Prediction of unfavourable response to checkpoint blockade in lung cancer patients through an integrated tumour-immune expression score

Abstract

Background: Treatment by immune checkpoint blockade (ICB) provides a remarkable survival benefit for multiple cancer types. However, disease aggravation occurs in a proportion of patients after the first couple of treatment cycles. Methods: RNA sequencing data was retrospectively collected. 6 tumour-immune related features were extracted and combined to build a lung cancer-specific predictive model to distinguish responses as progression disease (PD) or non-PD. This model was trained by 3 public pan-cancer datasets and a lung cancer cohort from our institute, and generated a lung cancer-specific integrated gene expression score, which we call LITES. It was finally tested in another lung cancer dataset. Results: LITES is a promising predictor for checkpoint blockade (area under the curve [AUC]=0.86), superior to traditional biomarkers. It is independent of PD-L1 expression and tumour mutation burden. The sensitivity and specificity of LITES was 85.7% and 70.6%, respectively. Progression free survival (PFS) was longer in high-score group than in low-score group (median PFS: 6.0 vs. 2.4 months, hazard ratio=0.45, P=0.01). The mean AUC of 6 features was 0.70 (range=0.61-0.75), lower than in LITES, indicating that the combination of features had synergistic effects. Among the genes identified in the features, patients with high expression of NRAS and PDPK1 tended to have a PD response (P=0.001 and 0.01, respectively). Our model also functioned well for patients with advanced melanoma and was specific for ICB therapy. Conclusions: LITES is a promising biomarker for predicting an impaired response in lung cancer patients and for clarifying the biological mechanism underlying ICB therapy.

Keywords

• Lung cancer
• Biomarker
• Checkpoint blockade
• Machine learning
• RNA sequencing