Deep learning-based CD8⁺ T cell model for predicting prognosis and‌ targeted immunotherapy benefit in ccRCC

Deep learning holds significant promise for improving clear cell renal cell carcinoma (ccRCC) diagnosis and treatment. However, studies concerning tumor microenvironment (TME) based on pathological images and artificial intelligence are very limited. In this multi-center cohort study, we developed a deep learning-based CD8⁺ T cell inflammation signature (DL-CD8T) for ccRCC from hematoxylin and eosin (H&E)-stained whole slide images through a clustering-constrained attention multiple-instance learning method. We developed the DL-CD8T for ccRCC in the training cohort, which could distinguish patients with higher level of CD8⁺ T cell inflammation through conventionally used H&E-stained whole slide image, with area under the curve (AUC) of 0.781 (0.741-0.817, p < 0.001). Verification of the DL-CD8T was also performed well in the independent CPTAC cohort, with AUC of 0.741 (0.599-0.853, p < 0.001). Further verification indicated that DL-CD8T(+) was significantly associated with high infiltration abundance of CD8⁺ T cell, higher expression levels of PD-1, PD-2, PD-L1, CTLA-4, TIM-3 and higher tumor mutation burden, which could distinguish ccRCC patients with high survival risk. In addition, DL-CD8T might predict combined targeted and immunotherapy benefit from digital pathology in ccRCC with hazard ratio of 0.27 (p = 0.029). Pending prospective validation, DL-CD8T might be used to improve risk stratification and inform personalized targeted therapy.