Monoclonal antibodies directed against protein antigens were obtained from many different commercial sources (listed in the AbMiner program itself), with no particular selection criteria except that antibodies directed against very small (~ 10 kDa) or very large (> 350 kDa) proteins were excluded. Species recognized included human, mouse, rat, dog, chicken, frog, and others. Each vial of antibody was assigned a unique AbMiner identification code number so that screening of the particular vial could be tracked. Pertinent information indicated in the schema in Table 1 was also recorded.

To validate antigen-specificity by Western blot against a wide range of human antigens, we harvested cell lysates from each of the NCI-60 cell lines as described previously 2324 and generated a standard pool (NCI-60 pool) containing approximately equal amounts of each of the 60 lines. Included in the pool were leukemias of several lineages, melanomas, and cancers of breast, ovary, prostate, colon, lung, kidney, and central nervous system. After electrophoresis and blotting, each gel yielded 16 four-mm wide nitrocellulose strips, each of which was used to screen a different antibody (Fig. 1a). To increase the likelihood of detecting low-abundance proteins, we loaded the gels quite heavily so that each small strip contained 0.6–1.3 μg of pool protein. Up to 64 antibodies were screened simultaneously by incubating each strip in its own miniature incubation chamber (BioRad, Hercules, CA, USA) (Fig. 1b).

The Western blot results were classified as follows: (a) single band: one predominant band, at the expected molecular weight; (b) multiple bands: extra bands remaining with a 5-second or longer film exposure; (c) wrong molecular weight: predominant band or bands at unexpected molecular weight(s); and (d) no band. Antibodies were Western-blotted up to three or four times if required to obtain clear results.

We focused on Western blot analysis and designed the screening process as we did because of the specific requirements for application of antibodies on reverse-phase lysate arrays. Antibodies that recognize an epitope from more than one protein (or isoform) can be used for detection and quantitation of proteins on a Western blot as long as any extraneous bands have different effective molecular weights and would show up as separate bands. The lysate array, in contrast, is effectively a multiplexed dot blot; the signal from each spot on the array is the summation of specific and non-specific binding of the antibody. Therefore, for the lysate arrays we used only screened antibodies that produced a single predominant band by Western at the expected molecular weight. However, antibodies that produced multiple bands may still be useful for other applications, so information on them is retained in AbMiner. Other types of quality control data (for example, based on immunoprecipitation, immunohistochemistry, or flow cytometry) may be most pertinent to other types of applications. AbMiner is extensible in that data fields can be added to accommodate and present such information.

As indicated in Fig. 2, approximately 70% of the commercially available antibodies we screened produced a single predominant band against the NCI-60 lysate pool; only 4% resulted in absent or questionable bands. Some antibodies that produced a single band by Western blot were expressed in only one or two cell lines out of the 60 on the NCI-60 lysate arrays (Fig. 1c), indicating that the validation method is quite sensitive. Since each cell line provides about 1/60^th of the pool's protein, we cannot rule out the possibility that a relatively faint extraneous band on the western would reflect a grossly incorrect target protein value for one or two cell types on the array. To check that possibility would require doing 60 times as many Westerns (or pools of small numbers of cell types perhaps). In principle, that would be desirable, but the uncertainties of the present quality control procedure are still considerably less than those inherent in many other transcriptomic and proteomic profiling technologies. As noted previously, the user will presumably reassess candidate antibodies on his or her particular experimental system in any case.

AbMiner's Gene Information database provides translation among different data platforms and makes it possible for the user to search by proteomic, transcriptomic, or genomic identifiers. To find the intersection between data sets from different platforms – such as cDNA 2526 and oligonucleotide microarrays 27– one generally must translate from one type of unique identifier to another 15. Finding an antibody that corresponds to a particular gene can be problematic because many commercially available antibodies do not have unique, universally used names that represent the target gene product.

To match antibodies names with the appropriate identifiers of the corresponding gene, we used MatchMiner 15. MatchMiner translates among various gene and protein identifier types, including HUGO name, common names, aliases, chromosome locations, GenBank accession numbers, Affymetrix identifiers, and IMAGE clone IDs (Fig. 3). For antibodies whose names could not immediately be translated with MatchMiner, other resources were used to cross-check and select the appropriate symbol 161718192021222829. The gene symbols were then matched with the other identifiers as described above. The final MatchMiner output included the gene name and symbol, the HUGO approval status of the symbol, mRNA and protein RefSeqs, UniGene Cluster IDs, and LocusIDs. Through a many-to-many relationship with the Antibody Information File in AbMiner, any antibody can be associated with its appropriate gene identifiers and vice versa. That feature of AbMiner (see Table 1) facilitates navigation among antibody, antigen, protein, DNA, and RNA data, allowing for integration of information from disparate data sets.

AbMiner is a relational database comprised of two major components: (i) a data entry module constructed using FileMaker Pro5.0™ (Santa Clara, CA USA) and used by our team for data entry as well as for detailed tracking of the antibody validation process.

(ii) a web application for sharing the various types of information on antibodies, antigens, and genes with the research community. The web application, written principally in Java, leverages a variety of available resources: MySQL as the database engine; Hibernate to map the objects into the database; JSP, Struts, and Tiles to render the user interface; and JUnit and HTTPUnit for testing individual programming units and the overall system. In addition to providing the web user interface, we have defined a simple HTTP specification that facilitates linkage of other applications into AbMiner. The web application was constructed under the "Agile Development" paradigm, which encourages close, iterative interaction between user/tester/motivators of the package (biologists) and software engineers 30. That interaction, and the continuing revision of specs that the agile process encourages, ensure that AbMiner will serve broad needs of biological researchers. It has been received enthusiastically in extensive beta-testing.

Both components of AbMiner have the same underlying model, which includes three main modules: Antibody, Screening, and Gene. AbMiner uses a relational database approach to manage the complex relationships among those elements. The relationships are generally not one-to-one. For example, a given gene often codes for different splice variants, which may or may not be recognized by the same antibody. Conversely, multiple antibodies from different vendors or hybridoma clones may target a protein encoded by a single gene. An additional complication is that, because of the continuing re-annotation of the human genome, some identifiers are not unique or constant. That dynamic process is exemplified by retired or relocated UniGene clusters that can sometimes result in more than one UniGene Cluster ID or LocusID entry for the same gene. By constructing AbMiner as a relational database, we have been able to organize and update those one-to-many, many-to-one, and many-to-many relationships.

Currently, AbMiner is populated with screening data generated by our own laboratory, but we plan to incorporate data from other studies and repositories when available. We also plan to put data entry pages on the web component for input of screening information from other investigators in the research community who wish to contribute (with appropriate attribution)

The identifiers described in the last section perform an additional function by serving as link-outs to their respective entries in LocusLink, UniGene, GeneCards, Entrez's RefSeq, and our MedMiner program. AbMiner also provides links to the Mammalian Phosphorylation Resource (MPR) 31, a web site that contains sequence information for phosphorylation sites recognized with specificity by commercial antibodies, and links to the Clinical Proteomics Databank 32, which provides a list of phospho-specific antibodies tested and used in the Clinical Proteomics Program of the NCI. Other public and commercial antibody databases, such as the Antibody Resource page 33and Abcam 34, are also linked. Finally, AbMiner will serve as the central public database for an antibody repository planned by the NCI Center for Cancer Research (CCR).

As already noted, development of AbMiner was motivated by our need to organize information on antibodies for lysate array studies, and it has proved itself an almost indispensable tool in that respect. Particularly important is the information on the correspondence between antibody names, antigen names, and the variety of gene identifier types. We were able, for example, to address the question of similarities and differences between mRNA and protein expression profiles across the NCI-60 12. Identifiers of proteins quantitated on lysate arrays were matched with identifiers of transcripts assessed on spotted cDNA arrays (i.e., Image Clone Ids) and Affymetrix oligonucleotide arrays (i.e., Affymetrix Ids) using MatchMiner 15 and AbMiner. A central, unexpected finding was that cell-structure-related proteins showed higher correlation between protein and mRNA levels across the 60 cell lines than did non-cell-structure-related proteins 12. Using the annotations and translation capabilities in AbMiner, those analyses have now been extended to 89 proteins detected on the lysate arrays by 154 different antibodies (Shankavaram et al. manuscript in preparation).

We have also applied the resources of AbMiner to the identification of molecular biomarkers at the protein level. For one such study, we developed a multi-step "integromic" protocol 1430, which included: (a) identification of candidate markers using cDNA microarrays; (b) re-sequencing of candidate clones; (c) corroboration of the candidates' expression patterns from the cDNA microarray using Affymetrix oligonucleotide chips; (d) protein expression analysis using reverse-phase protein lysate arrays; and (e) prospective validation of candidate biomarkers on tissue microarrays consisting of hundreds of tumor samples. With that algorithm we identified villin and moesin as molecular markers that distinguish between colon and ovarian adenocarcinomas. Those cancer types can be difficult to distinguish in a few percent of metastatic or disseminated lesions in the abdomen, and the differential diagnosis is important because it determines what drugs will be used for therapy. Our protocol was successful in that case, but it depended on the availability and effective screening of quality antibodies for identification of diagnostic markers at the protein level on the lysate and tissue arrays. AbMiner gene identifiers will help other investigators in similar searches for molecular markers at the protein level, even when the search has begun with genomic databases. Because AbMiner provides extensive information for over 600 validated antibodies, the transcriptional signature of a gene can often be corroborated directly at the protein level.

The protein-mRNA expression studies are continuing, with promise. More than 500 of the 635 antibodies currently in AbMiner have UniGene Cluster IDs that match those from four different NCI-60 microarray platforms (see Table 2 and 3), providing a foundation for cross-comparison, for validation studies, and for hypothesis generation. AbMiner is contributing both to the selection of antibodies and to the crucial link between protein and transcriptional data.

AbMiner is certainly not the most comprehensive in terms of numbers of antigens or antibodies covered. In that regard, the classic source is Linscott's Directory 35. The Advanced Type Culture Collection (ATCC) 36 keeps listings of large numbers of their available cell lines, including hybridomas, and the Antibody Resource Page 33 and Abcam 34 provide useful databases on antibodies. AbMiner includes links to all of those sources. A number of companies provide databases of the antibodies they sell, but those will not be reviewed here. AfCS Signaling Gateway provides 38 provides information on 138 proteins (principally in the signaling pathways) and antibodies against them. Western Blot quality control information, generally on one or a few cell lines, is included. Exactantigen 39 provides gene-specific and species-specific information on antibodies, with links to manufacturers' data sheets. The useful Human Protein Atlas 4041 features immunohistochemical images for a variety of newly generated and other antibodies, complementing the focus of AbMiner. There are also a number of specialized antibody collections (e.g., on 3-D structures of antibodies or on neurological or HIV-related reagents) 4243, but none that we have seen present ranges of information similar to that of AbMiner. It would be well beyond the scope of this article to review those databases, but a number of them are described, with outlinks, by Linscott's. AbMiner's database will continue to expand, but not with the intention of competing with Linscott's in coverage.

Overall, to the best of our knowledge, none of the other sources have the range of information types on the antibodies, the antigens, the vendors, and the antigen's genes that AbMiner does, and none of them give the type of multiple-tissue Western blot specificity data or protein microarray data that are compiled in AbMiner. Insofar as we have found, any Western blot results given in the other databases had been obtained against single cell types. The quality control criteria represented in AbMiner are stricter and more comprehensive in that we have validated the antibodies by Western blot against a pool representing a wide range of cancer cell lines from nine different organs of origin and from different cell lineages. Non-specificities showed up in that more rigorous testing when they didn't in testing against individual cell types.

AbMiner has unique relational characteristics for dealing with the one-to-many, many-to-one, and many-to-many relationships among antibody reagents, their antigens, and the genes of those antigens. Through the use of MatchMiner 15, supplemented by manual curation from additional bioinformatic resources, AbMiner gives a useful range of gene identifier types not otherwise easy for the casual user to find. The Antibody Resource Page provides a listing of "databases/software" on immunological reagents 33, but none of those listed have major overlap with AbMiner in terms of their program and search capabilities. We are currently using the structure of AbMiner as template for an analogous database on siRNA reagents.

We developed AbMiner as we did to provide the type of information needed for "integromic" 3044 studies of the type described above for biomarker identification – that is, for the integration of different types of molecular data at the DNA, RNA, protein, and functional levels. But the program is also being found useful (in beta-testing) by researchers with simpler aims: e.g., those who simply want to find the right antibody for an assay.