Direct prediction of genetic aberrations from pathology images in gastric cancer with swarm learning.

BACKGROUND: Computational pathology uses deep learning (DL) to extract biomarkers from routine pathology slides. Large multicentric datasets improve performance, but such datasets are scarce for gastric cancer. This limitation could be overcome by Swarm Learning (SL). METHODS: Here, we report the results of a multicentric retrospective study of SL for prediction of molecular biomarkers in gastric cancer. We collected tissue samples with known microsatellite instability (MSI) and Epstein-Barr Virus (EBV) status from four patient cohorts from Switzerland, Germany, the UK and the USA, storing each dataset on a physically separate computer. RESULTS: On an external validation cohort, the SL-based classifier reached an area under the receiver operating curve (AUROC) of 0.8092 (± 0.0132) for MSI prediction and 0.8372 (± 0.0179) for EBV prediction. The centralized model, which was trained on all datasets on a single computer, reached a similar performance. CONCLUSIONS: Our findings demonstrate the feasibility of SL-based molecular biomarkers in gastric cancer. In the future, SL could be used for collaborative training and, thus, improve the performance of these biomarkers. This may ultimately result in clinical-grade performance and generalizability.

Metabolic tumor constitution is superior to tumor regression grading for evaluating response to neoadjuvant therapy of esophageal adenocarcinoma patients.

The response to neoadjuvant therapy can vary widely between individual patients. Histopathological tumor regression grading (TRG) is a strong factor for treatment response and survival prognosis of esophageal adenocarcinoma (EAC) patients following neoadjuvant treatment and surgery. However, TRG systems are usually based on the estimation of residual tumor but do not consider stromal or metabolic changes after treatment. Spatial metabolomics analysis is a powerful tool for molecular tissue phenotyping but has not been used so far in the context of neoadjuvant treatment of esophageal cancer. We used imaging mass spectrometry to assess the potential of spatial metabolomics on tumor and stroma tissue for evaluating therapy response of neoadjuvant-treated EAC patients. With an accuracy of 89.7%, the binary classifier trained on spatial tumor metabolite data proved to be superior for stratifying patients when compared with histopathological response assessment, which had an accuracy of 70.5%. Sensitivities and specificities for the poor and favorable survival patient groups ranged from 84.9% to 93.3% using the metabolic classifier and from 62.2% to 78.1% using TRG. The tumor classifier was the only significant prognostic factor (HR 3.38, 95% CI 1.40-8.12, p = 0.007) when adjusted for clinicopathological parameters such as TRG (HR 1.01, 95% CI 0.67-1.53, p = 0.968) or stromal classifier (HR 1.86, 95% CI 0.81-4.25, p = 0.143). The classifier even allowed us to further stratify patients within the TRG1-3 categories. The underlying mechanisms of response to treatment have been figured out through network analysis. In summary, metabolic response evaluation outperformed histopathological response evaluation in our study with regard to prognostic stratification. This finding indicates that the metabolic constitution of the tumor may have a greater impact on patient survival than the quantity of residual tumor cells or the stroma. © 2021 The Authors. The Journal of Pathology published by John Wiley & Sons, Ltd. on behalf of The Pathological Society of Great Britain and Ireland.

Chlamydia-induced curvature of the host-cell plasma membrane is required for infection.

During invasion of host cells, Chlamydia pneumoniae secretes the effector protein CPn0678, which facilitates internalization of the pathogen by remodeling the target cell's plasma membrane and recruiting sorting nexin 9 (SNX9), a central multifunctional endocytic scaffold protein. We show here that the strongly amphipathic N-terminal helix of CPn0678 mediates binding to phospholipids in both the plasma membrane and synthetic membranes, and is sufficient to induce extensive membrane tubulations. CPn0678 interacts via its conserved C-terminal polyproline sequence with the Src homology 3 domain of SNX9. Thus, SNX9 is found at bacterial entry sites, where C. pneumoniae is internalized via EGFR-mediated endocytosis. Moreover, depletion of human SNX9 significantly reduces internalization, whereas ectopic overexpression of CPn0678-GFP results in a dominant-negative effect on endocytotic processes in general, leading to the uptake of fewer chlamydial elementary bodies and diminished turnover of EGFR. Thus, CPn0678 is an early effector involved in regulating the endocytosis of C. pneumoniae in an EGFR- and SNX9-dependent manner.

Mouse brain proteomics establishes MDGA1 and CACHD1 as in vivo substrates of the Alzheimer protease BACE1.

The protease beta-site APP cleaving enzyme 1 (BACE1) has fundamental functions in the nervous system. Its inhibition is a major therapeutic approach in Alzheimer's disease, because BACE1 cleaves the amyloid precursor protein (APP), thereby catalyzing the first step in the generation of the pathogenic amyloid beta (Aß) peptide. Yet, BACE1 cleaves numerous additional membrane proteins besides APP. Most of these substrates have been identified in vitro, but only few were further validated or characterized in vivo. To identify BACE1 substrates with in vivo relevance, we used isotope label-based quantitative proteomics of wild type and BACE1-deficient (BACE1 KO) mouse brains. This approach identified known BACE1 substrates, including Close homolog of L1 and contactin-2, which were found to be enriched in the membrane fraction of BACE1 KO brains. VWFA and cache domain-containing protein 1 (CACHD)1 and MAM domain-containing glycosylphosphatidylinositol anchor protein 1 (MDGA1), which have functions in synaptic transmission, were identified and validated as new BACE1 substrates in vivo by immunoblots using primary neurons and mouse brains. Inhibition or deletion of BACE1 from primary neurons resulted in a pronounced inhibition of substrate cleavage and a concomitant increase in full-length protein levels of CACHD1 and MDGA1. The BACE1 cleavage site in both proteins was determined to be located within the juxtamembrane domain. In summary, this study identifies and validates CACHD1 and MDGA1 as novel in vivo substrates for BACE1, suggesting that cleavage of both proteins may contribute to the numerous functions of BACE1 in the nervous system.

Preservation of Epstein-Barr virus status and mismatch repair protein status along the metastatic course of gastric cancer.

AIMS: Epstein-Barr virus (EBV) in-situ hybridisation and mismatch repair (MMR) protein immunohistochemistry identifies two subgroups of gastric cancer (GC) with high immunogenicity and likelihood for response to immune check-point inhibition. As tumour biology may change during the metastatic course, which can negatively influence the success of therapeutic decisions made on primary tissue, we investigated the consistency of GC EBV and MMR status within primary tumours and metastases. METHODS AND RESULTS: We investigated a cohort of 415 primary resected GC, including 111 cases with corresponding distant metastases and 297 cases with lymph node metastases. Tumours were analysed by EBV in-situ hybridisation and MLH1, PMS2, MSH2 and MSH6 immunohistochemistry using tissue microarray technique. Primary tumours were grouped as EBV-positive MMR-proficient, EBV-negative MMR-deficient and EBV-negative MMR-proficient. Eleven of 415 (2.7%) of primary tumours were EBV-positive MMR-proficient, whereas 49 of 415 (11.8%) of tumours were EBV-negative MMR-deficient. EBV and MMR protein status showed full concordance with that of the primary tumours. MMR-deficient tumours were of lower pT-category (P < 0.001), had fewer lymph node metastases [24 of 49 (49%) versus 273 of 361 (75.6%) cases; P < 0.001] and a lower rate of distant metastases [six of 49 (12.2%) versus 105 of 366 (28.7%) cases; P = 0.015]. CONCLUSION: We demonstrate a strong correlation of EBV and MMR status between primary tumours, lymph node and distant metastases in a large series of primary resected GC. The cases showed the expected frequency of EBV-positive MMR-deficient and EBV-negative MMR-proficient tumours. We conclude that tissue testing for molecular subtyping for therapeutic decision-making can be reliably performed on primary tumours and metastases in GC.

Favourable long-term survival of patients with esophageal cancer treated with extended transhiatal esophagectomy combined with en bloc lymphadenectomy: results from a retrospective observational cohort study.

BACKGROUND: Although considered complex and challenging, esophagectomy remains the best potentially curable treatment option for resectable esophageal and esophagogastric junction (AEG) carcinomas. The optimal surgical approach and technique as well as the extent of lymphadenectomy, particularly regarding quality of life and short- and long-term outcomes, are still a matter of debate. To lower perioperative morbidity, we combined the advantages of a one-cavity approach with extended lymph node dissection (usually achieved by only a two-cavity approach) and developed a modified single-cavity transhiatal approach for esophagectomy. METHODS: The aim of this study was to evaluate the outcome of an extended transhiatal esophageal resection with radical bilateral mediastinal en bloc lymphadenectomy (eTHE). A prospective database of 166 patients with resectable cancers of the esophagus (including adenocarcinomas of the AEG types I and II) were analyzed. Patients were treated between 2001 and 2017 with eTHE at a tertiary care university center. Relevant patient characteristics and outcome parameters were collected and analyzed. The primary endpoint was 5-year overall survival. Secondary outcomes included short-term morbidity, mortality, radicalness of en bloc resection and oncologic efficacy. RESULTS: The overall survival rates at 1, 3 and 5 years were 84, 70, and 61.0%, respectively. The in-hospital mortality rate after eTHE was 1.2%. Complications with a Clavien-Dindo score of III/IV occurred in 31 cases (18.6%). A total of 25 patients (15.1%) had a major pulmonary complication. The median hospital stay was 17 days (interquartile range (IQR) 12). Most patients (n = 144; 86.7%) received neoadjuvant treatment. The median number of lymph nodes resected was 25 (IQR 17). The R0 resection rate was 97%. CONCLUSION: In patients with esophageal cancer, eTHE without thoracotomy resulted in excellent long-term survival, an above average number of resected lymph nodes and an acceptable postoperative morbidity and mortality.

Her2-Targeted Therapy Induces Autophagy in Esophageal Adenocarcinoma Cells.

Esophageal adenocarcinoma (EAC) is a highly lethal cancer type with an overall poor survival rate. Twenty to thirty percent of EAC overexpress the human epidermal growth factor receptor 2 (Her2), a transmembrane receptor tyrosine kinase promoting cell growth and proliferation. Patients with Her2 overexpressing breast and gastroesophageal cancer may benefit from Her2 inhibitors. Therapy resistance, however, is well documented. Since autophagy, a lysosome-dependent catabolic process, is implicated in cancer resistance mechanisms, we tested whether autophagy modulation influences Her2 inhibitor sensitivity in EAC. Her2-positive OE19 EAC cells showed an induction in autophagic flux upon treatment with the small molecule Her2 inhibitor Lapatinib. Newly generated Lapatinib-resistant OE19 (OE19 LR) cells showed increased basal autophagic flux compared to parental OE19 (OE19 P) cells. Based on these results, we tested if combining Lapatinib with autophagy inhibitors might be beneficial. OE19 P showed significantly reduced cell viability upon double treatment, while OE19 LR were already sensitive to autophagy inhibition alone. Additionally, Her2 status and autophagy marker expression (LC3B and p62) were investigated in a treatment-naïve EAC patient cohort (n = 112) using immunohistochemistry. Here, no significant correlation between Her2 status and expression of LC3B and p62 was found. Our data show that resistance to Her2-directed therapy is associated with a higher basal autophagy level, which is not per se associated with Her2 status. Therefore, we propose that autophagy may contribute to acquired resistance to Her2-targeted therapy in EAC, and that combining Her2 and autophagy inhibition might be beneficial for EAC patients.

Expression patterns of programmed death-ligand 1 in esophageal adenocarcinomas: comparison between primary tumors and metastases.

Expression analysis of programmed death-ligand 1 (PD-L1) may be helpful in guiding clinical decisions for immune checkpoint inhibition therapy, but testing by immunohistochemistry may be hampered by heterogeneous staining patterns within tumors and expression changes during metastatic course. PD-L1 expression (clone SP142) was investigated in esophageal adenocarcinomas using tissue microarrays (TMA) from 112 primary resected tumors, preoperative biopsies and full slide sections from a subset of these cases (n = 24), corresponding lymph node (n = 55) and distant metastases (n = 17). PD-L1 expression was scored as 0.1-1, >1, >5, >50% positive membranous staining of tumor cells and any positive staining of tumor-associated inflammatory infiltrates and/or stroma cells. There was a significant correlation with overall PD-L1 expression between the full slide sections and the TMA (p = 0.001), but not with the corresponding biopsies. PD-L1 expression in tumor cells >1% was detected in 8.0% of cases (9/112) and 51.8% of cases (58/112) in tumor-associated inflammatory infiltrates and/or stroma cells of primary tumors. Epithelial expression in metastases was found in 5.6% of cases (4/72) and immune cell expression in 18.1% of cases (13/72), but did not correlate with the expression pattern in the primary tumor. Overall PD-L1 expression in the primary tumor did not influence survival. However, PD-L1 expression was correlated with the number of CD3+ tumor-infiltrating lymphocytes in the tumor center, and a combinational score of PD-L1 status/CD3+ tumor-infiltrating lymphocytes was correlated with patients' overall survival.

Control of Homeostasis and Dendritic Cell Survival by the GTPase RhoA.

Tissues accommodate defined numbers of dendritic cells (DCs) in highly specific niches where different intrinsic and environmental stimuli control DC life span and numbers. DC homeostasis in tissues is important, because experimental changes in DC numbers influence immunity and tolerance toward various immune catastrophes and inflammation. However, the precise molecular mechanisms regulating DC life span and homeostasis are unclear. We report that the GTPase RhoA controls homeostatic proliferation, cytokinesis, survival, and turnover of cDCs. Deletion of RhoA strongly decreased the numbers of CD11b(-)CD8(+) and CD11b(+)Esam(hi) DC subsets, whereas CD11b(+)Esam(lo) DCs were not affected in conditional RhoA-deficient mice. Proteome analyses revealed a defective prosurvival pathway via PI3K/protein kinase B (Akt1)/Bcl-2-associated death promoter in the absence of RhoA. Taken together, our findings identify RhoA as a central regulator of DC homeostasis, and its deletion decreases DC numbers below critical thresholds for immune protection and homeostasis, causing aberrant compensatory DC proliferation.

Label-free Quantitative Proteomics of Mouse Cerebrospinal Fluid Detects ß-Site APP Cleaving Enzyme (BACE1) Protease Substrates In Vivo.

Analysis of murine cerebrospinal fluid (CSF) by quantitative mass spectrometry is challenging because of low CSF volume, low total protein concentration, and the presence of highly abundant proteins such as albumin. We demonstrate that the CSF proteome of individual mice can be analyzed in a quantitative manner to a depth of several hundred proteins in a robust and simple workflow consisting of single ultra HPLC runs on a benchtop mass spectrometer. The workflow is validated by a comparative analysis of BACE1-/- and wild-type mice using label-free quantification. The protease BACE1 cleaves the amyloid precursor protein (APP) as well as several other substrates and is a major drug target in Alzheimer's disease. We identified a total of 715 proteins with at least 2 unique peptides and quantified 522 of those proteins in CSF from BACE1-/- and wild-type mice. Several proteins, including the known BACE1 substrates APP, APLP1, CHL1 and contactin-2 showed lower abundance in the CSF of BACE1-/- mice, demonstrating that BACE1 substrate identification is possible from CSF. Additionally, ectonucleotide pyrophosphatase 5 was identified as a novel BACE1 substrate and validated in cells using immunoblots and by an in vitro BACE1 protease assay. Likewise, receptor-type tyrosine-protein phosphatase N2 and plexin domain-containing 2 were confirmed as BACE1 substrates by in vitro assays. Taken together, our study shows the deepest characterization of the mouse CSF proteome to date and the first quantitative analysis of the CSF proteome of individual mice. The BACE1 substrates identified in CSF may serve as biomarkers to monitor BACE1 activity in Alzheimer patients treated with BACE inhibitors.

QARIP: a web server for quantitative proteomic analysis of regulated intramembrane proteolysis.

Regulated intramembrane proteolysis (RIP) is a critical mechanism for intercellular communication and regulates the function of membrane proteins through sequential proteolysis. RIP typically starts with ectodomain shedding of membrane proteins by extracellular membrane-bound proteases followed by intramembrane proteolysis of the resulting membrane-tethered fragment. However, for the majority of RIP proteases the corresponding substrates and thus, their functions, remain unknown. Proteome-wide identification of RIP protease substrates is possible by mass spectrometry-based quantitative comparison of RIP substrates or their cleavage products between different biological states. However, this requires quantification of peptides from only the ectodomain or cytoplasmic domain. Current analysis software does not allow matching peptides to either domain. Here we present the QARIP (Quantitative Analysis of Regulated Intramembrane Proteolysis) web server which matches identified peptides to the protein transmembrane topology. QARIP allows determination of quantitative ratios separately for the topological domains (cytoplasmic, ectodomain) of a given protein and is thus a powerful tool for quality control, improvement of quantitative ratios and identification of novel substrates in proteomic RIP datasets. To our knowledge, the QARIP web server is the first tool directly addressing the phenomenon of RIP. The web server is available at http://webclu.bio.wzw.tum.de/qarip/. This website is free and open to all users and there is no login requirement.

ADAM10 is the physiologically relevant, constitutive alpha-secretase of the amyloid precursor protein in primary neurons.

The amyloid precursor protein (APP) undergoes constitutive shedding by a protease activity called alpha-secretase. This is considered an important mechanism preventing the generation of the Alzheimer's disease amyloid-beta peptide (Abeta). alpha-Secretase appears to be a metalloprotease of the ADAM family, but its identity remains to be established. Using a novel alpha-secretase-cleavage site-specific antibody, we found that RNAi-mediated knockdown of ADAM10, but surprisingly not of ADAM9 or 17, completely suppressed APP alpha-secretase cleavage in different cell lines and in primary murine neurons. Other proteases were not able to compensate for this loss of alpha-cleavage. This finding was further confirmed by mass-spectrometric detection of APP-cleavage fragments. Surprisingly, in different cell lines, the reduction of alpha-secretase cleavage was not paralleled by a corresponding increase in the Abeta-generating beta-secretase cleavage, revealing that both proteases do not always compete for APP as a substrate. Instead, our data suggest a novel pathway for APP processing, in which ADAM10 can partially compete with gamma-secretase for the cleavage of a C-terminal APP fragment generated by beta-secretase. We conclude that ADAM10 is the physiologically relevant, constitutive alpha-secretase of APP.

Label-free quantitative analysis of the membrane proteome of Bace1 protease knock-out zebrafish brains.

The aspartyl protease BACE1 cleaves neuregulin 1 and is involved in myelination and is a candidate drug target for Alzheimer's disease, where it acts as the ß-secretase cleaving the amyloid precursor protein. However, little is known about other substrates in vivo. Here, we provide a proteomic workflow for BACE1 substrate identification from whole brains, combining filter-aided sample preparation, strong-anion exchange fractionation, and label-free quantification. We used bace1-deficient zebrafish and quantified differences in protein levels between wild-type and bace1 -/- zebrafish brains. Over 4500 proteins were identified with at least two unique peptides and quantified in both wild-type and bace1 -/- zebrafish brains. The majority of zebrafish membrane proteins did not show altered protein levels, indicating that Bace1 has a restricted substrate specificity. Twenty-four membrane proteins accumulated in the bace1 -/- brains and thus represent candidate Bace1 substrates. They include several known BACE1 substrates, such as the zebrafish homologs of amyloid precursor protein and the cell adhesion protein L1, which validate the proteomic workflow. Additionally, several candidate substrates with a function in neurite outgrowth and axon guidance, such as plexin A3 and glypican-1 were identified, pointing to a function of Bace1 in neurodevelopment. Taken together, our study provides the first proteomic analysis of knock-out zebrafish tissue and demonstrates that combining gene knock-out models in zebrafish with quantitative proteomics is a powerful approach to address biomedical questions.

Surgical pathology of adenocarcinomas arising around or within the gastroesophageal junction.

Classification of adenocarcinomas (AC) arising around or within the gastroesophageal junction (GEJ) is hampered by major morphologic and phenotypic overlaps. We reviewed the surgical pathology of esophagectomy specimens of 115 primary resected AC of the esophagus as defined by the 5th edition of the WHO classification regarding the anatomical site of the tumor, with corresponding categorization according to the Siewert AEG Classification and the preceding 4th edition of the WHO (discriminating esophageal adenocarcinomas/EAC and adenocarcinomas of the gastroesophageal junction/AdGEJ), and further histology findings. In addition, immunohistochemistry (IHC) for CDX2, CK7, CK20, MUC2, MUC5AC and MUC6 was performed. Sixty-eight cases were Siewert AEG type I and 47 cases Siewert AEG type II. Out of the AEG I tumors, 26 were classified as AdGEJ. Regardless of the classification system, more proximally located tumors showed less aggressive behavior with lower rates of lymph node metastases, lymphatic, venous and perineural invasion, better histological differentiation (p < 0.05 each) and were more frequently associated with pre-neoplastic Barrett's mucosa (p < 0.001). Histologically, the tumors displayed intestinal morphology in the majority of cases. IHC showed non-conclusive patterns with a frequent CK7+/CK20+ immunophenotype in all tumors, but also a gastric MUC5AC+ and MUC6+ phenotype in some proximal tumors. In conclusion, histology of the tumors and IHC failed to distinguish reliably between more proximal and more distal tumors. The presence of Barrett's mucosa rather than location alone, however, may help to further differentiating adenocarcinomas arising in this region and may be indicative for a particular biologic type.

Programmed death-ligand 1 (PD-L1) expression in primary gastric adenocarcinoma and matched metastases.

BACKGROUND: Combination of immunotherapy and chemotherapy is recommended for first line treatment of gastric adenocarcinoma (GC) patients with locally advanced unresectable disease or metastatic disease. However, data regarding the concordance rate between PD-L1 combined positive score (CPS) in primary GC and matched regional lymph node metastasis (LNmet) or matched distant metastasis (Dmet) is limited. METHODS: Tissue microarray sections from primary resected GC, LNmet and Dmet were immunohistochemically stained with anti-PD-L1 (clone SP263). PD-L1 expression was scored separately in tumour cells and immune cells and compared between matched primary GC, LNmet and/or Dmet. CPS was calculated and results for CPS cut-offs 1 and 5 were compared between matched samples. RESULTS: 275 PD-L1 stained GC were analysed. 189 primary GC had matched LNmet. CPS cut-off 1 concordance rate between primary GC and LNmet was 77%. 23 primary GC had matched Dmet but no matched LNmet, CPS cut-off 1 concordance rate was 70%. 63 primary GC had both matched LNmet and matched Dmet, CPS cut-off 1 concordance rate of 67%. CPS cut-off 5 results were similar. The proportion of PD-L1 positive tumour cells increased from primary GC (26%) to LNmet (42%) and was highest in Dmet (75%). CONCLUSION: Our study showed up to 33% discordance of PD-L1 CPS between primary GC and LNmet and/or Dmet suggesting that multiple biopsies of primary GC and metastatic sites might need to be tested before considering treatment options. Moreover, this is the first study that seems to suggest that tumour cells acquire PD-L1 expression during disease progression.

Specific amino acids in the BAR domain allow homodimerization and prevent heterodimerization of sorting nexin 33.

SNX33 (sorting nexin 33) is a homologue of the endocytic protein SNX9 and has been implicated in actin polymerization and the endocytosis of the amyloid precursor protein. SNX33 belongs to the large family of BAR (Bin/amphiphysin/Rvs) domain-containing proteins, which alter cellular protein trafficking by modulating cellular membranes and the cytoskeleton. Some BAR domains engage in homodimerization, whereas other BAR domains also mediate heterodimerization between different BAR domain-containing proteins. The molecular basis for this difference is not yet understood. Using co-immunoprecipitations we report that SNX33 forms homodimers, but not heterodimers, with other BAR domain-containing proteins, such as SNX9. Domain deletion analysis revealed that the BAR domain, but not the SH3 (Src homology 3) domain, was required for homodimerization of SNX33. Additionally, the BAR domain prevented the heterodimerization between SNX9 and SNX33, as determined by domain swap experiments. Molecular modelling of the SNX33 BAR domain structure revealed that key amino acids located at the BAR domain dimer interface of the SNX9 homodimer are not conserved in SNX33. Replacing these amino acids in SNX9 with the corresponding amino acids of SNX33 allowed the mutant SNX9 to heterodimerize with SNX33. Taken together, the present study identifies critical amino acids within the BAR domains of SNX9 and SNX33 as determinants for the specificity of BAR domain-mediated interactions and suggests that SNX9 and SNX33 have distinct molecular functions.

Interspatial Distribution of Tumor and Immune Cells in Correlation with PD-L1 in Molecular Subtypes of Gastric Cancers.

(1) Background: EBV-positive and mismatch repair-deficient (MMRd) gastric cancers (GCs) show higher levels of tumor-infiltrating lymphocytes (TILs) and PD-L1 expression and thus a more profound response to immunotherapy. However, the majority of GCs are EBV-negative (EBV−) and MMR proficient (MMRp). We analyzed PD-L1 expression and TILs in EBV-MMRpGCs in comparison to EBV-positive (EBV+) and MMRdGCs to identify an immunogenic phenotype susceptible to immunotherapy. (2) Methods: A next-generation tissue microarray of 409 primary resected GCs was analyzed by Epstein-Barr encoding region (EBER) in situ hybridization for MSH1, PMS2, MSH2, MSH6, PD-L1, and CD8 immunohistochemistry. PD-L1 positivity was defined as a combined positive score (CPS) of ≥1. CD8+ TILs and their proximity to cancer cells were digitally analyzed on the HALO™ image analysis platform. (3) Results: Eleven cases were EBV+, 49 cases MMRd, and 349 cases EBV-MMRpGCs. The highest rate of PD-L1 positivity was seen in EBV+GCs, followed by MMRdGCs and EBV-MMRpGCs (81.8%, 73.5%, and 27.8%, respectively). EBV+ and MMRdGCs also demonstrated increased numbers and proximity of CD8+ TILs to tumor cells compared to EBV-MMRpGCs (p < 0.001 each). PD-L1 status positively correlated with the total numbers of CD8+ TILs and their proximity to tumor cells in all subtypes, including EBV-MMRpGCs (p < 0.001 each). A total of 28.4% of EBV-MMRpGCs showed high CD8+ TILs independent of PD-L1. (4) Conclusions: PD-L1 and CD8 immunohistochemistry, supplemented by digital image analysis, may identify EBV-MMRpGCs with high immunoreactivity indices, indicating susceptibility to immunotherapy.

An unusual pseudolymphoma in the context of necrotizing fasciitis: A case report.

RATIONALE: The diagnosis of lymphoma in routine diagnostics can be challenging due to clinical, morphological and immunphenotypical overlap with unusual reactive processes termed "pseudolymphomas." PATIENT CONCERNS: 45-year-old male that underwent surgical debridement for a necrotizing fasciitis of the thigh with concomitant excision of a regional lymph node. DIAGNOSES: The lymph node demonstrated an architecture-effacing activation and proliferation of lymphoblasts and was initially misdiagnosed as an aggressive lymphoma. Only in consideration of the clinical context and with the help of additional immunohistochemical and molecular analyses the final diagnosis of a reactive lymphadenopathy could be made. INTERVENTIONS: No further therapy was required after the final diagnosis of a reactive lymphadenopathy was made. OUTCOMES: The clinical follow-up was unremarkable, with no evidence of residual disease after 6 months. LESSONS: This case report adds the parafollicular activation and proliferation of blasts and plasmablasts in the drainage area of an active infection to the spectrum of "pseudolymphomas" and reiterizes the importance of placing histopathological findings in the proper context.

Adversarial attacks and adversarial robustness in computational pathology.

Artificial Intelligence (AI) can support diagnostic workflows in oncology by aiding diagnosis and providing biomarkers directly from routine pathology slides. However, AI applications are vulnerable to adversarial attacks. Hence, it is essential to quantify and mitigate this risk before widespread clinical use. Here, we show that convolutional neural networks (CNNs) are highly susceptible to white- and black-box adversarial attacks in clinically relevant weakly-supervised classification tasks. Adversarially robust training and dual batch normalization (DBN) are possible mitigation strategies but require precise knowledge of the type of attack used in the inference. We demonstrate that vision transformers (ViTs) perform equally well compared to CNNs at baseline, but are orders of magnitude more robust to white- and black-box attacks. At a mechanistic level, we show that this is associated with a more robust latent representation of clinically relevant categories in ViTs compared to CNNs. Our results are in line with previous theoretical studies and provide empirical evidence that ViTs are robust learners in computational pathology. This implies that large-scale rollout of AI models in computational pathology should rely on ViTs rather than CNN-based classifiers to provide inherent protection against perturbation of the input data, especially adversarial attacks.

Benchmarking weakly-supervised deep learning pipelines for whole slide classification in computational pathology.

Artificial intelligence (AI) can extract visual information from histopathological slides and yield biological insight and clinical biomarkers. Whole slide images are cut into thousands of tiles and classification problems are often weakly-supervised: the ground truth is only known for the slide, not for every single tile. In classical weakly-supervised analysis pipelines, all tiles inherit the slide label while in multiple-instance learning (MIL), only bags of tiles inherit the label. However, it is still unclear how these widely used but markedly different approaches perform relative to each other. We implemented and systematically compared six methods in six clinically relevant end-to-end prediction tasks using data from N=2980 patients for training with rigorous external validation. We tested three classical weakly-supervised approaches with convolutional neural networks and vision transformers (ViT) and three MIL-based approaches with and without an additional attention module. Our results empirically demonstrate that histological tumor subtyping of renal cell carcinoma is an easy task in which all approaches achieve an area under the receiver operating curve (AUROC) of above 0.9. In contrast, we report significant performance differences for clinically relevant tasks of mutation prediction in colorectal, gastric, and bladder cancer. In these mutation prediction tasks, classical weakly-supervised workflows outperformed MIL-based weakly-supervised methods for mutation prediction, which is surprising given their simplicity. This shows that new end-to-end image analysis pipelines in computational pathology should be compared to classical weakly-supervised methods. Also, these findings motivate the development of new methods which combine the elegant assumptions of MIL with the empirically observed higher performance of classical weakly-supervised approaches. We make all source codes publicly available at https://github.com/KatherLab/HIA, allowing easy application of all methods to any similar task.