BIOINFORMATICS
Vol. 17 no. 1 2001
Pages 13–15
Pro-Frame: similarity-based gene recognition in
eukaryotic DNA sequences with errors
Andrey A. Mironov, Pavel S. Novichkov and Mikhail S. Gelfand
State Scientific Center for Biotechnology NIIGenetika, Moscow, 113545, Russia
Received on September 22, 1999; revised on June 5, 2000; accepted on July 27, 2000
ABSTRACT
Summary: Performance of existing algorithms for
similarity-based gene recognition in eukaryotes drops
when the genomic DNA has been sequenced with errors. A modification of the spliced alignment algorithm
allows for gene recognition in sequences with errors, in
particular frameshifts. It tolerates up to 5% of sequencing
errors without considerable drop of prediction reliability
when a sufficiently close homologous protein is available
(normalized evolutionary distance similarity score 50% or
higher).
Availability: The program is free for academic users and
available upon request at http://www.anchorgen.com
Contact: [email protected]
Analysis of sequence similarity is a powerful tool for
gene recognition. It is employed in a number of database
search programs, most notably BLASTX (Gish and States,
1993), and programs for exact prediction of exon–intron
structure, in particular, Procrustes (Gelfand et al., 1996;
Mironov et al., 1998), INFO (Hultner et al., 1994; Laub
and Smith, 1998), GeneWise (Birney and Durbin, 1997).
The common idea behind these algorithms is that among
numerous possible exon chains, an algorithm chooses the
chain having the highest similarity to a related protein
(target). This is done by modified dynamic programming
treating introns as a special case of gaps (GeneWise) or by
spliced alignment (Procrustes).
Testing of the similarity-based gene recognition programs demonstrated that given sufficiently close relatives,
they produce highly reliable predictions. In particular, the
correlation between predicted and real human genes is
96–99% when homologous vertebrate genes are available
(Mironov et al., 1998; Laub and Smith, 1998). However,
the quality of gene predictions when the genomic DNA
contains sequencing errors is much lower (Burset and
Guigo, 1996). One possibility to avoid this problem is to
use the DNA spliced alignment instead of aligning translated candidate exons with proteins (Sze and Pevzner,
1997). However, it is well known that protein alignments
are much more sensitive to distant similarities than
nucleotide alignments. Thus it is indicative that there exist
c Oxford University Press 2001
numerous protein–DNA alignment algorithms accounting
for frameshifts (Posfai and Roberts, 1992; Birney et al.,
1996; Guan and Uberbacher, 1996; Zhang et al., 1997;
Pearson et al., 1997). However, none of them handles
introns.
We have implemented a modified version of the spliced
alignment algorithm performing gene recognition in the
presence of frameshift errors. The algorithm treats introns
as non-penalized gaps that may start only at dinucleotide
GT and end at dinucleotide AG. Frameshifts and in-frame
stop codons in the genomic sequence are allowed, but
heavily penalized. There is an option for acceleration
of the dynamic programming, using the k-tuple alignment technique due to M.Roytberg (Nazipova et al.,
1995). Since sequencing errors can destroy the invariant
dinucleotides at splicing sites, the program has a postprocessing step. At this step the program identifies runs
of deletions at exon termini, and moves the exon–intron
boundary even if there are no suitable dinucleotides. More
exactly, observing more than 50% deleted positions in the
region (−30, +30) around the exon junction, the program
searches for the optimal position of the donor and acceptor
splicing sites allowing for a single deviation from the
invariant dinucleotide at each site. The program outputs
the exon positions before and after the correction and the
alignment of the predicted exons and the target protein.
Results of testing the algorithm on a sample of human
genes and related proteins from (Mironov et al., 1998) are
given in Figure 1. This sample consists of 256 genes. The
average length of genomic sequences is approximately
8100 nucleotides, with the longest sequence exceeding
180 000 nucleotides. The number of exons ranges from 1
through 54, the average number of exons per gene is 5.5.
The average length of exons in multi-exon genes is 140
nucleotides. The total number of protein targets is 731,
their average length is 575 amino acids. Five independent
rounds of mutations were performed for each sequence.
The 3655 predictions (five times 731 comparisons) were
done in about 5 h on a PC with Pentium II 400 MHz
processor under Windows NT.
The performance at different error levels is estimated
using the standard correlation coefficient measure (Burset
13
A.A.Mironov et al.
100
100
90
90
80
Pro
0%
70
1%
2%
60
0%
1%
70
2%
3%
60
3%
4%
50
4%
50
5%
5%
6%
40
6%
40
8%
10%
30
8%
30
10%
20
12%
12%
15%
20
15%
10
10
0
0
0
10
20
30
40
50
60
70
80
90
100
Fig. 1. Testing of Pro-Frame on a sample of human genes.
Horizontal axis: similarity between actual genes and related proteins
(in %, see the text for definition). Vertical axis: correlation
coefficient (in %). Each curve corresponds to a specific level of
sequencing errors (the percent of erroneous positions is given in the
legend on the right). ‘Pro’ corresponds to the original Procrustes
algorithm.
and Guigo, 1996; Mironov et al., 1998). For comparison
we also present the correlation coefficient demonstrated
by the original Procrustes algorithm and results of gapped
BLASTX (Altschul et al., 1997) using the target protein
as the query sequence (Figure 2). Since the performance
depends on the similarity between the gene and a target,
the figures feature plots of the correlation coefficient
at different similarity levels. The similarity measure is
the score of the alignment of the actual and target
proteins divided by the half-sum of the scores of (trivial)
alignments of the actual protein and the target protein
with themselves. Such normalization accounts for varying
protein length and amino acid composition. Sequencing
errors were modeled as random nucleotide substitutions
(80%), insertions (10%) and deletions (10%); the latter
two types of errors had length one through three with equal
probabilities. For instance, at the error rate 15% this means
that nucleotides at 12% of all positions have been changed,
3% of nucleotides were deleted (0.5% of single nucleotide
deletions, 0.5% of dinucleotide deletions, and 0.5% of
trinucleotide deletions), and 1.5% positions contained
insertions (with 0.5% being single nucleotides, 0.5%,
dinucleotides, and 0.5%, trinucleotides). Errors at GT-AG
invariant intron termini were not treated as a special case,
however, the fraction of mutated termini can be easily
estimated given the overall error rate.
At all similarity and error levels Pro-Frame provides
better recognition than straightforward BLASTX (cf.
Figures 1 and 2). It is noteworthy that in the absence of
sequencing errors Pro-Frame performs almost as well as
Procrustes when the target protein is close to the analyzed
gene, but the performance drops for distant relatives. This
agrees with our observations about importance of the
14
80
0
10
20
30
40
50
60
70
80
90
100
Fig. 2. Prediction of coding regions in the sample of Figure 1 using
BLASTX. Axes and notation as in Figure 1.
statistical filtering procedure implemented in Procrustes
(Mironov et al., 1998). On the other hand, up to 3% rate
of sequencing errors does not considerably influence the
reliability of predictions, and further, up to 6% of errors
are easily tolerated if the target protein is sufficiently close
to the analyzed gene.
The above results demonstrate that Pro-Frame may be
a useful tool for analysis of preliminary sequencing data,
e.g. phase I or II output of major sequencing projects, draft
human genome sequences, etc. The initial identification of
target proteins should be done by BLASTX, and then ProFrame can be used to exactly map the exon boundaries
and to find relatively short exons that could be missed by
the straightforward similarity analyses. Since Pro-Frame
does not rely on the statistical properties of the analyzed
genome (the only requirement is the GT–AG rule for
the intron termini), the program can be used for gene
recognition in invertebrate, plant, fungal, and even protist
sequences.
ACKNOWLEDGEMENTS
We are grateful to Drs V.Bafna, J.Fickett, P.Pevzner
and M.Roytberg for useful discussions. This work was
partially supported by grants from the Russian State
Scientific Program ‘Human Genome’, the Russian Fund
for Basic Research (99-04-48347 and 00-15-99362), the
Merck Genome Research Institute (244), and INTAS (991476).
REFERENCES
Altschul,S.F., Madden,T.L., Schaffer,A.A., Zhang,J., Zhang,Z.,
Miller,W. and Lipman,D.J. (1997) Gapped BLAST and PSIBLAST: a new generation of protein database search programs.
Nucleic Acids Res., 25, 3389–3402.
Birney,E., Thompson,J.D. and Gibson,T.J. (1996) PairWise and
SearchWise: finding the optimal alignment in a simultaneous
comparison of a protein profile against all DNA translation
frames. Nucleic Acids Res., 24, 2730–2739.
Birney,E. and Durbin,R. (1997) Dynamite: a flexible code generating language for dynamic programming methods used in sequence comparison. Proceedings of the 5th International Conference on Intelligent Systems for Molecular Biology AAAI
Press, Menlo Park, CA, pp. 56–64.
Burset,M. and Guigo,R. (1996) Evaluation of gene structure prediction programs. 34, 353–367.
Gelfand,M.S., Mironov,A.A. and Pevzner,P.A. (1996) Gene recognition via spliced sequence alignment. Proc. Natl. Acad. Sci.
USA, 93, 9061–9066.
Gish,W. and States,D.J. (1993) Identification of protein-coding
regions by database similarity search. Nature Genet., 3, 266–272.
Guan,X. and Uberbacher,E.C. (1996) Alignments of DNA and
protein sequences containing frameshift errors. Comput. Applic.
Biosci., 12, 31–40.
Hultner,M., Smith,D.W. and Wills,C. (1994) Similarity lanscapes: a
way to detect many structural and sequence motifs in both introns
and exons. J. Mol. Evol., 38, 188–203.
Laub,M.T. and Smith,D.W. (1998) Finding intron/exon splice
junctions using INFO, INterruption Finder and Organizer. J.
Comput. Biol., 5, 307–321.
Mironov,A.A., Roytberg,M.A., Pevzner,P.A. and Gelfand,M.S.
(1998) Performance-guarantee gene predictions via spliced
alignment. Genomics, 51, 332–339.
Nazipova,N.N., Shabalina,S.A., Ogurtsov,A.Yu., Kondrashov,A.S.,
Roytberg,M.A., Buryakov,G.V. and Vernoslov,S.E. (1995) SAMSON: a software package for the biopolymer primary structure
analysis. Comput. Applic. Biosci., 11, 423–426.
Pearson,W.R., Wood,T., Zhang,Z. and Miller,W. (1997) Comparison
of DNA sequences with protein sequences. Genomics, 46, 24–36.
Posfai,J. and Roberts,R.J. (1992) Finding errors in DNA sequences.
Proc. Natl. Acad. Sci. USA, 89, 4698–4702.
Sze,S.-H. and Pevzner,P.A. (1997) Las Vegas algorithms for gene
recognition: suboptimal and error-tolerant spliced alignment. J.
Comput. Biol., 4, 297–310.
Zhang,Z., Pearson,W.Ri. and Miller,W. (1997) Aligning a DNA
sequence with a protein sequence. J. Comput. Biol., 4, 339–349.
15