One of the more painful tasks in 'omic' research 12 is navigating among different gene or gene product identifiers. After a cDNA microarray experiment, for example, one usually must translate from IMAGE clone ids to GenBank accession numbers, HUGO names, common names, or chromosome locations for a list of genes. As we generate more and more data from diverse platforms and species, such translations will become increasingly complex but also more important to the synthesis of a coherent biological picture. Beyond simply looking up additional information about a list of genes, such synthesis will require the ability to find the intersection between two lists of genes that are designated by the same or a different identifier type.

Currently, the basic translations can be done on a gene-by-gene basis using public databases such as UniGene, LocusLink, OMIM (Online Mendelian Inheritance in Man), and the working draft of the human genome (from the University of California Santa Cruz (UCSC) or the National Center for Biotechnology Information) 345 or else in batch through Source 6 or GeneLynx 7. However, no single data source contains all the necessary information about every gene and, to complicate matters further, the relationships among identifiers are often not one-to-one. For example, there may be several GenBank accession numbers and multiple IMAGE clone ids for the same gene, and a single gene symbol may be an alias for multiple different genes. Therefore, any high-throughput solution to the problem must take these challenges into account and respond with an approach that minimizes the need for human intervention. At the same time, those instances when human intervention is necessary must be flagged and enough metadata must be provided for accurate decision-making without extensive further research.

Motivated by many days spent at the computer doing these tedious, time-consuming translations for our own experimental data, we developed MatchMiner 8 as a freely available public resource that automates the process for collections of genes. MatchMiner provides two primary functions. The first, LookUp, translates an input list of gene identifiers into a matching output list of identifiers of a different type; the second, Merge, combines two separate lists of either the same or different types of identifiers into one list that details all one-to-one, one-to-many, and many-to-many relationships between corresponding gene identifiers in the two lists.

As shown schematically in Figures 1 and 2, MatchMiner leverages information from the four public databases listed above, and from Affymetrix, by parsing them into relational tables for use in doing translations. The LookUp function can operate interactively on single identifiers or in batch mode on a list of identifiers in a file. When used interactively for one or a few genes, it saves the user the trouble of querying five different databases and collating the data. More important, however, is batch querying of a list file, for instance a list of the dozens or hundreds or thousands of genes that show interesting differences between samples in a microarray experiment. In this mode, the user specifies the input and output identifier types, as well as the search algorithms to be used in traversing the various data sources (Table 1). The program is context-sensitive in that it will search only the pertinent data sources (for example, only UniGene to identify IMAGE clone ids, which are not found in the other sources). An important feature is the optional output of diagnostic metadata that tell the user in which source (s) the identifier was found and whether an input identifier corresponds to more than one gene. This feature enables the user to judge the reliability of matches. The results can be displayed in HTML format or downloaded as tab-delimited text suitable for direct entry into a spreadsheet program. A summary indicates the number of successful and unsuccessful translations.

The Merge function, the most powerful function of MatchMiner, identifies which genes are common to two input lists of identifiers and gives detailed output of the one-to-one, one-to-many, or many-to-many relationships between corresponding identifiers in the two lists. This function is used, for example, to compare datasets of different experiment types (for example, transcript expression, protein expression, array-based comparative genomic hybridization (CGH)) by identifying the genes in common between them. The output includes summary tallies as well as a gene-by-gene listing of items matched, unmatched and not found. As with the LookUp function, diagnostic resource information is provided. Any identifier with an ambiguous gene assignment (for example, an IMAGE clone id that belongs to two different UniGene clusters) is flagged for user intervention, with all possible assignments returned.

In one illustrative case that motivated development of MatchMiner, we (X. Lee, K.J.B., F.G. Gwadry, W.C.R., G. Riddick, S. L. Pelletier, S.N., and J. N.W., unpublished data) had to match up as many as possible of 9,706 cDNA microarray clones 910 with HU6,800 Affymetrix chip oligonucleotide sets 11, having run both platforms on the same 60 human cancer cell samples (the NCI-60). To do so, we developed an early form of MatchMiner. The particular task was to identify all relationships between the 9,706 IMAGE clone ids and 7,129 GenBank accession numbers based on UniGene cluster membership. To complete the task manually, one gene at a time at maximum speed (about 30 seconds per gene) would take over 140 hours - even if one could keep accurate track of the results. In contrast, the current version of MatchMiner took 10 minutes on a 750 MHz Pentium III PC with 320 MB RAM to generate the merged list, specifying all possible matches between IMAGE clone ids and GenBank accession numbers. When we compared MatchMiner Merge results with those obtained using the LookUp function for a random sample of the genes, there were no discrepancies. The same task with Source required translating both lists into UniGene clusters and then further processing the data. After identification and reformatting of entries with multiple UniGene cluster associations, the resulting lists were imported into Microsoft Access and queried to create the appropriate matches. The entire procedure gave results similar to those of MatchMiner but took approximately one hour, most of that user time.

With the exception of MatchMiner, tools that can do some kind of translation are geared toward research dealing with expressed sequence, either at the RNA or protein level. However, many interesting questions can be asked from the perspective of genomic sequence. One example relates to the identification of genes represented in an array CGH experiment in which the targets on the chip are fluorescent in situ hybridization (FISH)- and site-tagged sequence (STS)-mapped bacterial artificial chromosome (BAC) clones. The challenge is to begin to interpret array CGH results in the context of the biological literature and of other classes of data. BAC clones are not generally annotated by the genes they span, but rather by their position in the cytogenetic and sequence-based maps. Therefore, an association between the BAC clones and genes must be made. MatchMiner provides this function with the ability to search on the FISH clone ids. Mapping of the FISH clones to genes is done by sequence alignment of the BAC ends during off-line construction of the overall MatchMiner database (Figure 3). MatchMiner takes 5 minutes to return the gene symbol for a list of 100 FISH-mapped BACs 12. Such a search is not possible using other tools. A summary of commonly used analogous tools and their capabilities can be found in Table 2.

As noted previously, identifiers are not always unique or uniquely assigned. For example, GenBank accession numbers are specific to a sequence, but the assignment of that sequence to a gene may change over time. Even more disconcerting, common gene names or aliases are often used by different investigators for different genes. Therefore, it is important to look in detail at the results of searches to check for correspondences other than one-to-one and to examine the data source tags to get a sense of the strength of the association between identifiers.

One non-obvious advantage of MatchMiner is that it can combine information from more than one of the data sources to show matches that could not be made on the basis of any single source. The gene ACVR2B, which has aliases ACTR-IIB and ACTRIIB, provides an example. LocusLink and OMIM both reference the HUGO symbol ACVR2B, but LocusLink does not reference ACTRIIB, and OMIM does not reference ACTR-IIB. Therefore, if one of the aliases were used as input, the success of any search outside of MatchMiner would be data-source dependent.

MatchMiner was written in Java and can be deployed as either a web or command-line application, the latter suitable for high-throughput pipeline purposes. In its design and implementation, we leveraged a variety of open-source tools and libraries, including jUnit (unit testing framework), CVS (configuration management), Xerces (XML parser) and Ant (build tool). Before run-time, data from UniGene, LocusLink, OMIM, UCSC and Affymetrix are downloaded and parsed to generate an integrated database implemented under MySQL. If an entry from the imported data matches a candidate gene that was already identified, it is assigned the same gene index. If an entry does not match any of the candidate genes, then a new gene index is generated. Import begins with data from the UCSC's 'Known Gene' table, followed by UCSC's EST (expressed sequence tag) table, LocusLink, UniGene, OMIM and Affymetrix. Different identifiers are stored in different tables and several tables are required to resolve the many-to-many relationships between identifiers (Figure 2a,b). The central algorithm for resolving identifiers uses an instantiation of the ChainOfResponsibility pattern 13, which combines different searches sequentially in a logical manner. In MatchMiner, it maximizes the likelihood of correctly translating back and forth from identifiers to gene indices using the different databases. The algorithm is non-trivial. For each identifier type, we establish a ChainOfResponsibility hierarchy of the data sources based on their respective abilities to match the user input (Table 3). The search algorithms then use this ranking. For example, when an input list of gene names is processed using the 'ALL (HUGO, Alias)' search algorithm, the list is scanned for HUGO names, and each one found is associated with the corresponding unique internal gene index. The remaining unmatched gene names are then scanned again, this time matching aliases (Table 1). The rationale is that an official HUGO name is more likely to be the desired match, but any match is better than none. A similar approach is used when going from the unique index to an output list. For instance, if the desired output is cytogenetic location, MatchMiner first scans the UCSC build of the human genome. If the location is not found there, LocusLink and Unigene are searched (Table 1). The ChainOfResponsibility approach enables us to combine the precision of highly focused algorithms with the greater coverage of more broadly based ones.

Although currently human-specific, MatchMiner will be expanded in the near future to incorporate data from other species, with emphasis on mouse. Additional features to be implemented include the ability to handle lists of mixed types of identifiers, the ability to request multiple types of identifiers within a single search, and the incorporation of additional public sources for use in making translations. We will continue to enhance and develop MatchMiner under a contract funded by the Center for Cancer Research of the US National Cancer Institute.

MatchMiner is available as a web-application or as a command line jar file at 8. The MatchMiner database is maintained on our server and updated at approximately 6-month intervals. Detailed documentation for both implementations is available at the site.

In summary, MatchMiner is an efficient application for navigating the complex world of gene and gene product identifiers. It can batch search publicly available databases to convert between identifier types and can determine the intersection of two gene lists with different identifiers. MatchMiner will greatly enhance the ability of the research community to annotate and compare 'omic' datasets.