Biology Reference
In-Depth Information
measure genetic interactions on a global scale in several
model organisms. Genetic interaction networks have
proved to be a powerful tool for dissecting gene function
and understanding the systems-level organization of
a genome. Beyond these important goals, large-scale
genetic interaction maps also have the potential to reveal
key insights into one of the most fundamental questions in
modern biology: how genomes specify phenotypes, a major
challenge towards understanding human disease.
Given increasingly scalable and affordable technology
for mapping genotypes in the human population, genome-
wide association studies (GWAS) for various disease traits
have become commonplace. These efforts have quickly and
dramatically increased our knowledge of genome variants
associated with various human traits or disorders, to date
producing nearly 1500 established associations (NIH
GWAS Catalog, http://
www.genome.gov/26525384
).
Despite this success, there are very few diseases, particu-
larly among those that commonly afflict the population, for
which the discovered variants are able to explain
a substantial portion of the heritable variance
[110]
.For
example, there are now 95 variants associated with LDL
and HDL cholesterol
than previously thought, but more work is required to define
the model by which they influence phenotypic variation.
Based on the large-scale reverse genetic screens in model
organisms discussed in this review, genetic interactions are
sufficiently common to support the plausibility of this
scenario. For example, among the 5.4 million double
mutants constructed in the latest yeast study, more than 3%
exhibit detectable positive or negative genetic interactions,
and the rate is substantially higher among genes that also
exhibit fitness defects as single mutants
[2]
.
Addressing the role of genetic interactions in the
unexplained heritability of GWAS is challenging because
these interactions are difficult to detect, given the param-
eters of most current association studies. In principle, each
pair of variants genotyped in a particular study can be
tested for interaction effects, and, in fact, there have been
many recent studies that take this approach
[113
115]
.In
general, the statistical power of such pairwise tests depends
on several factors, including the allele frequency, pop-
ulation structure, and sample size, but under reasonable
assumptions in a typical GWAS setting, recent estimates
have placed the required sample size near 500 000 indi-
viduals to realistically detect such interactions
[23]
, sug-
gesting that this may not currently be a practical question
without additional constraints.
One approach is to place additional constraints by only
considering a subset of locus pairs meeting some statistical
threshold, ormore generally, requiring some other knowledge
of association between the two genes. In some studies, the set
of genomic loci tested for interaction effectswas limited to the
subset that showed significant association as individual
factors
[116]
. Other studies have proposed to use literature-
based or experimentally derived networks to filter the set of
candidate pairs to be tested
[117]
. In a related study in yeast,
Hannum et al. analyzed the structure of pairwise interactions
derived from analysis of eQTL traits in a population of 112
segregants derived from an S. cerevisiae laboratory strain
crossed with a wild isolate
[118]
. They leveraged the fact that
they expected such interactions to exhibit modular structure
relative to known protein complexes and pathways, as had
been observed in genetic interactions derived from
reverse genetic mapping approaches. They indeed found
evidence that some pairwise interactions exhibited the same
between-complex/between-pathway structure observed in
reverse genetic interaction networks
[118]
.However,they
observed little overlap between the genetic interactions
derived from statistical analysis of eQTLs and those experi-
mentally measured from constructing double mutants, which
either suggests the types of interactions influencing expres-
sion variation are quite different from fitness-based reverse
genetic screens or potentially reflects the statistical challenges
associated with such an approach. The general strategy of
leveraging genetic interactions derived from reverse genetic
mapping approaches to narrow the search space for
e
levels, but
the combination only
explains 20
25% of the variation known to be heritable
[110]
. There have been a number of reasons proposed for
this discrepancy, including the presence of rare variants,
which may not be queried on current genotyping platforms,
or the presence of numerous small effect variants that are
below the level of detection given the sample size of
a typical study
[110,111]
. Another potential explanation
that we would like to consider is the influence of genetic
interactions between variants
[110,111]
.
Indeed, in a comparative study of closely related and
interbreeding yeast strains, Dowell et al.
[112]
examined all
genes for those that exhibit conditional essentiality, such that
deletion of a specific gene is essential in one strain (indi-
vidual) but not another. One possible explanation for
conditional essentiality was that it results from a synthetic
lethal interaction between two genes: the deleted gene and
another containing natural variation that leads to a strain-
specific LOF allele. However, in all cases of conditional
essentiality examined, it appeared that multiple modifier loci
were necessary to confer strain-specific essentiality
[112]
.
Thus, genetic networks involving natural variation may be
highly complex and often involve more than two genes.
The potential for genetic interactions to underlie
a significant proportion of inherited phenotypes would cause
an inflation in estimates of the total heritable variation,
because current models used in such calculations assume
a simple additive model
[23]
. Thus, the singly associated
variants may represent the set of all causal variants, but
interaction effects among them are also necessary to explain
the missing variation
[23]
. This may suggest we have actu-
ally captured variants that explain more heritable variation
e
