation in transcriptional burst size.

Gene expression noise varies with genomic position and is a driving force in the evolution of chromosome organization. Nevertheless, position effects remain poorly characterized. Here, we present a systematic analysis of chromosomal position effects by characterizing single-cell gene expression from euchromatic positions spanning the length of a eukaryotic chromosome. We demonstrate that position affects gene expression by modulating the size of transcriptional bursts, rather than their frequency, and that the histone deacetylase Sir2 plays a role in this process across the chromosome.

Identification of response-modulated genetic interactions by sensitivity-based epistatic analysis.

BACKGROUND: High-throughput genomics has enabled the global mapping of genetic interactions based on the phenotypic impact of combinatorial genetic perturbations. An important next step is to understand how these networks are dynamically remodelled in response to environmental stimuli. Here, we report on the development and testing of a method to identify such interactions. The method was developed from first principles by treating the impact on cellular growth of environmental perturbations equivalently to that of gene deletions. This allowed us to establish a novel neutrality function marking the absence of epistasis in terms of sensitivity phenotypes rather than fitness. We tested the method by identifying fitness- and sensitivity-based interactions involved in the response to drug-induced DNA-damage of budding yeast Saccharomyces cerevisiae using two mutant libraries - one containing transcription factor deletions, and the other containing deletions of DNA repair genes. RESULTS: Within the library of transcription factor deletion mutants, we observe significant differences in the sets of genetic interactions identified by the fitness- and sensitivity-based approaches. Notably, among the most likely interactions, only ~50% were identified by both methods. While interactions identified solely by the sensitivity-based approach are modulated in response to drug-induced DNA damage, those identified solely by the fitness-based method remained invariant to the treatment. Comparison of the identified interactions to transcriptional profiles and protein-DNA interaction data indicate that the sensitivity-based method improves the identification of interactions involved in the DNA damage response. Additionally, for the library containing DNA repair mutants, we observe that the sensitivity-based method improves the grouping of functionally related genes, as well as the identification of protein complexes, involved in DNA repair. CONCLUSION: Our results show that the identification of response-modulated genetic interactions can be improved by incorporating the effect of a changing environment directly into the neutrality function marking the absence of epistasis. We expect that this extension of conventional epistatic analysis will facilitate the development of dynamic models of gene networks from quantitative measurements of genetic interactions. While the method was developed for growth phenotype, it should apply equally well for other phenotypes, including the expression of fluorescent reporters.