A Database of Gene Expression Profiles of Korean Cancer Genome

Collection of datasets

We have collected and generated genomics data obtained from 1,304 Korean cancer patients collaborating with a number of hospitals in Korea, such as Seoul National University Hospital (liver), Chungbuk National University Hospital (bladder), Chonbuk National University Hospital (liver, bile duct, and colon), Yonsei University Severance Hospital (breast), Korea University Medical Center (stomach), and Kosin University Gospel Hospital (stomach). The data were partially deposited to and freely available from the Gene Expression Omnibus of National Center for Biotechnology Information. All datasets stored in the database were normalized using quantile normalization. Detailed repository status was illustrated in Supplementary Table 1.

Implementations

The system architecture consists of various software frameworks for robust activity. Our system was mainly implemented with JAVA-based environment. To provide user friendly and active interfaces, the ICEfaces (version 3.3.0, http://www.icesoft.org/) framework was used. To store and handle the datasets, the MySQL database management system was used (version 5.5.11, http://dev.mysql.com). Data queries on MySQL from JAVA are controlled by MyBatis, an XML-based SQL mapping framework (version 3.1.1, https://code.google.com/p/mybatis). All statistical analysis methods were implemented using R (version 3.0.1, http://www.r-project.org) with Bioconductor plugins (version 2.12, http://www.bioconductor.org). Calling R modules from JAVA is managed by the RCaller framework (version 2.1.1, https://code.google.com/p/rcaller). All services are hosted on an Apache Tomcat web server (version 6.0.26, http://tomcat.apache.org). A schematic diagram of the system architecture is shown in Supplementary Fig. 1.

Supported analysis methods

Our system currently contains statistical survival analysis methods for identifying a signature associated with cancer outcome and estimating its predictive value, i.e., the Cox proportional hazard model, log-rank test and Kaplan-Meier curves. Detailed methodologies are available in the previous methodology paper [11]. In addition, the system determines the significance of a molecule using bar plots illustrating the landscape of intensities among patients and two or more group box plots with a p-value obtained from a two sample t test or ANOVA methods.

Web-based bioinformatics tool

A platform to simply search and estimate statistical significance of molecules across various cancer types accessible to investigators without bioinformatics or statistics expertise are available at the KCGD website (Fig. 1). It supports dataset search, in which a user explores cancer genomics or epigenomics datasets in Korean stored in the database, and gene search, which determines statistical difference between cancer known subtypes or prognostic value of a molecule in which users are interested.

Fig. 1

Snapshot of the Korean Cancer Genome Database (KCGD) webpage. The system supports dataset search, in which a user explores cancer genomics or epigenomics datasets in Korean stored in the database (right upper panel), and gene search, which determines statistical difference between cancer known subtypes or prognostic value of a molecule (right lower panel).

When accessing the gene search module, a number of accession identifiers including gene symbol, RefSeq (NM_^* or XM_^*), Entrez and protein accession number (NP_^* or XP_^*) can be selected, then the user input keyword(s) for search and click the search button. Although the search criteria were arranged using small number of identifiers, most known molecule identifiers were incorporated in the database by full text indexing method, indicating any molecule identifiers were allowed to search by clicking "anything..." category. Clicking the search button, the measurement data of expression or methylation are sought by the keyword and matched list with molecule information and thumb nail image of measurement intensities across cancers are displayed. Next, by clicking a thumb nail image, the system try to estimate the significance of a molecule using the log-rank test with Kaplan-Meier curves and display bar plots to illustrate the landscape of intensities among patients and box plots with a p-value obtained from a two sample t test or ANOVA test (Fig. 1). With p-values by log-rank test or significance test with t test or ANOVA, the prognostic value of a molecule or its difference between cancer subtypes may be determined, respectively. The entire analysis procedures of the system were described in Supplementary Fig. 2.

Statistics of datasets

Currently, we have created a database containing a total of 1,403 patient samples, among which 1,304 samples are Korean and the remained 99 samples are non-Korean patients for comparison. Most data in the current database were created by gene expression profiling (n = 1,283, 91%) and the remained (n = 120, 9%) were by methylation profiling methods. The vast majority of cancer tissues (n = 1,363, 97%) consists of liver (n = 431, 31%), breast (n = 365, 26%), stomach (n = 307, 22%), and bladder (n = 260, 19%), among which highly ranked three cancer types (i.e., liver, breast, and stomach) were known to be most frequently occurred in Korean. In each cancer type, we also designated previously known sample subtypes to support assessment of difference of a molecule between cancer subtypes. In addition, survival data of liver (n = 99) and bladder cancer patients (n = 165) were involved in the database to instantly estimate the prognostic value of a molecule in which users are interested (n = 264, 19%). Supplementary Table 1 elucidated detailed statistics in KCGD database.