LCE: an open web portal to explore gene expression and clinical associations in lung cancer

Abstract

We constructed a lung cancer-specific database housing expression data and clinical data from over 6700 patients in 56 studies. Expression data from 23 genome-wide platforms were carefully processed and quality controlled, whereas clinical data were standardized and rigorously curated. Empowered by this lung cancer database, we created an open access web resource-the Lung Cancer Explorer (LCE), which enables researchers and clinicians to explore these data and perform analyses. Users can perform meta-analyses on LCE to gain a quick overview of the results on tumor vs non-malignant tissue (normal) differential gene expression and expression-survival association. Individual dataset-based survival analysis, comparative analysis, and correlation analysis are also provided with flexible options to allow for customized analyses from the user.

Fig. 1

Summary of lung cancer database variable distribution. This summary describes the datasets and features of the lung cancer database that feeds into the Lung Cancer Explorer. Gene expression data and clinical data were collected from 56 studies that include over 6700 patients. For each study and each variable, a pie chart is used to summarize the data. The color scheme for the pie chart sectors are provided below the gridded pie charts. Table S2 provides the specific sample sizes under each category

Fig. 2

Histology classification of samples collected in the lung cancer database. This tree diagram represents the hierarchical structure of the 2015 WHO classification system of lung tumors. Numbers on the red nodes denote the number of samples from the lung cancer database belonging to the corresponding histology type

Fig. 3

Examples of survival analysis with more significant results when cluster-based cutoff is used. a Bi-modal distribution of expression in Shedden_2008 dataset. The solid blue line marks the cutoff at the median, whereas the solid red line marks the cutoff determined by Gaussian mixture model. b Kaplan–Meier curves from the survival analysis of Shedden_2008 using groups defined by SMARCA4 gene expression with cutoff at median. P-value from the log-rank test is denoted at the bottom left corner of the plot. c Survival analysis of Shedden_2008 using groups defined by Gaussian mixture model of SMARCA4 expression. d Bi-modal distribution of KYNU expression in Schabath_2016 dataset. e Survival analysis of Schabath_2016 using groups defined by SMARCA4 gene expression with cutoff at median. f Survival analysis of Schabath_2016 using groups defined by Gaussian mixture model of KYNU expression

Fig. 4

High klotho expression has more significant association with positive survival outcome in males. For each of the six selected studies, survival analysis assessing prognosis association of KL gene expression was performed for male patients or female patients only. In each analysis, the median was used as a cutoff for dichotomizing patients. In all six studies, a more significant association with better prognosis was found in the male patients compared to the female patients

Fig. 5

Examples of different meta-analysis results in lung adenocarcinoma vs squamous cell carcinoma. a, b RORC tumor vs normal meta-analyses in lung ADC studies (a) and lung SCC studies (b). c, d CDCA2 survival meta-analyses in lung ADC studies (a) and lung SCC studies. Note that differential gene expression meta-analysis for RORC is only significant in lung SCC patients, whereas survival meta-analysis for CDCA2 is only significant in lung ADC patients. In each forest plot, the name of each study is followed by the number of tumor and normal samples (tumor vs normal meta-analysis) or total tumor samples (survival meta-analysis). SMD standardized mean difference, TE estimated treatment effect, seTE standard error of treatment effect, HR hazard ratio, CI confidence interval

Fig. 6

Meta-analysis estimates agree with qPCR measurements on tumor vs normal expression differences for 46 nuclear hormone receptor genes. Results from qPCR measurements of 30 tumor-normal pairs (x-axis values) and meta-analysis estimates from 21 studies (y-axis values) on gene expression differences between tumor and normal tissues for 46 nuclear hormone receptor genes were used to evaluate consistency between the two approaches. The values on the x-axis and y-axis are the standardized mean difference estimated by Hedges’ G method. The solid purple line represents a linear regression line, whereas the dashed gray line identifies where x equals y

Fig. 7

Assessment of result consistency by I² statistics in meta-analysis of survival-gene expression association. a Density estimation of I² distribution. Three genes with different I² statistics were selected as examples in (b), (c), and (d). A larger I² value suggests a larger degree of heterogeneity across studies, whereas a smaller I² value is reflective of a higher degree of consistency among studies. b, c, d Example forest plots of survival meta-analysis with different heterogeneity: large (b), intermediate (c), and small (d)

Fig. 8

Interaction between sample tissue type and smoking status in HBD gene expression. a, d Boxplots comparing HBD gene expression between two groups dichotomized on a single clinical variable: tissue type (a) or smoking status (d). b, c, e, f Boxplots comparing HBD gene expression between two groups defined by a combination of two clinical variables: different tissues in smoker (b), different tissues in non-smoker (c), tumor from patients with different smoking status (e), and normal tissues from patients with different smoking status (f)

Fig. 9

Different co-expression pattern between PARP2 and cycle genes. a, b, c, d Heatmaps of gene–gene correlation matrices from TCGA_LUAD_2016 for PARP2 and 10 selected cell cycle genes from tumor sample expression data (a) or normal sample expression data (b), and for PARP2 and 10 selected C2H2-type zinc finger genes (ZNF) from tumor sample expression data (c) or normal sample expression data (d). The highly positive correlation between PARP2 and cell cycle genes was seen only in tumor samples but not normal samples (a, b), whereas the high degree of positive correlation between PARP2 and ZNF genes was observed only in normal tissue samples but not tumor samples (c, d)