BodyMap-Xs: anatomical breakdown of 17 million animal ESTs for cross-species comparison of gene expression.

BodyMap-Xs (http://bodymap.jp) is a database for cross-species gene expression comparison. It was created by the anatomical breakdown of 17 million animal expressed sequence tag (EST) records in DDBJ using a sorting program tailored for this purpose. In BodyMap-Xs, users are allowed to compare the expression patterns of orthologous and paralogous genes in a coherent manner. This will provide valuable insights for the evolutionary study of gene expression and identification of a responsive motif for a particular expression pattern. In addition, starting from a concise overview of the taxonomical and anatomical breakdown of all animal ESTs, users can navigate to obtain gene expression ranking of a particular tissue in a particular animal. This method may lead to the understanding of the similarities and differences between the homologous tissues across animal species. BodyMap-Xs will be automatically updated in synchronization with the major update in DDBJ, which occurs periodically.

INTRODUCTION

Do homologous genes have similar expression patterns? On the one hand evolutionary theories predict that paralogous genes have complementary spatio-temporal expression patterns that are based on a model of the consequences of a gene duplication event (1,2). On the other hand, molecular biologists sometimes assume almost similar expression patterns among structurally similar genes because the cis-regulatory element is a part of a gene located next to the coding sequence on the genome (3). The fragmented picture emerging from a limited number of recent genome-wide comparisons (4,5) will become more coherent if a wider range of species are more accurately compared. Expressed sequence tag (EST) data are the best resources to carry out such a comparison because they cover a sufficiently wide range of species (currently, 54 species with >40 million ESTs). Further, it has a sufficient resolution in anatomy and nucleotide sequence.

For many years, EST databases have served to provide expression information of individual species (6–12). However, no attempt has been made to integrate expression information across species in a coherent manner despite its potential importance. There are two major obstacles in achieving this goal: (i) In the INSD format—the universal format of EST data available to the public—the RNA source is described in free text in various fields and (ii) biologically sound integration of anatomies across different species is difficult, if at all possible. We have overcome these obstacles by developing a program for identification and subsequent sorting of anatomical names at the organ level across species. By applying this to the entire set of animal EST data in DDBJ, we have developed a database for the comparison of gene expressions across different animal species. This database is automatically updated in synchronization with major updates in DDBJ.

METHODOLOGY

Resources

The data unique to BodyMap-Xs are the anatomical breakdowns of 17 million ESTs across animal species generated by a sorting program tailored for this purpose. Other data used include the EST data in DDBJ, UniGene (13,14) for EST-gene relationship and RefSeq (15) for the conversion of a UniGene entry to representative amino acid sequences. In addition, a gene cluster from the InParanoid database (16,17) was used to connect genes based on orthologous relationships. The BLASTP program was employed to collect similar genes within this database. The scheme for EST breakdown and database construction is outlined in Figure 1.

Figure 1

Outline of BodyMap-Xs construction with some relevant figures for the data. Starting from three public resources, construction of database is structured as a pipe-line.

Grouping of EST by the source library

Because INSD formats do not have an official identifier for the source library, EST records are first clustered based on the identity present in the descriptions in the fields for authors, citations and sources. Highly variable strings that are sometimes embedded in source field, such as clone ID, were suppressed in advance. From each cluster, hereafter referred to as the library, we arbitrarily selected one record to be processed during anatomical sorting.

Grouping of libraries at the organ level

In INSD, the description logic and terminology of the RNA used for library construction are not controlled. Therefore, it is advised that the procedure for material preparation can be written in free text as a ‘note’. However, in reality, the descriptions are scattered across multiple fields of the records. For the coherent categorization of libraries across species, we need to classify those libraries in some manner based on those free and scattered descriptions. Structured terminology or ontology for anatomical concepts have been proposed for some species (18,19); however, assignment of these concepts in ontologies, having mutually different description logics to the tissue descriptions of continuously growing EST records across species is a daunting task. Further, regarding the source information of gene expression studies, we valued the reproducibility and clarity of the assignment process more than the soundness of the resulting classification in a purely anatomical sense.

Based on this concept, we developed a sorting program that detects anatomical names in the EST records across species and sorts them according to explicit rules. First, we selected 200 most populated libraries and manually grouped them in a bottom-up manner by joining the less populated libraries to their closest neighbors until we obtained 40 groups of almost homogeneous libraries. These 40 groups, mostly representing individual organs, were labeled with the most concise organ name. Subsequently, the key patterns within the anatomical terms responsible for the grouping process were extracted and compiled into a pattern dictionary that maps each pattern to 1 of the 40 organ names. Patterns in this dictionary were then expanded manually to cover variations and synonyms to the key patterns and were applied back to the same 200 libraries. Based on the erroneous identification and conflicts found in this application, several rules that resolve these problems were added to the program. Subsequently, the sorting program was applied to rest of the human libraries. From the libraries that lacked matching patterns, 100 most populated libraries were selected, manually assigned to 1 of the 40 groups and the pattern dictionaries were expanded accordingly. By repeating this cycle several times, we obtained a program that can sort 95% of the human EST records including those that were positively identified as pooled tissue or whole body. The resultant program was then applied to the remaining animal ESTs and the same application-modification cycles were repeated. Although the rules based on human anatomy worked properly in most vertebrates as expected, some names of body parts or cells were unique to particular species. These names were added to the dictionary at the most homologous human library groups based on their function. For example, ‘head’ in insects was sorted as ‘cerebrum’ and ‘hemocyte’ in Ciona intestinalis was classified as ‘blood’. Using the present version of the program, 76–100% of the ESTs of each vertebrate species were anatomically classified. Even in C.intestinalis—the simplest chordate in the EST division—five different organs were identified (Supplementary Table 1).

In addition to the anatomical categorization, the libraries were further categorized with respect to two independent aspects—the condition of tissue and distortion in population prior to sampling—by the same sorting program based on the pattern dictionary. Based on the condition of the tissues, the libraries are divided into normal and tumor/cell lines. The distortion aspect discriminates normalization and other processes that distort the population in conventional libraries, which is sometimes employed to avoid redundant isolation of clones for the same gene.

EST–gene relation and gene–gene relation

In this database, an individual UniGene cluster is tentatively regarded as the smallest unit that is responsible for an expression pattern. According to the EST_ID:UniGeneID correspondence in NCBI, the data were organized in the UniGene ID × 40-organ matrix of EST frequency. In addition, orthologous genes were interconnected using InParanoid data (16,17) after translating an Ensembl ID (20) to an UniGene ID via LocusLink.

DATA PRESENTATION

Similar to the old BodyMap DB (6), users can navigate to observe the activity ranking of genes in a particular animal and tissue with their anatomical expression patterns at a selectable resolution in a concise tabular format (Figure 2). For each gene in the ranking, users can compare the expression patterns of its orthologous partners across species in the same format.

Figure 2

Browsing data in BodyMap-Xs. Starting from the index page, the expression ranking of relevant organs is shown (a). For genes with expression patterns of interest, the expression of orthologous partners can be shown (b). Similarity search allows users to compare expression patterns of genes structurally similar to the query (c). For validation of expression patterns, every figure in the table can be broken down to the library level. The library list shows why these libraries were categorized in a particular manner with KWIC format (d).

Genes can be retrieved in the same tabular representation of expression patterns by using an ID or a keyword and by similarity to the query sequence. Similarities are measured against the RefSeq peptides corresponding to the representatives for UniGene clusters using the BLASTP program. The results show the expression patterns of homologous genes. List of libraries can be retrieved either with exact matching of the keyword or after generalization of user's keyword as a tissue category name by the library sorting program.

Last but not least, we would like to emphasize that the automatic identification process is still error prone even after rounds of careful inspection and correction. This is mainly because the program does not consider the context where the relevant patterns are embedded. Further, a grouping scheme may not be obvious to some users or even inappropriate for some purposes. Therefore, we devised a method that allows the users to backtrack the automatic sorting process. Every number in the tables (either library count or EST count) can be broken down to the library level where the reasons as to why they were categorized in a particular manner are shown by the ‘keyword in context (KWIC)’ representation showing key patterns with short flanking strings in the original EST record. Using such information, users can verify the data or modify them according to their purpose.

FUTURE DEVELOPMENT

In light of the fact that as much as 40–60% of genes have alternatively spliced transcripts in mammals (21) and that a substantial portion of this splicing may be anatomically controlled (22), the resolution of expression should be enhanced to the transcript level from the gene level. Integration across different measurement platforms is planned through the incorporation of H-ANGEL function (23) to BodyMap-Xs. In the downstream of the expression comparison table for homologous genes, a tree view format that shows mutual evolutionary relations among the homologs is under construction. The BodyMap-Xs data will be updated immediately after major updates in the DDBJ are released; these updates take place four times a year. Further, the pattern dictionary will be updated after every update based on novel patterns found in the new release.

SUPPLEMENTARY DATA

Supplementary Data are available at NAR Online.