PANTHER protein library

The core of the PANTHER system is a collection of phylogenetically-defined protein families and subfamilies generated by computational algorithms, and curated by expert biologist using an extensive software system for associating ontology terms^1,3. The current release contains over 640K proteins from 82 genomes, of which 79 are from the Reference Proteome Project (http://www.ebi.ac.uk/reference_proteomes/) (Figure 1). UniProt IDs are used as primary protein identifiers. These proteins are representatives of their respective genes. Therefore, each gene is represented by only one protein. In addition, UniProt IDmapping (http://www.uniprot.org/mapping/) is used to map the primary protein IDs to other IDs from different databases and resources, which expands the capability of PANTHER to support a wider range of ID types (see Supported IDs in Box 1). The proteins are divided into 7729 families, each of which is represented by a phylogenetic tree, an HMM, and a multiple sequence alignment (MSA) (Figure 1). Protein family trees are constructed computationally from sequence data using a phylogenetic tree inference algorithm called GIGA²⁴. Nodes in the tree, corresponding to common ancestors of extant family members, are annotated by expert biologists with their inferred GO terms and PANTHER protein class terms, based on experiments performed on extant proteins. Subfamilies are determined based on the tree structure and often define the orthologous group, especially in organisms in the Deuterostomia superphylum. The functional annotations are propagated from the annotated ancestral nodes to the subfamily nodes, each of which is also represented by an HMM to allow classification of newly discovered protein sequences (Figure 2A). In addition, the HMM library and the scoring pipeline allow users to analyze genome-wide data from any organisms outside of the 82 within the PANTHER system (see below). Phylogenetic tree and functional annotation at the ancestral nodes have great advantage in annotation of previously unclassified genes. The PANTHER phylogenetic trees are currently used by the Gene Ontology Consortium in its curation pipeline ²².

Open in a separate window

Figure 2

Examples of PANTHER phylogenetic tree and pathway diagram

(A) A sample phylogenetic tree from PANTHER (PTHR11633, PLATELET-DERIVED GROWTH FACTOR). The family contains three subfamilies (blue arrows). SF1 is annotated as “PDGF/VEGF growth factor related protein 1” based on the annotation in the drosophila and c. elegans sequences (Q9VWP6 and Q9N143, respectively). There is a recent duplication that generates the PDGF A chain (SF3) and the PDGF B chain (SF2). Ontology terms are annotated to the node that represents the common ancestor if the extant family, in this case, AN0 (SF1). The classifications are propagated to all the descent nodes, including the AN4 (SF3) and AN33 (SF2).

(B) An example of PANTHER pathway diagram (P00047, PDGF signaling pathway). The diagram is shown in CellDesigner process diagram, which is similar to the SBGN-PD format, for example (blue circle), a transition of an input (e.g., ERK) to an output (eg., phosphorylated ERK) catalyzed by a modifier (eg., phosphorylated MEK). A pathway component (e.g, PDGF in red circle) is associated with the protein sequences in the protein library (red arrows in 2A) through expert curation. This association is supported by literature evidence. As a result, the pathway component of PDGF can be inferred to other orthologous protein sequences in the subfamilies in the library (SF2 and SF3 in 2A).

PANTHER pathway

The PANTHER pathway dataset uses controlled vocabulary and graphical notation to describe pathway knowledge⁷. The graphical representation of pathways is generated by CellDesigner, a pathway-editing tool⁸, and is compliant with the Systems Biology Graphical Notation Process Description standard (SBGN-PD) ²⁵ (Figure 2B). Currently, there are 176 expert-curated pathways in PANTHER. The scope of the pathways is similar to those described in textbooks or review articles, e.g., glycolysis, the PDGF signaling pathway or the p53 pathway. Each pathway contains three key classes of data (Figure 1). First is the pathway component (or molecule class), which represents a specific class of molecules that play the same mechanistic role within a pathway. It can be a protein (e.g., PDGF, Jak), a DNA region, or a simple molecule (e.g., ATP or glucose). If a pathway component is a protein, gene or transcribed RNA, it is associated with the protein sequences in the PANTHER protein library through manual curation (Figure 1 and ​and2).2). The individual protein sequences are instances of the pathway component. In these cases, a pathway component is typically a group of homologous/orthologous proteins across various organisms that are involved in the same biochemical reaction within the pathway. The second class is reaction, which represents biochemical relationships among different pathway components. Since all PANTHER pathways are in compliance with SBGN-PD, a typical reaction is a transition of an input (or reactant) to an output (a product) controlled by a modifier (Figure 2B). The last class is the cell/tissue type and cellular component, which provides the location where the reaction occurs.

There are two key features of the PANTHER pathway data model. First, the association between the molecule classes and protein sequences links the pathway to the phylogenetic relationships and statistical models (HMMs) of the protein families. As a result, if a protein is associated with a pathway component via experimental evidence, its orthologs can be inferred to the same component based on the PANTHER phylogenetic tree. Second, PANTHER pathway supports community standards, and all pathways are available in SBML²⁶, BioPAX²⁷ and SBGN²⁵ formats.

PANTHER tools

PANTHER provides a number of useful research tools, including the PANTHER HMM Scoring Tool, cSNP Analysis Tool, and Gene List Analysis Tool. Both the PANTHER HMM Scoring Tool and cSNP Analysis Tools have online and downloadable versions. The online versions allow only one protein sequence to be analyzed at a time, while the downloadable versions allow large batch analyses. A brief description of the downloadable version of PANTHER HMM Scoring Tool can be found in Box 2. The details of how to use these tools can be found in the PANTHER help page (http://www.pantherdb.org/help/PANTHERhelp.jsp) and the user manual. In this protocol, we will focus on the Gene List Analysis Tool. Sample gene list files can be found in the Supplement materials for testing.

The Gene List Analysis Tool can be accessed directly from the PANTHER home page (www.pantherdb.org) (Figure 3). There are several options for users to input a list of genes (and optionally quantitative data) for analysis. The PANTHER database uses some database identifiers as the primary IDs for each gene, and they are the most common option for input file (e.g., a UniProt identifier, an Entrez Gene identifier, or even a gene symbol). The PANTHER database also maps identifiers from a number of different sources for 82 different organisms using the UniProt IDmapping mechanism, and the identifiers in the user’s list are automatically mapped to the primary IDs in the PANTHER database (Figure 1). A total list of the supported IDs can be found in Box 1. A report is generated that lists not only the mapping of each gene, but also the identifiers that could not be mapped, if any. Since each gene in the PANTHER database belongs to a phylogenetic tree that is annotated with GO and PANTHER ontology terms and pathways, the mapped IDs from the gene list will inherit these annotations. It needs to be pointed out that, in a very rare case, a single gene symbol or its synonym can be mapped to more than one gene from the IDmapping. We are currently working with UniProt to improve such mappings. However, this is so rare that we don’t think it affects significantly to the results from the statistical tests. If the gene list is not from the 82 organisms, a user can still use the PANTHER tools, but they need to map each of their own identifiers to PANTHER identifiers first, using the downloadable PANTHER HMM scoring tool (Figure 1 and Box 2), and creating the Generic PANTHER Mapping File with two columns (the user’s ID, and the ID of the best PANTHER HMM hit) (see Box 1 for more details of the file format). Each gene in the list will inherit the same annotations as those stored in the PANTHER database for the given HMM.

Open in a separate window

Figure 3

The PANTHER home page with the Gene List Analysis Tools.

Once the gene list is classified with those functional terms, it can be analyzed in three different ways.

• Functional classification – This tool provides the functional classification results of the uploaded list, and displays them in either a gene list page or pie chart.

• Statistical overrepresentation test - This tool is based conceptually on the simple binomial test described previously²⁸. It compares a test gene list uploaded by the user to a reference gene list, and determines whether a particular class (e.g., a GO biological process or PANTHER pathway) of genes is over- or under- represented. A more detailed description of the tool can be found in Box 3.

  Box 3

  Statistic Overrepresentation Test

  The input (or test) list is usually a list of genes of your interest. It can be a list of genes that are up-regulated in the gene expression experiment, or have significant p-values from your GWAS experiment. The list is divided into groups based on GO or PANTHER classification (molecular function, biological process, cellular component, PANTHER Protein Class or PANTHER Pathway). As many as four test lists can be uploaded for each analysis. A reference list, which usually contains all the genes/proteins from which the list was drawn, is divided into groups in the same way. PANTHER provides reference proteome data set as default reference lists for all 82 genomes, so uploading a reference list is optional. Then, for each functional category, e.g., protein kinase for molecular function, cell proliferation for biological process, or apoptosis signaling pathway for pathway, the binomial test is applied to determine whether there is a statistical over- or under-representation of genes/proteins in the test list relative to the reference list.

  P-value calculation in the overrepresentation test

  The expected value is the number of genes you would expect in the test list for a particular PANTHER category, based on the reference list. For example, there are 20,000 genes in the reference list (ex: the human genome). 440 of these genes map to the GO term induction of apoptosis. Based on this, 2.2% (440 divided by 20,000) of the genes in the reference list are involved in induction of apoptosis. Now a test list that contains 500 genes is uploaded. Based on the reference list, it is expected that 11 genes (500 multiplied by 2.2%) in the test list would be involved in induction of apoptosis.

  If for this biological process more genes are observed in the test list than expected, you have an over-representation (+) of genes involved in induction of apoptosis. If fewer genes are observed than expected, you have an under-representation (−). A p-value is calculated then to determine whether the over- or under- representation is significant or not. For example, let’s assume that 21 genes are observed in the test list are involved in induction of apoptosis. Although this is almost twice as the expected value, the p-value is large and not significant (the p-value would be 0.722). Alternatively, if 35 genes are observed, this is very different than the expected value, so you would expect a small, significant p-value (the p-value would be 6.21e-7). This small p-value indicates that the result is non-random and potentially interesting, and worth looking at in closer detail. A p-value cutoff of 0.05 is recommended as a start point.

  The statistical method used in this test is the binomial test. In the binomial test we assume that under the NULL hypothesis, genes in the test list are sampled from the same general population as genes from the reference set, i.e. the probability p(C) of observing a gene from a particular category C in the test list is the same as in the reference list. We first estimate the probability p(C) from the reference set assuming that it is large and representative:

  p(C)=n(C)/N,

  where n(C) is the number of genes mapped to category C, and N is the total number of genes in the reference set.

  We then use the above estimate to find the p-value: the probability of observing k(C) genes (or a more extreme number) in the uploaded list of size K. Under the NULL hypothesis, the number of genes of mapped to C is distributed binomially with probability parameter p(C) and thus the p-value would be

  $$p-value=\sum \left(\begin{array}{c}K\\ k\end{array}\right)p{(c)}^k{(1-p(c))}^{K-k}$$

  where the sum runs from k(C) to K in the case of over-representation (i.e. when the number of observed genes k(C) is greater than expected p(C)*K under the NULL hypothesis), and 0 to k(C), in the case of under-representation (i.e. when k(C) is smaller than p(C)*K).

  When developing this analysis tool, we tested other statistic methods also. We decided to use the Binomial, since other methods tends to be not as accurate when the population sizes or the expect number is small.

• Statistical enrichment test - This tool, based on the work by the PANTHER group²⁹ and Lander’s group³⁰ for two different genomic data analyses at about the same time, uses the Mann-Whitney test³¹ to determine if any ontology class or pathway has numeric values that are non-randomly distributed with respect to the entire list of values. The numerical data can be normalized raw readouts from the microarray experiments or, more commonly, the fold-change value for each gene in a differential expression experiment, or calculated p-values from GWAS experiment. A more detailed description can be found in Box 4.

  Box 4

  Mann-Whitney Rank-Sum Test (U-Test)

  The statistical test is general enough to handle any numerical data, continuous or discontinuous, generated by experiments such as gene expression, proteomics or GWAS. First, a reference distribution is generated using all values from the input data. Then the entire list is divided into groups based on GO or PANTHER classification (molecular function, biological process, cellular component, PANTHER Protein Class or PANTHER Pathway), and the distributions for each group are generated. The probability that the functional category distribution was drawn randomly from the reference distribution is estimated using the Mann-Whitney Rank-Sum Test (U-Test) ²⁹.

  To perform the rank sum test, first the values of the genes that map to a given category are combined with the overall list of values that were input. Then, all the values are ranked from smallest to largest, with the smallest value getting a rank of 1. If multiple values are identical, the average of the ranks for these values is used.

  Then the rank sum is calculated for this category, by summing up the ranks for all of the genes that map to this category. The average rank, R1 is then calculated by dividing the rank sum by the number of genes, n1, that map to the category. Likewise, the rank sum is calculated for the list of all IDs uploaded, and the average rank, R2, is calculated by dividing the rank sum by the total number of genes uploaded, n2.

  Next, the Mann Whitney U statistic is calculated for both populations:

  U1 = n1* n2 + (n1 * (n1 + 1)) / 2 - R1

  U2 = n2* n2 + (n1 * (n2 + 1)) / 2 - R2

  The larger of these two values is the Mann Whitney U-statistic, U, whose distribution for small sample sizes can be found in most statistic books. In our case, our application is for large sample sizes, so we use the normal approximation:

  Z-score = (U- (n1* n2)/2)/sqrt(n1*n2*(n1+n2+1)/12).

  It follows that the p-value is the integral under the standard normal density.

It is worth pointing out that there are other similar tools to perform overrepresentation and enrichment tests, most notably, GSEA³² and DAVID³³. Although all three tools use very similar statistical algorithms in the back end, there are some differences among them. GSEA requires download and installation, so its targeted users are bioinformaticians who are more computer-savvy. Both DAVID and PANTHER are online tools and are more appealing to bench biologists. Compared to DAVID, PANTHER has a few advantages. First, PANTHER is currently part of the GO consortium, and integrates more updated GO curation data with the tools. In fact, DAVID downloads PANTHER data and integrate them in its analysis tools. Second, PANTHER allows users to analyze genome data from 82 organisms using the online tool, and any other organisms using the PANTHER scoring tool. Third, the phylogenetic trees in PANTHER protein library allow users to make more accurate ortholog prediction, and thus greatly enhance the capability of the tools. One weakness of PANTHER is that it cannot analyze data against resources outside of PANTHER, such as KEGG and Reactome.

B. Functional classification tool viewed in pie chart

This tool returns the results in a pie chart, which displays an overview of all ontology terms at the first (or most general) level within the same ontology (Figure 6). When a slice of the pie chart, which represents an ontology term, is clicked, a new pie chart will appear that contains its child ontology terms. Since one gene can be classified to more than one term, the pie chart is calculated based on the number of “hits” to the terms over the total number of class hits. Class hit means independent ontology terms. For example, if a gene is classified to 2 ontology terms that are not parent or child to each other, it counts as 2 class hits.

When you place the computer mouse pointer over a slice, the category name and a series of counts are displayed. In our example in Figure 6, the name is a GO term apoptosis (GO:0006915) followed by

1. the number of genes (70) from the uploaded list that are classified to the term apoptosis;

2. the percent (14%) of genes classified to apoptosis (70) over the total number of genes (500);

3. the percent (4.7%) of genes classified to this apoptosis (70) over total number of class hits (1494).

C. Statistical overrepresentation test

The results of this analysis tool are displayed in a table (Figure 7A). If one test gene list is uploaded, the table contains six essential columns of data:

1. The first column contains the name of the PANTHER classification category. If you are doing this analysis in pathways, you can click on the pathway name to view the corresponding pathway diagram.

2. The second column contains the number of genes in the reference list that map to this particular PANTHER classification category.

3. The third column contains the number of genes in your uploaded list that map to this PANTHER classification category.

4. The fourth column contains the expected value (see Box 3), which is the number of genes you would expect in your list for this PANTHER category, based on the reference list.

5. The fifth column has either a + or −. A plus sign indicates over-representation of this category in the test list: you observed more genes than expected based on the reference list (for this category, the number of genes in your list is greater than the expected value). Conversely, a negative sign indicates under-representation, i.e. fewer genes than expected.

6. The sixth column is the p-value as determined by the binomial statistic (see Box 3). This is the probability that the number of genes you observed in this category occurred by chance (randomly), as determined by your reference list. A small p-value indicates that the number you observed is significant and potentially interesting. A cutoff of 0.05 is recommended as a starting point.

If more than one test list is uploaded, columns 3 to 6 are repeated for each list.

From this result page, various statistics can be exported by using the drop-down menu next to the “Export results” button. The list of genes/proteins in any functional group can be viewed by clicking on the listed counts. When pathways are chosen as the functional categories, clicking on the pathway name brings up pathway diagrams colored according to preferences specified by the user (Figure 7B). The resulting pathway diagram can be exported as an image file (.png) by choosing “File -> Export image” from the Applet menu.

The authors would like to thank the Reference Proteome team, especially Craig McAnulla and Maria Martin, for their support to provide up-to-date Reference Proteome dataset, and Yukiko Matsuoka and Katoh Manami from the SBI Japan for their support on CellDesigner and pathway file update. This work is supported by NIH/NIGMS GM081084 to P.D.T.]. Funding for open access charge: University of Southern California.