Gene Set Enrichment Analysis (GSEA)

What is GSEA?

Gene Set Enrichment Analysis (GSEA) is a computational method that determines whether an a priori defined set of genes shows statistically significant, concordant differences between two biological states (e.g. phenotypes).
Reference: http://www.broad.mit.edu/gsea/

Why GSEA?

Traditional statistics use adjusted P-values with some arbitrary cutoff, treating genes with slightly different P-values as different entities. Also, small differences in mRNA abundance are often not detected, nor are large changes in just a few genes.

GSEA remedies this by using all the genes in your expression data for the analysis. GSEA also compiles per-gene statistics across genes within a gene set, allowing for the detection of small changes in many genes or large changes in few genes.

The GSEA algorithm implemented in MeV v4.3 is based on Zhen Jiang and Robert Gentleman's 2007 Bioinformatics paper (Jiang, Z., Gentleman, R., (2007). Bioinformatics. 2007 Feb 1; 23(3):306-13. Extensions to gene set enrichment analysis).

Brief Description of the GSEA Algorithm

The GSEA algorithm can be roughly divided in to three steps:

Calculate the per gene statistic. This is done by fitting a linear model to all the genes, separately and simultaneously.
Calculate the gene set statistic.
Estimate significance by:
- Permuting factor/phenotype/class labels
- Compute the per gene statistic for every permutation
- Compute the gene set statistic for every permutation
- Calculate and report Unadjusted p-values

How to Run GSEA

GSEA uses a set of parameter input dialogs that open sequentially to provide input options

that correspond to each step of the process. The first step of the process is data selection

which lets you assign phenotype/class labels to your samples.

Figure 1

The default assignment (Figure 1)is two groups (factors) with two levels per group. This can be changed to reflect the groups present in your data.

Figure 2

B-ALL and T-ALL are the two levels of this phenotype, so enter 2 in the “Number of levels of factor” textbox. If you have pre selected sample clusters and decide to use the “Cluster Selection” tab just assign group numbers using the Group Assignment drop down as shown in Figure 2.

Figure 3

You can also use the “Button Selection” tab shown in Figure 3 to achieve the same. Group1 and Group2 symbolize B-ALL and T-ALL. You can save these grouping using the “Save settings” button. To load saved groupings, use the “Load settings” button. Reset button will clear all your choices. Once you are done, click the Next > button.

This brings you to the “Parameter Selection” section shown in Figure 4.

Sample Probe Values:

	Sample1	Sample2	Sample3	St. Dev
Probe_1	10	20	30	3
Probe_2	20	5	10	4
Probe_3	20	15	10	2

Using Maximum Probe, the substituted values for Samples 1-3 would be 20, 20, 30. Using Median Probe, the values would be 20, 10, 15. With Standard Deviation (SD), the probe that has maximum SD across samples is used, so the values would be 20, 5, 10. NOTE: SD will be calculated by MeV on the fly. You do not have to do anything. This is just an example.

The ‘Browse’ button corresponding to “Select the directory containing your gene sets” lets you choose the directory containing gene set files. You can select the files you want to use from the “Available” panel.

“Selected” panel indicates the gene set files that you chose to use for this analysis. In addition to this, gene sets can also be downloaded from the MIT/Broad website http://www.broad.mit.edu/gsea/msigdb/downloads.jsp

If your gene set file is *.gmt or *.gmx format, “Select the identifier used to annotate genes in your selected gene set” drop down is automatically populated with “GENE_SYMBOL” as shown in figure above. In case of a custom gene set file, you must manually choose the gene identifier from the drop down.

The “Load Annotation Data” panel lets you upload annotations. Annotations are a MUST for running GSEA. Details on how to load annotations is described in Using the Annotation Feature in MeV manual.

The last step is to hit the Execute button. GSEA outputs besides the standard MeV viewers three new viewers namely “Test Statistic Graph”, “Leading Edge Graph Viewer” and “Geneset p-value graph”. “Significant Gene Sets” under “Table Views” lists gene sets sorted by their Over enriched (upper p values). Lower p values are the probability of seeing a test statistic lower than the observed one. Upper p values are the probability of seeing a test statistic higher than the observed one.

Figure 5

Right Clicking on the rows in the table as shown in Figure 5 lets you navigate to different viewers. "Excluded Gene Sets" table contains gene sets which do not pass the minimum genes per gene set criteria. These gene sets are not included in the analysis. "Probe to Gene Mapping" table shows all the probes mapping to a gene.

Figure 6

“Test Statistic Graph” shown in Figure 6 aims to show how genes within a gene set contribute to the overall gene-set-level metric. This metric is computed by summing the distance from the green line to the orange point and then normalizing this sum by the square root of the number of genes in the gene set.

Figure 7

“Leading Edge Graph” in Figure 7 shows which subset of genes within the gene set is contributing to the significance of the gene set level metric. The leading edge subset is calculated by first ranking the genes based on largest to smallest test statistics.

We then calculate the Jiang-Gentleman statistic for subsets of the gene set, starting with the first subset containing the gene with the largest t-statistic, and then incrementing the subset to include the next gene with the next largest t-statistic.

We iterate through until the final subset contains all the genes in the gene set. The subset which maximizes the Jiang-Gentleman statistic suggests that this group of genes contribute the most to the gene-set level metric.