KEGG spider 26 is a freely available web-based tool that implements a global metabolic network framework for the interpretation of gene lists. It has a simple interface: as input it accepts several types of gene or protein identifiers. For example, for the human genome, KEGG spider supports identifiers from 'Entrez Gene'27, 'UniProt/Swiss-Prot', 'Gene Symbol' 2728, 'UniGene' 27, Ensembl' 29, 'RefSeq Protein ID', 'RefSeq Transcript ID' 30, and'Affymetrix probe codes' 31. As output, the user gets a report on the statistical significance of the inferred network models (D₁, D₂,..), as well as a catalog of enriched KEGG pathways and Gene Ontology terms. For each model (D₁, D₂,..), a link is provided to obtain a graphical visualization. The visualization is performed by the Medusa package 32. In addition, the user can highlight genes from the model according to KEGG canonical pathways. The inferred network models can be downloaded as a text file and used with freely available packages for network analyses and visualization 3233.

Here, we present several examples of analysis of published experimental data by KEGG spider. To illustrate the advantages experimental researchers would get by using KEGG spider in comparison to commonly used pathway enrichment analyses, we provide a comparison between KEGG spider and GENECODIS 34, a tool recently published in Genome Biology that implements a possibility to perform enrichment analysis of KEGG pathways. The choice of GENECODIS was casual, as the results of enrichment analyses of KEGG pathways by other tools would be similar.

We also provide a comparison (Additional data file 1) of KEGG spider to KEGG atlas 23. KEGG atlas is a web tool that provides visualization of a gene list (converted into KEGG KO identifiers) in the context of the global metabolic network. As has been discussed above, KEGG atlas provides no quantitative or statistical analyses and, thus, supplies no criteria for the evaluation of the quality of provided graphical output. As demonstrated, the output of KEGG atlas for a random gene list looks similar to the experimentally derived gene lists.

In 35 a comparison of the expression profiles of cell populations from 20 diffuse-type gastric cancers with their corresponding non-cancerous mucosae was performed. The authors report in the paper the top 75 up- regulated and top 75 down-regulated genes. The 150 differentially expressed genes represent a variety of functions, including genes involved in various metabolic pathways. In total, 28 genes map to KEGG metabolic pathways. Enrichment analysis (Table 1) identified three pathways that are significantly overrepresented. For example, nine genes are from the 'metabolism of xenobiotics by cytochrome P450' pathway and five are involved in 'bile acid biosynthesis'.

The model D₁, containing directly connected genes, provided by KEGG spider covers 14 genes (p-value < 0.001). The model D₂, in which one intermediate gene is allowed, covers 24 genes (p-value < 0.001). Figure 2 presents a graphical visualization of the inferred D2 model, which spreads through five canonical KEGG pathways.

Therefore, in comparison to available analytical procedures, KEGG spider enhances our understanding of metabolism variation in gastric cancers. First, it demonstrates that deregulated genes do not split into independent groups (pathways) as may be concluded from standard enrichment analyses: almost all 24 (out of 28) genes form a non-interrupted (a maximum of one missing gene is allowed) network. Second, it provides not only information that 24 genes are mapped close to each other on the global metabolic network but also estimates the confidence of this event: the p-value reflects the probability of getting a non-interruptedly connected network that covers at least the same number of genes for a randomly sampled list of 28 genes (only genes mapped to KEGG metabolic pathways are used to generate the random lists).

Primary hepatolithiasis or intrahepatic calculi, which is characterized by the formation of gallstones in the intrahepatic bile duct, is an intractable liver disease and suspected to be one of the causes of cholangiocellular carcinoma. To obtain an insight into the disease, the proteomic analysis of liver tissue specimens was done (affected and unaffected hepatic segments from patients with primary hepatolithiasis) 36. For the specimens from the unaffected segments, 83 unique proteins were reported. For the specimens from the affected segments, 74 unique proteins were reported. Consequently, 12 up-regulated proteins and 21 down-regulated proteins were identified in affected versus unaffected hepatic segments.

For example, 17 out of 21 down-regulated proteins (unaffected versus affected hepatic segments) map to KEGG pathways. A standard enrichment analysis for the 21 down-regulated proteins found two pathways 'urea cycle' (five proteins) and 'glycolysis' (four proteins) to be enriched (Table 2). These results enable the conclusion that some characteristic metabolic pathways are violated in affected hepatic cells. Analysis with KEGG spider provides a comprehensive picture of the characteristic metabolic perturbations between normal and diseased cells. The model D₂, in which proteins are connected via one intermediate protein, covers all 17 proteins (p-value < 0.001) that are mapped to KEGG metabolic pathways. The model D₂ is presented in Figure 3. The KEGG spider model retrieves a comprehensive picture of the genetic basis of metabolic variations in comparison to standard enrichment analyses. As in the previous example, it demonstrates that deregulated genes are not independent (or split to independent pathways) and all 17 metabolism related proteins form non-interrupted (a maximum of one missing gene is allowed) network.

To support the practical significance of KEGG spider, we collected dozens of recently published experimental studies that reported lists of genes/proteins in various biological contexts. We reanalyzed them using KEGG spider and demonstrated that, in most cases, the models provided by KEGG spider improve our understanding of the genetic basis of metabolism variations. These results can be found at the KEGG spider web site 37.

Of particular interest are the studies that report differentially expressed genes/proteins between normal/disease cell states or treated/untreated cell states. We selected 17 such studies, which report at least eight genes/proteins that can be mapped to KEGG metabolic pathways and analyzed these genes/proteins using KEGG spider and GENECODIS. The comparative statistics is provided in Table 3. The 'GENECODIS' column reports results provided by GENECODIS, the 'k' column reports the number of pathways found to be enriched at a p-value < 0.05, and the 'max' column reports the number of input genes covered by the largest pathway. As can be seen, in all cases the interpretational power of enrichment analyses was quite limited. On average, from 10% to 40% of the input genes mapped to KEGG pathways could be interpreted by one canonical pathway. As can be seen, in all cases KEGG spider provided statistically valid models.

The KEGG REACTION database is a collection of chemical structure transformation patterns for substrate-product pairs (reactant pairs). We can build a global 'reaction network' (reactions are nodes, compounds are edges) by connecting with edges reactions that share the same compounds. In general, a reaction consists of multiple reactant pairs, and the one that appears on the KEGG metabolic pathway is called the main pair. To build a global reaction network, we used only compounds classified as main reaction pairs. Otherwise, many reactions will be connected only because they use or produce such compounds as H₂O, CO₂, and so on.

In KEGG, reactions are linked to orthologous groups of enzymes (KEGG ORTHOLOGY database) and orthologous groups are mapped to the genes (in most cases each orthologous group corresponds to ortholog genes from different genomes). Thus, reactions can be mapped to genes from a given genome, and the reaction network can be transformed into a global organism-specific gene metabolic network, where genes are nodes and compounds are edges, respectively. Some reactions are organism specific or are not annotated by an orthologous group. In this case, they are not present in the corresponding organism-specific gene network. Therefore, the resulting global gene metabolic network links by edges any two genes that are associated with reactions sharing common compounds (from the main reaction pair).

The distance between two arbitrary genes is computed as the minimum number of consecutive steps required to get from one gene to another by working through existing paths on the global gene metabolic network. Distance 1 means that two genes are directly connected. Distance 2 means that two genes are connected via one intermediate gene, distance 3 means that two genes are connected via two intermediate genes, and so on. Given a gene list, our purpose is to infer the network model that minimizes the distance between each connected gene pair according to pairwise distances between genes.

Initially, we map genes from the input list onto the global gene metabolic network. At this point all genes from the input list are disconnected. In the first step, we connect by edges gene pairs with distance 1 and look for connected subnetworks. The subnetwork with the maximal number of genes is referred to as an inferred network model D₁. We also refer to the number of genes in the maximal subnetwork as the size of the inferred model. In the second step, genes (from the input list) with distance 2 are connected by edges. The subnetwork with the maximal number of genes is inferred and is referred to as network model D₂. In a similar way, network models D₃, D₄,..., are inferred. Models D₂, D₃,..., incorporate genes that are not from the input list but are added to connect input genes in the network model. We refer to these added genes as intermediate genes.

The null hypothesis is that the input gene list has no bias in relation to the topology of the global gene metabolic network. A quality measure of the inferred network model can be its size, that is, the number of genes from the input list in the model. We have to estimate the probability to infer models with the same or bigger size from randomly generated gene lists of size N, where N is the number of input genes.

Let us assume that we have N genes in the input list. Using the network inference procedure described above, we infer the network models D₁, D₂, D₃. Let us denote S₁, S₂, S₃ to be the number of input genes in the inferred network models D₁, D₂, D₃. The values S₁, S₂, S₃ are used as statistics. To estimate the significance of the inferred model D₁, we compare the value S₁ with a distribution R_1j. In the same way, we estimate the significance of the inferred models D₂, D₃ by comparing the values S₂, S₃with distributions R_2j, R_3j, respectively.

The distributions R_1j, R_2j, R_3j are computed by a random simulation procedure 39. To generate the background distributions R_1j, R_2j, R_3j, we repeat the following simulation procedure k times. Index j = 1..k specifies the random simulation. Each time the random gene list B_j of size N (equal to the size of the input list) is generated. The network inference procedure described above is applied to the list B_j and the network models D_1j, D_2j, D_3j are inferred. Let us denote the number of genes from the random list B_j in the inferred network models D_1j, D_2j, D_3j as R_1j, R_2j, R_3j. Thus, after repeating k times the simulation procedure, we get the background distribution R_1j(j = 1..k) for model D₁, the background distribution R_2j(j = 1..k) for model D₂and the background distribution R_3j(j = 1..k) for model D₃.

To estimate the significance of the inferred network model D₁ for the input gene list, the value S₁ is compared to the distribution R_1j. Let n be the number of values from the distribution R_1j that are equal or greater than S₁. The estimate of the p-value p of the inferred network model D₁ is computed as p = (n + 1)/k. In the same way, the p-values for models D₂ and D₃ are computed using values S₂ and S₃ and background distributions R_2j and R_3j. In other words, the p-value is estimated as a share of random simulations where the size of the inferred models for a random gene list (size N) are equal to or greater than the size (S₁, S₂, S₃) of the inferred models for input gene list (size N).