PURPOSE Biomarkers on the basis of tumor-infiltrating lymphocytes (TIL) are potentially valuable in predicting the effectiveness of immune checkpoint inhibitors (ICI). However, clinical application remains challenging because of methodologic limitations and laborious process involved in spatial analysis of TIL distribution in whole-slide images (WSI). METHODS We have developed an artificial intelligence (AI)–powered WSI analyzer of TIL in the tumor microenvironment that can define three immune phenotypes (IPs): inflamed, immune-excluded, and immune-desert. These IPs were correlated with tumor response to ICI and survival in two independent cohorts of patients with advanced non–small-cell lung cancer (NSCLC). RESULTS Inflamed IP correlated with enrichment in local immune cytolytic activity, higher response rate, and prolonged progression-free survival compared with patients with immune-excluded or immune-desert phenotypes. At the WSI level, there was significant positive correlation between tumor proportion score (TPS) as determined by the AI model and control TPS analyzed by pathologists ( P < .001). Overall, 44.0% of tumors were inflamed, 37.1% were immune-excluded, and 18.9% were immune-desert. Incidence of inflamed IP in patients with programmed death ligand-1 TPS at < 1%, 1%-49%, and ≥ 50% was 31.7%, 42.5%, and 56.8%, respectively. Median progression-free survival and overall survival were, respectively, 4.1 months and 24.8 months with inflamed IP, 2.2 months and 14.0 months with immune-excluded IP, and 2.4 months and 10.6 months with immune-desert IP. CONCLUSION The AI-powered spatial analysis of TIL correlated with tumor response and progression-free survival of ICI in advanced NSCLC. This is potentially a supplementary biomarker to TPS as determined by a pathologist.

Objective The need to perform genetic sequencing to diagnose the polymerase epsilon exonuclease ( POLE ) subtype of endometrial cancer (EC) hinders the adoption of molecular classification. We investigated clinicopathologic and protein markers that distinguish the POLE from the copy number (CN)-low subtype in EC. Methods Ninety-one samples (15 POLE , 76 CN-low) were selected from The Cancer Genome Atlas EC dataset. Clinicopathologic and normalized reverse phase protein array expression data were analyzed for associations with the subtypes. A logistic model including selected markers was constructed by stepwise selection using area under the curve (AUC) from 5-fold cross-validation (CV). The selected markers were validated using immunohistochemistry (IHC) in a separate cohort. Results Body mass index (BMI) and tumor grade were significantly associated with the POLE subtype. With BMI and tumor grade as covariates, 5 proteins were associated with the EC subtypes. The stepwise selection method identified BMI, cyclin B1, caspase 8, and X-box binding protein 1 (XBP1) as markers distinguishing the POLE from the CN-low subtype. The mean of CV AUC, sensitivity, specificity, and balanced accuracy of the selected model were 0.97, 0.91, 0.87, and 0.89, respectively. IHC validation showed that cyclin B1 expression was significantly higher in the POLE than in the CN-low subtype and receiver operating characteristic curve of cyclin B1 expression in IHC revealed AUC of 0.683. Conclusion BMI and expression of cyclin B1, caspase 8, and XBP1 are candidate markers distinguishing the POLE from the CN-low subtype. Cyclin B1 IHC may replace POLE sequencing in molecular classification of EC.

Background: Activating mutations in the tyrosine kinase domain of epidermal growth factor receptor (EGFR) are predictive biomarkers for response to EGFR–tyrosine kinase inhibitor (TKI) therapy in lung adenocarcinoma (LUAD). Here, we characterized the clinicopathologic features associated with EGFR mutations via peptide nucleic acid clamping-assisted fluorescence melting curve analysis (PANAMutyper) and evaluated the feasibility of targeted deep sequencing for detecting the mutations.Methods: We examined EGFR mutations in exons 18 through 21 for 2,088 LUADs from July 2017 to April 2020 using PANAMutyper. Of these, we performed targeted deep sequencing in 73 patients and evaluated EGFR-mutation status and TKI clinical response.Results: EGFR mutation was identified in 55.7% of LUADs by PANAMutyper, with mutation rates higher in females (69.3%) and never smokers (67.1%) and highest in the age range of 50 to 59 years (64.9%). For the 73 patients evaluated using both methods, next-generation sequencing (NGS) identified EGFR mutation–positive results in 14 of 61 patients (23.0%) who were EGFR-negative according to PANAMutyper testing. Of the 10 patients reportedly harboring a sensitizing mutation according to NGS, seven received TKI treatment, with all showing partial response or stable disease. In the 12 PANAMutyper-positive cases, NGS identified two additional mutations in exon 18, whereas a discordant negative result was observed in two cases.Conclusions: Although PANAMutyper identified high frequencies of EGFR mutations, targeted deep sequencing revealed additional uncommon EGFR mutations. These findings suggested that appropriate use of NGS may benefit LUAD patients with otherwise negative screening test results.

2585 Background: TGF-beta activates fibroblasts within the TME and may be associated with poor response to immunotherapy.However, objective and reliable assessment of the TGFBs by gene expression profiling (GEP) is challenging. Using H&E whole-slide images (WSI) from The Cancer Genome Atlas (TCGA), we developed an artificial intelligence (AI) model to predict a normalized TGFBs, applying it to real-world datasets (RWDs) of advanced solid tumor patients treated with immunotherapy. Methods: The TGFBs was defined using the mean GEP values of the 'Hallmark TGF-beta signaling' gene set. H&E WSIs of 23 carcinoma types from TCGA (n=6,945) were used to discover the association of the TGFBs with various cell types, including fibroblasts and tumor-infiltrating lymphocytes (TILs) detected by an AI model, Lunit SCOPE IO (AI-1). Another custom AI model (AI-2) to predict TGFBs was developed by integrating self-supervised and supervised features and other semantic content extracted by AI-1 with multilayered perceptron classifiers. TGFBs prediction scores from the AI-2 were applied to 5 different RWDs containing 1,792 immunotherapy-treated patients (> 16 primary tumor types). The immune-excluded phenotype (IEP) was defined by high stromal TIL and low intratumoral TIL density. Results: In the TCGA dataset, TGFBs and fibroblast density detected by AI-1 were closely correlated (R=0.19, p<0.001) compared to other cell types, with both highest in pancreatic adenocarcinoma (PAC)(median [interquartile range, IQR] 0.73 [0.66-0.80], 1306 [985-1588]/mm2, respectively) and CMS4 colorectal cancer (0.61 [0.54-0.67], 868 [582-1128]/mm2) compared to other cancers (0.57 [0.47-0.66], 577 [295-890]/mm2). The AI-2 predicted TGFBs with area-under-the-curve of 0.781 on cross-validation within the TCGA dataset. The TGFBs of immunotherapy-treated RWDs were inferred from their corresponding H&E WSI, with scores comparable to those from the TCGA (0.42 [0.16-0.60] vs 0.58 [0.48-0.67]), with PAC having the highest TGFBs (0.65 [0.45-0.77]) in the RWD as well. Interestingly, TGFBs inferred scores and fibroblast density were significantly correlated (R=0.7, p<0.001) in the RWD and also significantly correlated with the IEP (0.52 [0.30-0.66]) compared to other immune phenotypes (0.42 [0.16-0.60], p<0.001). In the RWD, patients with inferred TGFBs in the upper 75% quartile had a poor clinical response to immunotherapy (objective response rate: 15.1% vs 20.7%, median progression-free survival 3.0 m vs 4.0 m, hazard ratio 1.21, 95% confidence interval 1.07-1.36, p=0.002). Conclusions: TGFBs is closely correlated with fibroblast density within the TME and may confer resistance to immunotherapy. Unique morphologic features of the TGFBs can be captured by AI, enabling scalable application to various datasets for clinical and translational research.