Prosecutor: parameter-free inference of gene function for prokaryotes using DNA microarray data, genomic context and multiple gene annotation sources

Abstract

Background: Despite a plethora of functional genomic efforts, the function of many genes in sequenced genomes remains unknown. The increasing amount of microarray data for many species allows employing the guilt-by-association principle to predict function on a large scale: genes exhibiting similar expression patterns are more likely to participate in shared biological processes.

Results: We developed Prosecutor, an application that enables researchers to rapidly infer gene function based on available gene expression data and functional annotations. Our parameter-free functional prediction method uses a sensitive algorithm to achieve a high association rate of linking genes with unknown function to annotated genes. Furthermore, Prosecutor utilizes additional biological information such as genomic context and known regulatory mechanisms that are specific for prokaryotes. We analyzed publicly available transcriptome data sets and used literature sources to validate putative functions suggested by Prosecutor. We supply the complete results of our analysis for 11 prokaryotic organisms on a dedicated website.

Conclusion: The Prosecutor software and supplementary datasets available at http://www.prosecutor.nl allow researchers working on any of the analyzed organisms to quickly identify the putative functions of their genes of interest. A de novo analysis allows new organisms to be studied.

Figure 1

Flowchart of Prosecutor. Flowchart of the functional prediction process in Prosecutor. First, the expression profiles from DNA microarrays (1A) are used to create a correlation matrix (1B). For every gene, the correlations with the remaining genes are retrieved from the correlation matrix and sorted (1B2). The sorted gene list is used to perform an iterative Group Analysis for every functional category (1B3). The resulting p-value is indicative for the prediction of a gene as a member of a functional category (1C). At this step, the regular iGBA process ends. However, to also assess the reliability of each prediction, the following steps are added. The complete list of p-values for every functional category is sorted (1C4), after which the positions of the members of the functional category are determined (1C5). These positions are used to create ROC curves (1D; see Results section for more information concerning ROC curves). The corresponding Area Under the ROC Curve (AUC) is then used as a measure of expression coherence value of a functional category.

Figure 2

Schematic overview of the additional information provided by Prosecutor. Various layers of information are supplied for the iGBA results (2A) from Prosecutor. Predicted functional assignments for genes whose operon members are already linked to the predicted function are indicated in the results (2B). In addition, this protocol is also followed for divergent genes that share the same upstream region (in this example pps and ydiA). The operon information that is used for the genomic context analysis is also used to detect known regulatory sequences for transcriptional modules (2C). Lastly, graph visualization is used to visualize the gene redundancy of the different functional assignments of Prosecutor (2D). Nodes in the graph represent functional categories and genes. Arrows represent membership of gene nodes to a functional category node as well as the putative functional prediction of the studied gene. The members of individual categories are placed in colored aggregates. In addition to the aggregates, a colored square is placed in each gene member of a category. The squares are colored using the colors of their matching aggregates. Members of different categories can easily be distinguished using the colored squares. An example of a functional prediction found by Prosecutor for ydiE from E. coli is shown. The expression of this gene was correlated with members of various functional categories involved in the uptake of iron. In addition to the functional association with the transcriptional module Fur, the upstream region of ydiE also contains a putative Fur DNA binding site.

Figure 3

Prediction ability of four annotation sources. Histograms of ROC areas (Area Under the Curve) for four annotation sources for E. coli based on 305 microarrays (3A) compared to randomized results (3B). The real data reveal a large amount of categories with AUC values larger than 0.8, which are almost absent in randomized results. These categories are the most promising candidates for which the iGBA approach will enable confident gene assignments functional predictions. Analysis of the AUC distribution across the annotation sources shows that the "transcription module" annotation source is the most informative, i.e., contains the largest amount of categories exceeding an AUC value of 0.9 (3A). This is intuitively very convincing as shared transcriptional regulation is the basis of coexpression. In addition to ROC areas for all GO terms, we have also analyzed the distribution of ROC areas for the GO annotation source using the "gold standard" [28]. This proposed "gold standard" (GS) consists of a specific trusted set of biological processes that maps proteins to well-defined functional classes to evaluate predictions. The authors supply a set of biological processes that is based on selection by a panel of biology experts. We have included AUC results for the GO annotation for E. coli using the GS. Analysis of the AUC distributions shows that the distribution of relative occurrences of the GS analysis and the analysis using a fixed member cutoff is comparable.

Figure 4

Prediction ability of two annotation sources for yeast. Histograms of ROC areas (Area Under the Curve) for two annotation sources (Gene Ontology and metabolic pathways) for S. cerevisae based on 1079 datasets from Stanford microarray database (4A) compared to randomized results (4B). The real data reveal a large number of categories with AUC values larger than 0.8, which are almost absent in randomized results. These categories are the most promising candidates for which the iGBA approach will enable confident gene assignments of functional predictions.