SIGMA: A System for Integrative Genomic Microarray Analysis of Cancer Genomes

Cell-line collection

Main functionalities of SIGMA

In order to increase the utility of this collection, a significant component of SIGMA is the web-based application which allows for the user-friendly mining of this dataset. Four major types of functionalities are offered by SIGMA: (1) interrogation of a single sample, (2) visualization and analysis of a single group of samples, (3) comparative analysis of two groups of samples, and (4) integration of data from multiple platforms (Figure 1B).

Visualization and interrogation of a single genome

The first function we discuss is the ability to view a single high resolution array CGH profile at multiple magnification levels. The major utility of this function is to display the underlying genomic architecture of a cell line, so that genetic features can be considered in experimental interpretation. For example, a whole genome karyogram of lung adenocarcinoma cell line H2087 profiled on the SMRT array platform (Figure 2A). From this image, we see many changes such as the loss of the 3p arm as well as segmental changes in chromosomes 8, 19 and 20. Specifically, we can select chromosome 8 (Figure 2B) and view that separately, then zoom into the region of interest and visualize it in finer detail (Figure 2C). Users can then highlight or place boundary lines in this region and query for which genes are located within the region of interest.

Subsequently, using the interval search option, users can retrieve the genes which are located in a desired region and have the option to query commonly used biological databases such as NCBI MIM, NCBI Gene, NCBI PubMed and the UCSC Genome Browser. For example, if we look at band 8q24.21 (Figure 2D), we can highlight the region and search the interval for which genes it contains. When we invoke the region search and retrieve only genes curated by RefSeq, we see there are 8 genes in the amplicon. If the user selects a particular gene, options to link out to the above mentioned biological databases become available. The utility of this function is to facilitate a direct connection from experimental findings to known, relevant information. Moreover, the ability to interrogate for specific genes and regions can be done for any types of the analysis outlined (Figure 1B).

Multiple genome comparison and mining across tumor types

A common research question is to look across a series of samples with common phenotype to identify recurrent genetic changes, for example comparison of lung adenocarcinomas [see Additional file 2]. With the spectrum of samples warehoused in SIGMA, such a query can be performed across multiple cancer types. For example, the alignment of a set of samples representing 8 different cancer types revealed common amplification of the MYC oncogene locus (Figure 3A), while the epidermal growth factor receptor (EGFR) locus is amplified only in a subset of samples (Figure 3B).

Recurrent alterations detected in one group of samples can be compared against those in another group, for example, the comparison of lung squamous cell carcinoma (SqCC) with cervical SqCC [see Additional file 3]. The strategy for comparison, for example the overlay of frequency plots derived from two groups of profiles, has been described elsewhere [7, 8].

Simultaneous visualization of data from multiple platforms

Cross platform comparison is essential to the multi-dimensional descriptions of a genome. Here, we have included a feature in SIGMA to allow users to view multiple platforms of data simultaneously. We use the breast cancer cell line, MCF7, to illustrate this functionality. Data from four different array CGH platforms were available publicly: SMRT array, Stanford cDNA microarray, Affymetrix 10K SNP array and the Affymetrix 100K SNP array. Figure 4A illustrates the cross platform display of chromosome 17, while Figure 4B shows the variable density of coverage by these four commonly used platforms.

Integrative visualization of DNA copy number and methylation

One of the novel features we have provided is the integrative visualization of copy number alterations with DNA methylation status. The major premise in studying copy number alterations at the DNA level is that these are the primary changes involved in driving changes in gene expression. Though gene dosage variation may be responsible for expression changes, alteration in DNA methylation pattern also contributes significantly to regulating gene expression. Recently, methods for global methylation analysis to measure aberrant DNA methylation status across tumor genomes have been developed [9–12]. Wilson et al (2006) compared methylation patterns with copy number profiles in lung cancer cells. Utilizing genomic and epigenetic data from this study for the H1395 lung cancer cell line, we illustrate a parallel display in SIGMA. Figure 5 shows a large segmental copy number gain spanning 1q21.2 to 1q23.1 with corresponding hypomethylation, localized precisely to 1q21.3 [13]. Significantly, both copy number gain and decreased methylation can elevate gene expression. The S100 calcium binding protein A10 (S100A10) gene within this region has been previously shown to be over-expressed in lung cancer [14]. It is readily apparent of the value of integrative studies examining aberrant DNA methylation and genomic copy number. With increased prevalence of studies of whole genome methylation, this feature will be of greater importance.

Sources of raw array CGH data

The raw data for 267 CGH profiles in the database was obtained from a variety of sources. They include both published and unpublished data [7, 8, 15–22]. Publicly available data were downloaded from NCBI GEO [4], Stanford Microarray Database [23] or websites affiliated with the author's laboratory. Data which were not publicly available were obtained by consent from the authors of the respective studies. The four array CGH platforms used for this study were the whole genome tiling path BAC (SMRT) array [24], the early access Affymetrix SNP 100 K array [16, 20], the Stanford cDNA microarray [22] and the Affymetrix SNP 10 K array [15]. In addition, 2 of the cell lines were profiled for whole genome DNA methylation status using MeDIP array CGH [10, 13]. For this launch version, we concentrated on available cell lines profiled on high density array platforms and did not include profiles from clinical specimens. A summary of the sources for the raw data is given in Table 1, while the detailed description of each of the cell lines in the collection is given in Additional file 1.

Application layout of SIGMA

There are three main components which comprise this application; a Java WebStart application interface allowing users to formulate queries and perform visualization, an Apache Web Service which facilitates the connection of the user application to the database and a relational database which is implemented using MySQL (Figure 1A). Utilization of the Java WebStart technology ensures that users will have the latest version of the application, without the need for manual updating. In addition, efficiency and speed of the application will largely be determined by the user's computer specifications. Hence, we have provided different versions of our application based on system resource utilization, allowing users with greater system resources to perform more analytical tasks per session.

Database implementation and structure

SIGMA is a continually expanding database of array CGH experiments. The launch version contains 267 genomic profiles generated from cancer cell lines, implemented using the MySQL relational database application. Each array CGH experiment is contained in a separate database table allowing for easy and seamless expansion of this database. Upon addition of a new profile, a database table which contains the information of each array CGH experiment is updated. This table stores a record of each experiment, with the name of the cell line, the American Type Culture Collection (ATCC) identification (if applicable), array platform used and the description of the cancer type as part of the schema. For two channel array-based profiles, the dye which was used to label the sample is also recorded. Lastly, mapping information pertaining to a clone or probe and its position in the genome is kept in file with a fixed format, such that subsequent improvements and updates to the genomic positioning of the array elements can be easily incorporated. Moreover, since individual microarray software platforms use their own map version, map information for all platforms were compiled based on data from the UCSC Genome Browser [25]. Currently, two genomic builds are supported: April 2003 (hg15) and May 2004 (hg17) assemblies.

Data processing

Data for each platform were processed as similarly as possible. SMRT array data were normalized using the stepwise framework for normalization with default parameters [26]. Similarly, Affymetrix 10 K and Affymetrix 100 K data were normalized and processed using dChip [27] with default settings. Specifically, the samples from the Affymetrix 10 K dataset of lung cancer cell lines were referenced against the group of matching blood lymphoblast lines and similarly, the breast cancer cell lines were referenced against their matching blood lymphoblast lines. Affymetrix 100 K data from Zhao et al. (2004) were referenced against 12 normal individuals and the NCI-60 profiles were referenced against 6 normal diploid controls. The gender of the profiles were not specified, hence data from the X chromosome may not be an accurate reflection. Segmentation of all data was performed using aCGH-Smooth [28], with data for the sex chromosomes removed prior to segmentation – as profiles were generated with sex matched or mismatched reference DNA in two channel hybridization experiments – and run with the settings of Lambda = 6.75 and "breakpoints per chromosome" = 100. Each element of the array is given a call with respect to normal: -1 if the element shows copy loss, 0 if the element shows no change in copy number and +1 if the element shows increased genomic content.