REVIEWS

MAXIMIZING THE POTENTIAL

OF FUNCTIONAL GENOMICS
Lars M. Steinmetz* and Ronald W. Davis‡
Geneticists have made tremendous progress in understanding the genetic basis
of phenotypes, and genomics promises to bring further insights at a rapid pace. The
progress in functional genomics has been driven primarily by the development of new
techniques that are used in a few dedicated research centres. Focusing on selected
advances in genomic technologies, we assess the results that have been obtained so far,
highlight the challenges faced by these new tools and suggest ways in which they can be
overcome. We argue that progress in functional genomics will depend on developing
high-throughput technologies that can easily be moved away from dedicated centres and
into individual laboratories.

COMPLEX TRAITS
A trait that is determined by
many genes, almost always
interacting with environmental
influences.
*European Molecular
Biology Laboratory,
Meyerhofstrasse 1,
69117 Heidelberg,
Germany.
‡
Department of Biochemistry
and Stanford Genome
Technology Center,
Stanford University,
855 California Avenue,
Palo Alto, California 94304,
USA.
Correspondence to L.M.S.
e-mail:
lars.steinmetz@embl.de
doi:10.1038/nrg1293
Biology is entering an exciting era brought about by the
increase in genome-wide information. Functional
genomics in particular is making rapid progress in
assigning biological meaning to genomic data. The
tools of functional genomics have enabled several
systematic approaches that can provide the answers
to a few basic questions for the majority of genes in a
genome, including when is a gene expressed, where is
its product localized, with which other gene products
does it interact and what phenotype results if a gene is
mutated. Functional genomics aspires to answer such
questions systematically for all genes in a genome in
contrast to conventional approaches that do so for one
gene at a time.
Several key biological challenges are central to con-
tinuing genome projects and are relevant to any
eukaryotic organism, from yeast to humans. One chal-
lenge is to understand how genes that are encoded in a
genome operate and interact to produce a complex liv-
ing system. A related challenge is to determine the
function of all the sequence elements in the genome. A
third challenge is to understand the contributions of
the multitude of sequence variants to phenotypic vari-
ation, both within and between species. One of the
most enduring challenges in genetics has been to find
the genetic variants that are responsible for COMPLEX
1
TRAITS . Current methods have mostly failed to meet
this challenge2, resulting in the need for new concepts
and genome-wide technologies if this complexity is to
be dissected.
Despite the unresolved issues, the power and poten-
tial of functional genomics is impressive. We illustrate
this here by discussing three core applications of genome
technology, using selected examples from different
organisms: genome-wide knock-out, gene expression
and genetic mapping studies. We go beyond these exam-
ples to point out the areas in which technological
improvements are possible.
As functional approaches and verification of their
accuracy often require genetic manipulation, many
technical advances in functional genomics have their
origin in model systems. Nonetheless, an effective tran-
sition of some of the technologies to humans is becom-
ing more attractive3. The utility of such a transition can
be maximized by careful evaluation of the power and
limitation of these approaches.
To obtain the most benefits from functional geno-
mics, we argue, the technology, which is at present
mainly carried out by a few dedicated centres, needs
to become integrated into individual laboratories.
Individual laboratories often have crucial expertise in
a specific biological problem, and although functional
genomics might provide approaches to address them, a
key discovery can often only be made by bringing the

190 | MARCH 2004 | VOLUME 5 www.nature.com/reviews/genetics

 REVIEWS

two together. We believe that for this to be achieved, two
goals should be met: experiments must be further
miniaturized and costs must be lowered.
Technological innovations
Efforts towards increased miniaturization and decreased
costs are exemplified by developments that originated
from genome sequencing. In many ways, functional
genomics was catalysed by the genome-sequencing pro-
jects: large-scale sequencing and the genome projects
created an increase in available DNA sequences, around
which new technologies that use this information were
developed. A result is one of the most widely recognized
and accessible genomics tools — the DNA microarray
— which allows parallel hybridization assays to be
carried out on an unprecedented, miniaturized scale.
The second, and often unrecognized, contribution of
the genome projects is the ~1,000-fold decrease in the
cost of DNA sequencing, which had to be achieved to
complete the Human Genome Project. The drop in
sequencing costs facilitated large-scale sequencing pro-
jects of other organisms and has contributed to the fact
that DNA sequencing is still the most frequently used
technology for detecting DNA variation. Today, the
comparison of genomes among several species allows
the study of numerous biological features, such as studies
of conserved sequences4–6.
Developments of genomic technology have until
now primarily focused on the generation of genome
sequence data, from the development of genome-
analysis technologies to the generation of physical and
genetic maps, the sequencing of model organism
genomes and the completion of the human genome
sequence. The next focus in genomics builds on the
genome sequences and heralds the beginnings of an
exciting phase of genome biology — the true genome era,
when deriving functional information from genome
sequences takes centre stage3. With this role in mind, we
evaluate three areas of functional genomics that have
been piloted in different model systems. We indicate
promising directions of research and suggest new
approaches that need to be designed.
Interfering with gene function. Phenotypic analysis of
mutants has been a powerful approach for determining
gene function. Gene function can be altered through gene
deletions, insertional mutagenesis and RNA INTERFERENCE
(RNAi) (BOX 1).
Few methods offer the experimental control that is
afforded by gene deletion. A true knock-out or null
mutation achieves complete functional reduction of the
encoded gene product. Because it is difficult to achieve in
many organisms, compromises have been made by gen-
erating incomplete knock-outs. Gene products can be
knocked-down or silenced as a result of point mutagene-
sis, insertional mutagenesis or RNAi. Although not yet
feasible on a large scale, proteins might be targeted using
drugs7, and it might eventually be possible to use drug
compounds to generate knock-downs for every gene
product in a genome and to apply them across species.
The power of systematic mutant analysis is well illus-
trated by an experiment in which an international con-
sortium systematically generated a gene deletion strain
for every gene in the yeast Saccharomyces cerevisiae
genome and analysed the phenotypes in a single tube
assay8,9 (FIG. 1). The quantitative fitness measurements
that are obtained for each gene with this tool enable
applications beyond determining whether a gene is
essential. This is an important advance because it opens
up a wide variety of applications based on quantitative
analysis, such as identifying functionally relevant genes
and drug targets, comparing function and expression,
defining candidate disease genes and studying molecular
evolution (BOX 2).

Box 1 | Comparison of knock-out approaches in model organisms

RNA INTERFERENCE
(RNAi). A process by which
double-stranded RNA silences
specifically the expression of
homologous genes through
degradation of their cognate
mRNA.
Targeted deletion by homologous recombination
Precise gene deletion can be readily achieved by homologous recombination in yeast 8,9 and mouse11. Because this
approach removes the targeted gene, functional reduction is complete. In organisms in which it works, this method
is the gold standard. Unfortunately, homologous recombination does not work efficiently in several model
organisms, including Arabidopsis and Caenorhabditis elegans. Although it has been shown to work in some cases,
as seen recently in Drosophila12, the efficiencies are still too low for systematic application.
Insertional mutagenesis
Disruption of gene sequences can be achieved by insertional mutagenesis using transposons or other
insertion sequences. Because the genome insertions are random, screening for disruption in a gene of interest
is required. The insertion can lead to complete, incomplete or no functional reduction, depending on where
the integration occurs. The insertion site and level of functional reduction therefore need to be determined
experimentally. The method has been used extensively in Arabidopsis16 and Drosophila77,78, yeast15, mouse79
and C. elegans 80.
RNA interference
RNA interference (RNAi) is the newest technology for reducing gene expression. It follows reports of gene silencing
in plants and other model organisms81, and is based on the observation from C. elegans that adding double-stranded
RNA (dsRNA) to cells often interferes with gene function in a sequence-specific manner17. In most cases, the level of
functional reduction is incomplete and the level of specificity is not entirely predictable24–26. Nevertheless, RNAi has
been shown to work in many model organisms. Current applications are primarily in C. elegans18, Drosophila19,
various plants 81, tissue culture cells of Drosophila 82 and mammals23.

NATURE REVIEWS | GENETICS VOLUME 5 | MARCH 2004 | 1 9 1

REVIEWS

a c
CP UPTAG CP KanMX CP DNTAG CP
Deletion cassette

ORF
Start Stop

Selection

F F
CP UPTAG KanMX DNTAG CP
F F
PCR amplification
F
F F
TAG
TAG TAG
F
F F F
TAG F F
F TAG TAG
F F

Hybridization Selection

Figure 1 | Assaying molecular barcode tags in yeast pools. a | Start-to-stop-codon deletion by double homologous
recombination with deletion cassettes. The 45-base pair (bp) sequences at each end of the PCR-amplified deletion cassettes are
identical to those found upstream and downstream of the targeted gene. On transformation, double homologous recombination
integrates the PCR product into the chromosome, displacing the target gene and generating a precise start-to-stop-codon
deletion. A dominant drug-resistance marker (KanMX) that is part of the PCR product serves to select for the integration event.
To allow pooling of the deletion strains and a parallel analysis, each PCR cassette also contains two unique 20-bp DNA
sequence tags that serve as molecular barcodes to uniquely identify each strain10 (UPTAG, 5′-end (upstream) tag sequence;
DNTAG, 3′-end (downstream) tag sequence). The tag sequences, flanked by two common priming (CP) sites, were designed to
be different to avoid cross-hybridization and to have optimal hybridization properties. b | Competitive growth of deletion pool,
genomic DNA extraction, PCR amplification and array hybridization. Because tag sets are flanked by two CPs, all UPTAGs
or DNTAGs can be amplified in a single PCR reaction from genomic DNA that is isolated from the mixed pool. The tags
can be quantitatively measured by hybridizing the PCR products to a high-density oligonucleotide array that contains the
complementary tag sequences. This microarray allows the signal from each tag to be distinguished. The signal quantity is
proportional to the abundance of strains in the culture. c | Measurement of quantitative deletion phenotypes in pools during an
experimental perturbation. By comparing the change in signal intensity during selection, change in abundance of strains (fitness)
can be calculated for all strains in parallel.

One of the strengths of the systematic deletion analy-
sis in yeast lies in the fact that deletion strains can be
pooled for parallel analysis. This is possible owing to the
presence of unique 20-base pair (bp) DNA sequence tags
in the deletion cassettes that serve as molecular barcodes
to uniquely identify each strain10.
Despite their power, genome-wide deletions that are
generated by homologous recombination have yet to be
reported in organisms other than yeast. In many cases,
such as mouse11, Drosophila12 and sheep13, in which the
homologous recombination system has been developed,
low targeting efficiencies and practical considerations
in maintaining deletion collections are important
concerns.
A targeted deletion approach is also not without
caveats — only annotated open reading frames (ORFs)
in a genome are generally targeted; the phenotypes of
overlapping ORFs cannot be easily distinguished; trans-
formation events might have introduced secondary
mutations14; essential genes are inviable as homozygous
deletions; and functionally redundant genes might not
show a detectable phenotype when deleted.
Apart from complete knock-outs achieved by gene
deletion, two other approaches that are more easily
applied to other organisms — insertional mutagenesis
and RNAi — might generate partial knock-outs. Partial
knock-outs are not as suited to interpreting fitness
effects, because the molecular nature of the mutation, its

192 | MARCH 2004 | VOLUME 5 www.nature.com/reviews/genetics


SATURATION MUTAGENESIS
A mutagenesis screen that has
reached a stage of saturation in
which additional mutagenesis
does not seem to recover
mutations in new genes.
HAPLOINSUFFICIENCY
A gene dosage effect that occurs
when a diploid requires both
functional copies of a gene for a
wild-type phenotype. An
organism that is heterozygous
for a haploinsufficient locus does
not have a wild-type phenotype.
STOICHIOMETRIC
The molar ratio of interacting
molecules.
TRANSFERRED DNA
(T DNA). The segment of DNA
in the Ti plasmid of
Agrobacterium tumefaciens that
is transferred to plant cells and
inserted into the chromosomes
of the plant.
NATURE REVIEWS | GENETICS
REVIEWS
Box 2 | Systematic mutant analysis in yeast
Knock-out analysis of the yeast genome is further along than that of any other organism, both in terms of the percentage
of the genome that has been deleted and the ease with which the quantitative phenotype of all the mutants can be
analysed. Its applications therefore serve to illustrate the power of genome-wide mutational analysis.
Identifying functionally relevant genes
By assessing quantitative reductions in fitness, a single study42 of yeast-homozygous diploid deletions doubled the
number of genes that are functionally implicated in sporulation, even though yeast sporulation has been studied
genetically for more than 30 years83. This study exemplifies the general finding that for every pathway that has been
studied so far using yeast deletions — including those in which SATURATION MUTAGENESIS has been reached — new genes
have been uncovered using systematic deletions.
Comparing function and expression
Intriguingly, comparing results from deletion and mRNA-expression studies showed that there is little overlap in the genes
that the two approaches identify. For sporulation data, only 16% of genes with changes in expression showed a significant
defect in sporulation when deleted42; for growth on non-fermentable carbon sources, the overlap was 7% (REF. 41); for
growth in galactose, high pH, high salt and sorbitol, the overlap was less than 7% (REF. 8), and in the case of DNA-
damaging agents, the number of genes with a fitness defect that showed differential expression in response to DNA
damage was no larger than expected by chance40. These consistent findings of low overlap indicated that it is not the result
of measurement error but that it is biologically significant, which means that change in gene expression might be less
functionally relevant than commonly anticipated.
Identifying drug targets
In an application of systematic deletions to define drug targets, heterozygous diploid strains were used for drug-induced
84,85
HAPLOINSUFFICIENCY profiling . The assay — in which sub-lethal concentrations of a drug reduce the fitness of strains
that carry heterozygous deletions of the drug target — was exploited genome-wide as a tool for drug target identification
and target-specificity evaluation.
Defining candidate disease genes
Deletion strains have specifically been used to study biological processes that have been conserved between yeast and
humans. A total of 466 genes in which deletions impaired mitochondrial respiration were identified and aligned against
the human genome to yield new mitochondrial candidate genes41. The candidates are of value in studies of human
mitochondrial disorders in which genomic intervals have been mapped but no responsible gene has been identified so far.
Recently, one gene (LRPPRC) was implicated in a human mitochondrial disorder, for which the yeast orthologue had
such a characteristic deletion phenotype52.
Studying molecular evolution
The quantitative knock-out analysis has catalysed new developments in molecular evolution. Using published genome-
wide fitness data in yeast, it was shown that highly conserved proteins tend to have larger fitness effects when deleted86, that
protein members of an interaction network tend to evolve at slower, yet similar, rates87, that duplicated genes tend to have
no, or weak, fitness effects, presumably as a result of functional redundancy between paralogoues88, and that heterozygous
deletions of protein subunits of protein–protein complexes result in a STOICHIOMETRIC imbalance that is often deleterious89.
effect on functional reduction and the degree of sec- into ORFs can lead to complete or partial reduction
ondary effects are often uncertain. Nevertheless, in some depending on exact location, whereas insertions into
cases — for example, when a partial reduction in the intergenic sequences often have no effect. The site of
function of essential genes is of interest — these integration and degree of functional reduction therefore
approaches can be beneficial. need to be determined for each mutant of interest.
One approach to insertional mutagenesis in yeast A more directed approach uses RNAi to silence
involves transposons that have been designed to measure the function of a target gene17. Although RNAi has
mutant phenotypes, gene expression and protein local- not been reported in S. cerevisiae, it is applicable in a
ization, all from the same collection of mutants15. In wide variety of model systems, including worms18,
Arabidopsis thaliana, 88,000 Agrobacterium TRANSFERRED- flies19, zebrafish20, mouse21, plants22 and human cul-
DNA (T-DNA) insertion lines — in which more than tured cells23. The primary advantages of RNAi are the
20,000 predicted genes have been mutated — have been ease of generating short double-stranded RNAs
generated, providing a powerful resource for future (dsRNAs) that mediate RNAi and the flexibility of inhi-
mutational analysis16. bition: the user can spatially and temporally control the
The main disadvantages of insertional mutagenesis interference reaction. Furthermore, it has been applied
in comparison to a targeted gene deletion approach are genome-wide18 by feeding worms Escherichia coli that
that full genome saturation is difficult and that some contain dsRNA-producing plasmids. The disadvantage
regions in the genome are not susceptible to insertion, in systematic genome-wide application, however, is that
whereas other regions are hot spots. Furthermore, the the level of functional reduction is unpredictable and
degree of functional reduction for a desired gene difficult to measure experimentally. A study showed that
depends on the location of the integration. Integrations small interfering RNAs (siRNAs) that target different
VOLUME 5 | MARCH 2004 | 1 9 3
REVIEWS
parts of the same gene achieve different levels of reduc-
tion and secondary off-target effects24. Two studies have
further shown that siRNAs mediate an interferon
response in mammalian cells as a secondary effect25,26.
Caution must therefore be exercised before attributing a
particular response to the targeted gene. Nevertheless, it
is hoped that these issues can be addressed, as RNA
silencing methodologies are likely to be applied genome-
wide in many organisms owing to their ease of use.
Gene-expression profiling. Gene-expression profiling is
the most widely used functional genomic technology
today, in part because it was one of the earliest to be
developed and in part because it achieves high through-
put in a single tube assay.
Applications of gene-expression studies can be
divided into three principal areas: those that generate
signature profiles, those that study transcription and its
regulation and those that determine the function of
gene products. Each is associated with different biologi-
cal concerns. Although in the first two applications the
Box 3 | Technical concerns about microarrays
Microarray analysis can detect the presence of individual target sequences in complex
mixtures because hybridization specificity is exploited. Nevertheless, important
technical concerns are associated with different types of microarray platform. Two types
of high-density microarray platform are most widely used and can be distinguished by
the method of probe placement onto the array surface. In the first type of microarray,
oligonucleotide probes are synthesized directly on the glass surface90,91.
Photolithography and photosensitive oligonucleotide synthesis chemistry90 are the most
common processes used and achieve a density that is greater than 1,000,000 spots/cm2.
In the second type of microarray platform, represented mainly by the cDNA
microarray92, probe synthesis is separated from array manufacture92–94. On printed
microarrays, probes are synthesized and then mechanically spotted or printed in an
arrayed format, with a density of ~10,000 spots/cm2.
The method of array manufacture places different demands on quality control. Probes
on synthesized microarrays are short and are synthesized directly on the glass surface. A
sequence database directs the synthesis of probes. Because synthesis is highly
reproducible, absolute levels of expression can be estimated, and because the probes are
short, SNPs can be detected. Probes can be specifically designed to distinguish splice
variants and members of gene families. Nevertheless, the short probes require accurate
databases, as errors in sequence databases translate to errors in probe sequence on the
arrays. However, as sequences in databases are updated and sequence quality improves,
the chance for errors decreases. Importantly, because there can be multiple probes to
each gene, the array designs provide internal redundancy.
For printed microarrays, probe generation is separated from array synthesis; therefore,
storage and tracking of probe samples is crucial. Errors occur when probes are spotted
from microtitre plates that are either wrong, out of order, in the wrong orientation or
contaminated. One study reported that only 62% of DNA samples from plates used for
microarray spotting were of the correct identity 95. Such errors would be further amplified
by errors during and after printing 96. Printing results in variations in spot size and probe
concentrations, and is at present addressed by co-hybridization of a reference to the
same array. The choice of reference is crucial for data interpretation as it affects the ratio
calculation, and the use of different reference samples by different laboratories makes
comparisons between experiments performed by different laboratories difficult44.
Another problem is with distinguishing splice variants and members of gene families
— the long sequences on cDNA microarrays allow cross-hybridization. This is an
important limitation because, to some extent, the molecular complexity in humans
arises from differential splicing 97. Array designs that can distinguish between splice
variants need to be implemented98.
194 | MARCH 2004 | VOLUME 5
measurement of gene expression is closely related to the
biological goal, more assumptions are required to infer
function from expression differences in the third type
of study.
Perhaps the most powerful application of gene-
expression data is its use as a signature profile, which
can be used as a detailed molecular phenotype. Although
the identity of the individual genes that change expres-
sion is, in the first case, irrelevant, they must have
accurate gene assignments (BOX 3). The profile across
thousands of genes gives a distinctive pattern for a sam-
ple. Such patterns have been used in model organisms
to classify genetic mutants by similarities in profile27 and
to evaluate secondary effects of drug treatments28.
However, the most powerful applications of signature
profiling are likely to be for characterizing human disease
populations (see below).
Another application of gene expression is to study
mRNA transcription and its regulation. Gene-expression
data can be used to search for regulatory sequences by
looking for common sequence elements in upstream
regions of genes that have similar expression profiles29–31.
To find target genes of specific transcription factors,
genes that are under expressed as a result of transcrip-
tion-factor deletions or overexpressed as a result of tran-
scription-factor gain-of-function mutations can be
identified32. A complementary approach identifies the
DNA that is bound to transcription factors by co-
immunoprecipitation33. Integration of the analysis of
DNA that is bound to transcription factors with co-
expression of genes has shown a complex network of
interactions that explains the difficulty of distinguish-
ing primary from secondary effects in the analysis of
transcription-factor mutations33. Together, these data
help to formulate hypotheses about transcription factors
and their targets. The approaches can be complemented
by knock-out analysis of predicted pathway members,
with the aim of dissecting the regulatory circuitry and
signalling pathways that regulate transcription.
The application that is at present least understood is
the use of gene expression to determine gene-product
function. Most mRNAs are not functional themselves;
they are intermediates and transmitters of information
from the genome to the proteome. Inferring function
from mRNA levels is indirect and rests on several
assumptions. One is that evolutionary selection has
been tight enough to ensure that mRNAs are pro-
duced and present only when the corresponding gene
products are needed. This assumes that no regulatory
changes occur at the protein level that are not
reflected by changes in amounts of mRNA. A second
assumption is that transcription of one gene is inde-
pendent of another and that there is no competition
for resources.
These assumptions are questionable. A cell functions
as a unit and therefore transcription-factor availability
and protein concentrations probably depend on their
use at other parts of the cell34. It is also evident that regu-
latory changes occur during mRNA translation, post-
translational modification and protein degradation. In
fact, experiments show that the correlation between
www.nature.com/reviews/genetics

mRNA changes and changes in protein level in a cell is
weak35–37. Although a genome-wide comparison is still
missing — mainly as a result of the technical difficulty
of accurately measuring protein levels under multiple
conditions for many proteins38,39 — the trend is appar-
ent: one cannot assume that there is a correlation at the
individual gene level. In addition, the poor correlation
between changes at the gene-expression level and fit-
ness effects of gene products8,9,40–42 (BOX 2) indicates that
there are many more genes that change expression
between any two experimental conditions in yeast than
there are genes that affect phenotype by gene deletion
during the same transition. In addition, they point out
that most genes with a deletion phenotype in yeast do
not show an expression change when measured for the
same perturbations.
There are two types of issue when we interpret
expression differences: if gene expression is changed or
remains unchanged between two conditions, how reli-
able is this result and what is its biological significance?
Whether the result is reliable depends to a large extent
on the technology that is used to measure the expression
difference and on the accuracy of the experimental exe-
cution; these concerns exemplify technical challenges
that are associated with high-throughput approaches. In
the case of gene-expression profiling, they indicate that
more rigid quality control and careful experimental
design are needed43,44 (BOX 3). This technical concern is
independent of the difficulty that is associated with the
functional interpretation of expression changes because
there are difficulties even when considering the most dif-
ferentially expressed genes for which different platforms
often agree.
Genetic mapping of quantitative trait loci. A third prin-
cipal area of high-throughput genomics is the identifi-
cation of the genetic factors that underlie complex traits.
Developments in this field can be divided into three
areas that correspond to the process of dissecting the
genetic basis of complex traits: defining genetic inter-
vals, identifying candidate genes and verifying an allele’s
contribution to the phenotype.
Genetic mapping has been successful in finding
important genes for some complex diseases such as
asthma45. However, in most cases, genetic mapping has
been problematic1, as evidenced by reports that previ-
ously mapped intervals failed to withstand significance
tests in subsequent studies46.
A few recently developed high-throughput genotyp-
ing technologies show promise for testing the feasibility of
high-resolution association studies. The rationale in these
studies is to test thousands of polymorphisms in carefully
selected populations to identify sequence variants that
are more frequent in individuals with one phenotype
than in individuals with other phenotypes. The tech-
nologies can achieve throughputs of 10,000’s of markers
(BOX 4). Unfortunately, genotyping the 50,000–1,000,000
markers that have been proposed as the minimum
needed by statistical predictions47 still represents a
tremendous challenge, especially because the markers
need to be genotyped in large numbers of samples. The
NATURE REVIEWS | GENETICS
REVIEWS
challenge is to develop methods that can accurately
genotype thousands of markers when only a small
amount of sample is available and at low costs. One
technology uses molecular barcodes to genotype
directly from genomic DNA48. It uses the same concept
of multiplexing as the yeast deletion project: the signal is
amplified using common primer sequences after the
genotyping reactions have occurred (BOX 4).
Even with the ability to identify map intervals, the
causative variants remain to be identified. In cases in
which intervals have been identified, several approaches
have been taken to locate the underlying genes. A sys-
tematic approach involves identifying all sequence vari-
ants in these intervals. This approach is challenging
because the sequencing of entire intervals is, in most
cases, impractical given the large sizes and, often, limited
DNA samples. Furthermore, mutations that are present
in a heterozygous state are difficult to identify. A tech-
nology that uses the mismatch-repair system of bacteria
holds promise for focusing attention on cloned frag-
ments that contain a polymorphism49,50 (BOX 5). The
approach can be combined with molecular barcodes for
high-throughput mutation scanning.
Rather than scanning an interval for all polymor-
phisms, candidate genes can be prioritized. Such
approaches can be applied either to genes in mapped
intervals or to candidate genes genome-wide, in the
absence of positional information. One study has used
data mining of the published literature to establish a
connection between disease terms and functional terms
associated with genes51. The approach scores a candidate
on the basis of how frequently it can be connected to the
disease description by terms that co-appear in the litera-
ture (using PubMed). Another approach is to use func-
tional genomic data from model organisms to rank
human candidates according to functional information
about their orthologues41,52.
Gene-expression profiling has also been proposed as
a general method for identifying candidate genes53. One
study integrated a disease interval with experimental
data of expression and proteomics to identify a muta-
tion in patients with Leigh syndrome52. Another study
successfully found a quantitative trait locus (QTL) gene
on the basis of the absence of mutant-gene expression
in hypertensive rats54. In the latter study, gene expression
was decreased as a result of partial deletion of the ORF.
However, for traits that are not caused by deletions that
result in expression differences or mutations that affect
mRNA levels, expression might not identify causative
variants.
A recent review47 assessed the types of mutation
found in Mendelian diseases and enables inferences
about possible implications for disease-gene identifica-
tion approaches in complex traits. Surprisingly, regula-
tory changes are found in fewer than 1% of the more
than 1,400 known Mendelian disease genes (see Human
Gene Mutation Database in online links box). The vast
majority of mutations are missense or nonsense changes
in the ORFs (58%), insertions or deletions (30%) and
splicing variants (10%). If similar patterns are true for
complex traits, it might be expected that genes identified
VOLUME 5 | MARCH 2004 | 1 9 5
REVIEWS

Box 4 | High-throughput concepts for genotyping

DNA hybridization to a microarray
PCR-amplified DNA99, or genomic DNA in the case of small genomes100, that contains the polymorphisms to be genotyped
is hybridized to an oligonucleotide array in one assay. If a target has a perfect match to a probe, a stronger hybridization
signal is obtained than if it has a mismatch. High throughput is achieved through a high density of probes on arrays.
Fluorescence polarization
In one version of fluorescence polarization genotyping101, the genomic DNA that contains a polymorphism to be assayed
is amplified by PCR. A primer is designed to end one base before the polymorphic position. The primer is extended in a
reaction with two different fluorescent dideoxy-nucleotides that are specific to the alleles of the SNP. The amount of each
fluorescent base that is incorporated can be determined by scanning the plate at two wavelengths. Fluorescent bases
incorporated into the extended primer rotate in solution more slowly than free-floating (unincorporated) fluorescent
nucleotides, leading to fluorescence polarization. This method can be carried out in multi-well plates, is economical and is
simple to use.
Mass spectrometry
Genotyping can be achieved by analysing primer extension products by mass spectrometry102. The increased mass of an
extended primer is detected by a shift in mass peak. The speed, throughput and accuracy of mass spectrometry allow for
fast and accurate genotype assignments.
Molecular barcodes
In a method termed ‘molecular inversion probes’, genotyping can be performed directly on genomic DNA and thousands
of SNPs can be analysed in one reaction by taking advantage of the multiplexing capability of DNA barcodes48. As
illustrated in the figure, for each SNP, a probe that contains a unique DNA barcode is designed to have two ends that are
complementary to the bases that flank the polymorphic site. The SNP base is filled in with a single base extension and
ligation closes the gap to form a circular piece of DNA (exonuclease digests away unreacted linear probes). The barcodes
from reacted probes are
detected by hybridization to Anneal Gap fill-polymerization Gap fill-ligation
a high-density array that Barcode tag
contains the tag complements.
A high level of multiplexing can Genomic Probe
be achieved by performing the CGGAGATGGCCCA CGGAGATGGCCCA CGGAGATGGCCCA
genotyping reactions first and GCCTCT CCGGGT GCCTCTACCGGGT GCCTCTACCGGGT
then amplifying the signal by
PCR. Figure modified with Exonuclease selection Probe release Amplification Hybridization
permission from REF. 48 ©
(2003) Macmillan Magazines
Ltd.

by altered expression will primarily be modulators of
the phenotype or will be involved in secondary effects
and not the causative variants themselves.
Meeting the challenges
The future of new genomic technologies is crucially
linked with bringing them into the laboratory of an
individual investigator because it is there that the key
biological expertise that is required to make a biologi-
cal breakthrough lies. Therefore, the technologies with
most promise are those that allow high-throughput
biology to be carried out in an individual laboratory,
instead of genome centres and well-equipped and
well-financed institutes. For this purpose, the app-
roaches that achieve high throughput with one
method performed many thousands of times in paral-
lel should be distinguished from those that achieve
high throughput in a single tube assay (FIG. 2). The lat-
ter approaches not only allow the transfer of high
throughput to individual laboratories but also allow a
further scale up, by several orders of magnitude,
through the use of robotics or high-throughput equip-
ment. This advance also enables the next cycle — the
reduction of multiple parallel high-throughput assays
back into a single tube. These development cycles allow
unprecedented exponential increases in throughput
and decreases in cost that would define a paradigm
shift in biological interrogation.
Molecular barcoding is among the most promising
technologies for miniaturizing high-throughput
approaches into single tube assays. Through a combi-
nation of molecular barcodes and the microarray-
based detection system, the yeast-deletion approach has
achieved an unparalleled level of throughput, and its
many applications are a testament to its power (BOX 2).
A molecular barcode assay might simply consist of
attaching tags to the biological entities that are to be
interrogated (for example, knock-out strains), collect-
ing mixed tag pools after selection, performing a single
PCR reaction, labelling it and hybridizing it to a
microarray to obtain the relative abundance of all bio-
logical entities in the selected pool. The primary advan-
tage of the barcode concept is that it is versatile and can
be applied to other knock-out approaches, such as
insertional mutagenesis and RNAi, as well as to other
biological assays (see below).

196 | MARCH 2004 | VOLUME 5 www.nature.com/reviews/genetics

 REVIEWS

Interfering with gene function. Although the deletion
approach for yeast is high throughput, its power can be
improved by complementing the assay with further
approaches; for example, temporally induced knock-
outs using repressible promoters55 or targeting pro-
teins for proteolysis56 for the study of essential genes
that are inviable as homozygous knock-outs. Further-
more, to uncover molecular differences among mutants
when no growth differences are apparent, the fitness
data of knock-outs could be complemented with
metabolite profiling57,58 and expression profiling of dele-
tion strains27. Finally, the systematic construction of
double deletions that has been initiated to assess syn-
thetic lethality59,60 could help to characterize functionally
redundant genes.
To ensure that molecular barcodes are associated
with the right ORFs, Shoemaker et al.10 incorporated the
molecular barcode tags into the primers that contain the
target-specific homologous sequences, making it
unlikely that one molecular barcode sequence will
associate with a deletion of a different ORF. It does
not, however, guarantee that all barcode sequences
are without errors. DNA-primer synthesis might
have introduced mutations in some of the barcode
sequences, which would affect the hybridization signal
on microarrays. The presence of two molecular bar-
codes in each mutant makes it unlikely that both con-
tain mutations; in addition, the comparative analysis of
samples before and after selection controls for such mis-
takes. Nevertheless, barcodes are being sequenced and
the hybridization effect should be corrected by remaking
the strains or by introducing the corresponding changes
into the microarray probes.
To advance the insertional mutagenesis approach,
barcodes could, in theory, also be incorporated into each
insertion sequence, enabling a pooled knock-out analy-
sis. Such pooled approaches using transposons have
been carried out in bacteria61,62. Technically, the RNAi
approach might also be aided by the use of molecular
barcodes, which can in theory be integrated into the
plasmids or the genome-integration cassettes that
express dsRNA molecules. Such an advance might sim-
plify mutant tracking and enable pooled analysis in
organisms for which this is practical.

Box 5 | Identifying polymorphisms by mismatch-repair detection

Mismatch-repair detection uses the repair Linear vector Pool of single-stranded
Pool of PCR (5-bp CRE deletion, standards cloned in a vector
system of bacteria to separate DNA fragments products methylated) (full length CRE, unmethylated)
that contain a polymorphism from those that do
not50. A pool of PCR products to be tested is
+ +
mixed with linear vector DNA. The vector DNA
is methylated and contains a 5-base pair (bp)
deletion of an incorporated CRE gene. A pool of Heteroduplex, Taq ligase + endonuclease
single-stranded, unmethylated control DNAs
(standards) is added to this mixture, which
consists of, for example, PCR products that are
amplified from a different individual and cloned Transformation
into a vector that is identical to the first but
contains the full-length Cre gene (no 5-bp strS
deletion). On denaturing and reannealing, the tetR

mixture forms a heteroduplex vector that lox

CRE
Cre encodes a site-specific
recombinase that recognizes and
binds to specific sites called lox.
Two lox sites recombine at nearly
100% efficiency in the presence
of Cre, allowing DNA that is
cloned between two such sites to
be removed by Cre-mediated
recombination.
EPISOME
An independent DNA element,
such as a plasmid, that can
replicate extrachromosomally
or that can be maintained by
integrating into the genome
of the host.
LOX SITE
A site to which Cre recombinase
binds to mediate recombination,
allowing DNA that is cloned
between two such sites to
be removed.
contains the methylated tester DNA and the Cre No variation lox Variation
deletion on one strand and the unmethylated
control DNA and the full-length Cre gene on the
tetR
other strand. After digesting away unligated strS strS
tetR
fragments, the heteroduplex molecules are
transformed into an Escherichia coli strain that CRE
carries on its EPISOME two LOX SITES that flank
tetracycline-resistance (tetR) and streptomycin-
sensitive (strS) genes. If there are no
polymorphisms in the test fragment relative to
the standard, no repair occurs (5 bp or more lox strS
insertions/deletion polymorphisms, as in the tetR
Cre gene, do not trigger the mismatch-repair
system). The heteroduplex replicates and
generates a plasmid with an active Cre gene,
which leads in turn to recombination of the tet sensitive tet resistant
tetR/strS/lox cassette. This event renders the cell str resistant str sensitive
tetracycline sensitive and streptomycin resistant
(the strain contains a streptomycin-resistance gene on its chromosome). If a polymorphism is present, repair occurs and
extends into the Cre gene. Repair yields an inactive copy of Cre, leaving the cell tetracycline resistant and streptomycin
sensitive. By growth on the respective selectable media conditions, variant and invariant pools can be separated.

NATURE REVIEWS | GENETICS VOLUME 5 | MARCH 2004 | 1 9 7

REVIEWS

Gene-expression profiling. Evaluating gene expression

1
today shows that, although very popular, the data and
the technology might be among the least understood.
The overlap between deletion phenotype and gene
A1 expression in yeast, for example, questions the functional
Amplify using relevance of expression changes.
robotics Applications that do not require an explanation for
the functional relevance of an expression change might
avoid this potential pitfall. One application is in medi-
cine, and might yield a true demonstration of the power
2 of expression analysis. By classifying disease samples
into groups63–65, signature profiling has immediate
promise for disease diagnosis and personalized medi-
cine. A comparison of expression profiles between
A1 A2 A3 Ai patients and controls has identified groups of genes that
Miniaturize with can be used as predictors of treatment outcome and
a new concept survival rates66. These predictors might be comple-
mented in the future with genetic predictors that are
identified at the DNA sequence level. In fact, it is possi-
ble that many biological assays, including identification
3
of sequence variants, can be carried out on arrays so that
millions of individual measurements can be made on an
individual, providing excellent resolution for disease
A1– i diagnosis by pattern recognition. Meanwhile, in the
Amplify using short-term, while the challenge is to classify the dis-
robotics ease samples rather than to identify the genetic causes,
expression analysis is already powerful on its own.
Furthermore, if, as has been suggested, gene expression
might reflect the environmental past to which a patient
4
has been exposed43, expression profiling of patients could
yield further useful information. In both cases, the expres-
sion signatures are easy to score and provide different data
for diagnosis from the data that are conventionally avail-
A 1– i B 1– i C 1– i Z 1– i
able through measures of clinical parameters and from a
Miniaturize patient’s description of symptoms.
again with a
new concept How can the mRNA-expression changes be inter-
preted to obtain functional information? It is likely that
understanding the functional relevance of the changes in
gene expression will require understanding the complex
5 networks that regulate expression, which in turn requires
understanding the regulation of transcription factors
and the function of the sequences to which they bind.
A1–i
Repeat Therefore, integrating expression profiles with binding
B1–i
sites of transcription factors and knock-out phenotypes
C1–i of pathway members might help with understanding
the significance of expression differences33.
It is also likely that inferring function from expression
Z1–i
change is most powerful when analyses are combined
Figure 2 | High-throughput development cycles. High throughput can be achieved by en masse across many conditions to obtain a molecular
amplifying the number of experiments using robotics and automation (transition from stage 1 readout of the regulatory patterns of a gene. Similar to
to stage 2). However, a true advancement does not occur until a new concept or approach
using expression profiles as signature patterns for differ-
allows high throughput to be achieved in a single experiment, illustrated here with a single tube
(stage 3). This development allows a further increase in throughput with multiple parallel high-
ent samples, this application uses signature patterns
throughput reactions (stage 4), which can in turn be reduced back to a single assay, completing obtained for each gene to group genes on the basis of
the next cycle (stage 5). An example for stage 2 might be performing individual Northern blots for similarity67,68. This application seems especially promis-
all genes in the yeast genome; for stage 3, it might be measuring genome-wide gene expression ing for studies of transcriptional regulation29–31,33, and a
in parallel using a microarray. Stage 5 might involve combining different whole-genome better view of the regulatory network that controls tran-
measurements, such as gene expression, genotyping, knock-out phenotypes, protein scription will help us to understand why so many genes
interactions, protein levels, splicing, and so on, in a single assay. Extending this approach to
include a dedicated unique molecular barcode for each element in each assay might make it
change expression in response to environmental pertur-
possible to make all measurements with a single high-density microarray hybridization. Further bations. Analysis across many conditions might also
developments could then allow these assays to be run not just for one strain, but for many help to define functional groupings30 and to assign
strains or individuals in parallel. probability-scored functional assignments to unknowns

198 | MARCH 2004 | VOLUME 5 www.nature.com/reviews/genetics

 REVIEWS

that have been assigned to groups69. Nonetheless, the a

co-expression of genes, even across thousands of condi- A1 A2
tions, is in many cases not enough to make a correct
functional association. Studies have begun to integrate B1 B2 B1 B2
A1 vs A2
other parameters, such as evolution70, to filter out gene
C1 C2 C1 C2
interactions that are not functionally relevant.
As the data sets for such global analyses are often
not collected by the same experimenter, these analyses A1 A2 A1 A2
still raise issues of data reliability. For example, print-
ing mistakes (BOX 3) that result in a partial randomiza- B1 B2
B1 vs B2
tion of spots on cDNA microarrays introduce errors.
C1 C2 C1 C2
Historically, errors were detected by repeating experi-
ments, and microarray experiments of expression pat-
terns under many conditions might also benefit from A1 A2 A1 A2
being repeated. Raw data need to be made available to
cross-validate experiments (for cDNA microarrays, that B1 B2 B1 B2
C1 vs C2
is the data before ratio measurements are calculated); it C1 C2
is especially important when gene expression is used as a
diagnostic indicator that the technical and biological Genetic Genetic Genetic Genetic
background background background background
error rates of microarray data are addressed. Consistent 1 2 1 2
with the diverse applications of expression data, the b
~6,000 ~6,000
concerns indicate that our understanding of the biologi-
cal significance of gene expression is still rudimentary
and can be much improved by insights from more
rigorous studies.

Genetic mapping of QTLs. New methods that allow the A1 – Z1 A2 – Z2

COMPLEMENTATION
One example is the use of partial
diploids to determine whether
two mutations affect the same or
different genes. If the mutations
are in the same gene, they
generally fail to complement
each other and the diploid
retains the mutant phenotype.
By contrast, mutations in
different genes usually
complement one another and
restore a wild-type phenotype to
the diploid. However, exceptions
to both cases abound.
HEMIZYGOUS
A diploid genotype that has only
one copy of a particular gene, as
in X-chromosome genes in a
male, or when the homologous
chromosome carries a deletion.
rapid discovery, scoring and functional analysis of
genetic variation could revolutionize QTL analysis.
Verifying and measuring the effect of an allele on the
phenotype is an area in complex trait dissection that
requires increased attention. Sequence-based approaches
might identify polymorphisms but in practice they can-
not distinguish between neutral polymorphisms and
causative mutations. Candidate gene approaches are
based on previous knowledge and cannot identify
mutations in previously unknown genes. The most sys-
tematic approaches would combine identification of
alleles with proof of their contribution to the pheno-
type. In model organisms, genes have often been identi-
fied and their role confirmed by COMPLEMENTATION71. Yet,
in most cases, it is impractical to test every gene in an
interval by complementation. To overcome this problem,
a promising functional tool of reciprocal HEMIZYGOSITY was
developed in yeast72. The technology can systematically
identify a QTL allele by determining the effects of a sin-
gle allele in an otherwise uniform genetic background.
The assay is based on a phenotypic comparison between
two strains in which a single allele is switched in a het-
erozygous diploid strain. The method can be adapted
for the whole genome — by uniquely tagging each
strain with molecular barcode tags, all reciprocal hem-
izygous strains could be grown as a pool. With this
approach, it might be possible to circumvent linkage
mapping and analyse all allelic variants between two
genomes in a single step (FIG. 3).
The dissection of the quantitative trait of high-
temperature growth in yeast using reciprocal hemizygos-
ity identified three genes that are located in a single QTL
interval72. The conventional approaches of sequence,
expression and marker-trait association analyses were
c
A1 A2 Deletion
effect
B1 Allelic switching B2
effect
C1 C2
Figure 3 | Allelic switching with reciprocal hemizygosity.
a | All pairwise allelic comparisons are performed between two
genomes in a heterozygous hybrid strain background. Two
strains are compared that differ in the allele of only one gene
and are heterozygous diploid for the rest of the genome. As
illustrated, allelic comparisons are achieved using reciprocal
hemizygous deletions that disrupt a single allele, although, in
theory, a similar comparison can be carried out with allele
replacements or disruptions of combinations of alleles. b | With
molecular barcodes, the comparisons for an entire genome
can be performed in pools. c | By hybridizing barcodes to
arrays, each experiment promises to identify the allelic variants
that contribute differentially to the phenotype under the tested
conditions. In the example shown, the strain that carries the B1
allele yields a greater hybridization signal than the reciprocal
strain that carries the B2 allele, indicating that the B1 allele
confers a fitness advantage over the B2 allele.
tested in this study but failed to identify the QTL alleles.
The tight linkage between three QTL alleles with different
levels of contribution indicated that linkage mapping —
the method of choice for single-gene Mendelian traits —
is intrinsically deficient when applied to quantitative
traits. Although narrowing a map interval in the hope of
approaching a single point might locate the main con-
tributor, neighbouring genetic factors could be missed.
Contributors might also go undetected when their effect
is small or when they are located in trans, opposite
another QTL allele.

NATURE REVIEWS | GENETICS VOLUME 5 | MARCH 2004 | 1 9 9

REVIEWS

EPISTASIS
An interaction between non-
allelic genes, such that one gene
masks, interferes with or
enhances the effect of the other
gene.
To speed the dissection of quantitative traits, new
tools, such as reciprocal hemizygosity, genome-wide
marker mapping and mutation-scanning methods, need
to be more broadly applied73,74. The ideal approaches
should be applicable to different genetic backgrounds, as
the alleles identified in one background might not
necessarily have the same effects in another background
or under different environmental conditions75. Further-
more, allele frequencies are likely to range from common
to rare72, highlighting the demand for approaches that
make no prior assumptions about allele frequency in
populations. Alleles are also likely to have additive and
EPISTATIC effects, requiring tools that can analyse single
alleles and higher-order combinations of alleles.
Classification of phenotypes also deserves attention76.
In cases in which samples do not come from crosses but
are sampled from populations, phenotypic heterogene-
ity could have detrimental effects on the ability to detect
association between a marker and disease. Phenotypic
classification might be aided by molecular phenotypes
based on expression profiles, signature patterns and
other molecular parameters. By integrating genetics
with high-throughput tools of functional genomics in
this way, it might be possible to advance the dissection
of complex traits and gain deeper insights into the
interaction of genes and the environment.
Conclusions
Functional genomics has changed the way biology is
done. And yet the field is still in its infancy in terms of
detailing the complexity that underlies biological sys-
tems, such as the complex network of genetic regula-
tion, protein interactions and biochemical reactions that
make up a cell. As the applications illustrate, systematic
knock-out analyses in multiple model organisms will aid
the definition of different cellular components and help
to describe the conservation of biological processes dur-
ing evolution. Detailed analysis of mRNA regulation will
help to clarify the functional significance of gene expres-
sion. Furthermore, the discovery of examples for the
genetic basis of complex traits in model organisms will
help to formulate more rigorous hypotheses for what to
expect in higher organisms and promises to provide
better approaches for dissecting human disorders.
Technological improvements need to maximize the
potential offered by functional genomics. False predic-
tion rates of high-throughput approaches need to be
eliminated, particularly as we advance towards applying
functional genomics in the clinical setting for diagnostics
and personalized medicine. Improvements in data analy-
sis, comparison and the integration of approaches that
combine different measurements for a genome need to
be improved to generate a global picture and to move
towards generating refined models of cellular processes.
Importantly, to accelerate discoveries, high-through-
put biology needs to be brought into individual laborato-
ries because that is where the biological expertise lies. The
use of molecular barcodes is one example of an innova-
tion that provides increased throughput through minia-
turization, accompanied by reduced costs. Incorporating
such technology into other assays would provide the
advance in efficiency that makes it possible to imagine
high-throughput experimentation as an integral part of
any individual investigator’s laboratory. It should also be
possible to combine different whole-genome measure-
ments, such as gene expression, genotyping, knock-out
analysis, protein interactions, protein levels and splicing
into a single assay — a further step towards miniaturiza-
tion — as proposed in FIG. 2. These developments would
in turn maximize biological discovery and take functional
genomics to the next level. Similar approaches would also
promise to revolutionize medical diagnostics for the
delivery of better health care and lower health care costs.
Although we believe that high-throughput experi-
mentation will become more widespread, we expect
single-gene studies to remain essential. They continue
to help to verify and evaluate high-throughput data
sets; high-throughput technologies are often not sensi-
tive enough to obtain accurate information for every
gene in a genome and, in some cases, detailed analysis
of individual genes is needed to establish a link between
the global picture and individual discoveries that
involve few sets of genes. Nevertheless, the integration
of functional genomics into individual laboratories will
provide the most accurate test of the power brought
about by integrating single-gene studies with high-
throughput approaches for faster and more efficient
biological discovery.

1. Risch, N. J. Searching for genetic determinants in the new
millennium. Nature 405, 847–856 (2000).
2. Flint, J. & Mott, R. Finding the molecular basis of quantitative
traits: successes and pitfalls. Nature Rev. Genet. 2, 437–445
(2001).
3. Collins, F. S., Green, E. D., Guttmacher, A. E. & Guyer, M. S.
A vision for the future of genomics research. Nature 422,
835–847 (2003).
4. Waterston, R. H. et al. Initial sequencing and comparative
analysis of the mouse genome. Nature 420, 520–562 (2002).
5. Kellis, M., Patterson, N., Endrizzi, M., Birren, B. & Lander, E. S.
Sequencing and comparison of yeast species to identify
genes and regulatory elements. Nature 423, 241–254 (2003).
6. Cliften, P. et al. Finding functional features in
Saccharomyces genomes by phylogenetic footprinting.
Science 301, 71–76 (2003).
7. Stockwell, B. R. Chemical genetics: ligand-based discovery
of gene function. Nature Rev. Genet. 1, 116–125 (2000).
8. Giaever, G. et al. Functional profiling of the Saccharomyces
cerevisiae genome. Nature 418, 387–391 (2002).
9. Winzeler, E. A. et al. Functional characterization of the
S. cerevisiae genome by gene deletion and parallel analysis.
Science 285, 901–906 (1999).
10. Shoemaker, D. D., Lashkari, D. A., Morris, D., Mittmann, M.
& Davis, R. W. Quantitative phenotypic analysis of yeast
deletion mutants using a highly parallel molecular bar-coding
strategy. Nature Genet. 14, 450–456 (1996).
The first use of PCR-amplifiable molecular barcodes
for high-throughput parallel biology.
11. Thomas, K. R. & Capecchi, M. R. Site-directed mutagenesis
by gene targeting in mouse embryo-derived stem cells. Cell
51, 503–512 (1987).
12. Rong, Y. S. & Golic, K. G. Gene targeting by homologous
recombination in Drosophila. Science 288, 2013–2018 (2000).
13. McCreath, K. J. et al. Production of gene-targeted sheep by
nuclear transfer from cultured somatic cells. Nature 405,
1066–1069 (2000).
14. Hughes, T. R. et al. Widespread aneuploidy revealed by
DNA microarray expression profiling. Nature Genet. 25,
333–337 (2000).
15. Ross-Macdonald, P. et al. Large-scale analysis of the yeast
genome by transposon tagging and gene disruption. Nature
402, 413–418 (1999).
16. Alonso, J. M. et al. Genome-wide insertional mutagenesis
of Arabidopsis thaliana. Science 301, 653–657 (2003).
A large-scale insertional mutagenesis screen in
Arabidopsis.
17. Fire, A. et al. Potent and specific genetic interference by
double-stranded RNA in Caenorhabditis elegans. Nature
391, 806–811 (1998).
18. Kamath, R. S. et al. Systematic functional analysis of the
Caenorhabditis elegans genome using RNAi. Nature 421,
231–237 (2003).
19. Kennerdell, J. R. & Carthew, R. W. Heritable gene silencing
in Drosophila using double-stranded RNA. Nature
Biotechnol. 18, 896–898 (2000).
20. Nasevicius, A. & Ekker, S. C. Effective targeted gene
‘knockdown’ in zebrafish. Nature Genet. 26, 216–220 (2000).
21. Wianny, F. & Zernicka-Goetz, M. Specific interference with
gene function by double-stranded RNA in early mouse
development. Nature Cell Biol. 2, 70–75 (2000).

200 | MARCH 2004 | VOLUME 5 www.nature.com/reviews/genetics


22. Ogita, S., Uefuji, H., Yamaguchi, Y., Koizumi, N. & Sano, H.
RNA interference: producing decaffeinated coffee plants.
Nature 423, 823 (2003).
23. Elbashir, S. M. et al. Duplexes of 21-nucleotide RNAs
mediate RNA interference in cultured mammalian cells.
Nature 411, 494–498 (2001).
24. Jackson, A. L. et al. Expression profiling reveals off-target
gene regulation by RNAi. Nature Biotechnol. 21, 635–637
(2003).
25. Bridge, A. J., Pebernard, S., Ducraux, A., Nicoulaz, A. L. &
Iggo, R. Induction of an interferon response by RNAi vectors
in mammalian cells. Nature Genet. 34, 263–264 (2003).
26. Sledz, C. A., Holko, M., de Veer, M. J., Silverman, R. H. &
Williams, B. R. Activation of the interferon system by short-
interfering RNAs. Nature Cell Biol. 5, 834–839 (2003).
27. Hughes, T. R. et al. Functional discovery via a compendium
of expression profiles. Cell 102, 109–126 (2000).
The use of expression signature profiles on yeast
knock-out mutants for achieving functional
groupings.
28. Marton, M. J. et al. Drug target validation and identification
of secondary drug target effects using DNA microarrays.
Nature Med. 4, 1293–1301 (1998).
29. Pilpel, Y., Sudarsanam, P. & Church, G. M. Identifying
regulatory networks by combinatorial analysis of promoter
elements. Nature Genet. 29, 153–159 (2001).
30. Ihmels, J. et al. Revealing modular organization in the
yeast transcriptional network. Nature Genet. 31, 370–377
(2002).
31. Segal, E. et al. Module networks: identifying regulatory
modules and their condition-specific regulators from gene
expression data. Nature Genet. 34, 166–176 (2003).
32. Holstege, F. C. et al. Dissecting the regulatory circuitry of a
eukaryotic genome. Cell 95, 717–728 (1998).
33. Lee, T. I. et al. Transcriptional regulatory networks in
Saccharomyces cerevisiae. Science 298, 799–804 (2002).
34. Brenner, S. Sillycon valley fever. Curr. Biol. 9, R671 (1999).
35. Ideker, T. et al. Integrated genomic and proteomic analyses
of a systematically perturbed metabolic network. Science
292, 929–934 (2001).
36. Griffin, T. J. et al. Complementary profiling of gene
expression at the transcriptome and proteome levels in
Saccharomyces cerevisiae. Mol. Cell. Proteomics 1,
323–333 (2002).
37. Washburn, M. P. et al. Protein pathway and complex
clustering of correlated mRNA and protein expression
analyses in Saccharomyces cerevisiae. Proc. Natl Acad.
Sci. USA 100, 3107–3112 (2003).
38. Patterson, S. D. & Aebersold, R. H. Proteomics: the first
decade and beyond. Nature Genet. 33 (Suppl), 311–323
(2003).
39. Ghaemmaghami, S. et al. Global analysis of protein
expression in yeast. Nature 425, 737–741 (2003).
40. Birrell, G. W. et al. Transcriptional response of
Saccharomyces cerevisiae to DNA-damaging agents does
not identify the genes that protect against these agents.
Proc. Natl Acad. Sci. USA 99, 8778–8783 (2002).
41. Steinmetz, L. M. et al. Systematic screen for human disease
genes in yeast. Nature Genet. 31, 400–404 (2002).
42. Deutschbauer, A. M., Williams, R. M., Chu, A. M. &
Davis, R. W. Parallel phenotypic analysis of sporulation and
postgermination growth in Saccharomyces cerevisiae. Proc.
Natl Acad. Sci. USA 99, 15530–15535 (2002).
43. Steinmetz, L. M. & Davis, R. W. High-density arrays and
insights into genome function. Biotechnol. Genet. Eng. Rev.
17, 109–146 (2000).
44. Yang, Y. H. & Speed, T. Design issues for cDNA microarray
experiments. Nature Rev. Genet. 3, 579–588 (2002).
45. Van Eerdewegh, P. et al. Association of the ADAM33 gene
with asthma and bronchial hyperresponsiveness. Nature
418, 426–430. (2002).
46. Sklar, P. et al. Association analysis of NOTCH4 loci in
schizophrenia using family and population-based controls.
Nature Genet. 28, 126–128. (2001).
47. Botstein, D. & Risch, N. Discovering genotypes underlying
human phenotypes: past successes for mendelian disease,
future approaches for complex disease. Nature Genet. 33
(Suppl), 228–237 (2003).
48. Hardenbol, P. et al. Multiplexed genotyping with sequence-
tagged molecular inversion probes. Nature Biotechnol. 21,
673–678 (2003).
49. Faham, M. & Cox, D. R. A novel in vivo method to detect
DNA sequence variation. Genome Res. 5, 474–482 (1995).
50. Faham, M., Baharloo, S., Tomitaka, S., DeYoung, J. &
Freimer, N. B. Mismatch repair detection (MRD): high-
throughput scanning for DNA variations. Hum. Mol. Genet.
10, 1657–1664 (2001).
51. Perez-Iratxeta, C., Bork, P. & Andrade, M. A. Association of
genes to genetically inherited diseases using data mining.
Nature Genet. 31, 316–319 (2002).
52. Mootha, V. K. et al. Identification of a gene causing human
cytochrome c oxidase deficiency by integrative genomics.
Proc. Natl Acad. Sci. USA 100, 605–610 (2003).
NATURE REVIEWS | GENETICS
53. Wayne, M. L. & McIntyre, L. M. Combining mapping and
arraying: an approach to candidate gene identification. Proc.
Natl Acad. Sci. USA 99, 14903–14906 (2002).
54. Aitman, T. J. et al. Identification of Cd36 (Fat) as an insulin-
resistance gene causing defective fatty acid and glucose
metabolism in hypertensive rats. Nature Genet. 21, 76–83
(1999).
55. Belli, G., Gari, E., Aldea, M. & Herrero, E. Functional analysis
of yeast essential genes using a promoter-substitution
cassette and the tetracycline-regulatable dual expression
system. Yeast 14, 1127–1138 (1998).
56. Kanemaki, M., Sanchez-Diaz, A., Gambus, A. & Labib, K.
Functional proteomic identification of DNA replication
proteins by induced proteolysis in vivo. Nature 423,
720–725 (2003).
57. Raamsdonk, L. M. et al. A functional genomics strategy that
uses metabolome data to reveal the phenotype of silent
mutations. Nature Biotechnol. 19, 45–50 (2001).
58. Allen, J. et al. High-throughput classification of yeast
mutants for functional genomics using metabolic
footprinting. Nature Biotechnol. 21, 692–696 (2003).
59. Tong, A. H. et al. Systematic genetic analysis with ordered
arrays of yeast deletion mutants. Science 294, 2364–2368
(2001).
60. Ooi, S. L., Shoemaker, D. D. & Boeke, J. D. DNA helicase
gene interaction network defined using synthetic lethality
analyzed by microarray. Nature Genet. 35, 277–286 (2003).
61. Hensel, M. et al. Simultaneous identification of bacterial
virulence genes by negative selection. Science 269,
400–403 (1995).
62. Karlyshev, A. V. et al. Application of high-density array-based
signature-tagged mutagenesis to discover novel Yersinia
virulence-associated genes. Infect. Immun. 69, 7810–7819
(2001).
63. Heller, R. A. et al. Discovery and analysis of inflammatory
disease-related genes using cDNA microarrays. Proc. Natl
Acad. Sci. USA 94, 2150–2155 (1997).
One of the first applications of gene-expression
profiling for disease sample classification.
64. Perou, C. M. et al. Distinctive gene expression patterns in
human mammary epithelial cells and breast cancers. Proc.
Natl Acad. Sci. USA 96, 9212–9217 (1999).
65. Golub, T. R. et al. Molecular classification of cancer: class
discovery and class prediction by gene expression
monitoring. Science 286, 531–537 (1999).
66. Bohen, S. P. et al. Variation in gene expression patterns in
follicular lymphoma and the response to rituximab. Proc.
Natl Acad. Sci. USA 100, 1926–1930 (2003).
67. Eisen, M. B., Spellman, P. T., Brown, P. O. & Botstein, D.
Cluster analysis and display of genome-wide expression
patterns. Proc. Natl Acad. Sci. USA 95, 14863–14868 (1998).
68. Tavazoie, S., Hughes, J. D., Campbell, M. J., Cho, R. J. &
Church, G. M. Systematic determination of genetic network
architecture. Nature Genet. 22, 281–285 (1999).
69. Wu, L. F. et al. Large-scale prediction of Saccharomyces
cerevisiae gene function using overlapping transcriptional
clusters. Nature Genet. 31, 255–265 (2002).
70. Stuart, J. M., Segal, E., Koller, D. & Kim, S. K. A gene-co-
expression network for global discovery of conserved
genetic modules. Science 302, 249–255 (2003).
71. Glazier, A. M., Nadeau, J. H. & Aitman, T. J. Finding genes
that underlie complex traits. Science 298, 2345–2349 (2002).
72. Steinmetz, L. M. et al. Dissecting the architecture of a
quantitative trait locus in yeast. Nature 416, 326–330 (2002).
The first report of the dissection of a complex trait
from a description of the phenotype to identification
of the genes published in a single study. Provides
evidence for complex QTL architecture and describes
a new functional assay.
73. Darvasi, A. & Pisante-Shalom, A. Complexities in the genetic
dissection of quantitative trait loci. Trends Genet. 18, 489–491
(2002).
74. Christians, J. K. & Keightley, P. D. Genetic architecture:
dissecting the genetic basis of phenotypic variation. Curr.
Biol. 12, R415–416. (2002).
75. Mackay, T. F. Quantitative trait loci in Drosophila. Nature Rev.
Genet. 2, 11–20 (2001).
76. Freimer, N. & Sabatti, C. The human phenome project.
Nature Genet. 34, 15–21 (2003).
77. Spradling, A. C. et al. The Berkeley Drosophila Genome
Project gene disruption project: single P-element insertions
mutating 25% of vital Drosophila genes. Genetics 153,
135–177 (1999).
78. Peter, A. et al. Mapping and identification of essential gene
functions on the X chromosome of Drosophila. EMBO Rep.
3, 34–38 (2002).
79. Zambrowicz, B. P. et al. Disruption and sequence
identification of 2,000 genes in mouse embryonic stem cells.
Nature 392, 608–611 (1998).
80. Martin, E. et al. Identification of 1,088 new transposon
insertions of Caenorhabditis elegans: a pilot study toward
large-scale screens. Genetics 162, 521–524 (2002).
81. Gura, T. A silence that speaks volumes. Nature 404,
804–808 (2000).
REVIEWS
82. Clemens, J. C. et al. Use of double-stranded RNA
interference in Drosophila cell lines to dissect signal
transduction pathways. Proc. Natl Acad. Sci. USA 97,
6499–6503 (2000).
83. Esposito, M. S. & Esposito, R. E. The genetic control of
sporulation in Saccharomyces. I. The isolation of
temperature-sensitive sporulation-deficient mutants.
Genetics 61, 79–89 (1969).
84. Giaever, G. et al. Genomic profiling of drug sensitivities via
induced haploinsufficiency. Nature Genet. 21, 278–283
(1999).
85. Lum, P. Y. et al. Discovering modes of action for therapeutic
compounds using a genome-wide screen of yeast
heterozygotes. Cell 116, 121–137 (2004).
86. Hirsh, A. E. & Fraser, H. B. Protein dispensability and rate of
evolution. Nature 411, 1046–1049 (2001).
87. Fraser, H. B., Hirsh, A. E., Steinmetz, L. M., Scharfe, C. &
Feldman, M. W. Evolutionary rate in the protein interaction
network. Science 296, 750–752 (2002).
88. Gu, Z. et al. Role of duplicate genes in genetic robustness
against null mutations. Nature 421, 63–66 (2003).
An example of the use of fitness data in addressing
fundamental questions in molecular evolution.
Incorporates published data sets without new bench
experiments.
89. Papp, B., Pal, C. & Hurst, L. D. Dosage sensitivity and the
evolution of gene families in yeast. Nature 424, 194–197
(2003).
90. Fodor, S. P. et al. Light-directed, spatially addressable
parallel chemical synthesis. Science 251, 767–773 (1991).
The first high-density microarray made by direct
synthesis.
91. Blanchard, A. P., Kaiser, R. J. & Hood, L. E. Synthetic DNA
arrays. Biosens. Bioelectron. 11, 687–690 (1996).
92. Schena, M., Shalon, D., Davis, R. W. & Brown, P. O.
Quantitative monitoring of gene expression patterns with a
complementary DNA microarray. Science 270, 467–470
(1995).
The first cDNA microarray for gene-expression
profiling made by printing.
93. Ferguson, J. A., Boles, T. C., Adams, C. P. & Walt, D. R.
A fiber-optic DNA biosensor microarray for the analysis of
gene expression. Nature Biotechnol. 14, 1681–1684
(1996).
94. Khrapko, K. R. et al. Hybridization of DNA with
oligonucleotides immobilized in a gel: a convenient method
for recording single base replacements. Mol. Biol. (Mosk)
25, 718–730 (1991).
95. Halgren, R. G., Fielden, M. R., Fong, C. J. &
Zacharewski, T. R. Assessment of clone identity and
sequence fidelity for 1189 IMAGE cDNA clones. Nucleic
Acids Res. 29, 582–588 (2001).
96. Knight, J. When the chips are down. Nature 410, 860–861
(2001).
97. Modrek, B. & Lee, C. A genomic view of alternative splicing.
Nature Genet. 30, 13–19 (2002).
98. Johnson, J. M. et al. Genome-wide survey of human
alternative pre-mRNA splicing with exon junction
microarrays. Science 302, 2141–2144 (2003).
99. Patil, N. et al. Blocks of limited haplotype diversity revealed
by high-resolution scanning of human chromosome 21.
Science 294, 1719–1723 (2001).
100. Winzeler, E. A. et al. Direct allelic variation scanning of the
yeast genome. Science 281, 1194–1197 (1998).
101. Kwok, P. Y. SNP genotyping with fluorescence polarization
detection. Hum. Mutat. 19, 315–323 (2002).
102. Jurinke, C., van den Boom, D., Cantor, C. R. & Koster, H.
Automated genotyping using the DNA MassArray
technology. Methods Mol. Biol. 187, 179–192 (2002).
Acknowledgements
We would like to thank L. David and T. Neklesa for helpful com-
ments on the manuscript.
Competing interests statement
The authors declare that they have no competing financial interests.
Online links
DATABASES
The following terms in this article are linked online to:
Entrez: http://www.ncbi.nlm.nih.gov/Entrez
LRPPRC | tetR
OMIM: http://www.ncbi.nlm.nih.gov/Omim
Leigh syndrome
FURTHER INFORMATION
Arabidopsis Mutations: http://www.arabidopsis.org/abrc
Human Gene Mutation Database:
http://archive.uwcm.ac.uk/uwcm/mg/hgmd0.html
Yeast Deletion Database: http://yeastdeletion.stanford.edu
Access to this interactive links box is free online.
VOLUME 5 | MARCH 2004 | 2 0 1