Artificial intelligence (AI) analysis of tumor-infiltrating lymphocytes (TILs) in hematoxylin and eosin (H&E) slides to explore immune phenotypes in papillary thyroid cancer

6100 Background: Different biomarkers have been discovered to predict response to immunotherapy but the variability of response exists. TILs are a potential biomarker as infiltration of CD8+ T cells in the tumor microenvironment(TME) was related to treatment response to PD-1 blockade. Advanced well-differentiated papillary thyroid cancer(PTC) has limited standard-of-care treatment options and can exhibit a poor prognosis when exhausted. Our study aimed to classify the TME of PTC according to TILs. Methods: Using the TCGA PTC data, samples were classified into three immune phenotypes (immune-’desert’, immune-’excluded’, and ‘inflamed’) based on peri-tumor lymphocyte density from analysis of H&E slides using an AI model(Lunit SCOPE). Immune and stromal cell infiltration was inferred using xCell. Cytolytic scores(CYT) and thyroid differentiation scores(TDS) were calculated using mRNA expression. Results: Total of 383 samples were included. Immune phenotype analysis showed 183(47.7%) ‘desert’, 133(34.7%) ‘excluded’, and 67(17.4%) ‘inflamed’. Median age at diagnosis was the lowest in ‘inflamed’(inflamed 41.5 vs excluded 47 vs desert 50, p = 002). TNM stages were as follows in the order of stages 1 to 4: inflamed 49/4/8/3, excluded 64/13/32/16, and desert 86/28/40/19. BRAF V600E mutation was most prevalent in ‘excluded’(97/133, 72%), followed by ‘inflamed’(32/67, 47.7%) and ‘desert’(51/183, 27.8%). CYT was the highest in ‘inflamed’(median: inflamed 186 vs excluded 53 vs desert 39, p < 0.001). TDS was the highest in ‘desert’(median: inflamed -0.3 vs excluded -0.69 vs desert 0.73, p < 0.001). The 6 cell types with the most statistically significant difference in infiltration were activated dendritic cells(aDCs), DCs, B cells, epithelial cells(EpCs), CD8+ central memory T cells, and CD4+ memory T cells. The ‘inflamed’ showed the highest median value for all of these cells except EpCs. The cells with the highest median value in ‘excluded’ included EpCs, sebocytes, and immature DCs. The cells with the highest median value in ‘desert’ were lymphatic endothelial cells(ECs), ECs, and osteoblasts. Three immune phenotypes were associated with differential outcomes in disease-free survival(log-rank for trend, p-value = 0.01). There was no difference in overall survival. Conclusions: Our results suggest that 15% of PTC can have immune features which were related to a higher rate of response to immunotherapy across different tumor types. A previous study showed that PTC with a low TDS or BRAF V600E mutation is related to increased immune response indices. In our study, TDS or BRAF V600E mutation status could not distinguish the ‘inflamed’ from other phenotypes. This suggests that immune phenotype classification through TILs can provide distinct information about the TME of PTC.