New gene prediction method capitalizes on multiple genomes

Researchers at Stanford University report in the online open access journal, Genome Biology, a new approach to computationally predicting the locations and structures of protein-coding genes in a genome. Gene finding remains an important problem in biology as scientists are still far from fully mapping the set of human genes. Furthermore, gene maps for other vertebrates, including important model organisms such as mouse, are much more incomplete than the human annotation. The new technique, known as CONTRAST (CONditionally TRAined Search for Transcripts), works by comparing a genome of interest to the genomes of several related species.

CONTRAST exploits the fact that the functional role protein-coding genes play a specific part within a cell and are therefore subjected to characteristic evolutionary pressures. For example, mutations that alter an important part of a protein's structure are likely to be deleterious and thus selected against. On the other hand, mutations that preserve a protein's amino acid sequence are normally well tolerated. Thus, protein-coding genes can be identified by searching a genome for regions that show evidence such patterns of selection. However, learning to recognize such patterns when more than two species are compared has proved difficult.

Previous systems for gene prediction were able to effectively make use of one additional 'informant' genome. For example, when searching for human genes, taking into account information from the mouse genome led to a substantial increase in accuracy. But, no system was able to leverage additional informant genomes to improve upon state-of-the-art performance using mouse alone, although it was expected that adding informants would make patterns of selection clearer. CONTRAST solves this problem by learning to recognize the signature of protein-coding gene selection in a fundamentally different way from previous approaches. Instead of constructing a model of sequence evolution, CONTRAST directly 'learns' which features of a genomic alignment are most useful for recognizing genes. This approach leads to overall higher levels of accuracy and is able to extract useful information from several informant sequences.

In a test on the human genome, CONTRAST exactly predicted the full structure of 59% of the genes in the test set, compared with the previous best result of 36%. Its exact exon sensitivity of 93%, compared with a previous best of 84%, translates into many thousands of exons correctly predicted by CONTRAST but missed by previous methods. Importantly, CONTRAST's accuracy using a combination of eleven informant genomes was significantly higher than its accuracy using any single informant. The substantial advance in predictive accuracy represented by CONTRAST will further efforts to complete protein-coding gene maps for human and other organisms.

Further information about existing gene-prediction methods and the advance CONTRAST brings to the field can be found in a minireview by Paul Flicek, which accompanies the article by Batzoglou and colleagues.

1) CONTRAST: A discriminative, phylogeny-free approach to multiple informant de novo gene prediction
Samuel S Gross, Chuong B Do, Marina Sirota and Serafim Batzoglou
Genome Biology (in press)

During embargo, article available at: http://genomebiology.com/imedia/2690900661452718_article.pdf?random=533533

After the embargo, article available at journal website: http://genomebiology.com/

2) A PDF of the minireview is available upon request at press@biomedcentral.com

3) Please name the journal in any story you write. If you are writing for the web, please link to the article. All articles are available free of charge, according to BioMed Central's open access policy.

Article citation and URL available on request at press@biomedcentral.com on the day of publication

4) Genome Biology publishes articles from the full spectrum of biology. Subjects covered include any aspect of molecular, cellular, organismal or population biology studied from a genomic perspective, as well as genomics, proteomics, bioinformatics, genomic methods (including structure prediction), computational biology, sequence analysis (including large-scale and cross-genome analyses), comparative biology and evolution. Genome Biology has an impact factor of 7.12.

5) BioMed Central is an independent online publishing house committed to providing immediate access without charge to the peer-reviewed biological and medical research it publishes. This commitment is based on the view that open access to research is essential to the rapid and efficient communication of science.