Machine Learning and Mendelian Randomization Reveal a Tumor Immune Cell Profile for Predicting Bladder Cancer Risk and Immunotherapy Outcomes.

Abstract

The objective of this study is to develop predictive models for bladder cancer (BLCA) using tumor infiltrated immune cell (TIIC)-related genes. Multiple RNA expression data and scRNA-seq were downloaded from TCGA and GEO databases. We calculated tissue specificity index (TSI) and developed a computational framework to identify TIIC signature score based on 3 algorithms. Univariate Cox analysis was performed and TIIC-related model was generated by 20 machine learning algorithms. A significant correlation between TIIC signature score and survival status, tumor stage, and TNM staging system was found. Patients with BLCA in high-score group had more favorable survival outcomes and enhanced response to PD-L1 immunotherapy. Our TIIC model shows better performance in predicting prognosis of BLCA. Diverse frequencies of mutations were observed in human chromosomes across groups categorized by TIIC score. There was no statistically significant correlation observed between non-cancerous bladder conditions and bladder cancer when examining the SNPs associated with the genes in the prognostic model. However, a statistically significant association was found at the SNP sites of rs3763840. There was no significant association between bladder stone and bladder cancer, but there was a significant association on the SNP sites of rs3763840. In conclusion, we constructed a novel TIIC signature score for the prognosis and immunotherapy for BLCA, which offers direction for predicting the OS of patients with BLCA.