# Predicting Immunotherapy Outcomes in NSCLC Using RNA and Pathology from Multicenter Clinical Trials

## TL;DR

A new machine learning model called LIRA predicts immunotherapy outcomes in lung cancer patients better than existing methods.

## Contribution

LIRA is a novel RNA-based model that improves immunotherapy outcome prediction and identifies resistance mechanisms in NSCLC.

## Key findings

- LIRA outperforms PD-L1 and tumor mutation burden in predicting immunotherapy responses.
- High LIRA scores correlate with increased T cells and reduced epithelial cells in tumors.
- LIRA enables independent risk stratification and identifies early progression risk in immunotherapy.

## Abstract

Immune checkpoint inhibitors (ICIs) are widely used to treat advanced non‐small cell lung cancer (NSCLC). However, it remains crucial to identify patients who are unlikely to benefit from immunotherapy and to explore potential combination treatment strategies. In this study, 1127 advanced NSCLC patients from multicenter randomized clinical trials (OAK, POPLAR, ORIENT‐11) and an in‐house cohort who received ICIs, ICIs combined with chemotherapy, or chemotherapy alone are analyzed. Using bulk RNA‐seq transcriptomic data, an RNA‐based model, named the Lung Cancer Immunotherapy Response Assessment (LIRA), is developed, utilizing interaction analysis and a random forest algorithm to predict immunotherapy outcomes. LIRA outperforms PD‐L1 expression and tumor mutation burden in predicting responses, particularly in identifying early progression risk during ICI monotherapy (HR: 0.15, 95% CI: 0.11–0.20). Tumor profile analysis reveals that LRP8 and HDAC4 are associated with immunotherapy outcomes. Additionally, scRNA‐seq analysis of NSCLC tumors indicates a higher prevalence of T cells and a reduced proportion of epithelial cells in samples with a high LIRA‐score. The deep learning model pinpointed critical high‐attention regions within whole‐slide images that contributed decisively to the LIRA predictions. In summary, these results demonstrate that LIRA enables independent risk stratification of NSCLC patients and provides insights into potential resistance mechanisms.

LIRA, a machine learning‐based model, is developed using transcriptomic data from 891 NSCLC patients in the OAK and POPLAR cohorts. Its predictive performance is validated in multiple external cohorts. Patients stratified by LIRA‐score exhibit distinct clinical characteristics and tumor microenvironment profiles. Additionally, a deep learning model enables LIRA‐score prediction directly from whole‐slide pathology images.