Constructing a Predictive Model for PD-1 Blockade Therapy in Pan-Cancer Based on Machine Learning

Programmed cell death protein-1 (PD-1) blockade therapy have shown significant efficacy in cancer immunotherapy. However, low response rates and individual variability remain challenges. Currently, a universal biomarker to assess immunotherapy efficacy across various cancer types is lacking. In this study, single-cell RNA sequencing was applied to samples from seven cancer types, alongside bulk RNA-seq data from eight additional cancer types. LASSO regression and 15 machine learning algorithms were employed to construct 152 predictive models for immunotherapy efficacy. The results indicated that CD8 + effector T cells (CD8_Teff) in responders exhibited high infiltration and an activated, exhaustion-like phenotype. A predictive model based on seven effector T cell immunotherapy response genes (ETIRGS) effectively distinguished between responders and non-responders. The high-predicted scoring group exhibited significantly higher infiltration of CD8 + T cells and M1 macrophages than the low-predicted scoring group, along with elevated stromal and immune scores. Macrophages in responders acquired a pro-inflammatory phenotype upon activation by CD8_Teff cells, thereby enhancing the immune response. This study provides potential cross-cancer predictive biomarkers for PD-1 blockade therapy.