Machine Learning Identification of Immunotherapy Targets in Low-Grade Glioma Using RNA Sequencing Expression Data

Abstract

Objective: Immunotherapy has revolutionized cancer treatment in the past decade, but significant hurdles remain. Human studies with immune checkpoint inhibitors targeting programmed cell death protein have demonstrated suboptimal efficacy in the setting of low-grade gliomas (LGGs). Identification of mechanisms leading to inadequate anti-tumor immunity is paramount. The current study evaluates and validates barriers to immunotherapy using a novel machine learning algorithm.

Methods: We utilized The Cancer Genome Atlas (TCGA) to generate expression levels of 28 immune genes related to known immunotherapeutic targets or lymphocyte cytolytic activity. We created training and testing groups and 3 machine learning models to determine the genes most highly correlated to cytolytic activity (CYT). The 3 models were run through multiple regression by exhaustive selection, LASSO, and random forest. We validated computational results by comparing expression of pertinent genes in patient-derived glioma samples.

Results: Our models demonstrated linearity, a low mean-squared error, and consistent results with respect to the most important variables. Expression of ICOS, IDO1, and CD40 were the most important variables in all models and demonstrated positive correlation with CYT. Other variables included TIGIT and CD137. Genetic analysis from 3 IDH-mutants (IDHm) and 3 IDH-wild type (IDHwt) patient-derived glioma samples validated TCGA data and demonstrated lower levels of CYT in IDHm gliomas compared with IDHwt.

Conclusions: This novel methodology has elucidated 3 potential targets for immunotherapy development in LGGs. We also demonstrated a novel method of analyzing data using advanced statistical techniques that can be further used in developing treatments for other diseases as well.

Keywords: Drug design; Exhaustive selection; Immunotherapy; LASSO; Low-grade glioma; Random forest.