[Link]

Meeting report
Gene prediction: the end of the beginning

comment
Colin Semple
Address: Department of Medical Sciences, Molecular Medicine Centre, Western General Hospital, Crewe Road, Edinburgh EH4 2XU, UK.
E-mail: [Link]@[Link]

Published: 28 July 2000

reviews
Genome Biology 2000, 1(2):reports4012.1–4012.3
The electronic version of this article is the complete one and can be
found online at [Link]
© [Link] (Print ISSN 1465-6906; Online ISSN 1465-6914)

Reducing genomes to genes

A report from the conference entitled Genome Based Gene

reports
All ab initio gene prediction programs have to balance sensi-
Structure Determination, Hinxton, UK, 1-2 June, 2000, tivity against accuracy. It is often only possible to detect all
organised by the European Bioinformatics Institute (EBI). the real exons present in a sequence at the expense of detect-
ing many false ones. Alternatively, one may accept only pre-
dictions scoring above a more stringent threshold but lose
The draft sequence of the human genome will become avail- those real exons that have lower scores. The trick is to try and
able later this year. For some time now it has been accepted increase accuracy without any large loss of sensitivity; this

deposited research
that this will mark a beginning rather than an end. A vast can be done by comparing the prediction with additional,
amount of work will remain to be done, from detailing independent evidence. For example, one may increase confi-
sequence polymorphisms to discovering the complexities of dence in a predicted coding exon by detecting the presence of
the transcriptome - the totality of sequences transcribed - a sequence within it which codes for a known protein domain.
and, ultimately, the proteome - all the proteins encoded by The patterns made in DNA sequences by such domains may
the genome. All of this work will, to a greater or lesser extent, be detected using probabilistic models known as hidden
depend on all the genes having been correctly identified. It Markov models (HMMs). Predictions for exons that contain
will be necessary to document not only the coding exons of non-coding sequence or untranslated regions (UTRs) can be

refereed research
each gene but also non-coding exonic sequence and regula- refined by comparison with ESTs (expressed sequence tags) -
tory sequences. As this conference made clear, however, the sequences representing fragments of mRNA sequence that
production of genomic sequence has outstripped our ability include coding and/or UTR sequence. The latest generation
to reliably predict such features computationally. of gene prediction programs take advantage of such similar-
ity-based approaches to complement ab initio predictions.
Traditionally, gene prediction programs that rely only on the For example, Ed Uberbacher (Oak Ridge National Labora-
statistical qualities of exons have been referred to as per- tory, USA) unveiled the latest incarnation of the Grail
forming ab initio predictions (from the Latin: from the program - GrailEXP [[Link] - which uses
beginning). Ab initio prediction of coding sequences is an EST data to find both UTR boundaries and short exons, typi-
undeniable success by the standards of the machine-learning cally problematic areas for ab initio predictions. Anders interactions
algorithm field, and most of the widely used gene prediction Krogh (Centre for Biological Sequence Analysis, Denmark)
programs belong to this class of algorithms. It is impressive also showed improvements in accuracy for his gene predic-
that the statistical analysis of raw genomic sequence can tion program HMMgene [[Link]
detect around 77-98% of the genes present, which was the HMMgene/] when similarity-based evidence was incorpo-
range of sensitivity reported at the conference. This is, rated.
however, little consolation to the bench biologist, who wants
information

the complete sequences of all genes present, with some cer- Regulatory regions in the human genome are estimated to
tainty about the accuracy of the predictions involved. As occupy ten times the sequence length of coding sequences.
Ewan Birney (European Bioinformatics Institute, UK) put it, Prediction of regulatory sequences remains troublesome as
what looks impressive to the computer scientist is often they are invariably short sequences matching a rather vague
simply wrong to the biologist. consensus pattern that arise frequently by chance in
genomic sequence. Thomas Werner (GSF - National
2 Genome Biology Vol 1 No 2 Semple

Research Centre for Environment and Health, Germany) Making ambiguity clear
outlined a novel approach to circumventing the low Because ab initio prediction is far from perfect, and adding
sequence conservation of functionally equivalent promoters. other evidence can improve gene prediction, there have been
He treats promoter regions as clusters of small, locally con- several efforts to develop graphical interfaces for the com-
served sequence motifs or ‘modules’. It would seem that parison of results. These interfaces allow simultaneous
there are specific restrictions on the spacing and ordering of examination of the plethora of results generated by gene pre-
modules within a promoter region, imposed by the require- diction programs along with sequence similarities. The idea
ments of the regulatory protein complexes that bind there. is to show explicitly where evidence from different sources is
His program PromoterInspector [[Link] contradictory or in agreement. Human intervention then
de/free_services/] uses such restrictions to increase the takes the form of ‘polishing’ annotation: making decisions
accuracy of promoter prediction, achieving impressive about the reliability of predicted features and designing
results over large genomic sequences such as human chro- experiments to support or refute them. A graphical interface
mosome 22, where it reached levels of specificity of more called Artemis [[Link]
than 98% (that is, it generated less than 2% false positives developed in the Sanger Centre (poster presented by Kim
when compared to the published annotation). As with gene Rutherford), is designed to allow users to edit the features
prediction, however, there is a trade-off with sensitivity, and displayed. Artemis has been used extensively for annotation
only around 33% of known promoters were found. Michael of Sanger Centre projects up to 4 Mb in size. These have
Zhang (Cold Spring Harbor Laboratory, USA) has begun the been mainly pathogen genomes, but the program is now
task of incorporating regulatory motif detection into his gene being used in the annotation of the genome of the fission
prediction software MZEF. As with many other programs, yeast Schizosaccharomyces pombe. Various attempts are
MZEF is most successful when predicting internal, coding, being made to automate the polishing process and reduce
exons, so the idea is to tailor new programs to other classes the amount of human intervention necessary. Richard
of exon. For instance, models of initial exons could incorpo- Durbin (Sanger Centre, UK) described a new program called
rate upstream promoter regions and final exon models could GAZE, which is an offshoot of his ACEDB database software.
include poly(A) addition sites. GAZE integrates evidence from multiple sources to come up
with graphical representations of gene predictions. We were
Comparative genomics may be the unexploded bomb in gene also introduced to Ensemble [[Link] by
structure prediction, capable of sweeping away many of the Ewan Birney. It takes exons predicted ab initio that are con-
ambiguities in human gene predictions. Regions of sequence firmed via similarity results and assembles the exons into
conserved between species can reveal novel coding predicted genes that are presented graphically. In this way it
sequences and, more importantly, non-coding features that provides a ‘base line’ annotation for many of the fragmen-
could not otherwise be detected. Mikhail Gelfand (Centre for tary, unfinished human genomic sequences in the EMBL
Biotechnology, Russia) outlined strategies for finding regu- database. Around 2.9 Gb of the existing draft and finished
latory regions by comparison of bacterial species. He showed sequence of the human genome have been processed by
how the discovery of regulatory elements for heat-shock Ensemble and the results are freely available from the
protein genes allowed the detection of such elements at Ensemble website. Ensemble gene identifiers will remain
other loci and consequently the detection of novel co-regu- stable during rearrangements and extension of draft
lated genes that may be involved in the heat-shock response. sequences on the way to a definitive human genome
The mouse genome sequence will be available within a year sequence, which is at least three years away. Similar ‘indus-
or two and will doubtless provide a popular resource for trial scale’ analyses are being run using the Genome Channel
comparisons with human sequence, particularly in detecting [[Link] according to Ed
regulatory elements and refining exon boundaries. But more Uberbacher; the genome channel is an analysis pipeline pro-
than one speaker concluded that comparisons between cessing draft human genome sequence with a different com-
mouse and human sequences may not be as instructive as bination of programs from Ensemble.
those between human and other species. For instance,
Roderic Guigo (Institut Municipal de Investigacio Medica,
Spain) found that chicken sequences may be more helpful in Automatic for the people
predicting coding exons, as they show good conservation in The Cold Spring Harbor Genome conference earlier this year
coding regions but diverge substantially elsewhere. Conser- saw the creation of Genesweep [[Link]
vation between mammals seems to be more widespread, and [Link]], a ‘gene sweepstake’ where participants can
so creates less clear distinctions between conserved and bet on the final number of human genes that will be found.
divergent sequences. Webb Miller (Pennsylvania State Uni- The spread of bets reflects the current uncertainty among
versity, USA) presented a new program, PipMaker workers in human genomics, ranging from 27,462 up to
[[Link] (Percentage Identity 200,000. Interestingly, the mean is currently 62,598, much
Plot Maker) for graphically viewing sequence conservation lower than the ballpark figure of 100,000 we have all become
along genomic sequences. accustomed to. When the winner of Genesweep is announced
 [Link]

in 2003, what will be the reward for the rest of us? Once there
is a complete set of known and predicted genes in which we
have high confidence, it is to be made publicly available via

comment
the internet, and the way these data are presented will be
influenced by the experiences and software taken from other
genome projects. Various speakers at the meeting discussed
archiving genomic annotation data for projects in the plant
Arabidopsis thaliana, the fruit fly Drosophila melanogaster
and completed human chromosomes. Michael Ashburner
(EBI) stressed the importance of consistency in genome
annotation across species and described the Gene Ontology

reviews
(GO) project [[Link] GO is an
attempt to rigorously describe all the genes in a genome
according to the molecular functions of their products and
the biological processes and cellular components with which
they are associated. The classification and standardized ter-
minology of GO were used in the annotation of the D.
melanogaster genome, and it is hoped that GO will become a
community-curated entity, providing a democratic but

reports
central vocabulary for annotation.

Perhaps the real measure of the success of computational gene

predictions is in their successful integration into the biologists’
toolbox. Tim Hubbard (Sanger Centre) described the strategy
for annotation of the first completed human chromosome
sequence, of chromosome 22 [[Link]

deposited research
HGP/Chr22/]. Here the strength of computational predictions
- their sensitivity - was exploited to generate a set of candidate
exons. These candidates could then be used as the starting
point for laboratory work to discover the actual mRNA
sequences. For chromosome 22, 94% of genes were found and
at least partially predicted by ab initio methods, but 16% of
real exons were not predicted at all computationally and only
20% of predicted gene structures were correct. So, computa-

refereed research
tional techniques give us invaluable clues to gene structure,
but in the end it will be the addition of work by bench scien-
tists that will provide the full picture. With this in mind,
moves have begun to establish a Distributed Sequence Anno-
tation System (DAS) [[Link] to democra-
tize genome annotation. The idea is to designate a central
‘reference server’ which stores essential mapping and
sequence data for the genome, and multiple ‘annotation
servers’ maintained elsewhere by a range of different groups.
Researchers interested in a given region of the genome would interactions
use a web browser-like application to download and integrate
different features from servers of their choice. Thus, a reliable
central annotation can be maintained in parallel with a diver-
sity of less confidently predicted features that may or may not
turn out to be useful. It seems inevitable that human beings
will have to take on the final tasks of gene prediction and
information

annotation once the machines have had a first pass. One way
or another, human intervention is still essential in
computational gene prediction and, it would seem, the more
humans involved, the better.