Volume 10 Supplement 9
Towards large-scale sample annotation in gene expression repositories
© Pitzer et al; licensee BioMed Central Ltd. 2009
Published: 17 September 2009
Large repositories of biomedical research data are most useful to translational researchers if their data can be aggregated for efficient queries and analyses. However, inconsistent or non-existent annotations describing important sample details such as name of tissue or cell line, histopathological type, and subject characteristics like demographics, treatment, and survival are seldom present in data repositories, making it difficult to aggregate data.
We created a flexible software tool that allows efficient annotation of samples using a controlled vocabulary, and report on its use for the annotation of over 12,500 samples.
While the amount of data is very large and seemingly poorly annotated, a lot of information is still within reach. Consistent tool-based re-annotation enables many new possibilities for large scale interpretation and analyses that would otherwise be impossible.
Quantitative gene expression experiments provide key information for elucidating biological pathways and understanding diseases. Many methods have been developed over time, from hybridization based, Northern blotting, real-time polymerase chain reaction, high throughput microarrays and serial analysis of gene expression (SAGE), up to modern synthesis based sequencing methods (454, ABI/SOLiD, Illumina/Solexa). Microarrays are a popular technology for large-scale measurement of gene expression. Substantial public repositories have been set up to capture the wealth of information on gene expression generated by researchers world-wide: the National Center for Biotechnology Information's Gene Expression Omnibus (GEO) , the European Bioinformatics Institute's ArrayExpress  and the DNA Data Bank of Japan's Center for Information Biology Gene Expression Database (CIBEX) . There are, however, several essential pieces of information that are needed to make these measurements useful for anyone other than the original researcher. First, a clear description of the process is necessary to make results reproducible: This includes description of the sample preparation, microarray platform and reporter (probe) sequences. The second essential piece of information is a description of the data processing methods, from the raw image to the expression level. Both steps have been subject to rigorous standardization efforts in the past [4–6]. While it is still challenging to compare measurements, another essential piece of information has received far less attention: Clinical data describing the origin and characteristics of the samples, which is often sparse, inconsistent or simply absent.
Due to the public availability of large microarray repositories and reasonable standardization of measurement values, many 'meta-analyses' and meta-analysis systems emerged [7, 8]. These approaches usually merged the lists of differentially expressed genes obtained from previous studies. While this led to emphasis and validation of former results, new insights were less frequent. In order to derive genuinely new results from previous experiments, utilization of raw measurement values is almost inevitable.
However, analysing the raw data is challenging, as it requires (1) consistent and high quality probe annotations for all involved platforms, (2) appropriate cross-platform normalization methods, and (3) detailed sample annotations. Only few were able to report success with this approach . While mapping reporter sequences to genes no longer poses a problem, with high quality gene transcripts and several updated re-mappings at hand, normalization is more difficult, as many different factors come into play . Furthermore, sample annotations in popular large collections are hardly structured or consistent across studies and often lack important details.
We addressed the problem of annotating gene expression samples with a consistent set of variables on a large number of existing studies. While this seems to be a daunting task, there were sufficient similarities between samples to make it possible to annotate many samples at once. Moreover, we hypothesized that it was possible to obtain high quality annotations by non-expert individuals who received proper training.
Previous manual or automatic annotation attempts relied or tried to rely on domain experts to do the annotation , which was costly and time consuming. However, it has also been shown previously that using students for annotation is a worthwhile alternative to employing experts  and that the students themselves quickly become experts for annotating a certain disease. We were confident that students could become adept for the task of annotating a specific set of variables in a specific field.
In the past, several groups proposed annotation tools for microarray experiments [13–18]. These approaches tried to either automatically standardize the existing information or create collaborative platforms together with controlled vocabularies or ontologies to manually create consistent and reusable sample annotations. While there have been some successful cases of automatic term normalization in a smaller scale , attempts to automatically curate GEO, the largest public repository, were of limited success, especially when expecting detailed clinical annotation . The large diversity of information and unguided annotation currently present in large repositories call for taking the best ideas from these previous approaches and combining them into a new tool.
Results and discussion
Our database currently contains the 45 most popular platforms from GEO, 2,445 studies, and 58,432 samples. We also imported a total of 1.6 billion raw measurement values that can be used for new analyses with the help of a cross-platform probe annotation and a cross-platform normalization tool.
We were able to efficiently annotate more than 12,500 samples. More than half of these samples have been redundantly annotated by at least two different annotators. Within four weeks of work of one full time and one part time annotator and a following five weeks of work with four full time annotators, a total of almost half a million variable assignments were made. On average, every sample received 32 annotations.
The most frequently available variables were 'tissue' with 24,602 assignments, followed by 'disease state' with 12,098 annotations, 'sample type' with 9,792 annotations, and 'cell line' with 11,925. Information about genetic modification is also readily available in 11,135 annotations. Other frequently available variables were 'treatment', 'time series', 'gender', 'patient age', 'lymphatic spread', 'estrogen receptor status' and 'tumor type'.
Many possibilities of further processing are within reach with a method at hand that facilitates an efficient and rich annotation of existing gene expression data. One obvious use case is the regrouping of existing samples to perform new virtual studies. Another possible extension is the employment of this or a similar system directly at the public repositories so the uploaded data is annotated by the original submitter.
We have shown the feasibility of collecting publicly available information about previous microarray experiments and subjecting them to a consistent and efficient annotation. With the current speed and quality of annotation and a total of ten full-time annotators we project the time to annotate all of GEO (currently almost 250,000 samples) with two redundant annotations per sample to be 50 weeks. In the future, we plan to assemble new large datasets and perform new differential expression analyses, avoiding the high cost of sample collection, preparation and hybridization by exploiting existing data.
The annotation process started with the identification of possible studies for a certain disease by doing a keyword search. This search yielded a list of studies that had the keyword somewhere in their description or in the description of their samples. Then, annotators proceeded study by study, looking through available information in the local database and following publication links and their respective supplements. After going through all this information, the annotators usually had a good idea of what kind of samples they would subsequently be looking at. Armed with this, they proceeded sample by sample through each study.
Most of the annotation was conducted by one graduate and three senior biology students. An initial pilot annotation was done by one senior biology student and a physician. From this pilot annotation we obtained a feasible set of variables for large scale annotation. In a separate report  we provide more details about the quality of the annotation.
Import and structuring
The first step towards structured annotation was importing large parts of GEO into a relational database by parsing the SOFT files . While these files all have the same format and lexical structure, the individual study and sample annotations contained therein differed greatly in depth and presentation. We developed a tool that makes it easy to import these files. We mapped section names into a consistent scheme and filtered out poorly annotated sections. We found many examples of poor annotation. Some samples merely contained the compulsory organism and source fields but left out other essential information. Around 25% of the imported samples had descriptions shorter than 30 characters, and many just contained words like 'NA', 'null', the organ or a cryptic identifier.
Our database has a straightforward structure and captures just the essentials, like title, description, and organism for sample and studies. This information is the same for all popular expression databases like ArrayExpress or CIBEX, and the importer can be adapted to pull data from these sources.
We created a custom web front-end that showed the most consistently present fields and linked back to the original GEO entries. This web-platform was subsequently extended as a platform for uniform sample annotation. Moreover, we re-annotated probe-to-gene mappings of most of our 45 locally available microarray platforms using AceView  as a universal reference. We have created an automatic microarray annotation tool including an up-to-date high quality gene-transcript database that will be published soon.
One of the major difficulties was finding suitable variables that were frequently available and uniquely iden-tified concepts, while keeping the number of variables low. A concise and complex medical nomenclature had to be clearly separated into distinct terms and predefined choices, enabling non-expert annotators to comprehend existing descriptions and correctly label the described samples.
The underlying idea of the annotation system was to freely attach variables to samples, allowing any variable to be used for any sample. New variables could also be created and assigned if necessary. However, after the initial pilot study, a predefined set of variables with a predefined range of values was established for every disease through discussion and evaluation by experts. This encouraged consistent use of variables and values while at the same providing some freedom to capture seemingly important additional information. To further promote consistency between annotators, newly created variables and non-predefined values were visible to all annotators, albeit less prominently displayed than predefined variables in order to avoid cluttering the interface. This helped us keep a consistent naming scheme even for less important variables in case this information was also available. This guided yet flexible system proved extremely useful as we extended our annotation efforts to cover more diseases.
To facilitate faster annotation of frequently used values, new values could later be added to the set of predefined values and were then directly visible. Predefined variables were grouped into the categories 'essential', 'patient', 'treatment', and 'sample'. Variables were categorized according to individual diseases they appeared in. So far we created these sets of variables and annotated samples for breast cancer, colon cancer, insulin dependent diabetes mellitus, rheumatoid arthritis and systemic lupus erythematosus.
Variable definition file
Colon Cancer:Sample:*genetically modified
Colon Cancer:Sample:*tumor size
Colon Cancer:Essential:tissue, normal, benign, cancerous, non-colon
The right side of Figure 2 shows the sample annotation interface: After the local information for each sample was displayed (not shown in screenshot), a form with all predefined variables and values prompted for annotation. Variables, that had already been filled out, did not show up again for the same user. Below the initial section with predefined variables and values the annotator could find a powerful free-form annotation tool that allowed users to find and assign any variable that had previously been used, regardless of whether it was originally predefined or not. It also enabled the annotators to create new variables or assign new values to variables that did not allow free text. This feature proved to be useful in the pilot phase of annotation for a new disease, when it was still uncertain what information would be available. Further below, existing annotations were displayed, showing variable name, assigned value, the author of the annotation, and the number of times the variable had been used for every variable assignment. If the annotated value was one of the predefined values for a certain variable, it was highlighted in green.
The internal format of the annotations was a list of tuples of the form (sample id, variable id, value) inside our relational database that can be used to check for concordant annotations, extract samples with certain properties and their associated gene expression values.
The students performing the annotation where Pierre Cornell, Karrie Du, Evelyn Pitzer, Lindy Su, and Anthony Villanova. This research was funded by grant FAS0703850 from the Komen Foundation and D43TW007015 from the Fogarty International Center, NIH.
This article has been published as part of BMC Bioinformatics Volume 10 Supplement 9, 2009: Proceedings of the 2009 AMIA Summit on Translational Bioinformatics. The full contents of the supplement are available online at http://www.biomedcentral.com/1471-2105/10?issue=S9.
- Barrett T, Suzek TO, Troup DB, Wilhite SE, Ngau WC, Ledoux P, Rudnev D, Lash AE, Fujibuchi W, Edgar R: NCBI GEO: mining millions of expression profiles-database and tools. Nucleic Acids Res 2005, (33 Database):D562-D566.Google Scholar
- Parkinson H, Sarkans U, Shojatalab M, Abeygunawardena N, Contrino S, Coulson R, Farne A, Lara GG, Holloway E, Kapushesky M, Lilja P, Mukherjee G, Oezcimen A, Rayner T, Rocca-Serra P, Sharma A, Sansone S, Brazma A: ArrayExpress-a public repository for microarray gene expression data at the EBI. Nucleic Acids Res 2005, (33 Database):D553-D555.Google Scholar
- Ikeo K, Ishi-i J, Tamura T, Gojobori T, Tateno Y: CIBEX: center for information biology gene expression database. C R Biol 2003, 326(10–11):1079–1082. 10.1016/j.crvi.2003.09.034View ArticlePubMedGoogle Scholar
- Brazma A, Hingamp P, Quackenbush J, Sherlock G, Spellman P, Stoeckert C, Aach J, Ansorge W, Ball CA, Causton HC, Gaasterland T, Glenisson P, Holstege FC, Kim IF, Markowitz V, Matese JC, Parkinson H, Robinson A, Sarkans U, Schulze-Kremer S, Stewart J, Taylor R, Vilo J, Vingron M: Minimum information about a microarray experiment (MIAME)-toward standards for microarray data. Nat Genet 2001, 29(4):365–371. 10.1038/ng1201-365View ArticlePubMedGoogle Scholar
- Whetzel PL, Parkinson H, Causton HC, Fan L, Fostel J, Fragoso G, Game L, Heiskanen M, Morrison N, Rocca-Serra P, Sansone SA, Taylor C, White J, Stoeckert CJ: The MGED Ontology: a resource for semantics-based description of microarray experiments. Bioinformatics 2006, 22(7):866–873. 10.1093/bioinformatics/btl005View ArticlePubMedGoogle Scholar
- Deutsch EW, Ball CA, Berman JJ, Bova GS, Brazma A, Bumgarner RE, Campbell D, Causton HC, Christiansen JH, Daian F, Dauga D, Davidson DR, Gimenez G, Goo YA, Grimmond S, Henrich T, Herrmann BG, Johnson MH, Korb M, Mills JC, Oudes AJ, Parkinson HE, Pascal LE, Pollet N, Quackenbush J, Ramialison M, Ringwald M, Salgado D, Sansone SA, Sherlock G, Stoeckert CJ, Swedlow J, Taylor RC, Walashek L, Warford A, Wilkinson DG, Zhou Y, Zon LI, Liu AY, True LD: Minimum information specification for in situ hybridization and immunohistochemistry experiments (MISFISHIE). Nat Biotechnol 2008, 26(3):305–312. 10.1038/nbt1391PubMed CentralView ArticlePubMedGoogle Scholar
- Severgnini M, Bicciato S, Mangano E, Scarlatti F, Mezzelani A, Mattioli M, Ghidoni R, Peano C, Bonnal R, Viti F, Milanesi L, Bellis GD, Battaglia C: Strategies for comparing gene expression profiles from different microarray platforms: application to a case-control experiment. Anal Biochem 2006, 353: 43–56. 10.1016/j.ab.2006.03.023View ArticlePubMedGoogle Scholar
- Rhodes DR, Kalyana-Sundaram S, Mahavisno V, Varambally R, Yu J, Briggs BB, Barrette TR, Anstet MJ, Kincead-Beal C, Kulkarni P, Varambally S, Ghosh D, Chinnaiyan AM: Oncomine 3.0: genes, pathways, and networks in a collection of 18,000 cancer gene expression profiles. Neoplasia 2007, 9(2):166–180. 10.1593/neo.07112PubMed CentralView ArticlePubMedGoogle Scholar
- Warnat P, Eils R, Brors B: Cross-platform analysis of cancer microarray data improves gene expression based classification of phenotypes. BMC Bioinformatics 2005, 6: 265. 10.1186/1471-2105-6-265PubMed CentralView ArticlePubMedGoogle Scholar
- Irizarry RA, Hobbs B, Collin F, Beazer-Barclay YD, Antonellis KJ, Scherf U, Speed TP: Exploration, normalization, and summaries of high density oligonucleotide array probe level data. Biostatistics 2003, 4(2):249–264. 10.1093/biostatistics/4.2.249View ArticlePubMedGoogle Scholar
- Shah NH, Rubin DL, Espinosa I, Montgomery K, Musen MA: Annotation and query of tissue microarray data using the NCI Thesaurus. BMC Bioinformatics 2007, 8: 296. 10.1186/1471-2105-8-296PubMed CentralView ArticlePubMedGoogle Scholar
- Buckner B, Beck J, Browning K, Fritz A, Grantham L, Hoxha E, Kamvar Z, Lough A, Nikolova O, Schnable PS, Scanlon MJ, Janick-Buckner D: Involving undergraduates in the annotation and analysis of global gene expression studies: creation of a maize shoot apical meristem expression database. Genetics 2007, 176(2):741–747. 10.1534/genetics.106.066472PubMed CentralView ArticlePubMedGoogle Scholar
- Fellenberg K, Hauser NC, Brors B, Hoheisel JD, Vingron M: Microarray data warehouse allowing for inclusion of experiment annotations in statistical analysis. Bioinformatics 2002, 18(3):423–433. 10.1093/bioinformatics/18.3.423View ArticlePubMedGoogle Scholar
- Navarange M, Game L, Fowler D, Wadekar V, Banks H, Cooley N, Rahman F, Hinshelwood J, Broderick P, Causton HC: MiMiR: a comprehensive solution for storage, annotation and exchange of microarray data. BMC Bioinformatics 2005, 6: 268. 10.1186/1471-2105-6-268PubMed CentralView ArticlePubMedGoogle Scholar
- Hancock D, Wilson M, Velarde G, Morrison N, Hayes A, Hulme H, Wood AJ, Nashar K, Kell DB, Brass A: maxdLoad2 and maxdBrowse: standards-compliant tools for microarray experimental annotation, data management and dissemination. BMC Bioinformatics 2005, 6: 264. 10.1186/1471-2105-6-264PubMed CentralView ArticlePubMedGoogle Scholar
- Butte AJ, Kohane IS: Creation and implications of a phenome-genome network. Nat Biotechnol 2006, 24: 55–62. 10.1038/nbt1150PubMed CentralView ArticlePubMedGoogle Scholar
- Lee HW, Park YR, Sim J, Park RW, Kim WH, Kim JH: The tissue microarray object model: a data model for storage, analysis, and exchange of tissue microarray experimental data. Arch Pathol Lab Med 2006, 130(7):1004–1013.PubMedGoogle Scholar
- Splendiani A, Brandizi M, Even G, Beretta O, Pavelka N, Pelizzola M, Mayhaus M, Foti M, Mauri G, Ricciardi-Castagnoli P: The genopolis microarray database. BMC Bioinformatics 2007, 8(Suppl 1):S21. 10.1186/1471-2105-8-S1-S21PubMed CentralView ArticlePubMedGoogle Scholar
- Dudley J, Butte AJ: Enabling integrative genomic analysis of high-impact human diseases through text mining. Pac Symp Biocomput 2008, 580–591.Google Scholar
- Lacson R, Pitzer E, Hinske C, Kim J, Galante PAF, Ohno-Machado L: Evaluation of Large-Scale Biomedical Data Annotation Initiative. AMIA Summit on Translational Bioinformatics 2009.Google Scholar
- Barrett T, Troup DB, Wilhite SE, Ledoux P, Rudnev D, Evangelista C, Kim IF, Soboleva A, Tomashevsky M, Edgar R: NCBI GEO: mining tens of millions of expression profiles-database and tools update. Nucleic Acids Res 2007, (35 Database):D760-D765. 10.1093/nar/gkl887Google Scholar
- Thierry-Mieg D, Thierry-Mieg J: AceView: a comprehensive cDNA-supported gene and transcripts annotation. Genome Biol 2006, 7(Suppl 1):S12.1-S12.14. 10.1186/gb-2006-7-s1-s12View ArticleGoogle Scholar
- Bolstad BM, Irizarry RA, Astrand M, Speed TP: A comparison of normalization methods for high density oligonu-cleotide array data based on variance and bias. Bioinformatics 2003, 19(2):185–193. 10.1093/bioinformatics/19.2.185View ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.