Nature Genetics
25, 235 - 238 (2000)
doi:10.1038/76118
Estimate of human gene number provided by genome-wide analysis using
Tetraodon nigroviridis DNA sequenceHugues Roest Crollius, Olivier Jaillon, Alain Bernot, Corinne Dasilva, Laurence Bouneau, Cécile Fischer, Cécile Fizames, Patrick Wincker, Philippe Brottier, Francis Quétier, William Saurin
& Jean WeissenbachGenoscope and CNRS FRE2231, Evry cedex,
France.
Correspondence should be addressed to Jean Weissenbach jsbach@genosocope.cns.frThe number of genes in the human genome is unknown, with estimates ranging
from 50,000 to 90,000 (refs 1,2), and to more than 140,000 according to unpublished sources. We
have developed 'Exofish', a procedure based on homology searches,
to identify human genes quickly and reliably. This method relies on the sequence
of another vertebrate, the pufferfish Tetraodon nigroviridis, to detect
conserved sequences with a very low background. Similar to Fugu rubripes
, a marine pufferfish proposed by Brenner et al.3
as a model for genomic studies, T. nigroviridis is a more practical
alternative4 with a genome also eight times more compact than
that of human. Many comparisons have been made between F. rubripes
and human DNA that demonstrate the potential of comparative genomics using
the pufferfish genome5. Application of Exofish to the December
version of the working draft sequence of the human genome and to Unigene showed
that the human genome contains 28,000−34,000 genes, and that Unigene
contains less than 40% of the protein-coding fraction of the human genome.
To determine the conditions that would generate alignments in coding regions
between human DNA and a pufferfish distant by 400 million years, we first
tested a large number of BLAST conditions on a small set of 13 annotated human-pufferfish
homologous genes (Table 1). We used F. rubripes
genes because no complete T. nigroviridis gene sequence existed
at the time of this work. We then applied the optimal conditions to a larger
set of 322 annotated human genes and the partial T. nigroviridis genome
sequence (33% of which has been determined), in which the positions of genes
are unknown. We found that the existing sequence of the T. nigroviridis
genome detects 26.5% of the 2,693 human exons in conditions in which
no alignments fall in introns (Fig. 1a). The
724 exons detected are distributed in 64.9% of the genes (209/322). To estimate
the influence of the amount of T. nigroviridis genome sequenced on
the sensitivity of this approach in detecting exons and genes in human DNA,
we represented the fraction of exons and genes identified with increasing
amounts of T. nigroviridis sequence (Fig. 1b).
The fraction of human exons detected increases at a rate proportional to the
amount of T. nigroviridis genome coverage generated. The probability
of identifying a gene by at least one of its exons is higher because genes
in general contain many exons, in addition to the fact that the random sequence
tag (RST) database represents approximately 170,000 random sequences in the
genome.
 | | |
| | |
To reflect the fact that different T. nigroviridis sequences may
generate overlapping alignments over the same exon and define a single, conserved
human region, we defined the contiguous assembly of the different overlapping
alignments as an 'ecore' (for evolutionary conserved region).
In the set of 322 reference genes, the 209 genes (or 724 exons) that were
detected by T. nigroviridis contained 831 ecores (2.58 ecores per gene).
This result (Fig. 1a) provides a means to decide
if new alignments between human and T. nigroviridis DNA overlap human
exons, based on their length and percentage of identity. This criterion is
the basis of the Exofish (for exon finding by sequence homology) selection
mechanism (Fig. 2). To confirm the sensitivity of Exofish
in detecting human genes, we performed a second comparison on a set of 4,888
complete human cDNA sequences extracted from Unigene version 105 (ref. 6). Using this set, 70% of the genes were identified, and
each gene contained an average of 3.18 ecores (including the 30% of undetected
genes). This ratio was used to derive a number of genes from a given number
of ecores detected by Exofish.
| | |
We analysed the sequence of chromosome 22 (ref. 7)
with Exofish to estimate its capacity to confirm existing annotations and
to detect new genes. We found 1,525 ecores over the complete length of the
chromosome (Fig. 3). The distribution of ecores among
the different types of annotated features showed important variations (Table 2). Related genes (based on homologies to protein and
genes from human and other species) and predicted genes (based on EST sequences)
contained less ecores than known genes. These two categories of annotations
also contained less annotated exons per gene, presumably because their respective
counterparts in sequence databases are only partially homologous or are incomplete.
In fact, 70 of 148 predicted genes consisted of a single exon. We estimate
that approximately 50% of the 181 ecores that fell outside of annotations
belonged to genes that are incompletely annotated or to pseudogenes. Therefore,
the remaining 90 ecores corresponded to approximately 30 novel genes on chromosome
22 (Fig. 3b,c). We thus estimate that
chromosome 22 contains less than 600 genes.
| | Figure 3. Examples of chromosome 22 results. | | |  | Open blue boxes linked by broken lines represent gene annotations from
ref. 7. Red boxes represent exons predicted by
Genscan. Green boxes represent ecores generated by Exofish that overlap gene
annotations, and dark blue boxes represent ecores that do not overlap annotations.
The scale in nucleotides is relative to the sequence described in ref. 7. a, Typical result in which a known gene with
19 exons (encoding carnitine palmitoyltransferase I) is partially predicted
by Genscan (17 exons) and Exofish (9 exons). b, On the 'Up'
strand (above the scale), five ecores indicate a new gene that is not predicted
by Genscan, whereas two genes on the top and the bottom strand (similar to
mouse Htf9c and Ranbp1, respectively) have several exons predicted
by Exofish, whereas none are correctly predicted by Genscan. c, On
the 'Up' strand, both Genscan and Exofish partially confirm a
known gene (HMG2L1), whereas on the same strand a new gene seems to
be predicted by both approaches. d, LIMK2 has 16 annotated
exons, of which 14 are predicted by Genscan and 6 by Exofish. Two additional
exons that are presumably alternatively spliced are predicted by Exofish (arrows),
one of which is also predicted by Genscan.
 Full Figure and legend (34K) |
| |
| | |
Of the 1,344 ecores that fell within the boundaries of the annotations,
1,197 (89%) corresponded to genes and 147 (11%) to pseudogenes. To estimate
the sensitivity of Exofish in detecting genes on chromosome 22, we considered
only the 247 known genes, because others are likely to be incomplete. Ecores
were found in 32.0% of the 2,298 exons and 66.8% of the 247 known genes. These
values are comparable to the 26.5% of exons and 64.5% of genes identified
in the reference set of 322 human genes, and to the 70% sensitivity obtained
on 4,888 full-length cDNA sequences. Exofish detects only 8% of the 325 genes
predicted by Genscan that are not confirmed by homologies (compared with 64.5%
for known genes), suggesting that most of these predictions are false positives.
It is possible to exploit the compactness of the T. nigroviridis
genome to confirm that several neighbouring ecores that fell outside of existing
annotations do belong to the same gene. For instance, the five isolated ecores
(Fig. 3c) were joined by three T. nigroviridis
RSTs. Subsequent to the release of the sequence of chromosome 22 (ref. 7), a human cDNA clone and a homologous gene in a Caenorhabditis
elegans cosmid clone have confirmed that these ecores define a true gene.
By contrast, ecores identified inside the boundaries of the 545 annotated
genes, but outside exons (that is, in introns), would correspond to exons
that remained undetected by other homology-based approaches, presumably because
of alternative splicing. We found 25 ecores in the introns of 21 annotated
genes, of which 19 were also predicted by Genscan (Fig. 3
d). Approximately 50% of ecores that fell either in introns or
outside of annotations have been confirmed as exons by the chromosome 22 annotation
team at the Sanger Centre (J. Collins, D. Beare and I. Dunham, pers. comm.).
To estimate the number of genes in the human genome, we analysed the human
working draft sequence with Exofish. In release 61 (December 1999) of the
EMBL database, the publicly available human working draft sequence contained
1,272.3 Mb of non-redundant human DNA. Analysis of this fraction of the human
genome ( 42.4%) by Exofish generated 42,066 ecores. Results on human chromosome
22 indicated that 89% of ecores fell in genes, whereas the remaining 11% fell
in pseudogenes. Based on the result that Exofish detects on average 3.18 ecores
per human gene, the human genome would contain (42,066 0.89)/0.424 =
88,299 ecores and 88,299/3.18 = 27,767 genes. We estimated the gene distribution
for each chromosome and compared the results with the EST gene map of the
human genome8 (Fig. 4). The gene-dense
chromosomes (17, 19, 22) have an excess of ecores compared with ESTs, as does
chromosome 16. To set an upper limit to our estimate, another calculation
was based on the lower ratio of ecores per gene found in the initial gene
test set and gave 88,299/2.58 = 34,224 genes. We therefore estimate that the
human genome contains 28,000−34,000 genes.
| | |
We used Exofish on the non-redundant set of human gene sequences represented
by Unigene6 to estimate more accurately the fraction of protein-coding
DNA present in publicly available databases. Release 105 of Unigene contains
10,501 clusters represented by known genes, whereas the remaining 82,430 clusters
only contain EST sequences. When matched to the selected sequences representing
Unigene clusters, Exofish detected 33,079 ecores, which identified 62% of
the 10,501 known genes and only 4.2% of the EST sequences. As the human genome
is estimated to contain 88,299 ecores, the 33,079 ecores found in Unigene
represent only 37.5% of the coding fraction of human genes. This result is
coherent with the very low number of matches obtained on the EST sequences.
Most selected ESTs representing Unigene clusters (87%) are 3' reads
of cDNA clones, and most likely correspond to untranslated regions.
Because a genome is a finite entity that contains all genes with all exons
necessary to express all the proteins required at any stage or in any tissue,
the sensitivity of Exofish is not biased by the traditional problems encountered
in cDNA databases, such as alternative splicing and varying gene-expression
levels. This is confirmed by the fact that Exofish identifies the same fraction
of genes (2/3) in three collections of human genes of diverse origins
and characteristics. Our finding that the human genome contains only 28,000−34,000
genes is unexpected, considering that it corresponds to just over twice the
number of genes in the fly or worm. It is therefore to be expected that organismal
complexity is not a direct consequence of gene number, but has its source
in other mechanisms that may include alternative splicing and multi-domain
proteins. As Unigene contains 92,000 clusters, and Exofish predicts 28,000−34,000
genes in the genome, Unigene is partially redundant and also contains mostly
non-coding sequences. It is likely, however, that Unigene contains a 'tag'
for most human genes and as such is an invaluable resource for gene identification.
Exofish still cannot detect one-third of human genes (false negatives), including
those for which the corresponding T. nigroviridis sequence is not yet
known, those that evolve rapidly and for which protein sequence similarity
is weak, and those that are strictly specific to mammals. It is likely, however,
that smaller protein domains also participate in the detection process and
enable Exofish to detect genes outside the limits of orthologous or paralogous
genes. As described here, two immediate applications for Exofish include the
annotation of genomic DNA and the estimation of the coding fraction in cDNA
collections. Exofish also enables comparison of the T. nigroviridis
genome with entire vertebrate genomes at the protein level in a few hours
of computation time, and as such it is a powerful tool to explore new avenues
in vertebrate genome research in a way so far only possible for bacteria or
unicellular eukaryotes.
Methods
Construction of Exofish.
A summary of the approach
used to select the optimal BLAST conditions for Exofish is shown (
Table 1) and a full description is available (see
Methods, http://www.nature.com/ng/supplementary_info).
Exofish is available as an annotation tool for human sequences (http://www.genoscope.cns.fr/exofish
). We constructed a set of 322 complete human genes by global BLASTN
alignments between a database of 10,067 human mRNA sequences and 3,930 genomic
clones. Details of the parameters and selections used, as well as a file of
the 322 human genes in fasta format, are available (see
Methods, http://www.nature.com/ng/supplementary_info).
T. nigroviridis genomic sequence.
BAC library
construction and insert-end sequencing are described elsewhere9
Genomic DNA from a male T. nigroviridis specimen (ascertained as
T. nigroviridis using morphological and mitochondrial DNA sequence characteristics)
was extracted to construct a plasmid library. DNA was mechanically sheared,
separated on a preparative agarose gel, and a size fraction corresponding
to 3 kb was excised, end-repaired and cloned in pcDNA2. After electroporation
in DH10B electrocompetent cells, clones were plated on 2YT agar plates containing
70  g/ml carbenicilin and 100,000 recombinants were robotically picked
and replicated in microtitre plates. We sequenced 127,229 insert ends as described9. Including the BAC ends database, 174,828 sequences of an average
useful length of 886 nt were produced, equivalent to 154.9 Mb of combined
DNA. T. nigroviridis repeats (rRNAs, transposable elements, satellites)
and microsatellite repeats were masked following BLASTN alignments against
a T. nigroviridis repeat database9 and microsatellite
database, respectively10. Minisatellites were identified by
Tandem Repeat Finder11 and subsequently masked. Microsatellites
were further masked based on TBLASTX alignments, and low-complexity regions
were identified and masked by RepeatMasker. In total, 11% of nucleotides were
masked in the T. nigroviridis genomic sequence database.
Human working draft sequence and Unigene.
We retrieved
the entire HTG1, HTG2 and HTG3 sections, and sequences larger than 35 kb from
the human DNA section of EMBL release 61, and for each sequence the highest
version number was retained to remove internal redundancy. Sequences were
distributed as follows: HTG1, containing genomic clone sequences in unordered
segments (726.9 Mb, 24.2%); HTG2, in which contigs were ordered within each
genomic clone (55.5 Mb, 1.8%); HTG3, in which genomic clones are represented
as a contiguous sequence (36.4 Mb, 1.2%); and HUM, in which sequences are
considered finished, with an error rate of less than one in 104
bases (480.9 Mb, 16.0%). All sequences were filtered to remove remaining cloning
vector sequences (0.21%) and stretches of 'N' used to separate
sequence contigs in HTG1 and HTG2 (1.87%). We used Unigene version 105 (January
2000) for comparisons with the T. nigroviridis genome.
Computing alignments.
All alignments were computed
with the suite of BLAST (ref. 12) algorithms
or with the SMITH-WATERMAN (ref. 13) algorithm,
implemented in LASSAP (Large Scale Sequence Comparison Package) version 1.1.5
(ref. 14). For all calculations, hardware consisted
of four Digital quadriprocessor (AXP 21264 (EV6) at 525 MHz) computers (Compaq
GS60) with 4 Go of memory each, except for comparison of the partial T.
nigroviridis genome with the human working draft sequence, for which a
SUN Enterprise 10000 server with 64 UltraSPARC-II (400 MHz) processors and
64 Go central memory were used with LASSAP version 1.2.0a.
Accession numbers.
T. nigroviridis sequences,
EMBL AL163976 to AL352938; human cDNA clone, AB033118.
Note: Supplementary information is available on the Nature Genetics web site (http://genetics.nature.com/supplementary_info/).
Received 10 March 2000; Accepted 2 May 2000
REFERENCES
- Fields, C., Adams, M.D., White, O. & Venter, J.C. How many genes in the human genome? Nature Genet. 7, 345-346 (1994). | Article | PubMed | ISI | ChemPort |
- Antequera, F. & Bird, A. Number of CpG islands and genes in human and mouse. Proc. Natl Acad. Sci. USA 90, 11995-11999 (1993). | PubMed | ChemPort |
- Brenner, S. et al. Characterization of the pufferfish (Fugu) genome as a compact model vertebrate genome. Nature 366, 265-268 (1993). | Article | PubMed | ISI | ChemPort |
- Crnogorac-Jurcevic, T., Brown, J.R., Lehrach, H. & Schalkwyk, L.C. Tetraodon fluviatilis, a new puffer fish model for genome studies. Genomics 41, 177-184 (1997). | Article | PubMed | ChemPort |
- Elgar, G. et al. Generation and analysis of 25 Mb of genomic DNA from the pufferfish Fugu rubripes by sequence scanning. Genome Res. 9, 960-971 (1999). | Article | PubMed | ISI |
- Schuler, G.D. et al. A gene map of the human genome. Science 274, 540-546 (1996). | Article | PubMed | ISI | ChemPort |
- Dunham, I. et al. The DNA sequence of human chromosome 22. Nature 402, 489-495 (1999). | Article | PubMed | ISI | ChemPort |
- Deloukas, P. et al. A physical map of 30,000 human genes. Science 282, 744-746 (1998). | Article | PubMed | ISI | ChemPort |
- Roest Crollius, H. et al. Characterization and repeat analysis of the compact genome of the freswater pufferfish Tetraodon nigroviridis. Genome Res. (in press).
- Jin, L., Zhong, Y. & Chakraborty, R. The exact numbers of possible microsatellite motifs. Am. J. Hum. Genet. 55, 582-583 (1994). | PubMed | ISI | ChemPort |
- Benson, G. Tandem repeats finder: a program to analyze DNA sequences. Nucleic Acids Res. 27, 573-580 (1999). | Article | PubMed | ISI | ChemPort |
- Altschul, S.F., Gish, W., Miller, W., Myers, E.W. & Lipman, D.J. Basic local alignment search tool. J. Mol. Biol. 215, 403-410 (1990). | Article | PubMed | ISI | ChemPort |
- Smith, T.F. & Waterman, M.S. Identification of common molecular subsequences. J. Mol. Biol. 147, 195-197 (1981). | Article | PubMed | ISI | ChemPort |
- Glemet, E. & Codani, J. LASSAP, a large scale sequence comparisons package. Comput. Appl. Biosci. 13, 137-143 (1997). | PubMed | ISI | ChemPort |
Acknowledgments
We thank the sequencing and template preparation team at Genoscope; Sun
Microsystems for access to the SUN benchmark centre; and F. Francis for critical
reading of the manuscript. This work would not have been possible without
the public availability of a large fraction of the sequence of the human genome,
and we thank all contributing genome centres.
|
) of human genes and those for which at least
one alignment falls in an intron (
). This provides robust selection
criteria to determine if any new alignment corresponds to a human exon, based
on its length and identity with a T. nigroviridis sequence. For convenience,
all alignments longer than 60 aa were arbitrarily drawn at 60, the longest
measuring 245 aa. b, Evolution of the theoretical T. nigroviridis
genome coverage (—) and observed sensitivity in gene detection
(+) and exon detection (
) by Exofish in the set of 322 human genes,
as a function of T. nigroviridis sequences produced (10% increments).
The dotted line is positioned at the current status of the sequencing project
(33% of genome coverage). The theoretical coverage is calculated on the basis
of a Poisson distribution of sequences of average size 886 bases on a genome
of 385 Mb.
