exploRase: Multivariate exploratory analysis and visualization for systems biology

The datasets being produced by high-throughput biological experiments, such as microarrays, have forced biologists to turn to sophisticated statistical analysis and visualization tools in order to understand their data. We address the particular need for an open-source exploratory data analysis tool that applies numerical methods in coordination with interactive graphics to the analysis of experimental data. The software package, known as explorase, provides a graphical user interface (GUI) on top of the R platform for statistical computing and the GGobi software for multivariate interactive graphics. The GUI is designed for use by biologists, many of whom are unfamiliar with the R language. It displays metadata about experimental design and biological entities in tables that are sortable and filterable. There are menu shortcuts to the analysis methods implemented in R, including graphical interfaces to linear modeling tools. The GUI is linked to data plots in GGobi through a brush tool that simultaneously colors rows in the entity information table and points in the GGobi plots.

There is a rapidly growing number of software tools for analyzing data from high-throughput biological experiments. In addition to commercial packages, such as GeneSpring, there are many free software packages for analyzing the data generated by microarrays and other high-throughput technologies. One of these is Bioconductor , a free collection of bioinformatics tools that build on the strong statistical foundation of the R platform [4]. There are packages in Bioconductor for analyzing and visualizing genomic, transcriptomic, proteomic, and metabolomic data, and more. R is controlled by a scripting language, which allows Bioconductor to easily extend R with bioinformatics functionality.
This exibility of the script-driven interface is a double-edged sword, however. Biologists unfamiliar with programming struggle to take advantage of R and Bioconductor.
The other critical element for high-throughput data analysis is the means for the biologist to visualize and explore the data. The biologist should be able to ask questions about the data without committing to a singular course through the analysis. There are several open-source software tools that enable multivariate exploratory data analysis through interactive graphics. One such tool is GGobi [7]. GGobi is designed to be very exible and open-ended, with the goal of supporting a wide range of analyses. Unfortunately, this generality means that the biologist receives no biology-specic guidance during data analysis and visualization tasks.
ExploRase is designed to bridge the gap between biologists and sophisticated statistical analysis tools. This paper will rst present an overview of exploRase and its software foundation. This will be followed by a review of related work and an in depth description of the exploRase graphical user interface and statistical functionality. Next, there will be a tutorial that guides the user through a hypothetical analysis. Finally, the paper will conclude by discussing the future of exploRase.
2 Overview of exploRase As its name suggests, exploRase is designed to facilitate the exploratory analysis of biological data using R and Bioconductor. ExploRase aims to leverage R in conjunction with GGobi to provide the biologist with the necessary tools and guidance for analyzing and visualizing high-throughput biological data. ExploRase is written purely in R, permitting easy integration with R analysis packages. This also enables other R packages to integrate with exploRase via its public API.
The central component of exploRase is its graphical user interface (GUI), shown in Figure 1. The primary design consideration for the GUI is simplicity. There is no attempt to completely map the features of Bioconductor packages and GGobi to the exploRase front-end. Rather, the GUI supports only a subset of the features provided by the underlying packages, while augmenting the subset with shortcuts and conveniences. In order to provide its GUI, exploRase relies on the RGtk2 package [3], a bridge from R to the GTK+ 2.0 cross-platform widget library [1]. RGtk2 allows exploRase to present, completely from within R, a visually pleasing, feature-rich GUI that is identical across all major computing platforms.
GGobi serves as the visualization component of exploRase. The rggobi package [8] links R with GGobi. With rggobi, R packages are able to load data from R into GGobi, retrieve GGobi datasets into R, get and set the color of observations, create and congure displays, and more. ExploRase uses rggobi to load high-throughput datasets and synchronize the color of observations in GGobi plots with the colors in the biological metadata table in the exploRase GUI. This provides the key visual link between the GUI of exploRase and the visualizations of GGobi.
There is a wide range of analysis methods available in exploRase. Most of them are based on functions available in the default installation of R, while the biology specic methods rely on Bioconductor. All of the methods are deliberately simple. The user is able to compare biological conditions and calculate similarities between biological entities, such as genes, based on the experimental data.

Related work
ExploRase is unique among open-source tools in its integration of interactive graphics with R statistical analysis beneath a GUI designed especially for the systems biologist. The commercial microarray analysis program GeneSpring links to R and Bioconductor and oers some interactive graphics. The free program Cytoscape [5] is designed for viewing and analyzing experimental data in the context of biological networks and is integrated with R via plugins. However, it lacks interactive graphics outside of its network diagrams.
There are many examples of controlling R with a GUI, including several in Bioconductor. The limmaGUI package [6] provides a GUI that leads the user from preprocessing microarray data to modeling it with limma and producing reports. Unfortunately, limmaGUI lacks the interactive graphics and breadth of analysis features of exploRase. The Bioconductor iSPlot package provides general interactive graphics using the R graphics engine but oers only a small subset of GGobi's functionality.
Rattle [9] is an RGtk2-based GUI that leverages R as it guides the user through a wide range of data mining tasks. There is a ltering component that the user may expose above the table, as shown in Figure 2.
This lters the table, as well as the GGobi plots, by any column in the table, as well as by entity list membership. Columns containing character data may be ltered according to whether a cell value equals, starts with, ends with, contains, lacks, or matches by regular expression the test value. Numeric values may be tested for being greater than, less than, equal to, or not equal to the test number. When ltering by color, the user may choose from the current palette of colors. The saved rules are displayed in a table below the rule editor. There is a checkbox in each row that toggles the activation state of the rule. Buttons allow the deletion of selected rules and the batch activation and deactivation of every rule. If a rule is saved, it is combined with all future rules, unless it is deactivated or deleted.  The results are added as a column in the entity metadata table and as a variable in GGobi. This allows the user to sort and lter according to the results, as well as visualize the results in GGobi. The rst set of methods is useful for nding entities with levels that dier greatly between two selected conditions.
The methods include subtracting one condition from the other, calculating the residuals from regressing one condition against the other, and nding the Mahalanobis distances across the conditions. The next set of methods are distance measures (Euclidean and Pearson correlation, centered and uncentered) for comparing a selected entity against the rest within a single sample. The nal two methods are hierarchical clustering and pattern nding. The cluster results are displayed in an interactive R plot, shown in Figure 4, where clicking on a branch point selects all of the child entities in the entity table.
The pattern nder calculates whether a gene is signicantly rising or dropping relative to the others for each sample transition. A change is called signicant if it is the upper or lower third of the changes.
The results are displayed as arrows embedded in the metadata table, as shown in Figure 5. The dialog in Figure 5 allows the user to query for specic patterns. The matching entities are selected in the The nal menu contains tools for processing experimental data. There is a convenience function for calculating replicate means and adding them to the data. The second option launches the dialog shown in Figure 7 that provides several simple rules for ltering out entities based on the experimental data. The cutos are based on minimum value, minimum fold-change, and maximum variance between replicates. This helps the user focus on entities with substantial levels that are changing more between treatments than within. The user may enter the test values directly or use the slider to get some idea of the range of values.

Getting started with exploRase
The rst step towards analyzing your data with exploRase is to load the data. One must note that exploRase is not designed for data preprocessing, so all preprocessing must be done before loading data 6  All of these format specications may sound intimidating, but, in practice, loading the data is a relatively simple task. The CSV format is output by many of the Bioconductor preprocessing tools, as well as Microsoft Excel. In our experience, many biologists already have spreadsheets that conform to the structures described above. The data loading process is further simplied by support for projects: all of the data les may be placed into an empty folder and loaded in a single step by choosing the folder in the open project dialog. An example project may be downloaded from the exploRase website [2].

ExploRase in action
In order to briey demonstrate the features of exploRase, we consider a microarray dataset from an experiment investigating the response of biotin-decient Arabidopsis mutants to treatment with exogenous biotin. The mutants were analyzed with and without biotin treatment. Wildtype plants were used as a control and there were two replicates for each set of conditions. Figure 3 summarizes the experimental design. The dataset was normalized using the RMA method.
The rst step, after launching exploRase, is to load the data. The les containing the experimental data and design matrices, along with a set of gene annotations from the MetNetDB, are placed into a new directory in the le system. The les are loaded into exploRase by clicking the Open Project button (see Figure 1) and selecting the directory containing the les.
The primary goal of this short analysis is to determine which genes appear to respond to biotin treatment in the mutant. Biotin treatment is not expected to have an eect in the wildtype, since wildtype plants are able to suciently produce their own biotin. In order to compare across conditions without having to consider each replicate individually, we add the replicate means to the data, assuming that there are no major inconsistencies within the replicate pairs. The means are added to the data by choosing the Average replicates option from the Tools menu.  One possible way to verify that those genes are indeed dependent on biotin treatment would be to t a linear model using limma, including eects for the genotype, biotin treatment, and their interaction. Figure 9 shows the F values for the interaction of biotin and genotype. The We have demonstrated that exploRase is an eective tool for analyzing and visualizing high-throughput biological data. Its direct access to R analysis packages, such as limma and others from the Bioconductor project, allow it to take advantage of the latest advances in statistical methods for bioinformatics.
The integration with GGobi, including synchronized brushing and the ability to add analysis results as GGobi variables, empowers exploRase to display a wide range of interactive multivariate graphics.
All of these advanced statistical features are enveloped within a simplied GUI that is tuned for a biologist. ExploRase has only scratched the surface of its potential. There are three axes of planned improvement. First, there is work towards biochemical network visualization that communicates the specic biological semantics of a network. An expansion of available analysis methods, with particular focus on clustering and metabolomic analysis, is also necessary. Finally, there is interest in enhancing the visualization of categorical data, such as GO terms and cluster assignments.