Biology Reference
In-Depth Information
Complementary assays that exploit modulations in gene
dosage also allow direct drug target identification. These
assays are discussed in further detail in the HIPHOP section
below.
been demonstrated that the same holds true for a complete
loss-of-function gene deletion. A later study demonstrated
the drug target of dyclonine to be Erg24
[16]
,and
importantly, when combined with an earlier study
[42]
,
provided compelling evidence that the dyclonine sensi-
tivity observed in the erg2
strain was due to the down-
stream inhibition of Erg24. It is possible that the bona fide
target of dyclonine might have been identified in this study
had the reference set included the ERG2 deletion strain;
however, as ERG24 is essential, a series of strains, each
expressing a different level of Erg24, would have been
required, and in addition, a similar series of strains rep-
resenting all essential genes, making the construction of
a comprehensive dataset impractical, and likely not even
possible.
The discussion above strongly suggests that one of the
conclusions of this study, that Erg2 is the protein target of
dyclonine, was incorrect. The message provided by this
re-examination of the compendium study clarifies the
reasons why the incorrect conclusion was made: (1) the
reference set was not sufficiently comprehensive; (2)
the incorrect assumption that a gene deletion would mimic
a decrease in protein activity due to inhibition by a small
molecule; (3) the drug target was identified based on
a downstream effect caused by the inhibition of the bona fide
target. The take-home messages are: (1) a reference dataset
can never be truly comprehensive; (2) the incorrect
assumption made was based on current understanding of
genetics and not on evidence; and (3) genomic technologies
used for a particular application should, whenever possible,
provide as direct a readout as possible, one best matched to
the molecular level affected by the perturbagen. In this case,
a transcriptional readout was used to measure the conse-
quences of inhibiting a protein. The assumption that was
made based on current understanding of genetics serves as
a cautionary reminder that such assumptions can result in
great cost, particular for large-scale genome-wide datasets.
For example, a similar genetic-based assumption was made
during the construction of the YKO collection. Specifically,
it was assumed that the three auxotrophic markers included
in the genetic background of the wild-type strain, from
which all deletion strains were derived, were 'benign'.
Recently, the auxotrophic markers have been shown to bias
data in some cases and to limit the application of the YKO
collection in others
[43
D
Compendium RNA Expression Approaches to
Predict the Drug Target/Mechanism of Drug
Action
The concept that large datasets could provide predictive
value was first introduced to practice by small molecule
screens performed at the National Cancer Institute (NCI),
known as the 'NCI-60' project. The concept involved using
profiles of cellular response to drug, measured across 60 cell
lines, to classify small molecules by MOA and toxicity.
Genome-wide expression studies borrowed many of
the experimental design principles and analytics from the
NCI-60 project
[41]
.
In an innovative study
[39]
published soon after the
introduction of RNA expression arrays, Hughes and
colleagues used the NCI-60 concept to demonstrate that
gene function and the mechanism of drug action could be
predicted by querying a reference set to find the 'best
match'. In this particular study
[39]
, a 'compendium' of
~300 genome-wide RNA expression profiles, of yeast
deletion strains and of a wild-type strain treated with small
molecules, served as the reference set. The individual
queries were expression profiles of either yeast strains
deleted for a gene of unknown function, or a wild-type
yeast strain treated with a small molecule of unknown
mechanism. After computing the similarity between the
query profile and the compendium profiles, in the example
described here the best matches in the compendium for
a query strain deleted for a gene with unknown function
are strains deleted for genes involved in the ergosterol
biosynthesis pathway, a common target of antifungal
agents. These results suggested that the uncharacterized
gene was also involved ergosterol biosynthesis, and
following confirmation and an additional follow-up study,
the gene was named ERG28. Eight other genes of
unknown function were similarly annotated, thus proving
the power of using large-scale datasets to predict gene
function. In an additional experiment, a profile of wild-
type yeast grown in the presence of dyclonine,
a compound of unknown MOA, was used as the query. In
this case, the query profile best matched the profile of the
deletion strain erg2
45]
. The details discussed here may
seem trivial; however, these same concepts have provided
the basis for much larger-scale reference datasets to be
generated
[46]
so that, while proven powerful in inferring
mechanism, future conclusions should be carefully consid-
ered to ensure that they are without bias. A second general
impact of this study, one that came at great expense, was the
futile pursuit of many genomic targets in drug discovery
based on expression data alone. This highlighted the need
to include, minimally,
e
, and it was inferred that Erg2 was
therefore the drug target. This conclusion was based on an
assumption that a decrease in protein activity due to
inhibition by a small molecule would mimic the complete
absence of gene activity. While genetic logic and the
literature generally support the view that a decrease in
protein activity due to inhibition by a drug would phe-
nocopy a genetic decrease-of-function allele, it has not
D
secondary evidence based on
