A key molecular driver of tumor-infiltrating lymphocytes in invasive breast cancer on machine learning–based meta-mining research

Abstract Background: The immune system affects all phases of tumor growth, from initiation to progression and dissemination. However, molecular mechanisms of tumor immunity remain unclear. Therefore, this study aimed to identify targets significantly associated with tumor-infiltrating lymphocytes (TILs) in early-stage breast cancer (BC) using a novel machine learning (ML) approach. Method: We identified 719 patients with early-stage BC from The Cancer Genome Atlas datasets having the available digital hematoxylin and eosin-stained whole slide images (WSIs) and transcriptomic data from tumor sites. The grades of stromal TILs in WSIs were evaluated using the criteria of the International Working Group for TILs in BC: low, intermediate, and high. Using our own artificial neural network ML methods, key genes were identified based on the differential mRNA expression between stromal-TIL grades. Results: In this study, the ML system identified 49 genes that demonstrated significantly different expressions between stromal-TIL grades. Clustering analysis with this gene set further divided patients into two molecular subtypes (subtypes 1 and 2), which were significantly associated with tumor aggressiveness. The 10-year overall survival of subtype 1 was significantly poorer than that of subtype 2 (hazard ratio: 2.27, 95% confidence interval: 1.11-4.64, p = 0.025). We also found that these 49 genes are strongly associated with inducible T-cell co-stimulator (ICOS). Conclusion: our findings suggest that TIL-related gene sets (ICOS-related genes) could interpret the complex molecular gene networks controlling tumor immunity in early-stage BC.