Combining hematoxylin and eosin (H&E) stained images and RNA sequencing (RNA-Seq) data to predict overall survival (OS) in patients with non-small cell lung cancer (NSCLC).

Abstract

1547 Background: IMpower150 was a phase 3 clinical trial that evaluated the efficacy of atezolizumab in patients with metastatic nonsquamous NSCLC; it demonstrated no significant OS benefit in the ACP (atezolizumab+carboplatin+paclitaxel) arm vs the BCP (bevacizumab+carboplatin+paclitaxel) control arm (hazard ratio [HR]=0.85; 95%CI, 0.71-1.03). The objective of this analysis was to identify a subpopulation of patients that benefits from ACP using H&E stained images and RNA-Seq data. Methods: Spatial statistics algorithms were applied to the coordinates of tumor cells and lymphocytes of the H&E stained images to capture spatial heterogeneity of the tumor microenvironment. The normalized and log-transformed RNA-Seq data underwent a nested feature selection procedure using a Cox proportional hazard model with L1 regularization and stability selection. Cutoffs for gene selection were determined using a permutation strategy with a false discovery rate <0.001. To investigate the association between the 41 derived spatial features, significant genes and OS, a Cox proportional hazard model with L2 regularization was fitted only for the ACP group. Survival groups were further identified using nested Monte Carlo Cross Validation to prevent over-fitting. Results: A total of 236 ACP and 235 BCP patients who had both H&E stained images and RNA-Seq data were analyzed. In the predicted long survival group, ACP patients had significantly longer median OS vs BCP patients based on H&E stained images (HR=0.61; 95%CI, 0.41-0.90; P=0.013) and RNA-Seq data (HR=0.64; 95%CI, 0.41-0.99; P=0.042). The combination of both modalities further improved the OS benefit between the arms (HR=0.44; 95%CI, 0.27-0.73; P=0.001). Data-driven selection of genes relevant for the prediction of OS included MAML3, AC024475.4, RGPD1, LCE3D and AC004156.1. Conclusions: Our approach was able to stratify a subpopulation of patients that significantly benefited from ACP compared with BCP treatment, particularly when integrating both H&E stained images and RNA-Seq data, which demonstrated the complementary value of both modalities. Our results could inform the development of a companion diagnostic that predicts individualized treatment response. Clinical trial information: NCT02366143.