To extract combinations of gene annotations, GENECODIS uses a modification to the methodology reported by Carmona-Saez and coworkers 6, which implements the apriori algorithm to extract associations among gene annotations and expression patterns.

The apriori algorithm was originally introduced by Agrawal and coworkers 7 and has been extensively used to extract association rules from transaction databases. This algorithm generates sets of elements that frequently co-occur in a database of transactions. Briefly, the procedure starts by determining the set of all single annotations ('itemset') that appear in at least x genes (also known as support threshold) from the list of interest and establish the frequent k itemsets, where k = 1. In the second iteration (k = 2), the set of frequent annotations found in the previous step is used to produce the new set of candidates of size 2 (2-itemset), and the database is scanned again to explore each gene and counting the frequency of each pair of annotations. However, if the set of annotations does not satisfy the minimum support constraint - that is, they do not occur in at least x genes - then they are not further considered to generate larger itemsets. The procedure continues until no more combinations are possible. At the end of this search all itemsets that contain the collection of annotations that co-occur in at least x genes are obtained (Additional data file 1).

In our previous work 6 we used the apriori algorithm to extract association rules among gene annotations and expression patterns. However, in this work we use it as the initial step in the methodology included in GENECODIS, namely the extraction of sets of annotation that frequently co-occur in a gene list.

It is important to note that increasing the number of different items (sources of annotation in this case) while decreasing the minimum support value can significantly multiply the number of concurrences and thus the computation time. Additional data file 1 contains a complete study of execution time and size of the itemsets for different support values in real datasets. Very extreme scenarios, such as extracting all possible combinations of terms that appear in at least one gene (support value of 1), is in many cases a computationally unfeasible task. For this reason we have provided the application with a minimum support value of 3, which is a reasonable threshold to extract significant biological information from gene lists.

Once all combinations of annotations that appear in at least x genes have been extracted, the method counts the occurrence of each set of annotations in the list of genes and in a reference list. Note that for each set of concurrent annotations its frequency is calculated as the number of genes that are simultaneously co-annotated with those terms. By default, GENECODIS uses as a reference set all genes from the corresponding genome at the NCBI Entrez Gene database 8, but users can upload their own reference set (for example, genes in a chip). Then, a statistical test is applied to identify categories, and their combinations, that are significantly enriched in the list of genes. Two statistical tests are implemented in GENECODIS: the hypergeometric distribution and the χ² test of independence. For a detailed description of these methods in the context of the ontological analysis of gene lists, see the work of Draghici and coworkers 9 and the online help for the program.

The p values can then be adjusted for multiple tests using a simulation-based correction approach 1011 or the false discovery method proposed by Benjamini and Hochberg 12. For the simulation-based correction, a gene list of the same size of the input list is generated by randomly selecting genes from those used as reference. The frequent itemsets are then extracted (as described above) from this random list and their corresponding p values are calculated. This process is repeated 10,000 times and the corrected p value for each k itemset is calculated as the fraction of simulations having any k itemset with a p value as good as or better than the p value for that k itemset.

Therefore, the result of the analysis performed by GENECODIS consists of a list of annotations or combinations of annotations with their corresponding p values. Annotations exhibiting p values below a certain threshold can be considered significantly associated with the list of genes under study and can be used to discern the biologic mechanisms relevant to the experimental system.

To illustrate GENECODIS, we show the results obtained using data generated by Smith and coworkers 15. They used oligonucleotide-based whole genome microarrays to measure gene expression levels in yeast during growth in oleate (peroxisome induction) and growth in glucose (peroxisome repression conditions). Using different clustering algorithms they identified 224 yeast genes whose expression patterns were similar to well known peroxisomal genes.

The list of these 224 genes was re-analyzed using GENECODIS, selecting biological process (BP) and cellular component (CC) GO Slim annotations. The simultaneous analysis of both categories provided a global picture of the biological processes associated with the experimental system linked to cellular localization information (Figure 2). As was expected, the most significant category associated with this gene list was 'peroxisome' (CC). Other single categories that were highly representative were 'generation of precursor metabolites and energy' (BP), 'carbohydrate metabolism' (BP), and 'lipid metabolism' (BP), which is consistent with the observation that the shift to growth in the presence of oleate activates genes encoding enzymes that are involved in fatty acid degradation, allowing efficient use of the new carbon source 16.

In addition to these single-category significant annotations, GENECODIS revealed a new set of associations with a strong biologic meaning. For example, taking a closer look at the second and third categories with the lowest p values, we can see that a significant set of genes were co-annotated with 'peroxisome' (CC) and 'lipid metabolism' (BP), and 'peroxisome' (CC) and 'organelle organization and biosynthesis' (BP), respectively. These findings allow us to easily identify the set of peroxisomal genes that are specifically involved in each one of these two different biological processes. Among the genes co-annotated as 'peroxisome' (CC) and 'lipid metabolism' (BP) are the genes involved in the fatty acid β-oxidation pathway, such as POX1, FAA2, ECI1, FOX2, POT1, and DCI1. Among genes co-annotated as 'peroxisome' (CC) and 'organelle organization and biosynthesis' (BP) are the PEX genes, which are involved in peroxisome assembly 15 and are required for the increase in the number of these organelles during growth on oleate 16.

Another interesting set of annotations that show the usefulness of the application are those categories related to mitochondrial genes. Forty-eight out of 887 yeast genes annotated as 'mitochondrion' (CC) were present in the list, and therefore this annotation exhibited a p value of 0.0248 (simulation corrected p value = 0.2; Additional data file 2). Consequently, based on the statistical test, this annotation is not considered significant. Nevertheless, GENECODIS was able to identify a set of significant co-annotations related to mitochondrial genes. For example, 6 out of 21 yeast genes that were simultaneously annotated with 'mitochondrion' (CC) and 'lipid metabolism' (BP) were present in the list, and this co-annotation exhibited a p value of 0.000162 (simulation corrected p value = 0.0086). Among these genes was, for example, the CRC1 gene, which is a mitochondrial inner membrane carnitine transporter that is required for carnitine-dependent transport of acetyl-coenzyme A from peroxisomes to mitochondria. In the same way, the co-annotation of 'mitochondrial membrane' (CC) and 'generation of precursor metabolites and energy' (BP) related to a subset of genes that are component of the mitochondrial respiratory chain was found to be significant, with a simulation corrected p value of 0.004.

Although fatty acid β-oxidation in Saccharomyces cerevisiae is restricted to peroxisomes, the association of mitochondrion related categories to this set of genes is highly consistent with the important role of these organelles in the metabolism of β-oxidation products. Acetyl-coenzyme A, the final product of the fatty acid β-oxidation pathway in peroxisomes is transported to the mitochondria for the final oxidation to CO₂ and H₂O 17. In this way, peroxisomal fatty acid β-oxidation demands a functional mitochondrial electron transport chain for energy production, and either functional peroxisomes and mitochondria are required for growth in the presence of oleate 15.

To provide a second example of the functionality of GENECODIS, we analyzed a set of 85 human genes expressed in testis reported by Su and coworkers 18. This dataset was also used by Zhang and colleagues 19 to illustrate the performance of the GOTree Machine (GOTM) software, and therefore it represents a good test case for our method. Zhang and colleagues, using GOTM, reported four main groups of GO biological process annotations related to the testis gene cluster: categories related to cell proliferation, cell cycle, mitosis, and meiosis; categories related to testis specific development; categories related to protein phosphorylation; and categories related to glycerolipid metabolism.

We used our tool to analyze this set of genes using the GO biological process categories and InterPro motifs that appear in at least three genes. The most significant concurrences are shown in Figure 3. Similar results to those reported by Zhang and coworkers 19 were obtained by GENECODIS, except for the case of categories related to glycerolipid metabolism, which were not extracted because they were present in only two genes. In addition, GENECODIS was able to provide new information for the functional interpretation of this set of genes. For example, the fifth association revealed that a significant set of genes in the analyzed list were co-annotated with 'protein amino acid phosphorylation' and 'cell cycle' GO biological process categories and contained protein kinase motifs. The importance of this observation is the explicit connection between 'protein amino acid phosphorylation' and 'cell cycle' categories.

In order to explain the 'protein phosphorylation' category in the context of the phenotypic feature of the gene cluster, Zhang and colleagues 19 remarked that, 'spermatozoa undergo a series of changes before and during egg binding to acquire the ability to fuse with the oocyte. These priming events are regulated by the activation of compartmentalized intracellular signaling pathways, which control the phosphorylation status of sperm proteins.'

Results provided by GENECODIS complement this finding and point out that, in this particular case, the 'protein phosphorylation' category is mainly related to proteins that are involved in cell cycle. Indeed, activation and inhibition of many key regulators of cell cycle are carried out by phosphorylation/dephosphorylation events.

This finding can be confirmed by examining the genes that were co-annotated with both categories: CDC2 (Entrez Gene ID: 983), aurora kinase A (Entrez Gene ID: 6790), NEK2 (Entrez Gene ID: 4751), BUB1 (Entrez Gene ID: 699), and BUB1B (Entrez Gene ID: 701). All of these have been associated with testis tissues and cell proliferation events in previous studies. For example, the NEK2 gene is predominantly expressed in spermatocytes and appears to be associated with meiotic chromosomes in these cells 20; expression of the gene BUB1B in testis decreases with advancing age, and it may play a role in regulating infertility 21.

These two examples illustrate the type of information provided by GENECODIS, which can be useful in helping researchers to interpret large lists of genes generated by high-throughput experimental techniques.