Nature Genetics
26, 135 - 137 (2000)
doi:10.1038/79839
Guilt by associationDavid Altshuler1, 2, 3, Mark Daly3
& Leonid Kruglyak4, 51 Departments of Genetics and Medicine, Harvard Medical
School, Boston, Massachusetts, USA.
altshul@genome.wi.mit.edu 2 Department of Molecular Biology and Diabetes Unit,
Massachusetts General Hospital, Boston, Massachusetts
, USA. 3 Whitehead Institute/MIT Center for Genome Research
, Cambridge, Massachusetts, USA.
mjdaly@genome.wi.mit.edu 4 Howard Hughes Medical Institute, Program in Genetics,
Fred Hutchison Cancer Center, Seattle, Washington,
USA.
leonid@fhcrc.org 5 Division of Human Biology, Program in Genetics, Fred
Hutchison Cancer Center, Seattle, Washington,
USA. Positional cloning of common disease genes is a central but elusive
goal of human geneticists. Progress is now reported by Bell and colleagues
in their study of NIDDM1, a locus implicated in type 2 diabetes. The complex
nature of the reported association illustrates the challenge of implicating
a specific gene and mutation in the causation of polygenic disease.In the United States criminal justice system, guilt must be proven "beyond
a reasonable doubt," a stringent standard reflecting the serious consequences
that follow a guilty verdict. Perhaps because of these high stakes, 'proof'
is frequently a subject of heated debate. With positional cloning, similar
issues arise when the 'guilt' of a specific gene suspected of
disease causation must be proven. The stakes are high because of the intense
interest 'disease' genes elicit in patients, clinicians and biologists
alike. Proving causation, however, is not necessarily a straightforward proposition,
even for relatively simple cases of mendelian diseases. In some instances,
a 'smoking gun' may be found in the form of different disrupting
mutations in a single gene. The case for the prosecution is more difficult
when only missense changes or alterations in non-coding regions are observed.
Nonetheless, culprits in well over one hundred mendelian diseases have been
snared by positional cloning.
For common diseases with complex inheritance, convicting a guilty party
poses a far greater challenge. Genome-wide linkage scans designed to localize
disease genes have yielded few significant findings, and failure to reproduce
published linkage results is endemic. The lack of strong linkage signals indicates
that, for most complex traits, no single locus confers a high degree of risk.
Because mutation of any single gene is neither necessary nor sufficient, the
correlation between genotype and phenotype is imperfect, and individual recombinants
cannot be trusted for fine-mapping. This means that implicated genomic intervals
are large, and yet do not include the gene with 100% confidence1,
2.
Finally, as mutations contributing to complex traits may be subtle alterations
of gene function or regulation, rather than disruptions of coding sequence,
they may be difficult, if not impossible, to recognize from primary sequence
data alone. For all these reasons, there are no examples of 'complex
disease' genes cloned purely by position (aside from highly familial
subtypes).
On page 163 of this issue,
Yukio Horikawa, Naohisa Oda, Nancy Cox, Graeme Bell and colleagues report
on their pioneering effort to positionally clone a gene that affects susceptibility
to type 2 diabetes in Mexican Americans. Through a combination of clever detective
work and brute force, they indict a gene encoding calpain-10 (CAPN10),
and show that certain combinations of polymorphisms in this gene are associated
with the risk of type 2 diabetes. By indicating a role for a calpain protease,
these findings propose a fundamentally new hypothesis for diabetes research.
But because no smoking gun is found, it is a great challenge to convict the
guilty party (that is, gene and mutation) beyond a reasonable doubt. The study
provides an unprecedented look at the cutting edge of positional cloning for
complex traits and teaches lessons of lasting importance about just how difficult
finding genes for common diseases may turn out to be.
A dragnet for suspects
In 1996, Hanis et al. reported significant evidence for linkage
of type 2 diabetes to the distal long arm of human chromosome 2 (the locus
was designated NIDDM1; ref. 3). The implicated
region was large, with the 1-lod support interval, which is expected to contain
the responsible gene in 80−90% of cases1, extending over
12 cM. The discovery of an interaction between NIDDM1 and a putative
locus on chromosome 15 (ref. 4) focused attention
on a sub-region of 7 cM. At this point, good fortune intervened—construction
of a physical map revealed that the 7 cM spans only 1.7 million base pairs,
rather than the expected 7 million (the average ratio of physical distance
to genetic distance across the human genome is approximately 1 million base
pairs per cM). This unusually high recombination rate is probably a consequence
of the telomeric location of NIDDM1.
But finding the causative gene(s) in 1.7 Mb of sequence is a formidable
task that has no generic solution. This particular interval contains at least
7 known genes and 15 expressed sequence tags (ESTs): none of these are obvious
candidates. Horikawa et al. chose to screen polymorphisms in the region
for association with diabetes, relying on linkage disequilibrium (LD) to lead
them toward their quarry. An initial examination of 21 SNPs detected a hint
of association between diabetes and multi-locus haplotypes at 3 consecutive
SNPs. This haplotype frequency difference prompted the identification and
typing of additional SNPs in the region, some of which were nominally associated
to diabetes. The authors then sequenced a 66-kb interval in the region (centred
on the associated SNPs) in 10 diabetics, revealing 3 genes (including
CAPN10) and 179 polymorphisms.
On the basis of distinct patterns of LD in the initial 10 individuals,
63 of the 179 polymorphisms were chosen for typing in a larger set of 100
diabetic cases and controls. Sixteen of these SNPs showed modest, nominally
significant association with diabetes. To narrow the number of suspects, the
authors developed a new statistical test, termed "partitioning of linkage."
The test assumes that the locus under study contributes to risk in only a
subset of patients, and that this risk is conferred by a single common allele.
The presence of this allele—or one in LD with it—should then selectively
tag families responsible for evidence of linkage to the locus. Conversely,
the remaining families (which lack the risk allele) should display no such
evidence. In effect, the test demands association of an allele not just with
disease but with those cases 'linked' to the locus under investigation.
Presumably, associations that fail this test are spurious or too weak to be
detected by linkage methods. Care must be taken to ensure that the test is
applied in an unbiased fashion, and a comprehensive examination of its properties
has yet to be published. Specifically, the appropriate statistical thresholds
and the extent to which the method may be confounded by population admixture
require investigation. In this case, however, the evidence for partitioning
was strong, and implicated a single SNP (UCSNP-43), an intronic G/A polymorphism
within CAPN10. The common G allele (present on 75% of normal chromosomes)
was associated with increased risk of diabetes and partitioned the evidence
for linkage to a greater degree than expected by chance, as assessed by simulation.
An unusual suspect
At first glance, UCSNP-43 seems an unlikely suspect for a polymorphism
that affects disease causation (or even gene function). CAPN10—a
ubiquitously expressed member of the calpain-like cysteine protease family—has
no known or previously proposed role in diabetes physiology. Moreover, UCSNP-43
lies in an intron of CAPN10, and in a region that is not conserved
in the mouse. It has no discernable effect on splicing. The authors are able
to document an effect of UCSNP-43 on transcription, but the relevance of this
effect to gene function or diabetes has yet to be established. Moreover, a
polymorphism with a high frequency and modest association with increased risk
(the G allele frequency is 75% in the general population and 80% in diabetics)
could not, acting alone, have produced the linkage signal observed in the
original genome scan of 440 sibpairs. Indeed, the expected lod score attributable
to such an allele is less than 0.05 and more than 100,000 sibpairs would be
required to detect linkage with a lod score of 4.03, as observed by Hanis
et al.3.
These considerations raised the possibility that SNP43 might not be the
sole perpetrator, but rather contributes modestly or is in LD with the actual
causal variant(s). This led the authors to examine haplotypes in the region.
Their data do not reveal a single variant or haplotype that is responsible
for diabetes risk. Rather, heterozygosity for two different, common haplotypes—both
of which carry the G allele at UCSNP-43—seems to be required to influence
diabetes. Surprisingly, homozygotes for either haplotype were not at significantly
altered risk. Critically, these findings were confirmed in an independent
sample of Mexican Americans, making it unlikely that they represent a statistical
fluctuation after a search over many haplotypes. It seems that the locus exists
in several functionally distinct forms, and that a combination of two different
forms in a patient may be required to influence disease.
Examination of genetic variation in CAPN10 in two European populations
offers a potential explanation why subsequent linkage studies (outside Mexican
Americans) had not identified linkage to NIDDM1. In both German and
Finnish patients, the same combination of two haplotypes shows a trend toward
increased susceptibility, comparable in magnitude to that seen in Mexican
Americans. However, although both haplotypes are common in Mexican Americans,
one of them is rare in the Europeans, reducing the overall effect of the locus
on risk of diabetes in these populations and potentially accounting for the
weaker linkage findings. Examination of additional patient samples from these
and other ethnic groups are critical to understanding the consistency and
magnitude of these effects (and, based on the content of recent conferences,
will probably be reported soon). In this regard, it is noteworthy that the
first follow-up study5 fails to replicate the association with
type 2 diabetes in Pima Indians (although it does report association with
diabetic sub-phenotypes).
Learning from NIDDM1
Horikawa et al. have worked as long and as hard as any other group
to clone a human polygene, and their report represents the state of the art
of genetic analysis of complex traits. It is therefore both sobering and instructive
that, despite their success in identifying associated haplotypes, answers
to key questions remain uncertain. Does the association with CAPN10
account for linkage to NIDDM1? Is CAPN10 itself the 'diabetes'
gene? If so, what specific polymorphism(s) is directly responsible for altering
risk of diabetes, and by what mechanism?
The double-edged sword of LD
Initial localization to the CAPN10 region relied on a scan of 21
SNPs across a 7-cM genetic interval. Is this density of markers adequate to
comprehensively search for common haplotypes contributing to disease? At 1
SNP every 0.3 cM (equivalent to a genome-wide survey of 10,000 SNPs), this
survey is significantly less dense than is expected to be necessary to search
typical genomic regions for association6,
7,
8. As with the
detection of linkage to Huntington disease in a screen of only 12 markers9, the positive finding of association with the sparse initial map
must have been a pleasant surprise, and it is possible that other associations
in the linked region have been overlooked. Just as linkage screens became
more systematic with the development of genetic maps, it is hoped that dense
SNP maps and efficient assays will similarly facilitate more comprehensive
association screens. Determining the optimal density of markers, as well as
appropriate thresholds for following up initial associations, are important
challenges for the field.
Of course, 1 SNP every 0.3 cM might suffice in cases where LD is extensive.
However, in such situations the burden of a search is not lessened, only shifted.
In cases where LD extends for tens to hundreds of kilobases, a sparse initial
map may be sufficient to detect association (as it appears to have done in
the present case); however, such an expanse of LD will require the exhaustive
examination of many more candidate variants across this larger region. Many
of these variants—potentially spanning several genes—may demonstrate
similar or identical patterns of association with disease, making it difficult
to determine which are causal variants and which happen to come along for
the ride on the risk haplotype. For example, protection against HIV infection
is conferred by inheriting a large, perfectly conserved, haplotype-block surrounding
the gene CCR5 (ref. 10). It is the biology
of CCR5 and its 32-bp deletion that proves which gene in the region
grants protection.
Along these lines, it is intriguing that Horikawa et al. find 16
SNPs displaying some association to diabetes; these span much of the 66-kb
sequenced region, and a number lie in GPR35, a gene adjacent to
CAPN10. How many of these are markers for the described, associated haplotypes,
and over what distance can the association be observed? If not attributable
to the implicated haplotypes, what explains the association of the others?
Additional work is required to understand the relationship between the three-marker
system used initially to identify the sub-region (markers 1, 2, 19), the three
markers for haplotype association in CAPN10 (markers 43, 19, 63), and
other positive markers throughout the region. For example, it is probably
not a coincidence that marker 19 is present in both the initial 3-marker system
and the haplotypes showing strongest association with disease. If so, could
the at-risk haplotypes extend to UCSNP-1 and 2, which reside outside the 66
kb sequenced region?
In the end, genetic evidence alone may never be sufficient to establish
the causal gene and mutation where LD is extensive. And even when the causal
variant is known, regulatory effects (such as those proposed for UCSNP-43)
can act over considerable distance. A regulatory element on human chromosome
5, for example, controls transcription of at least 3 genes spread over 120
kb (ref. 11). As Horikawa et al. point
out, ultimate proof in such situations must be biological rather than statistical.
Such proof can be obtained through the transfer of specific variants, complete
haplotypes or transcripts into cellular or animal models of disease. To observe
an effect in such transgenic animals will require an appropriate host genotype
and environment, however, which may be especially challenging in the case
of NIDDM1 and CAPN10. The current model requires compound heterozygosity
for two different haplotypes at CAPN10 and interaction with a locus
on chromosome 15, a combination that will be hard to recapitulate in a heterologous
system.
Closing arguments
The study by Horikawa et al. is a landmark in the decade-long effort
to clone genes for polygenic disorders. It endorses the value of positional
cloning for rounding up genes not otherwise suspected to have a role in a
disease (rather than the usual suspects), and generates an exciting hypothesis
for diabetes researchers to pursue. It also illustrates just how difficult
the endgame of positional cloning can be, and contributes valuable data to
the broader (often theoretical) debate about the utility of LD mapping and
the analysis of common disease. (See, for example, the Commentary12 on page 151
of this issue, and refs 13,
14,
15.)
There is encouragement in Horikawa et al.'s successful application
of SNPs and LD mapping to find common haplotypes displaying significant population-attributable
risk. However, the complicated nature of genetic effects uncovered by the
study indicates that we need to develop a better framework for connecting
genotypes with phenotypes, and to contemplate genetic complexity as something
more than 'several genes'. As other workers catch up with Horikawa
and colleagues, the generality of these lessons will become clearer, and more
surprises are sure to follow. As with any good detective story, the present
instalment leaves one eager for the next.
REFERENCES
- Kruglyak, L. & Lander, E.S. Am. J. Hum. Genet. 56, 1212–1223 (1995). | PubMed | ISI | ChemPort |
- Kruglyak, L. & Lander, E.S. Am. J. Hum. Genet. 58, 1092–1093 (1996). | PubMed | ISI | ChemPort |
- Hanis, C.L. et al. Nature Genet. 13, 161–166 (1996). | Article | PubMed | ISI | ChemPort |
- Cox, N.J. et al. Nature Genet. 21, 213–215 (1999). | Article | PubMed | ISI | ChemPort |
- Baier, L. et al. J. Clin. Invest. 106, R69–R73 (2000). | PubMed | ISI | ChemPort |
- Collins, A., Lonjou, C. & Morton, N.E. Proc Natl Acad. Sci. USA 96, 15173–15177 (1999). | Article | PubMed | ChemPort |
- Kruglyak, L. Nature Genet. 22, 139–144 (1999). | Article | PubMed | ISI | ChemPort |
- Boehnke, M. Nature Genet. 25, 246–247 (2000). | Article | PubMed | ISI | ChemPort |
- Gusella, J.F. et al. Nature 306, 234–238 (1983). | Article | PubMed | ISI | ChemPort |
- Stephens, J.C. et al. Am. J. Hum. Genet. 62, 1507–1515 (1998). | Article | PubMed | ISI | ChemPort |
- Loots, G.G. et al. Science 288, 136–140 (2000). | Article | PubMed | ISI | ChemPort |
- Collins, F.S., Guyer, M.S. & Chakravarti, A. Science 278, 1580–1581 (1997). | Article | PubMed | ISI | ChemPort |
- Weiss, K.M & Terwilliger, J.D. Nature Genet. 26, 151–157 (2000). | Article | PubMed | ISI | ChemPort |
- Lander, E.S. Science 274, 536–539 (1996). | Article | PubMed | ISI | ChemPort |
- Risch, N. & Merikangas, K. Science 273, 1516–1517 (1996). | PubMed | ISI | ChemPort |
- Risch, N.J. Nature 405, 847–856 (2000). | Article | PubMed | ISI | ChemPort |
|
