Biology Reference
In-Depth Information
adipose tissues, between the sexes and between species
[4]
,
but because it was supported as strongly causal for nearly
all of the metabolic traits scored in the cross (fat mass,
weight, plasma glucose, insulin, lipid levels, and aortic
lesions)
[3]
. Again, the causal relationship between this
subnetwork and the disease traits was established by
leveraging the changes in DNA in this population that were
simultaneously associated with disease and expression
traits. The entire subnetwork was shown to be under the
control of genomic loci associated with the metabolic traits,
while the predictive network modeling strongly indicated
that the module was causal for the disease traits, and was
not simply reacting to or acting independently of these
traits.
Of the more than 100 genes supported in this module as
causal for metabolic disease traits such as obesity and
diabetes, many, such as Zfp90, Alox5, C3ar1,and Tgfbr2,
had been previously identified and validated as causal for
metabolic traits
[20,22]
. In addition, three other genes were
selected for validation because they were independently
supported as causal for metabolic traits in other studies (Lpl
and Lactb), or because they were supported as causal for
such a wide variety of metabolic traits (Ppm1l)
[3]
. Inter-
estingly, the degree of connectivity in this causal metabolic
subnetwork was extreme. Perturbations to genes in this
module that were previously validated as causal for the
metabolic traits caused expression changes in many other
genes validated as causal for metabolic traits. For example,
over-expression of Zfp90 in mouse not only generated an
expression response that was significantly overlapping with
the causal metabolic module, but it caused changes in other
genes, such as Pparg, known to have an impact on meta-
bolic traits
[3]
.
the-art therapies in the future will be based on targeting
combinations of genes
[52,53]
, and for such applications
not only is it important to infer the direction of each
interaction (i.e., do you antagonize or activate a given
target), but one must be able to predict the degree to which
each gene should be knocked down or activated (in
a quantitative sense), only by generating accurate
predictive models of complex phenotypes can we most
efficiently search for such combinations to pursue for
experimental proof of concept.
The success of modeling complex systems in the future
will depend on constructing networks that are predictive of
complex behavior, not merely descriptive. In order to
achieve these more predictive models in complex systems
such as humans, we must expand existing networks so they
reflect relationships between cell types and tissues, not just
within a single cell type or tissue; capture a greater range of
molecular phenotypes to enhance understanding of rele-
vant functional units that define biological processes of
interest; and improve modeling capabilities, ideally
drawing on the expertise of other fields that have pioneered
causality-type reasoning. The complex phenotype-associ-
ated molecular networks we can construct today are
necessarily based on grossly incomplete sets of data. Even
given the ability to assay DNA and RNAvariation in whole
populations in a comprehensive manner, the information is
not complete, given rare variation, DNA variation other
than SNP/copy number, variation in non-coding RNA
levels, and variation in the different isoforms of genes, are
far from being completely characterized in any sample, let
alone in entire populations. Beyond DNA and RNA,
measuring all protein associated traits, interactions
between proteins and DNA/RNA, metabolite levels,
epigenetic changes, as well as other molecular entities
important to the functioning of living systems, are not yet
possible with existing technologies. Further, the types of
high-dimensional data we are able to routinely generate
today in populations represent only a snapshot at a single
time point, which may enable the identification of the
functional units of the system under study and how these
units relate to one another, but does not enable a complete
understanding of how the functional units are put together,
the mechanistic underpinnings of the complex set of
functions carried out by individual cells and by entire
organs and whole systems comprised of multiple organs.
Despite these and other advances required to more
routinely develop predictive models of living systems (we
really are only just scratching the surface), two of the more
critical developments I believe will most enable the reali-
zation of more accurate network models relate to unifying
different modeling approaches in a mathematically
coherent way and transforming the way in which
communities of researchers collaborate to build predictive
models.
CONCLUSION AND FUTURE DIRECTIONS
Thegenerationofeverhigherdimensionaldata(DNA
sequencing, RNA sequencing, epigenomic profiling, pro-
teomic profiling, metabolomicprofiling,andsoon)atever
higher scales demands sophisticated mathematical
approaches to integrate these data in more holistic ways to
uncover not only patterns of molecular, cellular, and
higher-order activities that underlie the biological
processes that define physiological states of interest, but to
uncover causal relationships among molecular and
cellular phenotypes and between these phenotypes and
clinical traits such as diseaseordrugresponse.Oneofthe
more successful frameworks for representing large-scale
high-dimensional data are networks. Here I have detailed
one particular approach to reconstructing predictive
network models of living systems that leverages DNA
variation as a systematic variation source and Bayesian
network reconstruction algorithms to take a top-down
approach to modeling complex systems. Because state-of-
