Prediction of genes is the most important part of genome annotation, connecting the DNA sequence with the wide array of experimental data. In eukaryotic organisms with large introns, ab initio predictions are useful but have a high false positive rate and often predict partially incorrect gene structures. Thus, incorporation of all available evidence for gene prediction is necessary.

The Ensembl gene build system incorporates a wide range of methods including ab initio gene predictions, homology and gene prediction HMMs. Genes are placed in the genome using a three step process. First, ‘best in genome’ positions for all known human proteins from SPTREMBL (6) are found using a fast protein to DNA matcher (pmatch, R. Durbin, unpublished software). These positions are refined using genewise (7) to provide an accurate gene structure. UTRs are also aligned to each gene structure using full-length cDNAs where known. Secondly, a similar process is used to align paralogous human proteins and proteins from other organisms to the genome to form a set of novel human genes. Finally, the ab initio program genscan (8) is run across the entire genome to create a set of genscan peptides. Exons from these predicted peptides that are confirmed by blast matches to proteins, vertebrate mRNA and UniGene clusters are assembled into genes.

The above process creates a set of transcripts and these are grouped into genes wherever an exon is shared. These ‘Ensembl genes’ are regarded as being accurate predicted gene structures with a low false positive rate, since they are all supported by experimental evidence of at least one form via sequence homology. Ensembl human genes are identified by numbers beginning ENSG (transcripts begin ENST, exons begin ENSE and translations begin ENSP). These identifiers are keep stable, as far as is possible, between assemblies of the human genome.

Ensembl is continuously refining and extending its gene building process, calibrating it against regions of the genome that have been hand annotated and experimentally investigated, such as human chromosome 22 (9). We are in the process of integrating EST data into Ensembl gene building. ESTs offer a considerable advantage in aiding the prediction of non-coding exons, especially those located within the 3′-UTR. Two EST/genome alignment algorithms, namely exonerate (G. Slater, unpublished) and EST_genome (10), have been integrated with the Ensembl gene-building pipeline to yield gene predictions incorporating EST alignments. Because EST data are notorious for their high sequence error rate, strict quality measures have been introduced such that only splicing ESTs are considered, and priority is given to those ESTs which align on the genome into clusters.

A view of a region of genomic sequence is shown in Figure ​Figure1.1. Ensembl contigview web pages feature the ability to scroll along entire chromosomes, while viewing the features within a selected region in detail. Features are integrated from external data sources such as HUGO gene names, genetic markers, disease genes and SNPs, with links to primary databases. The user can control which features are displayed and dynamically integrate external DAS data sources as well as their own annotation (see below). Matches between WGS mouse reads and the human genome from exonerate are also displayed. The individual mouse reads can be accessed via the EBI trace server which is also provided via Ensembl (http://trace.ensembl.org/). There is an integrated, context sensitive, searchable help system which can be accessed by selecting the help button on any page.

The Ensembl web site provides a variety of alternate views of the data. These include mapview web pages, which show relationships between cytogenetic bands and the genome sequence via markers, and displays feature distribution plots; geneview web pages showing information about individual Ensembl gene with its transcripts and gene structures and proteinview web pages, showing information about individual Ensembl translations with functional annotation from InterPro. Similarity searching is also integrated into the web site. BLAST (11) and SSAHA (http://www.sanger.ac.uk/Software/analysis/SSAHA/) search tools are available against the entire human genome sequence, predicted gene datasets and mouse genome trace and whole genome assembly datasets.

Ensembl can be accessed in a variety of ways apart from web pages. Ensembl annotation can also be viewed interactively using the Apollo Java viewer, which is being developed as a collaborative project between the Berkeley Drosophila Genome Project (http://www.bdgp.org/) and Ensembl. The Ensembl FTP site provides a variety of data download formats, e.g. FASTA files of gene and protein sequences; EMBL and GenBank formats containing annotation of the raw genomic sequence. This includes the full dumps of the MySQL database used by the web site (see below). Extensive data dumping tools are also available from the contigview web pages, allowing regions to be selected and dumped in many flat file formats. Regions can also be dumped as graphical images for printing in a variety of formats and layouts.

The architecture of the software is split into biologically meaningful objects (business objects) and database connectivity objects (adaptors). This separation makes it easier to evolve the schema to address new datatypes or analyses while maintaining code stability. New datasets can be added easily by providing the necessary adaptors and business objects.

One of the core design features of the system is the ‘Virtual Contig’ (VC) object, which allows access to genomic sequence and its annotation as if it was a continous piece of DNA in a 1–N coordinate space, regardless of how it is stored in the database. This is important since it is impractical to store large genome sequences as continuous pieces of DNA, not least because this would mean updating the entire genome entry whenever any single base changed. The VC object handles reading and writing of features and behaves identically regardless of whether the underlying sequence is stored as a single real piece of DNA (a single raw contig) or an assembly of many fragments of DNA (many raw contigs). Because features are always stored at the raw contig level, ‘virtual contigs’ really are virtual and as a result are less fragile to sequence assembly changes. It is this feature that allows Ensembl to handle draft genome data in a seamless way and makes it possible to change between different genome assemblies relatively painlessly. This feature should also put us in a good position to handle haplotype sequences efficiently as they become available.