Background

The use of high throughput genomic and proteomic technologies has resulted in the creation of large datasets of differentially expressed genes and proteins. Even after further statistical analysis these datasets may be sufficiently large such that the validation of all possibilities are outside the resources of the investigators. In these situations there is a need to efficiently triage and sort the dataset to identify the genes of highest interest to the scientists. In many situations the experimental design takes advantage of specific biological samples available to the investigator. Thus the investigator often has additional scientific data and personal insight that may be helpful in guiding the examination of the genomic output. Yet many tools developed to sort and add supplemental information to the genomic data use global processes such as metabolic pathway mapping [2–4], promoter binding [5], chromosomal location, or Gene function/GO terminology [6, 7], and thus may not leverage the additional knowledge of the investigator. This leaves the scientist with the time consuming task of manually sorting through the dataset with the appended data to identify genes that may provide useful information. Here we present an automated tool, BEAR GeneInfo, which allows a user to simultaneously query the biomedical literature with lists composed of multiple gene names while using additional tailored search terms. The associated output of biomedical references is provided for further review and subsequent query modification, allowing the user to follow-up on interesting trends in the data, thus maximizing the potential of the genomic data. This tool joins the list of additional tools including PubMatrix [8], MatchMiner [9], and XplorMed [10] that are enhancing the ability of scientists to perform integrated searches of large complex datasets, and by doing so identify new trends and associations within the scientific data.

Implementation

Interface and database design

BEAR GeneInfo consists of five components (figure 1); A web based interface for user data input and results display, a CGI script for user data processing and results display, an underlying database to store gene related information, Perl scripts for database maintenance and data updates, and link-outs to NCBI. The database architecture was created in Rational Rose using an object oriented design and implemented in Oracle 9 (figure 2). The database was populated through downloads and updates derived from Unigene [11], Locuslink [12] and the UCSC Genome Browser [13]. Additional gene names were identified through MatchMiner [14] based on queries of individual Unigene IDs. These MatchMiner queried gene names were designated as gene name aliases, and used to populate the Unigene (gene name) alias table (figure 2). In order to limit the number of GenBank accession numbers displayed in association with a given Unigene or Locuslink ID, SOURCE [15] was queried with the respective IDs and their defined "representative GenBank accession number" was downloaded [16]. The associated GenBank accession numbers, Unigene IDs, and Locuslink IDs were preserved within the tables for use in the querying of user uploaded gene lists.

Figure 1
figure 1

GeneInfo user interface

Full size image
Figure 2
figure 2

GeneInfo database in Rational Rose

Full size image

Pathways for database queries to derive gene name lists

BEAR GeneInfo allows for the querying of PubMed with gene names and gene symbols based on the entry of either Unigene ID, GenBank accession number, or LocusLink ID, and as such pathways for the identification of gene names based on each type of data entry were created. If the user enters in Unigene IDs which have been retired BEAR GeneInfo uses a CGI subroutine to query Unigene and updates the Ungiene ID, using the new Unigene ID to continue the query. When the list of genes is provided as either GenBank accession numbers or Unigene IDs the queries utilize gene names associated with the Unigene IDs, and also identify gene symbols associated with the corresponding Unigene and Locuslink IDs. After these have been identified BEAR GeneInfo adds the gene names derived from MatchMiner based on a query with the Unigene ID. All these gene names and symbols are used to search PubMed. When the list of genes is provided as LocusLink IDs the same process is used, only one uses the official LocusLink gene name instead of the official Unigene gene name. Due to the complexity of the gene naming conventions, we have sought to offer different methods for searching with different gene name options. The lists of gene names are filtered for general "user unfriendly" elements such as parentheses or descriptive words such as "ESTs weakly similar to" or "homo sapiens".

PubMed queries based on the gene name lists

The textual complexity of the gene names can lead to very different search results. For this reason we have incorporated a series of filters into BEAR GeneInfo for use in choosing the format of the gene name searches. There are two choices; a default filter that uses processes we have developed in the construction of the BEAR GeneSifter tool (X. Wen, G. Zhou, H. Liu and M. Datta, unpublished), and a gene search term splitter. The default filter takes a given gene name, and removes any parenthesis i.e. "(", "[", ")", or "]" or semicolons ";" and colons ":" from the text, replaces commas (,) with a Boolean AND term, and places quotation marks (") around the text groups for searching. The gene search term splitter takes the gene name and also removes any parenthesis i.e. "(", "[", ")", or "]" or semicolons ";" and colons ":" from the text, and uses comma replacement with the Boolean AND term, but then takes each search word in the gene name and separates them, and searches each individual term against the user added search terms. The results are then combined and reported for the gene name. In each case an additional general "nonsense" filter is applied, such that if a single search string returns more than 2,500 references, the term is deemed to be "nonsense" and its references are not included in the query results. Figure 3 demonstrates an example of the text filters for a given gene. Of note, different numbers of references can be retrieved based on the choice of search text filters. In addition to the gene names, the user may enter additional search terms based on their specific knowledge of the experimental system. These terms can be entered using Boolean logic, and will be searched with respect to each individual gene name that has been uploaded in an "AND" or "NOT" format. Queries of the NCBI PubMed database are restricted to nights (9 pm – 5 am EST) and parsed over time to limit the user time on the PubMed database. More effective queries for high throughput users would use a locally downloaded copy of PubMed. Results are then sent to the user by email, which directs the user to a web site for receipt of their output data. The output data is stored for ten days on the site before it is deleted. The resultant output is provided to the user in two formats, http and txt. In each case the user receives a list of the genes and the user's additional query specifications along with the number of associated biomedical references. In the http file the user receives hyperlinks to the references while in the txt file the user receives a list of the PubMed IDs for further analysis. In the http file additional hyperlinks are supplied for each individual search term within a gene ID query, allowing a user to examine the effect each search term has on the final compiled list of references for that gene ID (figure 4). The resultant output data can be examined by the user and used to refine subsequent queries that are performed.

Figure 3
figure 3

Example output based on the two different word filtering options. The default filter searches the gene names with the users search terms. The gene search term splitter takes each gene name and separates the terms, combining each with the user search terms, and then combining the results. The default filter results in a higher number of references than the gene search term filter.

Full size image
Figure 4
figure 4

Example output evaluated by individual search queries. In the example the gene ID generated six gene name queries linked to the user generated search term "antioxidant". The individual queries generated different numbers of references, the results of which were combined in the final result. Hyperlinks to each query link to PubMed, allowing the user to determine which queries contribute specific references to the final result.

Full size image

Results

Example use: Selenium treatment of prostate cancer cells

The experimental data presented here is a component of a published study described in the companion article by Schlicht et al. (ref-pending). The methods for the derivation of the data are present in that study. Portions of the data are reanalyzed here and presented not as new data, but as an example of how the BEAR GeneInfo tool can be used. A more detailed study of the data can be obtained from the Schlicht et al. reference (ref-pending). Selenium is currently being investigated as a potential chemopreventive in prostate cancer. While clinical trials have suggested some potential benefit, the mechanism of action of Selenium is unclear, and is an area of active research. Using gene expression microarray technology a dataset reflecting the differential expression of the human metastatic prostate cancer cell line PC3 was developed after treatment with Selenium for either 6 hours (1123 genes) or 5 days (1053 genes). Twenty-four of the genes demonstrated differential gene expression with respect to Selenium at both 6 hours and 5 days. Using this set of 24 genes, additional prioritization and sorting was attempted using the BEAR GeneInfo tool (table 1). A list of the 24 genes was uploaded into BEAR GeneInfo as either GenBank accession numbers, Unigene IDs, or Locuslink IDs, and run in comparison with additional user provided search terms including "prostate cancer", "Selenium", "metastasis", "apoptosis" and "antioxidant". The majority of the genes (average 13.8/24 genes, median 15/24 genes, range 10/24 to 16/24 genes) returned no references for the associated search terms. The average number of references was 5583 references per gene, although the median was 0 references per gene, with a range of 0 to 57,693. Of note, the ability to generate lists of gene names was similar for different gene IDs. The Unigene ID and GenBank accession number yielded similar results. Five genes returned more than 1,000 references for each search. Upon review these gene names identified common phrases or word groups, and as such were disregarded. The evaluation of the returned references was made possible through a search term query, in which each search term used for a given gene ID is hyperlinked to the related references (figure 4). Examining the associated references with the search terms allows the user to identify inappropriately identified references, and thus disregard specific search results. BEAR GeneInfo searches were repeated for the list of 24 genes using the two types of search filters (table 2). In addition, the genes names were individually searched with the associated query terms in PubMed using the EndNote reference program and then number of returned references counted. All the associated references were read to evaluate the relevance of the publications with regard to the search terms. This review confirmed that the return of large (greater than 1,000) references was due to misrepresented words or phrases in genes with complex naming features. A comparison of the search filters revealed different results for the default and gene search term splitter filters, with the latter returning larger numbers of references, but not affecting the most significant numbers of references. These differences did not affect the relevance of the references identified. The results were similar to that obtained by manually searching PubMed. In all three cases where BEAR GeneInfo returned over 1,000 references, no references were identified through manual searching, confirming the non-specific nature of the automated results. Thus it appears that the most interesting genes would be characterized as having limited, but not too much, associated data in PubMed. This would reflect genes where early studies had been attempted, but for which extensive studies have not already been completed (thus leaving more to examine). Alternatively they may represent genes where connections to the user's study have yet to be identified. There was no situation where BEAR GeneInfo failed to identify references while a manual search using Endnote was able to return references. In table 2 using BEAR GeneInfo three genes met the arbitrary criteria of having references for both prostate cancer and the majority of the additional search terms; peroxiredoxin-1, surfeit-5, and lysozyme homolog. Manual searching only identified two of these three genes. The third, the lysozyme homolog, was identified through the use of multiple gene name queries generated by BEAR GeneInfo, and as such identified the significance of this type of protein with respect to prostate cancer and selenium. Thus it appears that the use of an automated system for data mining of the biomedical literature was at least as accurate as manual data mining, and offers significant advantages in time savings. The identification of peroxiredoxin-1 was interesting as there are references studying this gene in prostate cancer in both the rat and the human, a design in the initial study [17, 18]. Peroxiredoxin-1 was also studied in Selenium, but had not been examined with respect to Selenium in prostate cancer. Lysozyme has been used to study cellular differentiation in tumor diagnosis for years, and has been described in variant forms of prostate cancer [19–21]. Its identification here may imply a role not just in tumor identification, but also in function. In a search across multiple terms in (table 1) the identification of an F-Box protein is intriguing as other F-Box proteins have already been implicated in both prostate and prostate cancer cellular functions [22–24]. F-Box proteins have yet to be implicated in Selenium action, but have been identified based on their roles in antioxidant protection [25]. Thus the resulting identification of 274 query-specific references associated with three genes allowed for the focused examination of these genes, and reduced the list of genes to be used in subsequent validation studies by one eighth (87.5%) and provided interesting genes for further validation studies.

Table 1 The Selenium Prostate Cancer Gene Table. Each Unigene ID was searched against PubMed using GeneInfo. Additional search terms were used and the number of references recorded. The default filter was used in all searches.
Full size table
Table 2 BEAR GeneInfo search results with different filters. Results are shown for the selenium prostate cancer gene list searched with the additional terms "prostate cancer" and "selenium". Results are presented using the default or the Gene search term splitter filters, and compared to hand searched results from EndNote. All numbers are presented as (number of references returned/number of relevant references).
Full size table

Discussion

The use of high throughput techniques in biology often results in large dataset outputs where data needs to be triaged and further analyzed. In the process of data triage one hopes to leverage as much information as possible to allow for the correct sorting of genes. This process needs access to comprehensive databases of information on genes so that opportunities, in particular poorly defined or ignored gene targets, are not missed. A rich source of gene data is the corpus of biomedical literature present in PubMed [26]. This corpus provides a large dataset of gene information that can be mined for supplementary information related to genomic data analysis. Approaches have sought to create comprehensive datasets that identify all the relationships present in the biomedical literature between genes, genes and drug targets, and along metabolic pathways [9, 27, 28]. These approaches have used either curated or automated data compilation. The curated databases have been of great value, but are limited by their level of coverage necessitated by the labor-intensive nature of the process. Data mining techniques have been used to identify associated information within the biomedical literature (XplorMed [10]) or between the biomedical literature and lists of gene and drug names (MedMiner [5] and PubGene [29]). The former tool does not allow for the use of large gene names lists while the latter two tools are limited in their inability to modify search terms based on user-specific knowledge, unpublished results, or the unique characteristics of the biological experiment. BEAR GeneInfo is a web based tool designed to allow a more flexible user-driven data mining process. A recent tool that does offer the ability to combine a list of gene names with additional search terms is the PubMatrix system for multiplex literature mining [8]. PubMatrix offers unique value in its integration of genes with user derived search terms, but is different from BEAR GeneInfo in its use of gene names. Both PubMatrix and BEAR GeneInfo can be used to identify references from NCBI PubMed, when queried by gene namelists, while BEAR GeneInfo can derive gene namelists from GenBank Accession numbers, Unigene or LocusLink ID lists. The use of compound search terms, as one uses in the PubMed query interface is common to PubMatrix and BEAR GeneInfo, although the integration of search terms is based on the Boolean "AND" operator in PubMatrix, while BEAR GeneInfo offers both "AND" or "OR" Boolean operators. Both tools are limited by the quality of reference curation and do not use statistical filtering of query results. BEAR GeneInfo will return the search result in a web page display of the PubMed records for review by investigators, while PubMatrix utilizes more graphical based data presentation. The information it provides include two link buttons, one for direct link to the PubMed references, and one for a separate page displaying the URLs for each matched PubMed reference. The search results also display the gene name, and hyperlinked GenBank accession number, Unigene ID and LocusLink ID. BEAR GeneInfo provides an interface for the user to modify the searches with user-driven Boolean search terms along with the gene so the result will yield more focused information for researchers to review. BEAR GeneInfo is available at: http://www.prostategenomics.org

Use of the BEAR GeneInfo system: Selenium treatment in prostate cancer

The true value in the use of a tool such as BEAR GeneInfo is the ability for the user to prioritize and sort genes identified through high throughput methods such as microarray studies. By allowing the user to define additional search terms based on the biological experiment focused searches can be undertaken that leverage the knowledge of the user and allow for the tracing of specific links not always identified through metabolic pathways, chromosomal location, or simple differential expression. This is demonstrated by our identification of differentially expressed genes of interest with respect to Selenium treatment of prostate cancer cells. A weakness of this process is the limitation of the use of known biological links within a system based on the scientific data available to the user. This does limit the possibility of unbiased gene discovery, but this process can still be exercised through the examination of differentially expressed genes by purely statistical methods, without associated gene data present in the biomedical literature and based on previous scientific data. One of the main values of this process may be the ability to link diverse scientific fields and disciplines, with the only limitation being the ability of the user to imagine a possible link when querying the tool.

Conclusions

Here we have presented a tool, BEAR GeneInfo, that can be used to further define a list of genes that have been identified through genomic or genetic studies. Through the use of targeted searches with additional search terms the investigator can filter a list of genes, in the process prioritizing the ones that match their specific research interests. In addition, BEAR GeneInfo provides an initial point for launching further refinements of text based query tools that examine the biomedical literature. The tool is freely available on the web at http://prostategenomics.org[1], and the authors will provide scripts and database components if requested

Availability and requirements

The BEAR GeneInfo database is currently available as a web based tool at http://prostategenomics.org[1] but the components can be obtained after contacting the author at mdatta@mcw.edu. The database is built in Oracle 9 and the scripts are all Perl based.