Locating Genes on Chromosomes

The Human Transcriptome Map: Clustering of Highly Expressed Genes in Chromosomal Domains Huib Caron,12
Barbera van Schaik,13 Merlijn van der Mee,3 Frank Baas,4 Gregory Riggins,6 Peter van Sluis,1 Marie-Christine
Hermus,1 Ronald van Asperen,1 Kathy Boon,1 P. A. Vote,2 Siem Heisterkamp,5 Antoine van Kampen,3 Rogier
Versteeg1Science 16 February 2001:Vol. 291. no. 5507, pp. 1289 1292 DOI: 10.1126/science.1056794 REPORTS
1

Department of Human Genetics, 2 Department of Pediatric Oncology, Emma Children's Hospital, Academic Medical
Center, University of Amsterdam, Post Office Box 22700, 1100 DE Amsterdam, Netherlands. 3 Bioinformatics
Laboratory, 4 Neurozintuigen Laboratory, 5 Department of Clinical Epidemiology and Biostatistics, Academic Medical
Center, University of Amsterdam, Amsterdam, Netherlands. 6 Department of Pathology and Department of Genetics,
Duke University Medical Center, Durham, NC 27710, USA.
The chromosomal position of human genes is rapidly being established. We integrated these mapping data with
genome-wide messenger RNA expression profiles as provided by SAGE (serial analysis of gene expression). Over
2.45 million SAGE transcript tags, including 160,000 tags of neuroblastomas, are presently known for 12 tissue types.
We developed algorithms to assign these tags to UniGene clusters and their chromosomal position. The resulting
Human Transcriptome Map generates gene expression profiles for any chromosomal region in 12 normal and pathologic
tissue types. The map reveals a clustering of highly expressed genes to specific chromosomal regions. It provides a
tool to search for genes that are overexpressed or silenced in cancer.
GeneMap'99 (1) gives the chromosomal position of 45,049 human expressed sequence tags (ESTs) and genes
belonging to 24,106 UniGene clusters. To obtain an expression profile of these genes, we made use of the SAGE
technology and databases. SAGE can quantitatively identify all transcripts expressed in a tissue or cell line (2). It is
based on the extraction of a 10-base pair (bp) tag from a fixed position in each transcript and the sequencing of
thousands of these tags. Software programs and databases support the identification of the mRNAs corresponding to
the tags in a SAGE library. However, this step is prone to errors, and tag assignment requires manual verification.
The National Center for Biotechnology Information (NCBI) SAGEmap database has electronically extracted tags
from mRNAs and ESTs in UniGene clusters. A manual check of 156 tags extracted from 30 UniGene clusters showed
that wrong tags mainly stemmed from sequence errors in ESTs and from errors in their 5' and 3' orientations. We
developed algorithms to select 3'-end clones of 713,489 ESTs assigned to UniGene clusters and identified their tags.
Sequence comparison algorithms discarded tags caused by sequence errors while preserving tags from alternative
transcripts or single nucleotide polymorphisms [see supplementary information for AMCtagmap details (3)]. We
identified reliable tags for 18,954 of the 24,106 UniGene clusters mapped on GeneMap'99. Manual analysis of
287 tags extracted from 86 UniGene clusters from intervals of chromosomes 1 and 22 showed an error rate of 6.2%
in our electronic tag identification algorithms. To check for errors in UniGene clustering, we verified tags on the
available sequenced P1-derived artificial chromosomes (PACs) of the mapped markers and annotated them accordingly
[see legend to Fig. 2 and
supplementary information
(3)].

Fig. 2. Extended interval

view of a chromosome 2p
region showing
neuroblastoma-specific
overexpression of the
neighboring genes N-myc
(UniGene Hs. 25960) and
DDX-1 (UniGene Hs.
78580). A small part of the
interval D2S287 to
D2S2375 is shown. The

left columns show the marker and centiray position as defined on GeneMap'99. The right side shows the UniGene
number, tag sequence, and the description of the UniGene cluster. Expression levels in the libraries are normalized
per 100,000 tags and shown by colored bars with a range from 0 to 15. Numbers give the tag counts per
100,000 tags. The tags are annotated by symbols. To identify tags produced by hybrid UniGene clusters, we analyzed
for each marker of GeneMap'99 the corresponding PAC sequenced in the Human Genome Project, as well as two
adjacent PACs. Tags that are present on these PACs are from ESTs belonging to the mapped marker and are marked
by P in a light green box. Tags not present on these PACs are probably derived from a contaminating EST not
belonging to the mapped marker and are marked by P in a red box [see Web site (4)]. This check is not yet available
for all markers. Tags belonging to more than one UniGene cluster are marked by 2/3 or >3 in a yellow box. The
expression levels of tags belonging to more than three clusters are not shown and are not used in the totals of the
concise interval maps and the whole chromosome maps. Tags from ESTs of opposite orientation in the UniGene
cluster are marked with AS in a purple box. [View Larger Version of this Image (20K GIF file)]

The Human Transcriptome Map [for Web site, see (4)] uses these tag assignments to relate 2.31 million tags in public
SAGE libraries (NCBI SAGEmap database) (5) and 160,000 tags in our neuroblastoma SAGE libraries to the UniGene
clusters mapped in GeneMap'99. The Human Transcriptome Map shows expression profiles for any chromosomal
region in 12 tissue types. SAGE libraries of a specific tissue were combined into tissue-specific libraries (e.g., normal
colon). We included tissues for which 100,000 or more tags were available, as most transcripts in a tissue are
represented in a library of this size (6). Five libraries represent normal tissues (colon epithelium, brain, mammary
gland, ovary, and prostate), and seven libraries represent tumor tissues (neuroblastoma, glioblastoma,
medulloblastoma, and carcinomas of colon, ovary, breast, and prostate). The Human Transcriptome Map has three
levels of resolution. The "whole chromosome view" shows gene expression per chromosome (Fig. 1). Each horizontal
blue or red bar represents the expression level of a UniGene cluster. UniGene clusters mapped by several markers
are shown only once, at the position of the highest reliability (1). The identity, map position, and precise expression of
the genes are shown in the "concise interval view." The highest resolution is given by the "extended interval view,"
where expression levels are shown for all individual tags of a gene (Fig. 2).

Fig. 1. Whole chromosome view of expression levels of the 1208 UniGene clusters mapped to chromosome 11 on the
GB4 radiation hybrid map of GeneMap'99. Each unit on the vertical axis represents one UniGene
cluster. UniGene clusters mapped by several markers are only shown once, at the position of the
highest lod score (the logarithm of the odds ratio for linkage). Only clusters for which we could
extract a tag with our algorithms are included. Expression is shown for SAGE libraries of 8 out of
the 12 available tissue types. Expression levels in the libraries are normalized per 100,000 tags.
Expression levels from 0 to 15 tags are shown by horizontal blue bars. Tag frequencies over 15 are
shown by red bars. The blue-only section to the right represents a moving median with a window size
of 39 UniGene clusters generated from the expression levels in "all tissues." Green bars indicate
RIDGEs. The boxed region shows the tissue-specific expression of a cluster of five
metalloproteinases and two apoptosis inhibitors in normal breast tissue and breast cancer tissue.
[View Larger Version of this Image (29K GIF file)]

The whole chromosome views reveal a higher order organization of the genome, as there is a strong clustering of
highly expressed genes. Chromosome 11 has several large regions of high gene expression, interspersed with regions
where gene expression is low (Fig. 1). This pattern is observed in all 12 tissues. An application of a moving median with
a window size of 39 genes to the chromosome 11 map even more clearly visualizes the expression differences (Fig. 1,
blue graph to the right). Most chromosomes show these clusters of highly expressed genes, which we call RIDGEs
(regions of increased gene expression) (Fig. 3). A quantitative definition of RIDGEs is not straightforward, as there

is a continuum from small to very large clusters. We analyzed whether RIDGEs can be explained by a random variation
in the distribution of highly expressed genes among the 18,954 genes of the Human Transcriptome Map. When
defined as regions in which 10 consecutive moving medians have a lower limit of four times the genomic median, we
identify 27 RIDGEs (green bars in Figs. 1 and 3). The probability of observing this number of RIDGEs under a random
permutation of the order of the 18,954 genes is very low [P = 10 12; see supplementary information (3)]. In addition,
Bayesian statistical modeling without prior cluster definition showed that a model of nonrandom distribution provided
the best fit with the observed clustering. These analyses show that RIDGEs most likely represent a higher order
structure in the genome.

Fig. 3. Regional expression profiles for 23 human chromosomes show a

clustering of highly expressed genes in RIDGEs. Expression levels are
shown as a moving median with a window size of 39 genes. There are
74 regions with one or more consecutive moving medians that have a lower
limit of four times the genomic median; 27 of them have a length of at
least 10 consecutive moving medians (indicated by green bars). [View Larger
Version of this Image (26K GIF file)]

Analysis of RIDGEs for physical characteristics suggests that many of them have a high gene density. Chromosome
18 is, on average, weakly expressed, and only 385 genes have been mapped to it on GeneMap'99. The equally large
chromosome 19 consists of a succession of RIDGEs and harbors 937 mapped genes (Fig. 3). Although many human
genes are still unmapped, the difference in gene density of chromosomes 18 and 19 is supported by CpG island density
analyses (7). The correlation between RIDGEs and gene density is even more suggestive for chromosomes 3 and
6 (Fig. 4). The RIDGE on chromosome 6 corresponds to the major histocompatibility complex (MHC) region. A
correlation between gene expression and density of mapped genes is found for 50 to 60% of the RIDGEs [Web fig.
1 (3)]. Typical RIDGEs count 6 to 30 mapped genes per centiray, compared to 1 to 2 mapped genes per centiray for
weakly transcribed regions. In RIDGEs, average expression levels per gene are up to seven times that of the genomic
average. This suggests that in RIDGEs, transcription per unit length of DNA is 20 to 200 times that in weakly
expressed regions. About 40 to 50% of the RIDGEs are not gene dense. These RIDGEs preferentially map to
telomeres, which is remarkable in light of the observed telomeric silencing in yeast (8, 9). Chromosomes 4, 13, 18, and
21 show an overall low gene expression and are devoid of RIDGEs (Fig. 3). The latter three chromosomes are
responsible for most constitutional trisomies, suggesting that the low expression and low gene density could limit the
lethality of an extra copy of them.

Fig. 4. Comparison of median gene expression levels and gene density for chromosomes 3 and 6. The left diagrams of
each chromosome show the expression levels as a moving median with a window size of 39 UniGene clusters. The right
diagram of each chromosome shows gene density. For each UniGene cluster, we calculated the average distance
between adjacent clusters in a window of 39 adjacent UniGene clusters. The inverse of this value is shown (inverse
centirays per gene). [View Larger Version of this Image (17K GIF file)]

The Human Transcriptome Map provides a tool to identify candidate genes that are overexpressed or silenced in
cancer tissue. Neuroblastomas frequently show amplification of the distal chromosome 2p region, which targets the
N-myc oncogene (10). Comparison of the whole chromosome views of chromosome 2p shows overexpression of two
adjacent genes in neuroblastoma SAGE libraries. The extended interval view identifies these genes as N-myc and the
often coamplified neighboring gene DDX-1 (Fig. 2). Therefore, global positional information of chromosomal defects is
sufficient to identify candidate oncogenes (11). Also, tumor-specific down-regulation can be detected. Examples are a
cluster of five matrix metalloproteinases on chromosome 11 [348 to 353 centirays (cR)] that are down-regulated in
breast cancer tissue (Fig. 1, box); the E-cadherin tumor suppressor gene on chromosome 16 (406 cR) that is downregulated in breast cancer tissue, as compared to normal breast tissue; and five carcinoembryonic antigen-related cell
adhesion molecule genes on chromosome 19 (238 to 244 cR) that are down-regulated in colon carcinoma tissue, as
compared to normal colon tissue (4).
Potential error sources in the Human Transcriptome Map are clustering errors in UniGene and the assignment of
wrong tags to UniGene clusters. Our algorithms assign ~6.2% erroneous tags to UniGene clusters. The influence of
these errors is probably attenuated. Assuming a total of 100,000 genes with 2 tags each, 200,000 tags would
represent all human genes. Because there are >1 million variants of a 10-bp tag sequence, ~80% of the erroneously
extracted tags will not match tags present in SAGE libraries and therefore will not influence overall expression
profiles. However, individual tags and expression levels of UniGene clusters may harbor errors and require
experimental confirmation. To test whether errors in UniGene clustering and mapping to GeneMap'99 may influence
our observation of RIDGEs, we constructed a sequence-based expression map for the annotated chromosome
21 sequence and for a 4.3-Mb annotated contig of the MHC region on chromosome 6 (12, 13). Also, these maps showed
that the MHC region is a pronounced RIDGE, whereas chromosome 21 is devoid of RIDGEs and has an overall weak
gene expression [see Web fig. 4 for maps (3)]. Therefore, the higher order structure of the genome observed with
the Human Transcriptome Map will largely be correct. The existence of RIDGEs is unanticipated, as a comparable
SAGE-based transcriptome map for yeast showed an even distribution over the genome of highly and weakly
expressed genes (8). Because the Human Transcriptome Map identifies different types of transcription domains, it
can now be analyzed as to how they relate to known nuclear substructures, such as nuclear speckles, PML bodies, and
coiled bodies (14-16). Definition of the position of tags to the full chromosomal sequences will further increase the
resolution of the transcriptome map. Incorporation of the growing number of SAGE libraries from different tissues
and various developmental stages will extend the overview of gene expression profiles in the human body.
REFERENCES AND NOTES
1.

P. Deloukas, et al., Science 282, 744 (1998) [Abstract/Free Full Text] .

2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.

V. E. Velculescu, L. Zhang, B. Vogelstein, K. W. Kinzler, Science 270, 484 (1995) [Abstract/Free Full Text] .
Supplemental Web material is available at www.sciencemag.org/cgi/content/full/291/5507/1289/DC1.
The Human Transcriptome Map is available at http://bioinfo.amc.uva.nl/HTM/.
A. Lal, et al., Cancer Res. 59, 5403 (1999) [Abstract/Free Full Text] .
V. E. Velculescu, et al., Nature Genet. 23, 387 (1999) [CrossRef] [ISI] [Medline] .
J. M. Craig and W. A. Bickmore, Nature Genet. 7, 376 (1994) [CrossRef] [ISI] [Medline] .
V. E. Velculescu, et al., Cell 88, 243 (1997) [CrossRef] [ISI] [Medline] .
D. E. Gottschling, O. M. Aparicio, B. L. Billington, V. A. Zakian, Cell 63, 751 (1990) [CrossRef] [ISI] [Medline]
M. Schwab, et al., Nature 305, 245 (1983) [CrossRef] [ISI] [Medline] .
N. Spieker et al., Genomics, in press.
The MHC Sequencing Consortium, Nature 401, 921 (1999) [CrossRef] [ISI] [Medline] .

13.
14.
15.
16.
17.

M. Hattori, et al., Nature 405, 311 (2000) [CrossRef] [ISI] [Medline] .

D. G. Wansink, et al., J. Cell Biol. 122, 283 (1993) [Abstract/Free Full Text] .
X. Wei, S. Somanathan, J. Samarabandu, R. Berezney, J. Cell Biol. 146, 543 (1999)
D. A. Jackson, F. J. Iborra, E. M. Manders, P. R. Cook, Mol. Biol. Cell 9, 1523 (1998)
We thank A. Luyf and A. Sha Sawari for their help in computational analyses, A. Lash for use of the
SAGEmap database and help in tag analyses, and E. Roos for expert digital imaging. Supported by grants from
the Stichting Kindergeneeskundig Kankeronderzoek, the A. Meelmeijer Fund, and the Dutch Cancer Society. H.C.
is a fellow of the Dutch Royal Academy of Sciences.

23 October 2000; accepted 11 January 2001 10.1126/science.1056794 Include this information when citing this
paper.

Clusters of Co-expressed Genes in Mammalian Genomes Are Conserved by Natural

Selection. Gregory A. C. Singer1, Andrew T. Lloyd2, Lukasz B. Huminiecki3 and Kenneth H. Wolfe
Department of Genetics, Smurfit Institute, University of Dublin, Trinity College, Dublin, Ireland Molecular Biology
and Evolution vol. 22 no. 3 Society for Molecular Biology and Evolution 2004; all rights reserved.
Research Article
E-mail: gacsinger@gmail.com
Abstract
Genes that belong to the same functional pathways are often packaged into operons in prokaryotes. However, aside
from examples in nematode genomes, this form of transcriptional regulation appears to be absent in eukaryotes.
Nevertheless, a number of recent studies have shown that gene order in eukaryotic genomes is not completely
random, and that genes with similar expression patterns tend to be clustered together. What remains unclear is
whether co-expressed genes have been gathered together by natural selection to facilitate their regulation, or if the
genes are co-expressed simply by virtue of their being close together in the genome. Here, we show that gene
expression clusters tend to contain fewer chromosomal breakpoints between human and mouse than expected by
chance, which indicates that they are being held together by natural selection. This conclusion applies to clusters
defined on the basis of broad (housekeeping) expression, or on the basis of correlated transcription profiles across
tissues. Contrary to previous reports, we find that genes with high expression are not clustered to a greater extent
than expected by chance and are not conserved during evolution.
Key Words: genome organization human genome mouse genome natural selection
Introduction
Prokaryotes use a simple yet elegant system to regulate the expression of their genes. Genes belonging to the same
functional pathways are often packaged into operons, which are transcribed into a single mRNA. Although this system
works well in the prokaryotic context, operons appear to be very rare in eukaryotes and have only been discovered in a
few organisms, most notably nematode worms (Zorio, et al. 1994; Blumenthal, 1998; Blumenthal et al. 2002) where it is
estimated that 15% of genes within Caenorhabditis elegans are contained in operons. However, the mechanisms
involved in the processing of polycistronic mRNAs are quite distinct in C. elegans compared to bacterial genomes, so it
is likely that nematode operons are independent innovations within their lineage (Blumenthal et al. 2002). Despite the
absence of operons, eukaryotes are still capable of a very fine level of control over gene transcription. However, this
is accomplished through the use of trans-acting factors that do not require the co-transcribed genes to be in close
proximity to each other (Niehrs and Pollet, 1999). Does this mean that the order of genes in the eukaryotic genome is
random? Certainly, if the positioning of genes within the genome is not important to transcriptional regulation then
the high rate of genome rearrangement events in eukaryotic genomes will lead to the complete randomization of gene
order in a short period of time (Huynen, Snel, and Bork, 2001). A number of studies, however, indicate that there is
some gene organization in eukaryotic genomes, and that cis-acting regulatory factors may play a larger role than
previously thought (Hurst, Pl, and Lercher, 2004).

In Saccharomyces cerevisiae, consecutive gene pairs in the genome show a higher level of co-expression than widely
separated genes (Cohen et al. 2000; Kruglyak and Tang 2000). These co-expressed genes cannot be operons, because
the two genes often occur on opposite strands of DNA, making polycistronic transcription impossible (Cohen et al.,
2000). The same co-expression of neighboring genes exists in C. elegans, which can largely be accounted for by
operons, but which is still present in gene pairs that are not part of the same operon (Lercher, Blumenthal, and Hurst,
2003a). Higher-order levels of gene organization have also been discovered. For example, muscle-specific genes in C.
elegans occur in blocks up to five genes in length (Roy, et al. 2002). In Drosophila melanogaster, even larger
structures of gene organization are present, with 20% of genes organized into clusters with similar expression
patterns ranging in size from 10 to 30 genes and up to 200 kb in length (Spellman and Rubin, 2002). In the mouse
genome, both housekeeping and immunogenic genes have been found in clusters (Williams and Hurst, 2002). Clusters
of housekeeping genes are also present in the human genome (Lercher, Urrutia, and Hurst 2002), in addition to
clusters of highly expressed genes (Caron et al. 2001; Versteeg et al. 2003) and muscle-specific genes (Bortoluzzi et
al. 1998). Thus, there is abundant evidence from a range of organisms that gene order in eukaryotic genomes is not
random. But the reasons for this non-random arrangement are still unclear.
The co-expression of closely spaced genes might be attributable to chromatin structure (Hurst, Pl, and Lercher
2004). For example, it is known that when chromatin is opened to facilitate gene transcription, the open region can
extend to neighboring genes (Stalder et al. 1980; Hebbes et al. 1994). Thus, the transcription of one gene could
influence the transcription of neighboring genes, even if such a relationship is unintended (Spellman and Rubin 2002).
Could natural selection be tolerating the co-expression of neighboring genes rather than actively promoting it? Two
alternative hypotheses can explain the co-expression of neighboring genes. On the one hand, neutralist hypothesis
might propose that the two genes are functionally unrelated but that cis-acting regulatory elements cause the
transcription of one gene to influence the transcription of its neighbor. A selectionist hypothesis, on the other hand,
might propose that co-regulation of these genes is required and that a chance rearrangement in the past brought
them together (and thus facilitated their co-expression), which proved advantageous enough for the new gene order
to reach fixation in the population. One way of evaluating these alternative hypotheses is to use means other than
expression data to define gene relationships. Lee and Sonnhammer (2003) have shown that genes involved in the same
biochemical pathways tend to be clustered together in a variety of genomes, including human. Because the genes in
these clusters were defined a priori as being co-regulated, the non-random grouping of these genes is hard to explain
under the neutral model. Another way of distinguishing selection from neutrality in the co-expression of gene
neighbors is to look for evidence of negative selection preserving the groups of genes over time. Indeed, this
approach has shown that co-expressed gene pairs in S. cerevisiae are twice as likely to be preserved in Candida
albicans as neighbors that are not co-expressed (Huynen, Snel, and Bork, 2001; Hurst, Williams, and Pl 2002),
providing evidence that the gene pairings are an adaptation and not chance events. However, no studies have yet
demonstrated that large blocks of co-expressed genes are preserved over the course of evolution.
Clusters of housekeeping genes are very prominent in both the mouse (Williams and Hurst 2002) and human genomes
(Lercher, Urrutia, and Hurst 2002), but the orthology of these clusters has not been shown, nor have any studies
measured the degree of preservation of these clusters relative to the rest of the genome. Here, we use microarray
expression data from the Gene Expression Atlas (Su et al. 2002) to identify clusters of co-expressed and broadly
expressed genes in the human and mouse genomes, confirming previous results based on expressed sequence tag
(EST) and serial analysis of gene expression (SAGE) expression data (Lercher, Urrutia, and Hurst 2002). We then
investigate whether human gene expression clusters remain chromosomal neighbors in mouse, and vice versa, and
demonstrate that the clusters have been conserved to a greater degree than expected by chance. This indicates that
natural selection is preserving the structure of these expression modules within each genome.
Methods
Expression Data
Gene expression data for mouse and human were taken from the Gene Expression Atlas (http://expression.gnf.org;
Su et al. (2002)), which contains Affymetrix chip expression data (U74A for mouse, U95A for human) for many
different tissues, 19 of which are common to both the mouse and human: adrenal gland, amygdala, cerebellum, cortex,
dorsal root ganglia, heart, kidney, liver, lung, ovary, placenta, prostate, salivary gland, spleen, testis, thymus, thyroid,

trachea, and uterus. Many of the expression experiments are replicated, and we took the mean expression for each
tissue among the replicates. We eliminated genes that did not reach an Affymetrix Average Difference (AD) value of
at least 200 in at least one tissue, and tissues for which the expression level was very low (AD values <100) were
dropped to zero.
Mapping
UniGene clusters corresponding to Affymetrix tags were extracted from the Affymetrix probe consensus sequence
file (http://www.affymetrix.com/support/technical/byproduct.affx?cat=arrays). In some cases, two or more
Affymetrix tags were targeted against the same UniGene cluster, and only the tag with the highest average
expression across all libraries from the human (or mouse) was retained.
UniGene clusters were mapped to the genome National Center for Biotechnology Information [NCBI] (build 31 and
NCBIM build 30 for human and mouse, respectively) using the LocusLink and Ensembl databases. First, UniGene to
LocusLink mapping was extracted from the UniGene release file Hs.data (human build U150) and Mm.data (mouse build
U160). Second, LocusLink to Ensembl gene id mapping was extracted from the Ensembl database (release 14.31.1)
using the EnsMart tool (http://www.ensembl.org/Multi/martview). LocusLink clusters mapping to multiple UniGene
clusters or multiple Ensembl genes were discarded to ensure that the resulting mapping was unique and nonredundant. This procedure resulted in 4,451 human Affymetrix tags, and 4,522 mouse tags being mapped to the same
number of unique locations on the human and mouse genomes. These sets of genes were used to infer the existence of
clusters of genes with similar expression patterns. However, in the text we report the total number of genes within
clusters, including those for which we have no expression data.
Removal of Duplicated Genes
Duplicated genes are expected to have similar expression patterns, and such genes are frequently located in physical
proximity to each other and could give rise to a trivial clustering effect of co-expressed genes. We therefore
removed all but one gene belonging to any gene family as determined by the TRIBE algorithm (Enright, Van Dongen,
and Ouzounis 2002) and located within 10 Mb on the same chromosome. TRIBE families were extracted from the
Ensembl database (release 14.31.1) using the EnsMart tool. After removal of duplicated genes, there were 4,114
human and 4,187 mouse Ensembl genes linked to Affymetrix tags.
Measures of Gene Expression Similarity
We used three measures of expression similarity in this study. First, for any two genes we measured the
"housekeepingness" of the pair by multiplying the proportion of tissues in which gene A is expressed (AD value 200)
by the proportion of tissues in which gene B is expressed. This measure has the advantage of being strongly skewed
to the right, only assuming a high score if both genes are broadly expressed. This measurement was used to identify
clusters of housekeeping genes. Second, we measured the height of expression for a pair of genes by taking the mean
of their AD values across the 19 tissues listed in the Expression Data section, above. Previous studies have also used
the median or maximum expression values (Caron et al. 2001; Versteeg et al. 2003), but these measures are all highly
correlated with each other, so the choice between them is arbitrary (Lercher et al. 2003b). The height measure was
used to search for clusters of highly expressed genes. Third, we used the Pearson correlation coefficient as a simple
measure of co-expression across the 19 tissues to search for clusters of co-expressed genes.
Sliding-Window Algorithm
To measure the clustering of gene expression patterns in the genome, we used a sliding-window analysis with a window
size of 10 genes and a step size of one gene, ignoring genes for which no expression data were available. We limited
the physical length of the windows by ignoring those that exceeded 0.5 Mb per gene in size. Within each window, the
similarity in expression pattern (using each of the methods in the previous subsection) was measured between
consecutive genes, and the scores for the consecutive pairs were then summed to form a score for the entire window.
This score was compared to scores from 100,000 windows containing genes randomly sampled from the genome.
Windows that had a similarity score exceeding that of 95% of the randomized windows were deemed significant. A
second set of windows, with a more stringent significance cut-off of 99% were also annotated for later comparison to

the more relaxed values. When the analyses were complete, overlapping significant windows were merged, forming
blocks of "clustered" and "unclustered" regions in the genome.
Measuring Cluster Conservation
We identified mouse and human orthologous gene pairs using the EnsMart utility (version 14.1). Within each clustered
(or unclustered) region of the genome, a chromosomal breakage "opportunity" (i.e., an intergenic region where an
interchromosomal rearrangement could potentially occur during evolution) was counted between every two consecutive
human genes whose mouse orthologs were known (any intervening human genes without known orthologs were ignored).
If the mouse orthologs on either side of the breakage opportunity were on different chromosomes, a break event
was counted. The number of breakage events relative to the number of opportunities was then compared between
clustered and unclustered parts of the genome. We also compared the amount of intergenic space per chromosomal
break within gene clusters to that in 1,000 randomized data sets, to control for the uneven spacing of genes within
Results
Pairs of Housekeeping Genes Are Maintained by Natural Selection
Previous reports based on SAGE and EST data have shown that housekeeping genes are clustered in the human
(Lercher, Urrutia, and Hurst 2002) and mouse genomes (Williams and Hurst 2002). Unfortunately, the correlation
between SAGE, EST, and microarray data tends to be very poor (Huminiecki, Lloyd, and Wolfe 2003), and it was
therefore not clear that these clusters would be detectable using microarray expression data. For this reason, we
began by using microarray data to confirm the existence of clusters of housekeeping genes in the human and mouse
genomes. We initially looked for the smallest possible gene cluster: a simple pair of consecutive genes that have
similar expression breadth across the 19 tissues. The number of physical gene pairs for which we have expression
data for both genes is small, but it is sufficient for statistical analysis. When compared to 100,000 random gene
pairs, it is clear that there is an excess of consecutive pairs with broad expression profiles in the human genome (fig.
1). The curve is shifted to the right in the consecutive gene pairs compared to the random pairs ( p = 1.75 x 105 in a
one-tailed Wilcoxon test), and there is a significant excess of gene pairs with an expression breadth exceeding 0.85
in the consecutive gene pairs (9.5% vs. 5.0% in the consecutive and random pairs, respectively; p = 9.5 x 105 in a onetailed Fisher's exact test). Our examination of the mouse genome found similar results, with consecutive gene pairs
having a higher similarity of expression breadth than the randomized gene pairs (p = 0.00367), and 6.9% of gene pairs
exceeding the 95% level of expression breadth in the random gene pairs (p = 0.029). These findings indicate a
significant level of organization of housekeeping genes in both the human and mouse genomes, at least at the level of
gene pairs.

FIG. 1. Distribution of expression breadth in random gene pairs (white bars) and
consecutive gene pairs (shaded bars) in the human genome. There is an excess of
broadly expressed (housekeeping) gene pairs in the consecutive pairs compared to
the random pairs (p < 0.0001).

Although both genomes appear to contain linked pairs of housekeeping genes, it is

possible that these pairings are an independent innovation within each lineage and
are not orthologous. Moreover, it is necessary to distinguish chance pairs that
exist in both the mouse and human genomes from those that have been maintained
by purifying natural selection over time. To determine whether the pairs of
housekeeping genes originated before the split of the mouse and human lineages
and were conserved over the course of evolution, we downloaded orthology data
from Ensembl to compare the proportion of conserved housekeeping gene pairs
versus the proportion of conserved non-housekeeping pairs in each genome. We found that a very high proportion
(29/30; 96.7%) of housekeeping pairs in human (defined as a breadth score of 0.85, corresponding to the upper
5% of random gene pairs) are also consecutive pairs in the mouse genome. This is significantly higher than the
conservation of all gene pairs in the genome (84% (253/301); p = 0.0175 in a one-tailed exact unconditional test

(Berger, 1996)). Similarly, 28/29 (96.6%) of housekeeping pairs in the mouse genome are within a one-gene distance
of each other in the human genome, which is higher than that among the other genes for which expression and
orthology data are known, where only 351/382 pairs (91.9%) are preserved (this difference is not significant,
however: p = 0.223). These findings are suggestive that the pairing of these housekeeping genes is an adaptation that
is maintained by purifying selection, and that they are not chance events.
Large Clusters of Housekeeping Genes Are also Maintained by Natural Selection
It is already known that clusters of housekeeping genes in the human genome extend beyond gene pairs (Lercher et
al. 2002, 2003b), and we therefore wondered whether these larger structures were conserved over time. We
performed a sliding-window analysis on both the human and mouse genomes to find clusters of housekeeping genes
(see Methods). This procedure identified 30 housekeeping clusters in the human genome, containing between 33 and
203 genes, with a median of 86 genes (fig. 2). To assess whether these clusters are the product of natural selection
or chance events, we examined the degree to which the arrangement of human genes in the clusters is conserved in
the mouse genome. Of 1,567 opportunities for chromosomal breakage events (see Methods), 7 breaks were measured
within the housekeeping clusters (0.44%). By comparison, 42 breakages of 4,303 opportunities were recorded outside
the clusters (0.98%). This difference is small but statistically significant (p = 0.031 in a one-tailed Fisher's exact
test). Our analysis of the mouse genome yielded similar results, where 30 clusters of housekeeping genes ranging in
size from 39 to 127 genes were found, with a median of 62 genes (see figure 2 in the Supplementary Material online).
As in the human genome analysis, we found a lower proportion of chromosomal breakpoints within the mouse
housekeeping gene clusters than elsewhere in the genome (0.52% vs. 1.1%; p=0.049). We note that our test for cluster
conservation only tests for interchromosomal breaks and does not test for conservation of local gene order within the
clusters. It also does not require that the orthologs of clustered genes in one species should themselves form a
cluster in the other species (the expression data are too sparse to allow for such stringency).

FIG. 2. Maps of clusters of housekeeping (red), highly expressed (green), and

co-expressed (blue) genes in the human genome. Tick marks indicate the locations
of genes for which we have expression data. A larger version of this figure, as
well as an equivalent diagram for mouse, is available in the Supplementary
Materials online (figs. 1 and 2).

Previous studies have shown that clusters of co-expressed genes tend to be more tightly spaced than unclustered
genes (Hurst, Williams, and Pl 2002), and that housekeeping genes in particular are short and closely spaced
(Eisenberg and Levanon 2003). Indeed, we find that intergenic distances within our housekeeping clusters are much
shorter than the average intergenic distance outside of clusters (the median intergenic distances within housekeeping
clusters are 8,615 bp in human and 8,284 bp in mouse, less than half the values for the whole genomes). Because
chromosomal breakage events are more likely between genes that are widely separated, the short distance between
neighboring genes in housekeeping clusters could bias the outcome of our cluster conservation analysis. Therefore, we
performed a randomization experiment in which gene locations were held constant but expression profiles for each
gene were assigned randomly (without replacement). In that process, 1,000 randomized genomes were created, and
the clustering analysis was performed on each one. It is clear that the number of genes involved in housekeeping
clusters in the real genomes exceeds the numbers found in the randomized genomes (fig. 3a and 3b). Moreover, the
real clusters are located in areas of the genome where interchromosomal rearrangements are relatively rare.
FIG. 3. Distribution of the number of genes involved in housekeeping clusters in 1,000
randomized genomes for (a) human and (b) mouse. The numbers of genes found in
housekeeping clusters in the real (non-randomized) genomes are indicated by arrows. Both
are significantly higher than the means for the randomized datasets ( p < 0.04).

Considering the case of the human genome first, figure 4a shows a plot of the number of
chromosomal breakages within housekeeping clusters versus the total intergenic spaces

within those clusters within the 1,000 randomized genomes. The point for the real human genome (triangle in fig. 4a)
clearly lies outside of the distribution of randomized points. With 18 chromosomal breakage events, the human
housekeeping clusters should have a total of 38 mb of intergenic space based on a linear regression of the
randomization data. However, they nearly double that size, at 68.3 mb, so these blocks are extraordinarly large
considering the small number of chromosomal breakage events within them (p = 0.001). The equivalent analysis on the
mouse genome reveals similar results (see figure 3a in the Supplementary Material online), where again the actual
intergenic space within housekeeping clusters is greater than expected, given the number of chromosomal breakage
events observed: with 11 chromosomal breaks, the clusters should contain 48.0 mb of intergenic space, but they
actually contain 60.8 mb. This difference is not significant (p = 0.078), because the magnitude of deviation from
expectation is less than observed in the human genome, although there is the suggestion of a tendency toward cluster
conservation in the mouse genome.
FIG. 4. Number of chromosomal break events versus the amount of intergenic space
within expression clusters for 1,000 randomized human genomes. The triangle indicates the
numbers for the nonrandomized human genome. Y-axis positions have been "jiggled" to
reduce overlap between points. The true clusters are larger than expected ( p < 0.001)
based on the number of chromosomal breakage events observed for genes clustered by (a)
expression breadth and (c) expression correlation, but clusters of highly expressed genes
(b) lie well within the randomized data.

Clusters of Highly Expressed Genes Are Not Maintained by Natural Selection

A number of studies have shown that highly expressed genes also form clusters within the
genome (Caron et al. 2001; Versteeg et al. 2003). However, it is unclear whether this is
simply an artifact of detecting clusters of housekeeping genes, because housekeeping genes
tend to have above-average expression strength (Eisenberg and Levanon 2003; Lercher,
Urrutia, and Hurst 2002). Indeed, in our data set there is a strong correlation between the
height and breadth of expression for neighboring genes in the human genome (Spearman's
= 0.71), as well as in the mouse ( = 0.69). Nevertheless, we re-ran the sliding-window
analysis, this time using expression height instead of breadth as the measure of expression
distance between neighboring genes. We identified 14 clusters in the human genome,
ranging in size from 29 to 176 genes. Of the genes in these clusters, 66% overlap with the
housekeeping clusters identified previously (fig. 2). In mouse, only 12 clusters were found,
with 24 to 131 genes in each cluster; 22% of the genes in these clusters are also found in
mouse housekeeping clusters. We repeated the genome randomization procedure on both
the mouse and human genomes, using expression height as our clustering attribute. In
neither genome did the actual number of genes in clusters exceed that expected by chance.
In human, the randomized genomes had 900 genes (standard deviation ( ) = 210) within highly expressed clusters on
average, whereas the actual genome had 1,104 genes, a difference that is not statistically significant (p = 0.17 in a
one-tailed Z-test). In the randomized mouse genomes, 1,210 genes ( = 226) were placed within highly expressed
clusters on average, whereas the actual genome had only 892 genes within highly expressed clusters (p = 0.32). Thus,
we failed to find evidence of significant clustering of highly expressed genes at all. When we analyzed the degree of
conservation of these clusters in both genomes using the methods described in the previous section, in neither case
was the degree of conservation of the highly expressed gene clusters greater than what we expected by chance (see
figure 4b for human and figure 3b in the Supplementary Material online for mouse).
Co-expressed Genes Are also Clustered in the Human and Mouse Genomes
Another measure of gene expression similarity used for detecting gene expression clusters is the simple Pearson
correlation coefficient (Spellman and Rubin 2002; Stuart et al. 2003). An analysis of EST data has shown that linked
genes in mouse tend to be co-expressed (Williams and Hurst 2002), although that study concluded that the effect
was very weak. Our own analyses based on microarray data identified large clusters of co-expressed genes in both
mouse and human, and the number of genes involved in co-expression clusters greatly exceeds the numbers found in
randomized genomes. In the human genome, our clustering algorithm placed 3,046 genes inside co-expression clusters,

whereas the mean number in 1,000 randomized human genomes was only 1,281 ( = 291). This difference is highly
significant (
in a one-tailed Z-test). In mouse there were 2,455 genes in co-expression clusters,
compared to an average of only 1,704 genes ( = 329) in randomized genomes, which is also a significant excess (p =
0.011).
As with the housekeeping clusters, we compared the number of chromosomal breaks within the actual clusters
identified in mouse and human to the number of chromosomal breaks within the clusters identified in randomized
genomes. In both genomes, the number of breakage events is less than expected, given the amount of intergenic
space within the clusters (see figure 4c for human and figure 3c in the Supplementary Material online for mouse).
Thus, there is strong evidence that these co-expression clusters are being maintained by purifying selection.
Overlap Between Co-expression Clusters and Housekeeping Clusters
It is necessary to test whether co-expression and housekeeping clusters are independent of each other, or if there
is significant overlap between the two. Visual inspection of figure 2 shows that, at least in the human genome, there
does appear to be a significant overlap between the two measures. In fact, housekeeping and co-expression clusters
share 37% of their genes in human and 20% in mouse. To see what effect these overlaps had on our previous analyses,
we re-analyzed the housekeeping and co-expression clusters, this time removing genes that appeared in both clusters.
In human, 1,656 genes are found within housekeeping clusters but not co-expression clusters, a number that is
significantly higher than we observed when the same analysis was applied to randomized genomes (944, = 229; p =
0.00094). We also observed a higher degree of conservation of these clusters than we observed among the clusters
from the randomized genomes (fig. 5a; p = 0.012). We observed the same trends in mouse, where again there was an
excess of genes within housekeeping clusters (but outside co-expression clusters) compared to the numbers of genes
found in the randomized mouse genomes (1,553 vs. 1,288, = 259). However, this result was not significant (p = 0.15).
Moreover, these mouse clusters cannot be said to be conserved to a significantly higher degree than the randomized
clusters, although again, the trend points in that direction (p = 0.10; see figure 4a of the Supplementary Material
online).
FIG. 5. Number of chromosomal break events versus the amount of intergenic
space within expression clusters for 1,000 randomized human genomes. Symbols
and interpretation are the same as in figure 4, although in this case we have
measured the degree of conservation of housekeeping gene clusters that do not
overlap with co-expression clusters (a), and vice versa (b). In both cases, the true
clusters are larger than expected (p = 0.012 and p = 0.0080, respectively).

We next did the reverse analysis, by re-analyzing the clusters of co-expressed

genes after removing genes that also appeared within housekeeping clusters.
Again, a significantly higher number of genes remained within co-expressed
clusters than expected by chance (p=0.00082 for human and p=0.043 for mouse).
The remaining clusters were found to be significantly conserved relative to
randomizations in human (fig. 5b; p=0.0080), and mouse (see figure 4b of the
Supplementary Material online; p=0.048).
Together, these two analyses show that housekeeping genes and co-expressed
genes are independently clustered in the human and mouse genomes, and that
these clusters are maintained by negative selection.
These Observations Are Robust to More Stringent Definitions of Gene
Expression Clusters
It is worth noting that we used a very liberal significance cut-off when defining
our gene expression clusters: a window of genes is considered significant if their similarity of expression (in terms of
breadth, height, or correlation) exceeds the 95th percentile of 100,000 randomized gene windows (see the Methods).
Because of the large number of windows scanned in the genome, this procedure will report many false positives. For

this reason, we have repeated all of our analyses, only this time accepting gene windows whose measure of expression
similarity exceeded the 99th percentile of randomized windows. As expected, the number of genes involved in
expression clusters dropped significantly: only 966 genes are involved in housekeeping gene clusters in human, versus
the 2,754 found using the less stringent cluster definition. Similarly, in mouse only 618 genes are found within
housekeeping gene clusters using the strict cluster definitions, versus the 2,007 found when the relaxed definition
was used. Nevertheless, these numbers still exceed the number of genes found in clusters in 1,000 randomized
genomes, where the mean number of genes found within housekeeping gene clusters is 268 in human and 342 in mouse.
Analogous results are found in the expression height clusters and co-expression clusters, where the results from the
strict cluster definitions exactly mirror those from the loose definition, although the absolute number of genes is
lower in all cases (table 1).

Table 1 Number of genes in clusters in real and randomized genomes, calculated using a
stringent 99th-percentile definition of clusters
Organism

Human

Mouse

Cluster type

Houskeeping
Highly expressed
Co-expressed
Housekeeping
Highly expressed
Co-expressed

No. of genes in clusters

No. in bootstrapped genomes S.D.

Significance

966
74
879
618
268
926

268 137
197 107
305 144
342 146
234 99
411 156

0.0010
0.88
0.0010
0.046
0.39
0.0030

We also analyzed the degree of conservation of these stringently defined clusters and found that again, the patterns
mirrored those of the liberally defined clusters: housekeeping gene clusters are maintained to a greater degree than
clusters found within randomized genomes (p = 0.011 and p = 0.036 in human and mouse, respectively). This is also true
of clusters of co-expressed genes (p = 0.050 and p = 0.0020 in human and mouse, respectively), and as expected,
there is no evidence for conservation of clusters of highly expressed genes in either genome (p = 0.96 and p = 0.55 in
human and mouse, respectively). Thus, the results are qualitatively identical regardless of how stringent we are in
defining similarly expressed gene clusters: there is strong evidence for housekeeping gene and co-expressed gene
clusters in both the mouse and human genomes, but not for highly expressed gene clusters.
Discussion
We have found clusters of housekeeping genes in the human genome by using microarray expression data, which
confirms previous findings based on EST and SAGE data (Lercher, Urrutia, and Hurst 2002). Interestingly, there are
conspicuously few clusters on the X chromosomes in both mouse and human. We cannot rule out the possibility that
this an artifactual result of the sparse data we have analyzed (the density of available data for chromosome X is
lower than average in both organisms), but it is tempting to think that this is a real phenomenon, perhaps related to
the decreased recombination on sex chromosomes which might result in a decreased opportunity for cluster
formation. Although housekeeping gene clusters were previously known to exist in the mouse genome (Williams and
Hurst 2002), the orthology of the clusters in the two genomes had not been demonstrated. We have shown that
housekeeping clusters in the human genome tend not to be broken up in the mouse genome (and vice versa) relative to
other groups of genes, lending support to the hypothesis that these clusters are advantageous and are therefore
being preserved by purifying selection. It is notable that we found lower statistical support in all our analyses of the
mouse genome than in the human genome. However, given the spotty nature of our data set (covering only about 20%
of the genes in each genome), it is possible that this trend is a sampling artifact.
Our analyses suggest that reports of the clustering of highly expressed genes in the human genome (Caron et al.
2001; Versteeg et al. 2003) may, in fact, be indirectly detecting some of the clusters of housekeeping genes because
there is a high degree of correlation between the two measures (see figure 2 and figure 2 of the Supplementary

material online). This agrees with previous findings (Lercher, Urrutia, and Hurst 2002), but we have also
demonstrated that, unlike clusters of housekeeping or co-expressed genes, clusters defined by expression height are
not conserved to a greater degree than expected by chance, and are therefore probably not maintained by natural
selection.
We have presented results indicating that co-expressed genes are clustered in both the mouse and human genomes,
and that these clusters may be conserved by natural selection. Interestingly, previous reports have found that the
clustering of co-expressed genes is a weak effect in the mouse (Williams and Hurst 2002) and human (Lercher,
Urrutia, and Hurst 2002) genomes. Whether this disagreement is due to the expression data used (EST and SAGE
data in other studies versus microarray data here), the genes analyzed (no study of this sort has sampled more than
about 20% of known mouse and human genes), or the method used to define gene co-expression is unclear.
Our results show that purifying selection is preserving clusters of both housekeeping genes and co-expressed genes
in the human genome, and that the same forces may be at work in the mouse genome. In most of our analyses, gene
cluster conservation was observed to be weaker in mouse than in human. Although the trends in mouse were identical
to those in human, they often did not reach statistical significance. Unfortunately, we cannot rule out the possibility
that this is an artifact of the data we have analyzed. However, another possibility is that this is a real biological
phenomenon. Given that the mouse genome has undergone a large amount of gene rearrangement (Mullins and Mullins
2004), perhaps gene clusters are being eroded over time in mouse, while at the same time they are being actively
preserved in human. Although this scenario seems unlikely, it is consistent with our results, and further study will be
needed to rule it out.
What is the advantage of gene clustering in the first place? Presumably, the close proximity of the genes is an
adaptation that facilitates the co-regulation of their transcription. Eisenberg and Levanon (2003) reported that the
Gene Ontology annotations (Ashburner et al. 2000) for human housekeeping genes show a high proportion of
metabolism-related and RNA-interacting proteins (such as ribosomal proteins). These types of genes play a
fundamental role in the operation of every eukaryotic cell, and thus, if there is any benefit to arranging co-expressed
genes together in the genome, housekeeping genes will likely be subject to the strongest selection coefficients to
form such clusters. Moreover, housekeeping genes tend not only to be broadly expressed but also highly expressed,
which is a pattern that probably requires little regulation in comparison to genes with very specific expression
patterns. Perhaps housekeeping genes are more amenable to being controlled by broadly acting cis-regulatory
elements than other genes, or perhaps they are subject to repression of transcription through chromatin
modification in particular circumstances.
In conclusion, we have shown that there appears to be a selective benefit to the clustering of co-expressed and
broadly expressed genes in the human and mouse genomes. We believe this is the strongest evidence to date that the
non-random arrangement of genes in mammalian genomes is the product of natural selection.
Acknowledgements
This study was supported by a Science Foundation Ireland grant to K.H.W. We thank the two referees of this paper
for their constructive comments.
Footnotes
1

Present address: Human Cancer Genetics Program, The Ohio State University, Columbus, OH.
Present address: University College Dublin, Dublin 4, Ireland.
3
Present address: Center for Genomics and Bioinformatics, Karolinska Institutet Campus, Berzelius vg 35, SE-171
77 Stockholm, Sweden.
William Martin, Associate Editor
2

References
Ashburner, M., C. A. Ball, J. A. Blake et al. 2000. Gene ontology: tool for the unification of biology. The Gene
Ontology Consortium. Nat. Genet. 25:2529.[CrossRef][ISI][Medline]

Berger, R. L. 1996. More powerful tests from confidence interval p values. Am. Statisti. 50:314318.[CrossRef][ISI]
Blumenthal, T. 1998. Gene clusters and polycistronic transcription in eukaryotes. Bioessays 20:480487.[CrossRef]
[ISI][Medline]
Blumenthal, T., D. Evans, C. D. Link, et al. 2002. A global analysis of Caenorhabditis elegans operons. Nature 417:851
854.[CrossRef][ISI][Medline]
Bortoluzzi, S., L. Rampoldi, B. Simionati, et al. 1998. A comprehensive, high-resolution genomic transcript map of
human skeletal muscle. Genome Res. 8:817825.[Abstract/Free Full Text]
Caron, H., B. van Schaik, M. van der Mee, et al. 2001. The human transcriptome map: clustering of highly expressed
genes in chromosomal domains. Science 291:12891292.[Abstract/Free Full Text]
Cohen, B. A., R. D. Mitra, J. D. Hughes, and G. M. Church. 2000. A computational analysis of whole-genome expression
data reveals chromosomal domains of gene expression. Nat. Genet. 26:183186.[CrossRef][ISI][Medline]
Eisenberg, E., and E. Y. Levanon. 2003. Human housekeeping genes are compact. Trends Genet. 19:362365.[CrossRef]
[ISI][Medline]
Enright, A. J., S. Van Dongen, and C. A. Ouzounis. 2002. An efficient algorithm for large-scale detection of protein
families. Nucleic Acids Res. 30:15751584.[Abstract/Free Full Text]
Hebbes, T. R., A. L. Clayton, A. W. Thorne, and C. Crane-Robinson. 1994. Core histone hyper-acetylation co-maps with
generalized DNase I sensitivity in the chicken beta-globin chromosomal domain. EMBO J. 13:18231830.[ISI]
[Medline]
Huminiecki, L., A. T. Lloyd, and K. H. Wolfe. 2003. Congruence of tissue expression profiles from Gene Expression
Atlas, SAGEmap and TissueInfo databases. BMC. Genomics 4:31.[CrossRef][Medline]
Hurst, L. D., C. Pl, and M. J. Lercher. 2004. The evolutionary dynamics of eukaryotic gene order. Nat. Rev. Genet.
5:299310.[CrossRef][ISI][Medline]
Hurst, L. D., E. J. Williams, and C. Pl. 2002. Natural selection promotes the conservation of linkage of co-expressed
genes. Trends Genet. 18:604606.[CrossRef][ISI][Medline]
Huynen, M. A., B. Snel, and P. Bork. 2001. Inversions and the dynamics of eukaryotic gene order. Trends Genet.
17:304306.[CrossRef][ISI][Medline]
Kruglyak, S., and H. Tang. 2000. Regulation of adjacent yeast genes. Trends Genet. 16:109111.[CrossRef][ISI]
[Medline]
Lee, J. M., and E. L. Sonnhammer. 2003. Genomic gene clustering analysis of pathways in eukaryotes. Genome Res.
13:875882.[Abstract/Free Full Text]
Lercher, M. J., T. Blumenthal, and L. D. Hurst. 2003 a. Coexpression of neighboring genes in Caenorhabditis elegans is
mostly due to operons and duplicate genes. Genome Res. 13:238243.[Abstract/Free Full Text]
Lercher, M. J., A. O. Urrutia, and L. D. Hurst. 2002. Clustering of housekeeping genes provides a unified model of
gene order in the human genome. Nat. Genet. 31:180183.[CrossRef][ISI][Medline]
Lercher, M. J., A. O. Urrutia, A. Pavlcek, and L. D. Hurst. 2003 b. A unification of mosaic structures in the human
genome. Hum. Mol. Genet. 12:24112415.[Abstract/Free Full Text]
Mullins, L. J., and J. J. Mullins. 2004. Insights from the rat genome sequence. Genome Biol. 5:221.[CrossRef][Medline]
Niehrs, C., and N. Pollet. 1999. Synexpression groups in eukaryotes. Nature 402:483487.[CrossRef][ISI][Medline]
Roy, P. J., J. M. Stuart, J. Lund, and S. K. Kim. 2002. Chromosomal clustering of muscle-expressed genes in
Caenorhabditis elegans. Nature 418:975979.[CrossRef][ISI][Medline]
Spellman, P. T., and G. M. Rubin. 2002. Evidence for large domains of similarly expressed genes in the Drosophila
genome. J. Biol. 1:5.[CrossRef][Medline]
Stalder, J., M. Groudine, J. B. Dodgson, J. D. Engel, and H. Weintraub. 1980. Hb switching in chickens. Cell 19:973
980.[CrossRef][ISI][Medline]
Stuart, J. M., E. Segal, D. Koller, and S. K. Kim. 2003. A gene-coexpression network for global discovery of conserved
genetic modules. Science 302:249255.[Abstract/Free Full Text]
Su, A. I., M. P. Cooke, K. A. Ching, et al. 2002. Large-scale analysis of the human and mouse transcriptomes. Proc.
Natl. Acad. Sci. USA 99:44654470.[Abstract/Free Full Text]
Versteeg, R., B. D. Van Schaik, M. F. Van Batenburg, M. Roos, R. Monajemi, H. Caron, H. J. Bussemaker, and A. H. Van
Kampen. 2003. The Human Transcriptome Map reveals extremes in gene density, intron length, GC content, and
repeat pattern for domains of highly and weakly expressed genes. Genome Res. 13:19982004.
[Abstract/Free Full Text]

Williams, E. J., and L. D. Hurst. 2002. Clustering of tissue-specific genes underlies much of the similarity in rates of
protein evolution of linked genes. J. Mol. Evol. 54:511518.[CrossRef][ISI][Medline]
Zorio, D. A., N. N. Cheng, T. Blumenthal, and J. Spieth. 1994. Operons as a common form of chromosomal organization
in C. elegans. Nature 372:270272.[CrossRef][ISI][Medline]
Accepted for publication November 25, 2004.