BioGraph is based on the data integration of 21 publicly available curated databases containing biomedical relations (Table 1; Additional materials and methods in Additional file 1) between heterogeneous biomedical entities such as genes, diseases, compounds, pathways, ontology terms, protein domains, disease and gene families, and microRNAs.

The integrated databases were selected based on their quality of relations with respect to curation methods and peer-reviewed references to the literature. Curated database producers employ domain experts to read and extract proven knowledge from the peer-reviewed scientific literature. Such processes of indexing, albeit time-consuming, ensure that the collected knowledge is accurate and complete, allowing for the successive establishment of new relations, for example, with BioGraph or related prioritization algorithms. We did not integrate databases constructed from high-throughput experiments with statistical or computational inferences where no manual curation of the indexed relations was performed. Such databases may include information of lower quality and consequently impair the predictive quality of consecutive data mining. We provide an assessment of each database's quality in the Results section.

The integrated databases in BioGraph consist of three types. (1) Curated databases (for example, Online Mendelian Inheritance in Man (OMIM) and various protein-protein interaction databases) constructed by manual extraction of published, peer-reviewed information about a specific type of information, guaranteeing the quality of the relations in these databases. (2) Curated ontology databases (for example, Gene Ontology (GO) and Medical Subject Headings) using hierarchical classifications of subjects. (3) Curated annotation databases (for example, GO Annotations (GOA) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway database) that relate biomedical entities or concepts to ontology terms.

With regard to the integration of diverse databases with diverse identifiers for the concepts, each concept is provided with a distinct accession number, based on the Unified Medical Language System (UMLS) 9, to guarantee each concept's uniqueness. It should be noted that some of the integrated concepts (especially microRNAs and pathways) are underrepresented in UMLS. In these cases, we have extended the index of UMLS identifiers by these concepts' originating identifiers (for example, by adopting miRBase and KEGG pathway accession numbers). Relations between concepts are extracted from the knowledge resources, represented in a common format, annotated with semantic relation types (denoting the meaning of the relations, for example, 'protein interaction' or 'disease drug') and references to supporting literature, as provided by the integrated databases. All relations in the network are equally weighed independent of their support in the databases or the literature. We have experimented with weighing relations differently, dependent on the quality of the resource database, semantic type or references in the literature, but have not noticed a significant effect of such weights on test benchmarks, as discussed later. To sanitize the resulting network for the subsequent data mining algorithms, disconnected concepts from the largest connected network are removed and dangling concepts (that is, concepts connected to only one other concept) are pruned. As a result, the integrated network comprises 54,567 biomedical entities representing unique biomedical concepts and 425,353 unique relations among these entities, supported by 244,258 references to 52,866 items from the biomedical literature. The integrated network is frequently updated with updates of its dependent resources and the list of integrated databases may be appended with additional resources.

Provided with the integrated network, one can intuitively conjecture that nearby concepts in the integrated network are related. Indeed, since functionally related concepts are connected in the graph, we may assume that concepts that are close but only indirectly related in the network may also be functionally related in the real world. However, empirical analysis of the network shows that most of the concepts in the network are interconnected in only a few steps. This indicates that the network shows so-called small-world properties. Indeed, there is a considerable abundance of highly connected nodes. For example, interactions of proteins with water and ATP compounds or functional annotations such as the location of a protein in membranes or protein binding are prevalent. These unspecific hubs serve as ubiquitous connections mediating short path lengths between functionally unrelated concepts. This characteristic of the network prevents successful prioritization using simple shortest path methods. Still, our prioritization technique relies on the detection of nearby concepts in the network with respect to a source concept, but we correct the ranking of concepts for their global accessibility in the network.

We provide a short technical summary of the methods here, but refer the interested user to the full implementation details in the Additional materials and methods in Additional file 1. We utilize stochastic random walks (trajectories on the network that consist of taking successive steps from one entity to a random related entity) on the knowledge network to measure the a priori importance or accessibility of concepts in a graph. This technique determines the global centrality of concepts in our integrated network. For this purpose, we compute the limit distribution that yields the probability of visiting the concepts when performing an infinite random walk on the integrated network. Google's PageRank algorithm 10 adopts a similar link analysis algorithm to rank web pages by their relative importance. Network hubs (top ranked concepts with a high prior probability) are generic and unspecific target concepts in the network (Additional Table 1 in Additional file 1). These hubs indicate important concepts for diverse biomedical processes, but should be avoided when trying to find relevant and non-obvious links between seemingly unrelated concepts.

For computing the vicinity of targets to a source concept in similarity to the prior probabilities, we compute the limit distribution of a stochastic model of random walks with restarts in the source concept (with probability 0.25 at each step). As such, we compute the a posteriori accessibility of each concept from the source concept, measuring the probability of visiting each target concept from the source disease, pathway, and so on. Concepts are scored by their posterior probability, divided by the square root of their respective prior probabilities and ranked with respect to this resulting score. In practice, for a gene prioritization query, a user of the web application provides a 'research subject' (for example, a disease, but also a pathway, a GO annotation or a gene may represent a research subject) and a list of 'research targets' (for example, putative genes or compounds) that need to be ranked in relation to the research subject. Our algorithm then assesses and ranks the relations between the source concept and each of the target concepts as above. Since any type of concept can be provided as the subject or target of a prioritization, our method does not require prior domain knowledge from the user, that is, there is no need to define a gene set of known disease-causing genes for the identification of related genes, which results in a more reproducible and robust user experience.

The method of performing random walks to determine the accessibility of target concepts implicitly generates ensembles of indirect paths between source and target concepts, which may serve as functional hypotheses for highly ranking targets. We can heuristically determine highly probable simple paths, that is, paths that do not contain cycles, of the random walk that starts in the source concept and ends in the target concept by adopting backtracking (Figure 2). The backtracking heuristic incrementally builds partial candidate paths, starting from the target to the source, while abandoning least likely paths along the way, leading to valid and specific paths that offer incentives for further functional research. A detailed description of the heuristic is available in the Additional materials and methods in Additional file 1.

The resulting set of paths is presented to the user as a network with putative hypotheses linking the source to the target. Each directed edge represents a supporting relation among intermediate concepts, with annotated semantic meanings and literature references intelligibly supporting the relation for evaluation by the user. In cases where the target is highly ranked, specific and relevant connections and concepts are included in the constructed hypotheses. If the functional hypotheses linking concepts is limited to visiting general hub concepts, this is usually a sign that the linked source and target concepts can be considered unrelated, reflected by a bad ranking score.

As a first application, we analyzed the performance of our platform in prioritizing known disease genes among all genes in our integrated knowledge base. For testing a known disease-gene association, we first removed the link between the disease and its susceptibility gene from the knowledge base. We then ranked all genes in the network in relation to the disease and evaluated the ranking of the test gene. If the test genes ranked high, BioGraph allows retrieval of these genes as valid disease genes based on integrated information linking the disease to its susceptibility gene. For this test, we adopted published benchmarks and compared our prediction performance to that of Endeavour.

Endeavour is a related and mature technology for gene prioritization 6, which adopts a data fusion method to build statistical models of known disease-causing genes with respect to various data sets. Using order statistics, genes are prioritized by measuring the matching quality of test genes to these training profiles.

We have computed the performance of BioGraph's prioritization method for the disease-gene prioritization benchmark initially published to evaluate the performance of Endeavour. This benchmark consists of 627 genes known to cause 29 diseases, selected from the OMIM database, of which 609 disease genes are present in our integrated knowledge base 6.

Benchmarking a disease gene with BioGraph requires that each disease gene is evaluated by first removing the direct relations between the gene and the disease from the integrated network to ensure that the relation to be prioritized is not already in the network. Moreover, variants of the disease (for example, subtypes or syndromes that have the disease as one of its symptoms) are also disconnected from the disease gene. In order to identify these related diseases, we have selected the diseases for which at least one of its UMLS synonyms has the original disease's name as its substring. For example, we identify Charcot-Marie-Tooth disease, type 4C as a related disease to Charcot-Marie-Tooth or Alport's disease as related to Deafness, since a synonym of Alport's disease is Nephritis with nerve deafness. This method provides an objective interpretation of the benchmark by guaranteeing that no prior direct information can be exploited by our prioritization algorithm. Subsequently, for the prioritization of a disease gene, we perform a ranking of the integrated network's 16,912 known human genes with respect to the disease concept.

For benchmarking Endeavour, each disease-gene relation was tested by removing the gene from the disease's known gene set, by training Endeavour on the remaining disease genes and by ranking the gene among a set of 99 random test genes. For both platforms, we adopt the AUC for analyzing the quality of these prioritizations.

The mean AUC for BioGraph's prioritization of disease genes among all human genes is 92.92%, where the reported AUC for Endeavour in prioritizing disease genes among 99 random genes is 86.6% 6. Additional Table 2 in Additional file 1 lists the AUC scores for the prioritization results per disease. Of the 609 disease genes in the benchmark, 181 prioritizations (29.72%) are ranked in the top 1% of the test set of all genes and 449 (73.73%) are ranked in the top 10%. In other words, in an experimental application where a causative gene is among a set of 99 random genes, BioGraph is consequently expected to rank the defecting gene as the top gene in 29.72% of the cases and in the top 10 with probability 73.73%.

The benchmark indicates that our prioritization approach yields a considerable improvement over mature technologies. There are two noteworthy differences in the experimental benchmarking design. Our platform does not require a training set of known disease-causing genes since it will implicitly base prioritizations on integrated disease-gene associations in addition to other heterogeneous types of integrated knowledge of the disease. This has a major advantage for the user since no prior knowledge of the disease is required. Secondly, our platform provides a ranking of the disease gene in relation to all known genes, where Endeavour ranks disease genes among a random set of 99 non-disease genes.

As a quality control of the integrated databases, we have assessed the effect of each database on the benchmarking results by leaving out one database at a time and by assessing the prioritization algorithm on the Endeavour benchmark. This experiment showed that none of the included databases significantly harms the overall prediction capabilities. Conversely, it should be noted that some databases (most specifically CTD gene-disease, GOA and Medical Subject Headings) are essential for successful prioritization, since leaving out these databases has a significantly negative impact on the benchmarks. More information on these quality checks is available in the Additional materials and methods in Additional file 1.

In the above benchmark tests, well-known disease genes are expected to rank high. Indeed, important susceptibility genes usually become the subject of intensive research efforts. Consequently, a literature and database bias may exist toward indirect evidence linking a gene to a disease in the integrated databases. Since BioGraph is capable of using this indirect evidence, the literature bias of important disease genes may strengthen the predictive power of our algorithm. To remove this bias, we can more objectively evaluate the platform by ranking recently discovered disease-gene relations that are not present in the knowledge base.

Provided with the integrated network for which the resource datasets were frozen in March 2010, we identified all recently curated additions of human disease-gene relations from the July 2010 releases of the OMIM Morbid Map (15 new disease genes) and CTD (830 direct, non-inferred relations) that are not present as direct relations in the knowledge base from March 2010. This yielded 845 recent disease-gene relations for which the ranks in the disease's genome-wide prioritization have been determined based on the integrated network of March 2010.

Figure 3 shows the ROC curve of the combined results, with AUC 86.14%. Of the 845 curated disease genes, 189 prioritizations (22.73%) are ranked in the top 1% of the test set consisting of all genes for its corresponding disease and 524 (62.01%) are ranked in the top 10%. The median rank of a disease gene is in the top 6.04%.

To illustrate possible applications of the framework, we have employed the platform to predict candidate genes for schizophrenia (SZ) and substantiate the top predictions with support adopting the automatically generated functional hypotheses.

SZ is a common neuropsychiatric genetic disorder with approximately 1% prevalence and with 64% heritability. It is characterized by a constellation of symptoms, including hallucinations and delusions, and symptoms such as severely inappropriate emotional responses, disordered thinking and concentration, erratic behavior, as well as social and occupational deterioration 1314.

The newly identified genes are indirectly inferred from the integrated knowledge, but not directly associated in our gene-disease resource databases. The predictions in this section are based on a dataset freeze of the integrated databases from September 2009. This data freeze allows us to test if the predicted genes have been observed in genetic studies since the data freeze. Table 2 shows the top 20 inferred BioGraph genes with respect to the SZ concept, as designated by its UMLS accession ID [UMLS:C0036341], and a short summary of the hypotheses of their relatedness with SZ by our platform.

PRL, the top inferred gene that is not a known disease-causing gene for SZ, encodes the prolactin hormone, of which the most commonly known function is to stimulate lactogenesis. Prolactin's relation to SZ is important, especially due to the effects of dopamine-regulating drugs aripiprazole and risperidone on the expression of prolactin and their adverse hyperprolactinemia-associated side effects 15 where the secretion of prolactin is regulated by dopamine, following the current dopamine hypothesis of SZ. We did not find published association studies of prolactin with SZ, although an association with autistic spectrum disorder was reported 16. Additionally, PRL is located on chromosome 6p22.3, which is linked to SZ through DTNBP1 17 and ATXN1 18. Although no causal associations have been shown between prolactin function and SZ, BioGraph hypothesizes PRL as a likely candidate gene for SZ.

The automatically inferred hypotheses by BioGraph that support the high ranking of PRL for SZ are along the lines of current understandings and are schematically shown in Figure 4. The most likely indirect links between SZ and PRL are through the antipsychotic compounds aripiprazole and respiredone, which are both dopamine antagonists affecting the expression of prolactin. This hypothesis also shows that both compounds are adopted as drugs for attention deficit disorder, Asperger syndrome and autistic disorders. Additionally, PRL is associated with autistic disorder, strengthening the importance of it for psychiatric disorders. Additional paths from SZ visit SZ-associated genes and commonalities among these genes with PRL; TAAR6 and PRL are both genes in the neuroactive ligand-receptor interaction pathway; CCL2 and PRL are both regulated by 8-bromo cAMP, a derivative of cyclic AMP; DRD3 and PRL share the GO annotation 'Regulation of multicellular organism growth'. These relations may serve as indicators for determining the putative functional involvement of PRL in the etiology of SZ.

Figure 5 provides hypothetical evidence for the involvement of the second inferred candidate gene HTR1A (serotonin receptor 1A) with SZ. The main hypothesis is driven by the receptor's interaction with the antipsychotic drugs aripiprazole and chlorprothixene. HTR1A is additionally linked to its paralog HTR2A, a known susceptibility gene for schizophrenia, via GO annotations on serotonin binding activity. Although our integrated disease-gene databases (OMIM Morbid Map and CTD) have not indexed HTR1A as a schizophrenia susceptibility gene, variants in the gene have previously been shown to be associated with schizophrenia and other psychopathologies 19. This example shows that BioGraph is capable of identifying known disease genes, even if these gene-disease associations are not in the integrated resources.

Significant associations with SZ and polymorphisms in 4 of the top 20 ranked genes, namely HTR1A 19, DRD2 20, DNMT3B 21 and ATXN1 18, have previously been shown. These known disease-gene relations are not indexed by our integrated databases, but were successfully prioritized by our data mining platform. Most notably, significant association of polymorphisms in DNMT3B with SZ was only reported in October 2009, where the data for our predictions are based on integrated databases from September 2009, demonstrating the usefulness of the currently proposed prioritization technique. Additionally, the highly ranking OMG gene has been shown to be associated with SZ, warranting replication studies for confirmation 22. SYN3 is the only gene in the top 20 for which several association studies have been performed, but where the findings show no support for SYN3 as a SZ susceptibility gene 23.

For the remaining 15 of the top 20 genes, we did not find published association studies to support or contradict possible roles of these genes in SZ, although for some genes associations with SZ-like symptoms or related psychiatric and various neurological disorders have been shown, supporting the putative role of these genes in SZ (Table 2).