An EGFR signature predicts cell line and patient sensitivity to multiple tyrosine kinase inhibitors

Abstract

EGFR is an oncogene with a high frequency of activating mutations in nonsmall cell lung cancer (NSCLC). EGFR inhibitors have been FDA-approved for NSCLC and have shown efficacy in patients with certain EGFR mutations. However, only 9% to 26% of these patients achieve objective responses. In our study, we developed an EGFR gene signature based on The Cancer Genome Atlas (TCGA) RNA-seq data of lung adenocarcinoma (LUAD) to direct the preselection of patients for more effective EGFR-targeted therapy. This signature infers baseline EGFR signaling pathway activity (denoted as EGFR score) in tumor samples, which is associated with tumor sensitivity to EGFR inhibitors and other tyrosine kinase inhibitors (TKIs). EGFR score predicted sensitivity of lung cancer cell lines to Erlotinib, Gefitinib and Sorafenib. Importantly, EGFR score calculated from pretreated samples was associated with patient response to Gefitinib and Sorafenib in lung cancer. Additionally, integration of the EGFR signature with TCGA LUAD data showed that it accurately predicted functional effects of different somatic EGFR mutations, and identified other mutations affecting EGFR pathway activity. Finally, using cancer cell line and clinical trial data, the EGFR score was associated with patient response to TKIs in liver cancer and other cancer types. The EGFR signature provides a useful biomarker that can expand the application of EGFR inhibitors or other TKIs and improve their treatment efficacy through patient stratification.

Keywords: EGFR; EGFR-targeted therapy; biomarker; tyrosine kinase inhibitor.

FIGURE 1

EGFR score predicts EGFR mutation status in lung tumor and cell lines. A, Boxplots depicting EGFR score differences between EGFR-mutant and wild-type lung cancer patients. B, Boxplots depicting EGFR score differences between EGFR-mutant and wild-type lung cancer cell lines. C, Receiver Operating Characteristic (ROC) curves for EGFR mutation status prediction in lung cancer patients using EGFR score as the predictor. D, Receiver Operating Characteristic (ROC) curves for EGFR mutation status prediction in lung cancer cell lines using EGFR score as the predictor. In all boxplots, P values were calculated by Wilcoxon rank-sum test. EGFR mutation status was based on the original publications of the corresponding dataset

FIGURE 2

EGFR score predicts sensitivity of lung cancer cell lines to Erlotinib and other tyrosine kinase inhibitors. A-B, Scatterplot presenting the correlation between EGFR scores and phosphorylated EGFR protein (activating phosphorylation) abundance in TCGA lung adenocarcinoma tumor samples (A) and CCLE lung cancer cell lines (B). C-D, Scatterplot presenting the correlation between EGFR scores and drug sensitivity of Erlotinib (C) and Gefitinib (D) in lung cancer cell lines. E, Scatterplot presenting the correlation between EGFR scores and drug sensitivity of Erlotinib in lung cancer cell lines. F-G, Boxplots depicting Erlotinib sensitivity differences between EGFR wild-type lung cancer cell lines with high (EGFR-Hi) and low EGFR scores (EGFR-Lo) in GSE32989 (F) and CCLE (G). Median EGFR score was used as the cutoff for dichotomizing wild-type EGFR lung cancer cell lines into EGFR-Hi and EGFR-Lo groups. H-J, Scatterplot presenting the correlation between EGFR score and drug sensitivity to ZD-6474 (H), Lapatinib (I) and Sorafenib (J) in lung cancer cell lines. K-L, Scatterplot presenting the correlation between EGFR scores and phosphorylated Her2 protein (activating phosphorylation) abundance (K) and phosphorylated B-Raf (activating phosphorylation) protein abundance (L) in lung cancer cell lines. In all scatterplots, ρ indicates Spearman correlation coefficient. In all boxplots, P values were calculated by Wilcoxon rank-sum test. Drug sensitivity is represented as Drug Activity Area (DAA) in CCLE and IC50 in GSE32989, with higher values of DAA and lower values of IC50 indicating higher sensitivity

FIGURE 3

EGFR score predicts lung cancer patient response to Gefitinib and Sorafenib. A, Kaplan-Meier plot depicting the post-operative survival probability over 5 years for samples with high (red) and low (blue) EGFR scores in patients treated with Gefitinib. B, Forest plot showing the hazard ratios (HR) and P values of EGFR score and several clinical variables estimated by a multivariate Cox regression model. C, Boxplot depicting EGFR score difference between responders and nonresponders in lung cancer patients treated with Sorafenib. All patients are EGFR wild-type. The P value was calculated by Wilcoxon rank-sum test. D, Receiver Operating Characteristic (ROC) curves for Sorafenib response prediction using EGFR score as the predictor. E, Kaplan-Meier plot depicting progression free survival probability over 5 years for samples with high (red) and low (blue) EGFR scores in patients treated with Sorafenib. F, Forest plot showing hazard ratios (HR) and P values of EGFR score and several clinical variables estimated by a multivariate Cox regression model. In all Kaplan-Meier plots, median EGFR score was used as the cutoff to dichotomize patients into EGFR-Hi and EGFR-Lo groups. Numbers in parenthesis indicate the number of patients in each group. Hazard ratio (HR) is calculated from univariate Cox regression model. P value between survival curves was calculated by log-rank tests. In all forest plots, HR was presented as the 95% confidence interval, the dotted lines indicate the null association and the Wald’s test was used to determine statistical significance

FIGURE 4

Various mechanisms for deregulating the EGFR pathway. A, EGFR scores of different EGFR somatic mutations in TCGA. The dashed line indicates the average EGFR score in EGFR wild-type patients. Synonyms mutations were labeled blue. B-C, Boxplots depicting EGFR score differences in EGFR amplified and nonamplified patients. Patients are all EGFR wild-type in C. The P value was calculated by Wilcoxon rank-sum test. D, Scatterplot presenting the correlation between EGFR scores and the average methylation level of CpG sites at the EGFR gene promoter region. E, Scatterplot presenting the correlation between EGFR mRNA expression and the average methylation level of CpG sites at the EGFR gene promoter region. F, Scatterplot presenting the correlation between EGFR mRNA expression and the EGFR score. G, Association of EGFR scores with LUAD clinical features and genomic features. EGFR score was ranked from high to low and associated with gender, smoking status, stage, EGFR mutation status, EGFR amplification status and gene somatic mutation status. P values were calculated by Wilcoxon rank-sum test. In all scatterplots, ρ was calculated by Spearman correlation. EGFR-mutant samples and wild-type samples are labeled by red and blue dots. In all boxplots, P values were calculated by Wilcoxon rank-sum tests

FIGURE 5

The EGFR signature is predictive of patient response to Sorafenib in liver cancer. A, Boxplot depicting EGFR score differences between Sorafenib responders and nonresponders in liver cancer. The P value was calculated by Wilcoxon rank-sum test. B, Receiver Operating Characteristic (ROC) curves for Sorafenib response prediction in liver cancer using EGFR score as predictor. C, Scatterplot presenting the correlation between EGFR scores and Sorafenib sensitivity in liver cancer cell lines. D, Barplot illustrating the correlation between EGFR scores and Erlotinib sensitivity in cancer cell lines across 11 cancer types. ρ indicates Spearman correlation coefficient. The significant correlation was labeled by a yellow bar. E, Scatterplot presenting the correlation between EGFR scores and Erlotinib sensitivity in large intestine cancer cell lines. In all scatterplots, ρ indicates Spearman correlation coefficient. Drug sensitivity is represented as Drug Activity Area (DAA) in CCLE