TissueInfo: high-throughput identification of tissue expression profiles and specificity

Abstract

We describe TissueInfo, a knowledge-based method for the high-throughput identification of tissue expression profiles and tissue specificity. TissueInfo defines a set of tissue information calculations that can be computed for large numbers of genes, expressed sequence tags (ESTs) or proteins. Tissue information records that result from the TissueInfo calculations are used to generate tables suitable for data mining and for the selection of genes according to a given expression profile or specificity. When benchmarked against a test set of 116 proteins and literature information, TissueInfo was found to be accurate for 69% of identified tissue specificities and for 80% of expression profiles. The accuracy of the identifications can be increased if query sequences for which little information is available from dbEST are ignored. Thus, with 80% coverage, TissueInfo achieves an accuracy of 76% for specificity and 89% for expression. For the same set of proteins, the curated tissue specificity offered in SWISS-PROT was accurate in 78% of cases. TissueInfo can be useful for the selection of clones for custom microarrays, selection of training sets for ab initio identification of tissue information, gene discovery and genome-wide predictions. Further information about the program can be found at http://icb.mssm.edu/tissueinfo.

Figure 1

The components of TissueInfo are illustrated and their relation to a local infrastructure. Arrows represent either a flow of data or relationships among components that provide or use services. Information Management imports raw data from dbEST, stores it in a relational database and supports the curation of such data. Curated information is made available as objects empowered with methods necessary to implement the algorithms described in Table 1. See the text for a description of the other components. TissueInfo requires access to dbEST, MegaBLAST and BLAST, which must be provided by the local bioinformatics infrastructure. TissueInfo records are generated for each query sequence going through the pipeline. A suitable data mining environment helps navigate data sets produced when large numbers of query sequences are annotated.

Figure 2

Reliability of predictions. A summary of the number of times the ‘correct’ or ‘agree’ tissue specificity or tissue expression is found when predictions are compared to SWISS-PROT and the literature. First line, accuracy of specificity prediction. Second line, accuracy of expression prediction. Each column represents a protein, an X indicates an incorrect prediction and a period represents a ‘correct’ or ‘agree’ prediction. (See text for evaluation details.) Third and following lines, adjusted number of hits, i.e. number of hits for which information is available in TissueInfo. The adjusted number of hits is given reading downwards at every point. Predictions are sorted according to increasing adjusted hit number.