This document discusses using gene expression data from the Gene Expression Omnibus to identify relationships between genes and phenotypic concepts like diseases, environmental exposures, and experimental conditions. It describes extracting concepts from sample annotations using the Unified Medical Language System and relating differential gene expression to these concepts. This establishes a network of relationships between genes, phenotypes, and environmental contexts. Identifying such phenome-genome and envirome-genome relationships could help discover new disease-associated genes.

1. A
Data
Perspective
on
Autonomy,
Human
Rights,
and
the
End
of
Normality
Isaac
S.
Kohane,
MD,
PhD

2. JAMA

3. We found gene-concept relations for other phenotypic concepts,
including diseases. The only gene relating to leukemia is DDX24; mean
normalizedexpressionof DDX24dropsfrom0.71inthesixdatasetsmea-
suring this gene not associated with leukemia to 0.44 in the four data sets
with the annotation (SupplementaryFig.1conline,P = 0.007,Q = 0.01).
Interestingly, this gene was first cloned from a leukemia cDNA library24.
increase in mean normalized expression levels in the four data sets asso-
ciated with injury compared with over 100 other data sets measuring
these genes without this annotation (Supplementary Fig. 1d,e online,
GPX3 increases from 0.56 to 0.97; MAPK14 increases from 0.52 to 0.89;
for both, P < 1 × 10−15 and Q < 0.0002). Both genes have been shown
to be related to injury of various forms. Increased expression of plasma
Slc2a1(2287)
Slc2a1 (20525)
Slc2a1 (24778)
Pnliprp1(21253
Pnliprp1 (18946)Pnliprp1 (84028)
Hmgcs2 (15360)
Cox5b (12859)(94194)
9195)
TNNI1(2462)
TNNI1 (7135)
Tnni1 (21952)
Tnni1 (29388)
PGAM
MYH6(20098)
MYH6 (4624)
Myh6 (29556)
MYBPC1(1846)
MYBPC1 (4604)
Mybpc1 (362867)
C0026845:Muscle
C0596981:Muscle Cells
C0242695:Muscle, Skeletal
C0521324:Skeletal
0
0.2
0.4
0.6
0.8
1 420
390
478
261
477
266
400
330
476
493
260
386
486
479
340
87
363
358
221
470
357
8
365
12
181
286
177
88
85
449
145
367
76
213
214
461
215
157
472
287
198
156
Meannormalizedexpressionlevelof
Hs.PDLIM3
Meannormalizedexpressionlevelof
MM./Pdlim3
GDS annotated with one or more of four muscle concepts
GDS annotated with none of four muscle concepts
0
0.2
0.4
0.6
0.8
1
419
491
255
458
364
273
463
63
466
253
332
468
165
174
64
366
62
399
276
489
238
254
278
256
GDS annotated with one or more of four muscle concepts
GDS annotated with none of four muscle concepts
c
ba
d
Figure 3 Network of relations between 46 biomedical concepts extracted from the annotations of data sets in Gene Expression Omnibus and 444
genes with differential expression associated with the presence or absence of the concept. (a) Light blue nodes are UMLS concepts. Pink nodes are
genes with higher expression levels in data sets annotated with their related concept; light green nodes are genes with lower expression levels in
annotated data sets. Pink and green nodes are contained within gray squares indicating ortholog families. Edges (dashed) between an ortholog family
©2006NaturePublishingGrouphttp://www.nature.com/naturebiotechnology
NATURE BIOTECHNOLOGY VOLUME 24 NUMBER 1 JANUARY 2006 55
Creation and implications of a
phenome-genome network
Atul J Butte1 & Isaac S Kohane2
Although gene and protein measurements are increasing in
quantity and comprehensiveness, they do not characterize
a sample’s entire phenotype in an environmental or
experimental context. Here we comprehensively consider
associations between components of phenotype, genotype
and environment to identify genes that may govern
phenotype and responses to the environment. Context from
the annotations of gene expression data sets in the Gene
Expression Omnibus is represented using the Unified Medical
Language System, a compendium of biomedical vocabularies
with nearly 1-million concepts. After showing how data sets
can be clustered by annotative concepts, we find a network
of relations between phenotypic, disease, environmental
and experimental contexts as well as genes with differential
expression associated with these concepts. We identify novel
genes related to concepts such as aging. Comprehensively
identifying genes related to phenotype and environment is a
step toward the Human Phenome Project5.
In analyzing a cancer sample,such as one extracted from a lung tumor,a
plethora of factors,such as phenotype and clinical history (for example,
chief complaint of hemoptysis, family history or tumor size), environ-
mental exposures (for example,duration of exposure to asbestos or ciga-
rette smoke) and experimental conditions (for example, anesthesia or
sample preparation) have to be considered besides the more basic aspects
of its gene expression and proteomic pattern. Though these snapshots
of genomic and physiological states have been used to determine thera-
peutic action1, they cannot solely represent either the entire‘envirome,’
defined in an extended version of the initial definition by Anthony et
al. for the special case of mental disorders, as the totality of equivalent
environmental influences contributing to all disorders and organisms5,
or the ‘phenome’, the physical totality of all traits of an organism as
defined by Mahner and Kary5–7, of the sample and organism.
Relations between enviromic concepts and phenomic concepts have
been invaluable to medicine.For example,one such relation is the asso-
ciation of environmental exposure to cigarette smoke with the phe-
notype of lung cancer development. Comprehensively relating specific
concepts in the envirome and phenome to specific genes could thus lead
to the identification of new disease-associated genes5. Though some
phenomic data are available8, these are greatly overshadowed in size by
the >60,000 microarray measurements in repositories such as the Gene
Expression Omnibus (GEO)9. Even for microarray data stored using
standards like Minimum Information About a Microarray Experiment
(MIAME) and Microarray Gene Expression Markup Language (MAGE-
ML)10,11, contextual annotations are represented by unstructured nar-
rative text; determining the phenotype and environmental context is no
longer a tractable manual process.A question we have sought to answer
is whether prior investments in biomedical ontologies can provide lever-
age in finding phenome-genome and envirome-genome relations.
We show here that a large set of phenome-genome and envirome-
genome relations can be found within a public repository of transcrip-
tome measurements, if the phenotypes and environmental context can
be ascertained for each experiment,along with the expression measure-
ments.We accomplished this by creating a system that extracts contex-
tual concepts from the sample annotations in GEO, represents these
concepts using the Unified Medical Language System (UMLS), unifies
the gene expression measurements across data sets using NCBI Gene
identifiers and finally relates the gene expression measurements to the
contextual concepts (Fig.1).UMLS is the largest available compendium
of biomedical vocabularies, containing >130 biomedical vocabularies
with ~1-million interrelated concepts12. UMLS already unifies vocab-
ularies used in molecular biology and genomics, such as the Medical
Subject Headings (MeSH), NCBI Taxonomy and the Gene Ontology,
with medical vocabularies including the International Classification of
Diseases and SNOMED International9,13,14.
Establishing a phenome-genome network
After manual elimination of incorrectly assigned concepts (Methods
and Supplementary Note online), mappings to 4,127 UMLS concepts
remained (from 296,843 mappings to 5,115 strings). Concepts were
from 18 source vocabularies, with MeSH (23%), Read Codes (17%)
and SNOMED International (14%) contributing the most. The GEO
series description annotation was the most information rich, as it
provided unique concepts (Supplementary Table 1 online). This was
likely because GEO series descriptions are often dissimilar to each
other, compared to sample descriptions, which are often repeated. As
expected, the concepts mapping to the most annotations are cells and
RNA (Table 1).
Parsing failed on too short annotations, containing only laboratory
identifiers and few recognizable words, or too long for parsing to com-
plete. Regardless, over 99% of GEO samples were successfully directly
1Stanford Medical Informatics, Departments of Medicine and Pediatrics,
Stanford University School of Medicine, 251 Campus Drive, Room X-215,
Stanford, California 94305-5479 USA. 2Informatics Program and Division of
Endocrinology, Children’s Hospital Boston and Harvard Medical School, 300
Longwood Avenue, Boston, Massachusetts 02115 USA. Correspondence should
be addressed to A.J.B. (abutte@stanford.edu).
Published online 10 January 2006; doi:10.1038/nbt1150
ANALYSIS
©2006NaturePublishingGrouphttp://www.nature.com/naturebiotechnology
a.k.a.
The
unreasonable
effectiveness
of
gene
expression

4. Blood
Gene
Expression
Detection

5. Invited
to
HLS
Meeting
“I
think
when
Ari
[Ne’eman]
talks
about
autism
and
I
talk
about
autism,
we’re
talking
about
people
with
different
clusters
of
autism.
I
know
he
doesn’t
like
the
word
‘cure.’
If
my
daughter
could
function
the
way
Ari
could,
I
would
consider
her
cured,”
says
Singer.
“I
have
to
believe
my
daughter
doesn’t
want
to
be
spending
time
peeling
skin
off
her
arm.”

6. Patterns
across
tens
of
thousands
of
patients…
6
Preprocessing: (1) We grouped the 6905 distinct (non-procedure) ICD9 codes in the dataset int
802 PheWAS categories (dimensionality reduction). (2) We only considered PheWAS codes with
at least 5% prevalence and patients with less than 50 of any particular code in 6-month period.
This preprocessing step left us with 4927 individuals with 45 common category codes.
Clustering: For each patient, count the number of occurrences of each code in each 6-month
window from age 0 to age 15. We then applied standard hierarchical clustering with Euclidean
distance, Ward's linkage, and a minimum cluster size of 2% of the population.
Analysis: Significant elements of clusters were assessed by creating 15,000 permutations of
random cluster assignments and creating an empirical chi-squared statistic distribution for the
observed vs. expected number of code occurrences in each time window in each cluster.
Basic Cluster Characteristics
patients
code counts
0-6 months
code counts
6-12 months
code counts
12-18 months
patient
clustering

7. 0 5 10 15
020406080
Years
Cases
●
●
● ●
●
●
●
● ●
●
●
●
● ●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
PDD
CP
Epilepsy
Autism or
Autisms?

8. 0 5 10 15
020406080
Years
Cases
●
●
● ●
●
●
●
● ●
●
●
●
● ●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
PDD
CP
Epilepsy
Otitis m.
Specific DD
Viral/Chlam

9. 0 5 10 15
020406080
Years
Cases
●
●
● ●
●
●
●
● ●
●
●
●
● ●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
● ● ● ● ● ●
● ● ●
● ●
●
●
●
●
●
●
● ●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
PDD
CP
Epilepsy
Otitis m.
Specific DD
Viral/Chlam
PDD
Hyperkinetic
Anxiety

10. Genetics
and
Embryology
of
CAH

12. Dwarfism
and
GH-­‐Deficiency

13. GIANT
study
A further possible source of missing heritability is allelic heterogen-
eity: the presence of multiple, independent variants influencing a trait
at the same locus. We performed genome-wide conditional analyses in
a subset of stage 1 studies, including a total of 106,336individuals. Each
study repeated the primary GWA analysis butadditionallyadjusted for
SNPs representing the 180 loci associated at P , 5 3 1026
(Sup-
plementaryMethods).Wethenmeta-analysedthesestudiesinthesame
way as for the primary GWA study meta-analysis. Nineteen SNPs
within the 180 loci were associated with height at P , 3.33 1027
(a
Bonferroni-corrected significance threshold calculated from the ap-
proximately 15% of the genome covered by the conditioned 2 Mb loci;
Table 1, Fig. 2, Supplementary Methods and Supplementary Figs 1
and 3). The distances of the second signals to the lead SNPs suggested
that both are likely to be affecting the same gene, rather than being
coincidentally in close proximity. At 17 of 17 loci (excluding two
contiguous loci in the HMGA1 region), the second signal occurred
within 500 kilobases (kb), rather than between 500 kb and 1 Mb, of
this lead SNP (binomial test P 5 2 3 1025
). Further analyses of allelic
heterogeneity may identify additional variants that increase the pro-
portion of variance explained. For example, within the 180 2-Mb loci,
a total of 45 independent SNPs reached P , 1 3 1025
when we would
expect less than 2 by chance.
Although GWA studies have identified many variants robustly asso-
ciated with common human diseases and traits, the biological signifi-
cance ofthesevariants,andthegenes on which they act, isoften unclear.
We first tested the overlap between the 180 height-associated variants
and two types of putatively functional variants, non-synonymous (ns)
SNPs and cis-expression quantitative trait loci (cis-eQTLs, variants
strongly associated with expression of nearby genes). Height variants
were 2.4-fold more likely to overlap with cis-eQTLs in lymphocytes
than expected by chance (47 variants: P 5 4.73 10211
) (Supplemen-
taryTable 7) and 1.7-fold morelikelytobecloselycorrelated (r2
$ 0.8in
the HapMap CEU sample) with nsSNPs (24 variants, P 5 0.004) (Sup-
plementary Methods and Supplementary Table 8). Although the
presence of a correlated cis-eQTL or nsSNP at an individual locus
does not establish the causality of any particular variant, this enrich-
0.18
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
Proportionofvarianceexplained
Proportion of variance
explained by 180 SNPs
5.00
×
10–8
5.00
×
10–7
5.00
×
10–6
5.00
×
10–4
5.00
×
10–5
5.00
×
10–3
5.00
×
10–2
P-value threshold
FINGESTURE 0.08 ± 0.02
RS2 0.11 ± 0.01
RS3 0.11 ± 0.01
GOOD 0.09 ± 0.02
QIMR 0.11 ± 0.02
Average
Lower 95% confidence intervals
Upper 95% confidence intervals
a
b
15
10
5
0 0
79,000
134,000
235,000
487,000
0 100 200 300 400 500 600 700
Cumulativeexpectedvarianceexplained(%)
Samplesizerequired
Cumulative expected number of loci
Figure 1 | Phenotypic variance explained by common variants. a, Variance
explained is higher when SNPs not reaching genome-wide significance are
included in the prediction model. The y axis represents the proportion of
LETTER RESEARCH
100’s
of
genes
implicated..

14. Criteria
for
Treatment
• “Growth
hormone
deficiency
(GHD)”
• “Idiopathic
short
stature
(ISS),
defined
by
height
standard
deviation
score
≤-­‐2.25”
associated
with
growth
rates
unlikely
to
result
in
normal
adult
height,
in
whom
other
causes
of
short
stature
have
been
excluded
and
a
little
story
from
25
years
ago

16. HCM

17. expected
100 individuals
HCM
Prevalence
=
1:500
HCM
Inheritance
=
Autosomal
Dominant
NHLBI
ESP
6503
individuals
observed
NHLBI
ESP
6503
individuals

18. Genetics-­‐induced
Health
Disparities
18
• Hypertrophic
Cardiomyopathy
• Classical
Autosomal
dominant
• 1:500
prevalence
• Grounds
for
eliminating
athletes
from
teams
• Affecting
entire
families
ntial:
ForReview
Figure 1B
Confidential: Destroy when review is complete.
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

19. Survival 3 Years After a WBC Test
(White, Male, 50-­69 Years;; Using Last WBC Between 7/28/05 and 7/27/06)

20. But
over
most
of
medicine…
• Even
the
most
basic
of
autonomy,
taking
your
data
with
you,
is
not
the
status
quo.

22. What
does
data
tell
us
about
human
rights
and
autonomy
• There
is
no
“normal”
but
there
are
desired
outcomes.
• Utilities
are
not
shared
across
parents,
patients,
providers
and
payors.
• Autonomy
makes
the
data-­‐sharing
broader.
• Broader
data
sharing
highlights
distinct
utility
functions.
• Activist-­‐level
data
sharing
today
– Less
energy
required
with
#OpenData
• Much
to
be
done
in
getting
data
analyses
done
“right”
• In
healthcare:
Recognize
and
harness
patients
as
collaborators.

23. Thank
you