Comparative Gene Prediction in Human
Genı ´s Parra,1Pankaj Agarwal,2Josep F. Abril,1Thomas Wiehe,3James W. Fickett,4
and Roderic Guigo ´1,5
1Grup de Recerca en Informa `tica Biome `dica. Institut Municipal d’Investigacio ´ Medica / Universitat Pompeu Fabra / Centre de
Regulacio ´ Geno `mica 08003 Barcelona, Catalonia, Spain;2GlaxoSmithKline, King of Prussia, Pennsylvania 19406, USA;
3Freie Universita ¨t Berlin and Berlin Center for Genome Based Bioinformatics (BCB), 14195 Berlin, Germany;4AstraZeneca
R&D Boston, Waltham, Massachusetts 02451, USA
The completion of the sequencing of the mouse genome promises to help predict human genes with greater
accuracy. While current ab initio gene prediction programs are remarkably sensitive (i.e., they predict at least a
fragment of most genes), their specificity is often low, predicting a large number of false-positive genes in the
human genome. Sequence conservation at the protein level with the mouse genome can help eliminate some of
those false positives. Here we describe SGP2, a gene prediction program that combines ab initio gene prediction
with TBLASTX searches between two genome sequences to provide both sensitive and specific gene predictions.
The accuracy of SGP2 when used to predict genes by comparing the human and mouse genomes is assessed on
a number of data sets, including single-gene data sets, the highly curated human chromosome 22 predictions,
and entire genome predictions from ENSEMBL. Results indicate that SGP2 outperforms purely ab initio gene
prediction methods. Results also indicate that SGP2 works about as well with 3x shotgun data as it does with
fully assembled genomes. SGP2 provides a high enough specificity that its predictions can be experimentally
verified at a reasonable cost. SGP2 was used to generate a complete set of gene predictions on both the human
and mouse by comparing the genomes of these two species. Our results suggest that another few thousand
human and mouse genes currently not in ENSEMBL are worth verifying experimentally.
After the genome sequence of an organism has been obtained,
the very first next step is to compile a complete and accurate
catalog of the genes encoded in this sequence. For higher
eukaryotic organisms, however, the accuracy of currently
available gene prediction methods to perform such a task is
limited (Guigo ´ et al. 2000; Rogic et al. 2001; Guigo ´ and Wiehe
2003). The increasing availability of genome sequences from
different organisms, however, has lead to the development of
new computational gene finding methods that use sequence
conservation to help identifying coding exons, and improve
the accuracy of the predictions (Fig. 1; Crollius et al. 2000;
Wiehe et al. 2000; Miller 2001; Rinner and Morgenstern
2002). Indeed, three such comparative gene prediction pro-
grams, SLAM (Pachter et al. 2002), SGP2, and TWINSCAN
(Korf et al. 2001) have been used for the comparative analysis
of the human and mouse genomes. These analyses lead to
more accurate gene predictions, and to the verification of pre-
viously unconfirmed genes. In this paper, we describe the
program SGP2. Typical computational ab initio gene predic-
tion methods rely on the identification of suitable splicing
sites, start and stop codons along the query sequence, and the
computation of some measure of coding likelihood to predict
and score candidate exons, and delineate gene structures (see
Claverie 1997; Burge and Karlin 1998; Haussler 1998; Zhang
2002 and references therein for reviews on computational
Similarity between the query sequence and known cod-
ing sequences (amino acid or cDNA) can also be used to infer
gene structures. When the query sequence encodes a protein
for which a close homolog exists, a special type of alignment
can be used between the DNA sequence and the target pro-
tein/cDNA sequence, in which gaps in the target sequence
corresponding to introns in the query sequence must be com-
patible with potential splicing signals. This is the approach in
GENEWISE (Birney and Durbin 1997) and PROCRUSTES
(Gelfand et al. 1996). Alternatively, the results of searching
the query sequence against a database of known coding se-
quences, using for instance BLASTX (Altschul et al. 1990,
1997; Gish and States 1993), can be incorporated more or less
ad hoc into the scoring schema of an ab initio gene prediction
method. The program GENOMESCAN (Yeh et al. 2001),
which incorporates BLASTX search results into the predic-
tions by the GENSCAN program (Burge and Karlin 1997), is an
example of a recent development in that direction.
Recently developed comparative gene prediction pro-
grams further exploit sequence similarity. Instead of compar-
ing anonymous genomic sequences to known coding se-
quences, anonymous genomic sequences are compared to
anonymous genomic sequences from the same or different
organisms, under the assumption that regions conserved in
the sequence will tend to correspond to coding exons from
homologous genes. The approach taken by the different pro-
grams to exploit this idea differs notably.
In one such approach (Blayo et al. 2002; Pedersen and
Scharl 2002), the problem is stated as a generalization of pair-
wise sequence alignment: Given two genomic sequences cod-
ing for homologous genes, the goal is to obtain the predicted
exonic structure in each sequence maximizing the score of the
E-MAIL email@example.com; FAX 34 93 224-0875.
Article and publication are at http://www.genome.org/cgi/doi/10.1101/
13:108–117 ©2003 by Cold Spring Harbor Laboratory Press ISSN 1088-9051/03 $5.00; www.genome.org
alignment of the resulting amino acid sequences. Both Blayo
et al. (2002) and Pedersen and Scharl (2002) solve the problem
through a complex extension of the classical dynamic pro-
gramming algorithm for sequence alignment.
In a different approach, the programs SLAM (Pachter et
al. 2002) and DOUBLESCAN (Meyer and Durbin 2002) com-
bine sequence alignment pair hid-
den Markov Models (HMMs;
Durbin et al. 1998) with gene pre-
diction generalized HMMs
(GHMMs; Burge and Karlin 1997)
into the so-called generalized pair
HMMs. In these, gene prediction is
not the result of the sequence align-
ment, as in the programs above;
gene prediction and sequence
alignment are obtained simulta-
A third class of programs adopt
a more heuristic approach, and
separate clearly gene prediction
from sequence alignment. The pro-
grams ROSSETA (Batzoglou et al.
2000), SGP1 (from ‘syntenic gene
prediction’; Wiehe et al. 2001), and
CEM (from ‘conserved exon
method’; Bafna and Huson 2000)
are representative of this approach.
All these programs start by aligning
two syntenic sequences and then
predict gene structures in which the
exons are compatible with the
alignment. The programs described
thus far rely on the comparison of
fully assembled (and when from
different organisms, syntenic) ge-
nomic regions. This limits their
utility when analyzing complete
large eukaryotic genomes, and in
particular when the informant ge-
nome is in nonassembled shotgun
form. To overcome this limitation,
the programs TWINSCAN (Korf
et al. 2001) and SGP2 take still
a different approach. The approach
is reminiscent of that used in
GENOMESCAN (Yeh et al. 2001) to
incorporate similarity to known
proteins to modify the GENSCAN
scoring schema. Essentially, the
query sequence from the target ge-
nome is compared against a collec-
tion of sequences from the infor-
mant genome (which can be a
single homologous sequence to the
query sequence, a whole assembled
genome, or a collection of shotgun
reads), and the results of the com-
parison are used to modify the
scores of the exons produced by ab
initio gene prediction programs. In
TWINSCAN, the genome sequences
are compared using BLASTN, and
the results serve to modify the un-
derlying probability of the potential exons predicted by
GENSCAN. In SGP2, the genome sequences are compared us-
ing TBLASTX (W. Gish, 1996–2002, http://blast.wustl.edu),
and the results are used to modify the scores of the potential
scores predicted by GENEID. TWINSCAN and SGP2 have been
successfully applied to the annotation of the mouse genome
Mouse orthologous gene
Human HLA class II alpha-chain gene
for the HLA class II alpha chain. Black boxes indicate the coding exons, while black diagonals indicate
the conserved alignments. The score of the conserved alignments (divided by 10) is given in the lower
panels. Although conserved regions between the human and mouse genomic sequences coding for
these genes fully include the coding exons, a substantial fraction of intronic regions is also conserved.
The TBLASTX outptut was post-processed to show a continuous non-overlapping alignment.
Pairwise comparison using TBLASTX of the human and mouse genomic sequences coding
Comparative Gene Prediction in Human and Mouse
(Mouse Genome Sequencing Consortium 2002), and have
helped to identify previously unconfirmed genes (Guigo ´ et
In the next section, we describe the algorithmic details of
SGP2, and its implementation. We also describe the sequence
sets used to benchmark SGP2 accuracy. Results based on these
data sets indicate that SGP2 is an improvement over pure ab
initio gene prediction programs, even when the informant
genome is only in shotgun form. We have found that 3x
coverage will generally suffice to achieve maximum accuracy.
Finally, we describe the application of SGP2 to the compara-
tive analysis of the human and mouse genomes.
SGP2 is a method to predict genes in a target genome sequence
using the sequence of a second informant or reference genome.
Essentially, SGP2 is a framework to integrate the ab initio
gene prediction program GENEID (Guigo ´ et al. 1992; Parra et
al. 2000) with the sequence similarity search program
TBLASTX. The approach is conceptually similar to that
used in TWINSCAN to incorporate BLASTN searches into
GENEID is a genefinder that predicts and scores all po-
tential coding exons along a query sequence. Scores of exons
are computed as log-likelihood ratios, which are a function of
the splice sites defining the exon, and of the coding bias in
composition of the exon sequence as measured by a Markov
Model of order five (Borodovsky and McIninch 1993). From
the set of predicted exons, GENEID assembles the gene struc-
ture (eventually multiple genes in both strands), maximizing
the sum of the scores of the assembled exons, using a dynamic
programming chaining algorithm (Guigo ´ 1998).
When using an informant genome sequence to predict
genes in a target genome sequence, ideally we would like to
incorporate into the scores of the candidate exons predicted
along the target sequence, the score of the optimal alignment
at the amino acid level between the target exon sequence and
the counterpart homologous exon in the informant genome
sequence. If a substitution matrix, for instance from the
BLOSUM family, is used to score the alignment, the resulting
score can also be assumed to be a log-likelihood ratio: infor-
mally, the ratio between the likelihood of the alignment
when the amino acid sequences code for functionally related
proteins, and the likelihood of the alignment, otherwise. In
principle, this score could be added to the GENEID score for
the exon. TBLASTX provides an appropriate shortcut to often
find a good enough approximation to such an optimal align-
ment, and infer the corresponding score: The optimal align-
ment can be assumed to correspond to the maximal scoring
high-scoring segment pairs (HSP) overlapping the exon. How-
ever, when dealing in particular with the informant genome
sequence in fragmentary shotgun form, often different re-
gions of a candidate exon sequence will align optimally to
different informant genome sequences. Thus, in the approach
used here, we identify the optimal HSPs covering each frac-
tion of the exon, and compute separately the contribution of
each HSP into the score of the exon. In the next section, we
describe in detail how this computation is performed.
Scoring of Candidate Exons
Let e be one of the candidate exons predicted by GENEID
along the query DNA sequence S. In SGP2, the final score of e,
s(e), is computed as
s?e? = sg?e? + wst?e?
where sg(e) is the score given by GENEID to the exon e, and
st(e) is the score derived from the HSPs found by a TBLASTX
search overlapping the exon e. Both scores are log-likelihood
ratios (and we compute both base two). Assuming that both
components are independent, they can be summed up into a
single score. However, the assumption of independence is not
realistic, sg(e) depends on the probability of the sequence of e,
assuming that e codes for a protein, while st(e) depends on the
probability of the optimal alignment of e with a sequence
fragment of the mouse genome, assuming that both se-
quences code for related proteins. Obviously, these two prob-
abilities are not independent. Their joint distribution could
only be investigated—at least empirically—if the Markov
Model of coding DNA used in GENEID, and the substitution
matrix used by TBLASTX were inferred from the very same set
of coding sequences. Since this is quite difficult, if not unfea-
sible, we use an “ad hoc” coefficient, w, to weight the contri-
bution of TBLASTX search, st(e) into the final exon score.
We compute st(e) in the following way. Let h1···hqbe the
set of HSPs found by TBLASTX after comparing the query
sequence S against a database of DNA sequences (Fig. 2A).
First, we find the maximum scoring projection of the HSPs
onto the query sequence. We simply register the maximum
score among the scores of all HSPs covering each position,
and then partition the query sequence in equally maximally
scoring segments (bounded by dotted lines in Fig. 2A) x1···xr,
with scores sp(x1)···sp(xr) (Fig. 2B).
Then, for each predicted exon e (Fig. 2C), we find Xe, the
set of maximally scoring segments overlapping e
Xe= ?xi: xi∩ e ? ??
where a ∩ b denotes the overlap between sequence segments
a and b, and ? means no overlap. We compute st(e)in the
sp?x?|x ∩ e|
where ?a? denotes the length of sequence segment a.
That is, each exon gets the score of the maximally scor-
ing HSPs along the exon sequence proportional to the frac-
tion of the HSP covering the exon. In other words, st(e) is the
integral of the maximum scoring projection function within
the exon interval.
Once the scores s have been computed for all predicted
exons in the sequence S, gene prediction proceeds as usual in
GENEID: The gene structure is assembled maximizing the
sum of scores of the assembled exons.
In practice, we run SGP2 in the following way. Given a DNA
query sequence and a collection of DNA sequences, we com-
pare the query sequence against the collection using TBLASTX
2.0MP-WashU [23-Sep-2001]. The query sequence can be a
genomic fragment of any size, including complete eukaryotic
chromosomes, whereas the collection of sequences may be
almost anything from just a homologous region or a partial
collection of genomic sequences from the same or another
species to the whole genome sequence of a second species,
either completely assembled or in shotgun form at any degree
of coverage. In particular, two different regions of the same
genome coding for homologous genes can be used within
SGP2; in this case the same genome acts as target and infor-
In all the analyses reported here, we used BLOSUM62 as
the amino acid substitution matrix, but changed the penalty
for aligning any residue to a stop codon to ?500. This helps
to get rid of a large fraction of HSPs in noncoding regions.
Because of TBLASTX limitations, large query sequences may
need to be split in fragments before the search, and the results
reconstructed afterwards. Results of TBLASTX search are then
Parra et al.
parsed to obtain the maximum scoring projection of the HSPs
onto the query sequence. The parsing includes discarding all
HSPs below a given bit score cutoff, subtracting this value
from the score of the remaining HSPs, weighting the resulting
score by w (see above), and collapsing the HSPs in to the
maximum scoring projections. In all analyses described here,
the bit score cutoff was set to 50, and w to 0.20. These values
were chosen to optimize the gene predictions in sequence sets
of known homologous human and mouse genomic sequences
(see the Results section).
The maximum scoring projection is given to GENEID in
general feature format (GFF; R. Durbin and D. Haussler,
http://www.sanger.ac.uk/Software/GFF/). GENEID uses it to
rescore the exons predicted along the query sequence as ex-
plained, and assembles the corresponding optimal gene struc-
ture. GENEID was already designed to incorporate external
information into the gene predictions, and no changes were
required in the program to accommodate it into the SGP2
context, only a small adjustment in the parameter file to cope
with the change in scale of the exon scores.
We have written a simple PERL script which, given a
query DNA sequence and the results of the TBLASTX search,
performs all the components of the SGP2 analysis transpar-
ently: the parsing of the TBLASTX search results, and the
GENEID predictions. In the case wherein both the query and
the informant sequence are single genomic fragments, the
gene predictions can be obtained in both sequences (without
the need for a second TBLASTX search). The script, as well as
the individual components, can be found at http://www1.
GENEID has essentially no limits to the length of the
input sequence, and deals well with chromosome size se-
quences. Limits to the length of the input query sequence that
can be analyzed by SGP2 are, thus, those imposed by
TBLASTX. GENEID is quite fast; given the parsed TBLASTX
results, it takes 6 h to reannotate the whole human genome in
a MOSIX cluster containing four PCs (PentiumIII Dual 500
Accelerating TBLASTX Searches
TBLASTX searches, although efficient, are much slower. Its
default usage may become computationally prohibitive when
comparing complete eukaryotic genomes. In the context of
SGP2, however, a number of TBLASTX options can be
changed to speed up the search, without significant loss of
sensitivity in the predictions (see the Results section). Thus,
results in human chromosome 22 and whole-genome com-
parisons have been performed using the following set of pa-
rameters: W = 5, -nogap, -hspmax = 150,000, B = 200, V = 200,
E = 0.01, E2 = 0.01, Z = 30,000,000, -filter = xnu + seg, and
S2 = 80. In these cases, the query sequences have been broken
up in 5 MB fragments, and the database sequences in 10 MB
fragments. In all cases, stop codons are heavily penalized
(?500) in the alignments. After the search is completed, lo-
cations of the resulting HSPs are recomputed in chromosomal
coordinates. Results in the single-gene sequence benchmark
data sets were obtained with default TBLASTX parameters.
Sequence Data Sets
Benchmark Sequence Sets
To optimize some of the parameters in SGP2 and to test its
performance, we used a set of known pairs of genomic se-
quences coding for homologous human and rodent genes.
The set is built after the set constructed by Jareborg et al.
(1999). This is a set of 77 orthologous mouse and human gene
pairs. We considered only the 33 pairs of sequences in this set
explanation of the figure.
Rescoring of the exons predicted by GENEID according to the results of a TBLASTX search. See the “SGP2” section for a detailed
coding for single complete genes. In addition, we discarded
six additional pairs, when we suspected that one of the mem-
bers could be wrongly annotated. Orthology in the Jareborg et
al. (1999) data set is based on sequence conservation. This
could bias the set towards the more highly conserved human/
mouse orthologous genes. To compensate for this bias, we
obtained an additional set of pairs of human/rodent ortholo-
gous genes through an approach which does not involve se-
quence conservation: We obtained the set of pairs of human/
mouse sequences from the SWISSPROT database sharing the
prefix (indicating the gene) in their locus names. We kept
only those pairs for which it was possible to find the corre-
sponding annotated genomic sequence—including the map-
ping of the transcript, and not only of the coding regions—in
the EMBL database. Fifteen additional genes were found this
way. Three of them were discarded because we suspected
wrong annotation in at least one of the members of the pair.
We believe that orthology in the remaining cases is highly
likely because of the absolute conservation of the exonic
structure (number and length of exons, and intron phases)
that we observed. We will call the resulting concatenated set
of 39 pairs of human/mouse homologous genes the SCIMOG
dataset (from Sanger Center IMim Orthologous Genes). The
data set and the detailed protocol used to obtain it can be
accessed at http://www1.imim.es/datasets/sgp2002/.
To test the accuracy of SGP2, we used the data set con-
structed by Batzoglou et al. (2000) of 117 orthologous human
and mouse genes. We discarded those pairs in which in at
least one of the sequences contained multiple genes, and
those in which the coding region started in position 1 in one
of the sequences of the pair. This resulted in 110 genes. We
will call this set the MIT data set. There is some overlap be-
tween the SCIMOG and MIT data sets, and thus the latter
cannot properly be called a test set. However, we decided not
to eliminate the redundant entries, so that the results could be
compared to those published for the ROSSETA program (Bat-
zoglou et al. 2000).
Finally, we tested SGP2 in the complete sequence of hu-
man chromosome 22 (Dunham et al. 1999). The masked se-
quence was obtained from http://genome.cse.ucsc.edu/
goldenPath/22dec2001/. Chromosome 22 is probably the best
annotated human chromosome. We used the gene annota-
tions at http://www.cs.columbia.edu/∼vic/sanger2gbd/. The
CDS set contains 554 genes. This is a conservative set that
only contains the coding region of genes and does not include
pseudogenes. This may lead to an underestimation of the
specificity of the predictions.
Mouse and Human Genome Sequences
We used versions MGSCv3 of the mouse genome
(2,726,995,854 bp, http://genome.cse.ucsc.edu/goldenPath/
mmFeb2002/) and NCBI28 of the human genome
(3,220,912,202 bp, http://genome.cse.ucsc.edu/goldenPath/
22dec2001/). Both masked and unmasked sequences were ob-
tained from these locations. ENSEMBL gene annotations for
these genomes were obtained from http://genome.cse.
the human genome, and from http://genome.cse.ucsc.edu/
goldenPath/mmFeb2002/database/ensGene.txt.gz for the
mouse genome. ENSEMBL predicts 23,005 and 22,076
nonoverlapping transcripts genes on the human and mouse
The measures of accuracy used here are extensively discussed
in Burset and Guigo ´ (1996). We will restate them briefly. Ac-
curacy is measured at three different levels: nucleotide, exon,
and gene. At the nucleotide and exon levels, we compute
essentially the proportion of actual coding nucleotides/exons
that have been correctly predicted—which we call sensitivity—
and the proportion of predicted coding nucleotides/exons
that are actually coding nucleotides/exons—which we call
specificity. To compute these measures at the exon level, we
will assume that an exon has been correctly predicted only
when both its boundaries have been correctly predicted. To
summarize both sensitivity and specificity, we compute the cor-
relation coefficient at the nucleotide level, and the average of
sensitivity and specificity at the exon level. At the exon level,
we also compute the missing exons, the proportion of actual
exons that overlap no predicted exon, and the wrong exons,
the proportion of predicted exons that overlap no real exons.
At the gene level, a gene is correctly predicted if all of the
coding exons are identified, every intron–exon boundary is
correct, and all of the exons are included in the proper gene.
In addition, we compute the missed genes (MGs), real genes
for which none of its exons are overlapped by a predicted
gene, and the wrong genes (WGs), predictions for which none
of the exons are overlapped by a real gene. In general, gene
finders predict the initial and terminal exons very poorly.
This often leads to so-called chimeric predictions—one pre-
dicted gene encompassing more than one real gene—or to
split predictions—one real gene split in multiple predicted
genes. Reese et al. (2000) developed two measures, split genes
(SG) and joined genes (JG), to account for these tendencies.
SG is the total number of predicted genes overlapping real
genes divided by the number of genes that were split. Simi-
larly, JG is the total number of real genes that overlap pre-
dicted genes divided by the number of predicted genes that
We evaluated the accuracy of SGP2 using a number of differ-
ent data sets. The lack of a gold standard of gene prediction
makes it difficult to get accurate assessments from any single
data set. We primarily used three data sets as described earlier.
To benchmark SGP2, we constructed BLAST databases
from the mouse and human sections of SCIMOG and MIT,
and each mouse/human sequence to the entire human/
mouse database, respectively. This enabled us to predict genes
in both the mouse and human databases. The results from
Gene Prediction in the SCIMOG Data Set
SnSp CC SnSp(Sn+Sp)/2ME WE
SGP2 (single complete genes)
SGP2 (multiple genes)
Parra et al.
112 Genome Research
comparing SGP2, GENSCAN, and ROSSETA accuracy values in
this case are taken from Batzoglou et al. (2000), and the results
of a simple TBLASTX search on the MIT data set are in Table
2 (below). For the TBLASTX searches, the maximum scoring
projection of the HSPs (see the above section titled “SGP2”) was
assumed to be the gene prediction. The score cutoff for the
HSPs was chosen to maximize the correlation coefficient (CC)
between the projected HSPs and the coding exons. In Table
1,2, we report the accuracy of GENSCAN, SGP2, and TBLASTX
on the SCIMOG dataset. The accuracy values for SGP2 are
reported under two scenarios: assuming a single complete
gene and assuming multiple genes. Both GENEID and SGP2
allow the external specification of a gene model (i.e., a small
number of rules specifying the legal assemblies of exons into
gene structures). These rules can be used to force SGP2 to
predict a single complete gene to make the results comparable
to those of ROSSETA. Without such a restriction (i.e., making
no assumptions about the number and completeness of the
genes potentially encoded in the query sequence), the results
are more directly comparable to those of GENSCAN (although
GENSCAN also has a tendency to start a prediction in any
sequence with an initial exon, and to terminate it with a
The accuracy of SGP2 is comparable to that of ROSSETA,
and is significantly higher than that of GENSCAN. SGP2 also
improves substantially over a simple TBLASTX search. The
relative low specificity of the TBLASTX search—even after the
large penalties for stop codons—reflects the fact that a sub-
stantial fraction of the conservation between the human and
mouse genomes extends into the noncoding regions (Mouse
Genome Sequencing Consortium 2002). At the nucleotide
level, SGP2 accuracy is almost equal in the MIT data set and
the SCIMOG data set (even though the SGP2 was trained on
SCIMOG). The accuracy at the exact exon level, however, de-
creases, in particular when prediction of multiple genes is
allowed. This is a problem inherited from GENEID, which
tends to replace short initial and terminal exons with longer
Accuracy of SGP2 as a Function of the Coverage
of the Mouse Genome
To investigate the utility of partial shotgun data as informant
sequence in our approach based on TBLASTX, we simulated
shotgun mouse sequence data at different levels of coverage
(1.5x, 3x, and 6x) from the mouse genes in the SCIMOG data
set, and used them to compare the human sequences in
SCIMOG using TBLASTX. The mouse genomic sequences was
shredded with uniformly distributed length between 500 and
600 bp with random starting points. No sequencing errors
were introduced. At each coverage, we measured the CC be-
tween the TBLASTX hits projected along the human genome
sequence, and the coding exons (choosing the TBLASTX score
cutoff resulting in the optimal CC). With 1.5x coverage, a
substantial fraction of the human coding region is not iden-
tified by TBLASTX, whereas with 3x, the results are quite simi-
lar to those obtained with 6x, which are identical to those
obtained with the fully assembled syntenic regions (Table 3).
This indicates that even with 3x coverage of the informant
genome, our method will produce results nearly identical to
those obtained with fully assembled regions. Assembled ge-
nomes, however, result in faster TBLASTX searches.
Accuracy of SGP2 in Human Chromosome 22
Human chromosome 22 was the first human chromosome
fully sequenced (Dunham et al. 1999), and it is quite the best
annotated thus far, due to a number of experimental fol-
lowups (Das et al. 2001; Shoemaker et al. 2001). Therefore, it
provides an excellent data set to validate any gene prediction
technology. Human chromosome 22 was searched using
TBLASTX against the masked whole-genome assembly from
the mouse genome (MGSCv3). The HSPs in chromosomal co-
ordinates resulting from the TBLASTX search were used in
GENEID to perform SGP2 gene prediction. Although the HSPs
had been computed on the masked sequence, in this case the
SGP2 predictions were obtained on the unmasked one. SGP2
predicted 729 genes on human chromosome 22. Table 4
shows the comparative accuracy of the SGP2, GENSCAN,
GENOMESCAN, and pure ab initio GENEID predictions (with-
out TBLASTX data). GENSCAN predictions on the masked se-
quence were taken from the USCS genome browser http://
genome.cse.ucsc.edu/. GENOMESCAN predictions were ob-
tained from ftp://ftp.ncbi.nih.gov/genomes/H_sapiens/
build28_chr_genomescan.gtf.gz. Pure ab initio GENEID
predictions were obtained on the masked sequence, and
can also be downloaded from http://www1.imim.es/
Although SGP2 is not more sensitive than GENSCAN, it
appears to be more specific (as it utilizes the mouse genome).
Gene Prediction Accuracy in the MIT Data Set
Sn SpCCSnSp(Sn+Sp)/2 ME WE
SGP2 (single complete genes)
SGP2 (multiple genes)
the Degree of Coverage in the SCIMOG Data Set
Accuracy of TBLASTX Predictions as a Function of
SnSp CCME WE
Comparative Gene Prediction in Human and Mouse
Fifty percent of the GENSCAN-predicted exons do not overlap
annotated chromosome 22 exons; this number is only 31%
for SGP2. Overall, SGP2 appears to be more accurate than
GENSCAN in human chromosome 22: GENSCAN’s CC at the
nucleotide level is 0.64, whereas that of SGP2 is 0.73. Al-
though accuracy decreases for both programs when going
from single-gene sequences (Tables 1, 2) to an entire chromo-
some, SGP2 retains more accuracy. GENSCAN overall shows
higher sensitivity than SGP2, but there were 45 real genes not
predicted by GENSCAN on human chromosome 22, and
SGP2 was able to predict, at least partially, 15 of them. This
suggests that SGP2 and GENSCAN may play complementary
roles. GENOMESCAN, on the other hand, did not appear to be
superior to GENSCAN in human chromosome 22.
Mouse matches (TBLASTX HSPs) covered 11% of the hu-
man chromosome 22. Though they covered 85% of the cod-
ing nucleotides, 74% of the HSPs fell outside annotated cod-
ing regions. This illustrates the difficulties of using genome
sequence conservation even at the protein level between hu-
man and mouse genomes to infer coding genes.
Prediction of Genes in the Human and
We used SGP2 to predict the entire complement of human
(NCBI28) and mouse (MGSCv3) genes. The masked sequences
of these two genomes were compared using TBLASTX. The
TBLASTX HSPs were used within SGP2. SGP2 predicted 44,242
genes in the human genome, and 44,777 genes in the mouse
genome. Obviously, it is difficult to accurately assess these
predictions. We used ENSEMBL genes as the set of reference
annotations and compared both GENSCAN and SGP2 predic-
tions to it. Figure 3 shows summaries of the accuracy of SGP2
at the chromosome level in the human and mouse genomes.
When compared against ENSEMBL, SGP2 is more accurate
than GENSCAN.GENSCAN. It is more specific at the nucleo-
tide level: the average SGP2 specificity is 0.60 for human and
0.61 for mouse, whereas these values for GENSCAN are 0.43
and 0.44. SGP2 is also equally sensitive at the nucleotide level:
The average SGP2 sensitivity is 0.82 for human and 0.85 for
mouse; these values for GENSCAN are 0.82 and 0.84. Overall,
the average SGP2 CCs are 0.70 for human and 0.72 for mouse,
and for GENSCAN, the respective averages are 0.59 and 0.61.
The accuracy of the SGP2 predictions, moreover, appears to
be more consistent across chromosomes than that of the
GENSCAN predictions. Interestingly, human chromosome Y
is an outlier, with genes in this chromosome being poorly
predicted. Genes in chromosome Y appear to be more difficult
to predict than genes in other chromosomes for pure ab initio
gene prediction programs, because chromosome Y is also an
outlier for GENSCAN. SGP2 suffers, in addition, on human
chromosome Y because the mouse chromosome Y has yet to
be sequenced, and thus there was no comparative informa-
Overall, 23,913 of the human predictions and 24,203 of
the mouse predictions overlapped ENSEMBL genes, whereas
95% of the mouse and 93% of the human ENSEMBL genes
were among the genes predicted by SGP2. Of the remaining
putative novel 20,570 mouse SGP2 genes and 20,193 human
SGP2 genes, 10,456 mouse and 9,006 human predictions were
found to be similar at P < 10?6to a prediction in the coun-
terpart genome. Of these, 5,960 and 4,909 have multiple ex-
ons and are longer than 300 bp. A significant fraction of these
putative homologous predictions are likely to correspond to
real genes (Guigo ´ et al. 2003). The predictions are interac-
tively accessible through the USCS genome browser (http://
genome.cse.ucsc.edu/) and through the DAS server at
ENSEMBL (http://www.ensembl.org, under “DAS sources”).
The complete set of prediction files is available at http://
Speeding Up TBLASTX Searches
Using TBLASTX to compare human and mouse whole-
genome sequences, even in masked form, is quite expensive
computationally because of the 6-frame translation on both
query and target. To substantially reduce the search time, we
used a word size of 5 and sacrificed some sensitivity (see the
section above titled “Accelerating TBLASTX Searches” for de-
tails). We also penalized stop codons heavily and did not per-
mit gaps. The computation took an estimated 500 CPU days
on a farm of Compaq Alphas.
Accuracy in Tables 1 and 2 was computed using default
TBLASTX parameters. Table 5 shows the comparative accu-
racy of TBLASTX and SGP2 predictions, under the default and
the speed-up configuration of TBLASTX parameters on the
SCIMOG data set. The sensitivity of speed-up TBLASTX
searches drops from 0.89 to 0.72, but specificity increases
slightly. SGP2 is more robust, and it compensates for some of
the sensitivity lost in the TBLASTX search. Overall accuracy
for SGP2, as measured by the CC, drops only from 0.95 to
Predictions on human chromosome 22 and the whole
human and mouse genomes have been obtained with this
speed-up configuration of parameters.
We have described the program SGP2 for comparative gene
finding, and presented the results of its application to the
human and mouse genome sequences. Results in controlled
benchmark sequence data sets indicate that, by including in-
Accuracy of Gene-finding Programs on Human Chromosome 22
SnSpCCSn Sp(Sn+Sp)/2 MEWE SnSp (Sn+Sp)/2 MG WGJG SG
Parra et al.
formation from genome sequence conservation, predictions
by SGP2 appear to be more accurate than those obtained by
pure ab initio programs, exemplified here by GENSCAN and
GENEID. Although there is not a significant gain in sensitiv-
ity, the specificity of the predictions appears to increase sub-
stantially, and a smaller number of false positive exons are
Indeed, one the major obstacles towards the completion
of the catalog of human (mammalian) genes is our inability to
assess the reliability of the large number of computational
gene predictions that have not been verified experimentally.
Whereas the ENSEMBL pipeline produces about 25,000 hu-
man and mouse genes, the NCBI annotation pipeline predicts
predicts close to 55,000 genes in this species. Although a large
fraction of the ENSEMBL genes correspond to computational
predictions without experimental verification, the method is
quite conservative, and recent ex-
periments suggest that essentially
all ENSEMBL genes are indeed real
(Guigo ´ et al. 2003). The problem
remains with the tens of thousands
of additional computational predic-
tions that are not included in
ENSEMBL. A fraction of them are
likely to be real, but the question is
how large this fraction is. The re-
sults obtained here in human chro-
mosome 22 seem to indicate that it
may not be very large. Although the
existence of hundreds of unidenti-
fied genes in this chromosome can-
not be completely ruled out, the re-
sults strongly suggest that a sub-
stantial fraction of these additional
computational gene predictions are
In this regard, the results pre-
sented here demonstrate that
through the comparison of the hu-
man and mouse genomes using
SGP2 (or another available com-
parative gene prediction tool), the
false-positive rate can be reduced
significantly, and the catalog of
mammalian genes better defined.
SGP2 predicts a few thousand can-
didate genes not in ENSEMBL that
we believe are worth verifying ex-
perimentally. Indeed, the experi-
mental verification of a subset of
these provides evidence of at least
1000 previously nonconfirmed
genes (Guigo ´ et al. 2003).
The predictions by SGP2 ob-
tained here are, of course, still far
from definitively setting this cata-
log. For one thing, the mouse may
be too close a species to human: A
large fraction of the sequence has
been conserved between the ge-
nomes of these two species. Indeed,
most sequence conservation be-
tween human and mouse does not
correspond to coding exons (Mouse Genome Sequencing Con-
sortium 2002), compounding gene prediction. This suggests
that the genome of another vertebrate species evolutionarily
located between fish and mammals could be of great utility to-
wards closing in the vertebrate (and mammalian) gene catalog.
SGP2 is flexible enough so that it can be easily accom-
modated to analyze species other than human and mouse.
The fact that it can deal with shotgun data at any level of
coverage means that as the sequence of a new genome starts
becoming available, it can be used to improve the annotation
of other already existing genomes. Particularly relevant in this
context is a feature of SGP2 (and GENEID) that we have not
explored here. SGP2 can produce predictions on top of pre-
existing annotations. For instance, we could have given to
SGP2 the location and exonic coordinates (in GFF format) of
known REFSEQ genes (or ENSEMBL), and SGP2 would have
predicted genes only outside the boundaries of these genes of
measured in the entire chromosome sequences using the standard accuracy measures: SN, (sensitiv-
ity); SP, (specificity); CC, (correlation coefficient); SNe, (exon sensitivity); SPe, (exon specificity); and
SNSP, (average of sensitivity and specificity at exon level). Predictions from both programs were
compared against the human and mouse ENSEMBL annotations. Each dot corresponds to the accuracy
measure of one chromosome. Chromosome labels are shown for outlier values. The boxplots (Tukey
1977) were obtained using the R-package (http://cran.r-project.org/).
Accuracy of the human and mouse SGP2 and GENSCAN predictions. The accuracy was
Comparative Gene Prediction in Human and Mouse
already well known exonic structure. Preliminary results in-
dicate that this approach improves gene prediction outside of
the preassumed genes, and reduces the rate of chimeric pre-
dictions (i.e., predictions encompassing multiple genes).
Moreover, we believe that SGP2 can be substantially im-
proved. The flexibility of the SGP2/GENEID framework makes
it quite easy to integrate additional information that can con-
tribute to the accuracy of the predictions: synonymous versus
nonsynonymous substitution rates in the alignments by
TBLASTX, conservation of the splice signals in the informant
genome, amino acid substitution matrices specific to the phy-
logenetic distance between the species compared, etc.
In this regard, the reasons to use the default BLOSUM62
matrix are not obvious. Given the expected sequence similar-
ity between mouse–human orthologs, BLOSUM80 appears to
be a better choice. However, we intended to also detect diver-
gent families. Towards that end, the superiority of BLOSUM80
is less clear. We have compared TBLASTX search results on
human chromosome 22 against the whole mouse genome.
Whereas the HSPs resulting from the BLOSUM62 search cover
84% of the chromosome 22 coding nucleotides, BLOSUM80
HSPs cover 88% of them. However, BLOSUM80 is much less
specific than BLOSUM62: 60% of the nucleotides in the
BLOSUM62 HSPs fall outside coding regions, compared to
88% for BLOSUM80. It is thus clear that the optimal matrix or
combination of matrices for comparative gene-finding using
TBLASTX requires further investigation.
Although a large fraction of the human genome se-
quence has been known for more than a year, the exact num-
ber of human genes and their precise definition remain un-
known. Gene specification in higher eukaryotic sequences is
the result of the complex interplay of sequence signals en-
coded in the primary DNA sequence, which is only partially
understood. Without an exhaustive catalog of human genes,
however, the promises of genome research in medicine and
technology cannot be completely fulfilled. The work pre-
sented here, in which it is shown that human–mouse com-
parisons can contribute to the completion of the mammalian
(human) gene catalog, underscores the importance of the
comparisons of the genomes of different organisms to fully
understand the phenomenon of life, and in particular to de-
ciphering the mechanism, central to life, by means of which
the genome DNA sequence specifies the amino acid sequence
of the proteins.
We thank the Mouse Genome Sequencing Consortium for
providing the mouse genome sequence as well as support
throughout the analysis process. We especially thank Fran-
cisco Câmara for arranging the data listed in the gene-
prediction page on our group Web site, and for setting up and
taking care of our DAS server. We also thank Ian Korf for
inspiring discussions regarding the parameters to use in the
TBLASTX search. We thank Enrique Blanco, Sergi Castellano,
and Moisés Burset for helpful discussions and constant en-
couragement. This work was supported by a grant from Plan
Nacional de I+D (BIO2000-1358-C02-02), Ministerio de Cien-
cia y Tecnologia (Spain), and from a fellowship to J.F.A. from
the Instituto de Salud Carlos III (99/9345).
The publication costs of this article were defrayed in part
by payment of page charges. This article must therefore be
hereby marked “advertisement” in accordance with 18 USC
section 1734 solely to indicate this fact.
Altschul, S.F., Gish, W., Miller, W., Myers, E.W., and Lipman, D.
1990. Basic local alignment search tool. J. Mol. Biol.
Altschul, S.F., Madden, T., Schaffer, A., Zhang, J., Zhang, Z., Miller,
W., and Lipman, D. 1997. Gapped BLAST and PSI-BLAST: A new
generation of protein database search programs. Nucleic Acids Res.
Bafna, V. and Huson, D.H. 2000. The conserved exon method. Proc.
Int. Conf. Intell. Syst. Mol. Biol. 8: 3–12.
Batzoglou, S., Pachter, L., Mesirov, J.P., Berger, B., and Lander, E.S.
2000. Human and mouse gene structure: Comparative analysis
and application to exon prediction. Genome Res. 10: 950–958.
Birney, E. and Durbin, R. 1997. Dynamite: A flexible code generating
language for dynamic programming methods used in sequence
comparison. Proc. Int. Conf. Intell. Syst. Mol. Biol. 5: 56–64.
Blayo, P., Rouzé, P., and Sagot, M.-F. 2002. Orphan gene
finding—An exon assembly approach. Theoretical Computer
Science (in press).
Borodovsky, M. and McIninch, J. 1993. GenMark: Parallel gene
recognition for both DNA strands. Comput. Chem. 17: 123–134.
Burge, C.B. and Karlin, S. 1997. Prediction of complete gene
structures in human genomic DNA. J. Mol. Biol. 268: 78–94.
Burge, C.B. and Karlin, S. 1998. Finding the genes in genomic DNA.
Curr. Opin. Struct. Biol. 8: 346–354.
Burset, M. and Guigo ´, R. 1996. Evaluation of gene structure
prediction programs. Genomics 34: 353–357.
Claverie, J.-M. 1997. Computational methods for the identification
of genes in vertebrate genomic sequences. Hum. Mol. Genet.
Crollius, H.R., Jaillon, O., Bernot, A., Dasilva, C., Bouneau, L.,
Fischer, C., Fizames, C., Wincker, P., Brottier, P., Quetier, F., et
al. 2000. Estimate of human gene number provided by
genome-wide analysis using Tetraodon nigroviridis DNA sequence.
Nat. Genet. 25: 235–238.
Das, M., Burge, C.B., Park, E., Colinas, J., and Pelletier, J. 2001.
Assessment of the total number of human transcription units.
Genomics 77: 71–78.
Dunham, I., Hunt, A.R., Collins, J.E., Bruskiewich, R., Beare, D.M.,
Clamp, M., Smink, L.J., Ainscough, R., Almeida, J.P., Babbage, A.,
et al. 1999. The DNA sequence of human chromosome 22.
Nature 402: 489–495.
Durbin, R., Eddy, S., Crogh, A., and Mitchison, G. 1998. Biological
sequence analysis: Probabilistic models of protein and nucleic acids.
Cambridge University Press, Cambridge.
Gelfand, M.S., Mironov, A.A., and Pevzner, P.A. 1996. Gene
recognition via spliced alignment. Proc. Natl. Acad. Sci.
Gish, W. and States, D. 1993. Identification of protein coding
regions by database similarity search. Nat. Genet. 3: 266–272.
Accuracy of TBLASTX and SGP2 Predictions Using “Default” versus Speed-Up Parameters
SnSp CCSn Sp(Sn+Sp)/2 MEWE
Parra et al.
Guigo ´, R. 1998. Assembling genes from predicted exons in linear
time with dynamic programming. J. Comp. Biol. 5: 681–702.
Guigo ´, R. and Wiehe, T. 2003. Gene prediction accuracy in large
DNA sequences. In Frontiers in computational genomic (eds. M.Y.
Galperin and E.V. Koonin), Caister Academic Press, Norfolk, UK.
Guigo ´, R., Knudsen, S., Drake, N., and Smith, T.F. 1992. Prediction
of gene structure. J. Mol. Biol. 226: 141–157.
Guigo ´, R., Agarwal, P., Abril, J.F., Burset, M., and Fickett, J.W. 2000.
Gene prediction accuracy in large DNA sequences. Genome Res.
Guigó, R., Dermitzakis, E.T., Agarwal, P., Pontig, C.P., Parra, G.,
Reymond, A., Abril, J.F., Keibler, E., Lyle, R., Ucla, C., et al. 2003.
Comparison of mouse and human genomes followed by
experimental verification yields an estimated 1,019 additional
genes. Proc. Natl. Acad. Sci. (in press).
Haussler, D. 1998. Computational genefinding. Trends in biochemical
sciences, supplementary guide to bioinformatics, pages 12–15.
Jareborg, N., Birney, E., and Durbin, R. 1999. Comparative analysis
of noncoding regions of 77 orthologous mouse and human gene
pairs. Genome Res. 9: 815–824.
Korf, I., Flicek, P., Duan, D., and Brent, M.R. 2001. Integrating
genomic homology into gene structure prediction. Bioinformatics
17 Suppl 1: 140–148.
Meyer, I.M. and Durbin, R. 2002. Comparative ab initio prediction
of gene structures using pair HMMs. Bioinformatics
Miller, W. 2001. Comparison of genomic DNA sequences: Solved
and unsolved problems. Bioinformatics 17: 391–397.
Mouse Genome Sequencing Consortium 2002. Initial sequencing
and comparative analysis of the mouse genome. Nature 420:
Pachter, L., Alexandersson, M., and Cawley, S. 2002. Applications of
generalized pair hidden Markov models to alignment and gene
finding problems. J. Comp. Biol. 9: 389–400.
Parra, G., Blanco, E., and Guigo ´, R. 2000. Geneid in Drosophila.
Genome Res. 10: 511–515.
Pedersen, C. and Scharl, T. 2002. Comparative methods for gene
structure prediction in homologous sequences. In Algorithms in
Bioinformatics (eds. R. Guigo ´, and D. Gusfield), Springer-Verlag,
Reese, M.G., Hartzell, G., Harris, N.L., Ohler, U., Abril, J.F., and
Lewis, S.E. 2000. Genome annotation assessment in Drosophila
melanogaster. Genome Res. 10: 483–501.
Rinner, O. and Morgenstern, B. 2002. Agenda: Gene prediction by
comparative sequence analysis. In Silico Biol. 2: 0018.
Rogic, S., Mackworth, A.K., and Ouellette, F. 2001. Evaluation of
gene-finding programs on mammalian sequences. Genome Res.
Shoemaker, D.D., Schadt, E.E., Armour, C.D., He, Y.D.,
Garrett-Engele, P., McDonagh, P.D., Loerch, P.M., Leonardson,
A., Lum, P.Y., Cavet, G., et al. 2001. Experimental annotation of
the human genome using microarray technology. Nature
Tukey, J.W. 1977. Exploratory data analysis. pp. 39–41.
Addison-Wesley, Boston, MA.
Wiehe, T., Guigo ´, R., and Miller, W. 2000. Genome sequence
comparisons: Hurdles in the fast lane to functional genomics.
Brief. Bioinform. 1: 381–388.
Wiehe, T., Gebauer-Jung, S., Mitchell-Olds, T., and Guigo ´, R. 2001.
SGP-1: Prediction and validation of homologous genes based on
sequence alignments. Genome Res. 11: 1574–1583.
Yeh, R., Lim, L., and Burge, C. 2001. Computational inference of
homologous gene structures in the human genome. Genome Res.
Zhang, M.Q. 2002. Computational prediction of eukaryotic
protein-coding genes. Nat. Rev. Genet. 3: 698–709.
WEB SITE REFERENCES
http://www.sanger.ac.uk/Software/formats/GFF/; GFF format
http://genome.cse.ucsc.edu/goldenPath/22dec2001/; Human genome
sequence goldenpath from Dec. 22, 2001 (hg10) equivalent to
genome sequence goldenpath from Feb. 2002 (mm2) equivalent
http://www.cs.columbia.edu/∼vic/sanger2gbd; Victoria Haghighi,
Human chromosome 22 curated annotations.
gtf.gz; Genomescan predictions from NCBI.
ensGene.txt.gz; Mouse ENSEMBL annotations file.
http://blast.wustl.edu; Washington University BLAST Archives
ensGene.txt.gz; Human ENSEMBL annotations file.
http://genome.cse.ucsc.edu; UCSC genome browser.
http://www.ensembl.org; ENSEMBL genome browser.
http://www1.imim.es/genepredictions/; GENEID and SGP2 full data
http://www1.imim.es/software/sgp2/; SGP2 home page.
http://www1.imim.es/datasets/sgp2002/; SGP2 training data sets
Received November 4, 2002; accepted in revised form November 15, 2002.
Comparative Gene Prediction in Human and Mouse