University Department of Oncology, Wellcome Trust Centre
for Molecular Mechanisms in Disease, Cambridge CB2 2XY
, UK saparici@hgmp.mrc.ac.uk
Biology occasionally mirrors human activity with unnerving irony. This
year has seen the spectacular rise and fall of biotechnology stock values
as at first exuberance and then sanity swept through the investor community.
Based on reports1,
2,
3 presented on pages 232, 235 and 239,
similar sentiments should now apply to estimates from some organizations of
ever-increasing values for the total number of human genes. With the near
completion of the human draft sequence, mere gene counting may seem a sterile
exercise—the 'real' answer will surely be known soon? The
analyses in this issue throw into sharp focus the question of what should
be counted as a gene. They indicate that, not only should our expectations
for the full number of human genes be revised downwards, but also, that existing
EST databases may contain as little as 40% of the protein-coding fraction
of the human genome.
EST clustering versus direct sampling Previous attempts at estimating the number of human gene loci have been
predicated on approaches such as measuring the complexity of cellular RNA,
reassociation kinetics, CpG island determination, evolutionary 'rules
of thumb' or assuming that cDNA sequences represent genes4,
4,
6.
This latter approach is quite popular, and employed by a group3
from The Institute for Genomic Research in one of the present analyses. The
authors make use of the extensive public cDNA sequence databases to estimate
the total number of genes. They carefully clustered the cDNA sequences (to
eliminate sequence redundancy), excluding 'singleton' sequences
(as these are probably artefactual), and then estimated the fraction of the
clustered sequences that might represent known genes. They estimate that there
are between 120,000 and 140,000 human genes. A key assumption is that clustering
is sufficient to remove artefacts that arise consequent to false poly(A) priming,
clone amplification, DNA contamination and other factors that would spuriously
inflate the total estimate. Crucially, however, one should challenge the assumption
that almost any (clustered) transcribed sequence represents a gene.
Two-sample comparison method for estimating gene numbers. The schematic
represents a body of sequences (grey spirals) for which a homogeneously representative
sample, n1, is taken (red spirals). The fraction of the total sequences, G,
in set 1 is given by f=n1/G. A second sample is taken, which may be
biased, redundant or incomplete (green spirals). Providing the sequences are
of sufficient quality to correctly identify orthologous matches of the first
set, then the fraction of matches for the first set in the second set will
approximate to the fraction of sequences from the first set, to the total
number of sequences. In other words, n1/G m/n2 where m is the
number of set 1 sequences matching set 2 sequences and n2 is the total number
of set 2 sequences. Thus G=n1n2/m.
Through alternative and independent processes, groups lead by Philip Green
and Jean Weissenbach arrive at a quite different estimate: they conclude that
the number of protein-coding genes is approximately 35,000 in total. Green
and colleagues1 use a modified form of the method applied to
the genome of Caenorhabditis elegans, a simple and elegant method involving
two samples (see figure). This approach
requires a small but homogeneously sampled collection of genes from the genome
and a second comparison sample which is larger but can be biased, redundant
and incomplete, as long as the sequence is of sufficient quality to reliably
match sequences of the first set. For the first set, the authors used either
the curated annotated genes from chromosome 22 (ref. 4)
or a filtered set of near full-length mRNA sequences from GenBank. Comparison
of these sets with a reduced set of clustered EST sequences indicates that
there are either 34,700 genes (based on chromosome 22) or 33,600 genes (based
on mRNA sequences) in the human genome.
Weissenbach and colleagues2 used a different approach that
exploits pufferfish genomic sequence. They describe the sequencing of BAC
clone ends from the compact genome of a pufferfish7,
8; these
collectively represent approximately one-third of the genome sequence of this
vertebrate. By comparing this sequence set with a known, curated set of protein-encoding
genes, they were able to calibrate an algorithm that detects orthologous coding
sequences (called ecores) between pufferfish and human genomes. On measuring
the mean number of ecores found per human gene across the calibration set
of genes and then running the same analysis across the human draft sequence
(approximately 60% of the human genome at the time of analysis), they arrived
at an upper limit of 34,000 genes. Note that the pufferfish sequence will
not pick up some human orthologues, owing to evolutionary distance or because
they are not represented in the sequence set. But so long as the percentage
of undetectable sequences (around 30% in this case) is not disproportionately
represented in the pufferfish or human sequence sets, the calculation holds
true.
They also showed that, whereas known, related and pseudogenes are easily
detected, applying the algorithm to chromosome 22 pulled out only 4% of the
300 gene sequences predicted by Genscan—indicating that Genscan significantly
overpredicts human genes. A similar detection sensitivity was observed with
the Unigene set: the pufferfish sequence identified about 65% of the 10,500
protein-coding transcripts as ecores, but only matched 4% of EST clusters.
The comparison indicates that the clustered EST set is redundant and probably
contains no more than 40% of the coding fraction of the human genome.
Bursting the bubble In principle, the estimates by Green and Weissenbach are subject to biases
that could lead them to an underestimation of gene number. As these biases
are methodologically independent, however, it seems unlikely that these estimates
will prove wildly inaccurate. Agreeing with them are estimates9,
10
of 40,000 and 45,000 genes based on chromosomes 21 and 22. As mammals have
probably experienced two genome doublings since their divergence from multicellular
invertebrates, the total is very unlikely to exceed about 60,000 genes in
any event. So why should calculations based on EST data have resulted in such
large estimates?
At the core of these findings are issues of definition and recognition—that
is, how does one define a gene, and how does one recognize it? According to
classical genetics, genes are the heritable units responsible for an associated
phenotype. Although in some cases this relationship derives from mutation
of regulatory elements or other non-coding DNA elements, in most cases it
is synonymous with mutation of protein-coding DNA sequences. Although the
tendency (especially in a pay-per-sequence access mode) is to assume that
any transcript represents a gene, classical genetics demands some evidence
of associated function. Crucially, what is not yet established (but is implied
to be relatively abundant by these studies) is the extent of biological "noise"
in the transcriptome of any given cell. In other words, what fraction of transcripts
which can be isolated have any meaningful function? What fraction might be
mere by-products of spurious transcription, spuriously fired off, perhaps
on the antisense strand from promoters or CpG islands associated with protein
coding genes (as seems to be the case with a number of imprinted genes)? The
good news therefore, is that the human draft sequence will be a goldmine for
protein-coding sequences not represented in the EST collections.
Clearly, the task of annotating genes in the human sequence will take time,
and a comparative approach has much to offer. Beyond this lies the challenge
of understanding gene regulation. The ability to make adequate comparisons
of non-coding sequences of different species should be a rapid means by which
to obtain a regulatory element 'framework' for the human genome.
Evolution is certainly more powerful and has more to teach us here than any
extant computer algorithm. Our ability to compare the genomes of many species
may yet turn in silico biology into a true science.
m/n2 where m is the
number of set 1 sequences matching set 2 sequences and n2 is the total number
of set 2 sequences. Thus G=n1n2/m.
