A systems survey unveils mechanisms of survival and death in community acquired sepsis

sugarannoyedΠολεοδομικά Έργα

16 Νοε 2013 (πριν από 3 χρόνια και 8 μήνες)

115 εμφανίσεις

1


A systems survey

unveils mechanisms of
survival and death in
community acquired sepsis


Raymond J. Langley
1
, Jennifer C. van Velkinburgh
1

[Co
-
first author]
,

Seth W. Glickman
4
, Ephraim L. Tsalik
2
,
Brandon J. Rice
1
, Arturo Suarez
3
, Robert P. Mohney
8
,
Debra

H. Freeman
2
, Christine Oien
2,5
, Mu Wang
9
,
Jacob
Wulff
8
,
Steph
anie

Reisinger
10
, Brian T. Edmonds
6
, Brian Grinnell
6
, David R. Nelson
6
, Alita A. Miller
7
, Ronald P.
Gladue
7
,
Darrell L. Dinwiddie
1
,
Lawrence Carin
11

, Bo Chen
11
, Chunping Wang
11
,
Geoffrey S.
Ginsburg
2
,
Charles
B. Cairns
4
, Ronny M. Otero
3
, Ralph C. Corey
2
, Vance G. Fowler Jr
2
, Emanuel P. Rivers
3
, Christopher W. Woods
2,5
,
Stephen F. Kingsmore
1


1
National Center for Genome Resources, S
anta Fe, NM
;

2
Duke

Institute for Genome Sciences and

Policy
,
and

Department of Medicine, Duke University School of
Medicine, Durh
am, NC
;

3
Departments

of Emergency Medicine, Henry Ford Ho
spital, Detroit MI
;

4
Department of Emergency Medicine, University of North Carolina School of Medi
cine, Chapel Hill, NC
;

5
Depart
ment of Medicine, Durham Veterans Administration Medi
cal Center, Durham, NC;

6
Eli Lilly and Company,

Indianapolis, IN;

7
Pf
izer Inc., Groton, CT;

8
Metabolon Inc., Durham, NC;

9
Monarch Life

Sciences, Indianapolis, IN;

10
Prosanos Inc., PA;

11
Department
of Electrical & Computer Engineering
, Duke University, Durham, NC.




2


A
n assumption

of systems
surveys

is that holism unveils molecular mechanisms in complex traits that are
undisclosed by reductionist approaches
1
. Infections
induce

over 10 million emergency department (ED) visits in
the United States annually, while sepsis (infection
with

systemic inflammatory response

syndrome
, SIRS)
causes
~
750,000 deaths per year and is
the tenth leading cause of death
2
-
9
.
Despite

decades of study, the
molecular events presaging

sepsis death
or

survival are

unclear. We undertook mass spectrometry (MS)
-
based
proteome

and
metabolome

analysis of
venous
plasma from over 200 well characterized
individuals

with
community
-
acquired sepsis

or controls (
with SIRS but un
infected
)
in discovery and validation studies
upon

ED
arrival

and 24 hours later
10
-
13
. Sepsis survivors

differ
ed

from non
-
infected, SIRS
-
positive patients by

altered
plasma metabolites consistent with
mobilization of diverse
energetic substrates and elevated

aerobic
metabolism. Sepsis survivors
and deaths differed

by antithetical changes

from controls

that were present at

arrival
and increas
ed in magnitude after 24 hours
.
S
epsis

patients who would die

had
elevation of many

met
abolic intermediates in

plasma and
lacked

fever
, suggesting

aborted catabolism of energetic substrates
.
Plas
ma metabolites and proteins

were

influenced minimally by
progression

to severe sepsis
or septic shock,
etiologic ag
ent or therapy, suggesting
a
dichotomy in

host
res
ponse

to bacterial

infection is predominant in

determination of

patient outcome
.
Global c
orr
elations of plasma protein and metabolite
values recapitulated
known enzyme

substrate,

product and

allosteric
regulator

associations, uncove
red additional putative
network perturbations in sepsis, and enabled

substantive

imputation of
putative, novel regulation of
primary

metabolism
. Predictive biomarker
models

were
developed

that

ha
d sufficient accuracy for utility

as
objective
adjuncts for
diagnosis, prognostic
differentiation and

individualized

treatment of sepsis
14
-
16
. Systems surveys
appear useful

for predictive differentiation
with
in
a complex pathologic state

and
nomination of potential,
novel interventions
.


3


1,152 individuals
who
arrived at

three urban, tertiary
-
care

EDs

between

2005
and

2009 with suspected,
community
-
acquired sepsis (acute infection and
>
2

SIRS criteria
17
) w
ere enrolled in the Community Acquired
Pneumonia and Sepsis Outcome Diagnostics (CAPSOD) study
[
NCT00258869
]
.
Medical history, p
hysical
examination

and
plasma and mRNA samples
were
obtained at enrollment (t
0
) and 24 hours later (t
24
). Patient
outcomes were

assessed

for 28 days

and

medical

records were

independently

adjudicated
18
. 150 enrollees

were selected for
MS
-
based
, venous

plasma metabolom
e and proteome

profiling at t
0
and

t
24
. They comprised

five

groups
:
day 28 sepsis survivors with uncomplicated courses (uncomplicated sepsis,
n=27)
, sepsis survivors
who

developed severe sepsis (n=25) or septic shock by day 3 (n=38)
,
d
ay 28 sepsis

deaths (n=31),
and non
-
infecte
d, SIRS
-
positive

patients [
enrolled

with presumed

sepsis but

not corroborated and

discharged wi
th a
different diagnosis; n=29]

(Supplementary Figure 1, Supplementary Table 1).
G
roups were matched for race,
sex, enrollment site, renal function and co
-
morbidity but differed

i
n

traits such as fever and

acute illness score

(
Acute Physiology a
nd Chro
nic Health Evaluation Score II,
APACHE II,
Supplementary Tables 2 and 3).
Sepsis
patients had definite infections with pathogenic bacteria. Sepsis groups
were
enriched for
Streptococcus
pneumoniae

(n=31),
E
scherichia

coli

(n=16) and
Staph
ylococcus

aureus

(n=27)
, to enable comparisons,

but

matched for

proportion of
etiologic agents and sites of infection
.

Fif
ty two additional sepsis survivors and
deaths

were
selected

and matched
for
replication

t
0

and

t
24

studies
(
Supplementary Table 4).

Of approximately 4,413 biochemicals

detectable in human biofluids
19
, 439

were measured
in
venous
plasma

from the 150
discovery
patients
using

label
-
free
,

combined
liquid and gas chromatography and
MS

at
t
0

or

t
24
, of which

332 were measured at both time
s
.
A total of
215

and 224

biochemicals,

at
t
0

and
t
24
,
respectively,

were annotated

human

metabolites

(Supplementary Figure 2).
Signal intensities

were normalized
to the
median for each
MS batch to minimize analytical
variation. Residual

median relative standard deviation
of repeated measurements of biochemical standards was 10%.
C
linical

assays of
venous
serum
creatinine,
capillary

lactate
and
serum
glucose (n = 149, 115, 149, respectively) correlated
reasonably
well
w
ith

log
2
-
transformed
MS
signal intensities

in plasma at
t
0

(r
2

=
0.90, 0.58, 0.56, respectively; Supplementary Figure 3),
indicating MS
-
measurements to be

semi
-
quantitative.
Z
-
score plots

showed right skewed
metabolite
4


distributions

at
t
0
, with increased skewing in severe sepsis and
sepsis death

(Supplementary Figure 4),
indicative of
more pronounced
biochemical variance
in these groups
.

P
rincipal c
omponent

analysis
,

Bayesian
factor analysis
20

and normalized energy plots

showed

renal function,
liver disease

and
sepsis
group
membership

to
largely
define
,
in descending order
,

variation

in the

plasma

metabolome
(
Supplementary
Fig
ure
s
5

and 6)
.

The

magnitude of
sepsis group effects
,

but not renal function or liver disease
,

was greater

at
t
24

than at t
0

(
Supplementary Fig.
6

-

8
)
, Supplementary Table 5),

indicating
metabolic perturbations
associated
with
sepsis
to
increase

after
seeking
medical attention
. Since the median time to death was

10.7 days,
increased effect size with time

was not due to

imminent death

and likely reflects enrollment at an early stage
in sepsis progression
.

Differences between sepsis groups were
sought

by analysis of variance (ANOVA) and
β process factor
analysis and group probit regression (BPFA
-
GPR)
20
. N
on
-
hypothesis
-
related

effects were

minimized
,

in the
former
,

by inclusion of
renal function and liver disease

as fixed effects
and, in the latter,
by
separation of renal
and sepsis group effects
.

In
survivor/
death

comparisons

this
was

too
conservative

since
acute
renal
dys
function
partly co
-
segregated with sepsis death

(
Supplementary Table 6
).

There were n
o

significant

metabolic

differences

among sepsis

survivor groups (
uncomplicated
sepsis
, severe sepsis and
septic
shock
)
, nor
unique to
sepsis due to
S
. pneumoniae
,
S
. aureus

and
E. coli

(
Supplementary Fig. 9 and 10
)
. These data suggest

sepsis survivors to represent a metabolic continuum at presentation, irrespective of imminent clinical course

or
etiology
.

Sepsis survi
vors differed from
non
-
infected
controls
in plasma levels of 49

and 42
metabolites
at
t
0

and
t
24
, respecti
vely (stratifi
ed ANOVA, FDR 5%).

Sixty

of 63 metabolites detected at both times had

concordant
changes,
suggesting

a
persistent

metabolic response to

sepsis (Supplementary Figure
s 8 and

11
;
Supplementary Tables
7 and 8
)
.
Sepsis survivors exhibited s
ignifica
nt

decreases

in
many

carnitine and
glycero
lphosphocholine

esters, amino acid
s and their

catabolites,

complex sugars, citrate, malate
, glycerol
,

glycerate

and phosphate
, relative to controls

(ANOVA and
BPFA
-
GPR
; Figure 1, Supplementary Figure 12,
Supplementary Tables 8

and 9).

S
everal free fatty acids

and steroids

were increased
. Lactate and

ketones

were
5


unchanged
.


In addition, sepsis survivors exhibited an average elevation in temperature of 1.3
°
C. Cumulatively,
these alterations att
est
unequivocally
to
elevated mobil
ization and aerobic metabolism

of lipid, protein and
sugar sub
strates in sepsis survivors
17, 21
-
25
.

Sepsis survivors and
day 28
sepsis
deaths diffe
red in levels of 76

and 128 plasma
metabolites
at
t
0

and
t
24
, respectively
(stratified ANOVA, FDR 5%; Figure 1; Supplementary Figures 8, 12 and 13
;
Supplementary
Tables 7
-
9
)
. Fifty five of these

were significant and 84
were concordant

at
t
0

and
t
24
,
indicating

sustained
metabolic
differentiation of

sepsis survivors and
future
deaths. Strikingly, p
atients who

would die exhibited

antithetical
metabolome

changes

from
survivors

that

were presen
t upon seeking medical care

and

burgeoned

with time:

Thus,
carnitine esters, amino acid catabolites, modified nucleosides, citrate, malate
,

pyruvate,

and
fatty acids were significantly elevated

in sepsis
patients who would

die

both by ANOVA and
BPFA
-
GPR
.

Also
elevated was l
actate
, an established

severity
marker in

sepsis.

Interestingly, anabolic steroids were
decreased

in sepsis deaths
, suggesting their antecedence

to

some changes in intermediary metabolism
.
A
cyl
-
GPCs

were
an exception to the trend of reversed direction of change in sepsis death
,
showing

additional decreases in
sepsis deaths
,
in agreement with previous studies
26, 27
.

Replication

of
findings

was sought
by identical

profiling of plasma from

52 additional
,
sepsis survivors
and deaths

at t
0

and

t
24
.

The replication set

had a longer median time to death than the discovery set (18.5
days versus 10.7 days, respectively) and, consequently, e
xhibited s
maller effect sizes and longitudinal increases
in differences between

sepsis

outcomes

(
Supplementary Fig.
9). As a result,
fe
wer significant differences

were
detected

between sepsis survivors and deaths (18

at
t
0

and

20 at
t
24
; Figure 1; Supplementary Fig. 11;
Supplementary Tables 7,8,10). However,
10

of these

were concordant and significant at both time points
and

all
of the
major
metabolic perturbations
found in

sepsis deaths

in the discovery set

were replicated (
elevated
carnitine esters, amino acid catabolites, modified nucleosides, citrate, malat
e and fatty acids,

de
creased

anabolic steroids and GPC esters
)
.

Together with
absence of fever in sepsis deaths, these changes

signify

a
remarkable dichotomy of host metabolic respo
nse to sepsis long before death:

Patients with s
epsis
who will
6


die show

similar
mobilization of lipid, protein and sugar energetic substrates

to those who survive, but
fail to
fully catabolize

these aerobic
ally
, in stark contrast to sepsis survival
.


Heterologous s
upport

for association of

sepsis outcome with
energetic homeostasis was sought by
MS
-
based
proteome

analysis of the same

samples.
Of

approximately 9,504 proteins detectable in plasma
28
,
4,536

were measured

using a linear ion
-
trap MS coupled with HPLC
,

of which 924 were common to

150

t
0

and

t
24

samples
(Supplementary Figure 14).

P
eptides were quantified by

log tr
ansformed,

quantile normalized,

area
s

under the curve

of ion chromatograms after background noise removal
29
.
Twenty
-
nine percent

o
f
peptide
s were identified with

high confidence

and

17%

were
support
ed by multiple
sequences

(Supplementary
Table 11
)
.
Despite a high

median coefficient of variation of peptide measurements
(
38.8
%
), c
linical

assays of
serum C reactive protein

and albumin
correlated

reasonably

well with log
-
transformed MS values in plasma
(
Supplementary Figure 1
5
).
Plasma
proteome

m
Score
s (
average
s

of the
absolute values of Z
-
scores
) at
t
0

showed

an identical
group
progression to that
of

plas
ma metabolites

(Supplementary Figure 1
6
)
.

A
lso a
kin to
metabolomics

were the major determinants of variation in the plasma proteome (
liver disease,
immunosuppressant

use

or neoplasia, and sepsis group membership
, in descending order
)

and
an increase in
sepsis group effects

from

t
0

to

t
24

(
Supplementary Figure 17
)
.

Unlike
metabo
lomics
, however, plasma protein

variation

was un
influenced by renal function

and a considerable proportion
was
inexplicable by clinical,
laboratory or analytic
parameters
.

Differences between sepsis groups were sought by ANOVA, with inclusion of non
-
hypothesis
effects as
fixed effects.

Akin to metabolites, there were few significant protein differences among sepsis survivor groups
(uncomplicated sepsis, severe sepsis and septic shock), nor differences unique to sepsis due to
S. pneumoniae
,
S. aureus

and
E. coli

(Supplementary Fig. 18 and 19),
reinforcing the suggestion that

sepsis survivors represent
a molecular

continuum
, irrespec
tive of day three sepsis severity or

bacterial etiology.

In contrast to metabolites,
alterations in
plasma protein
s

initia
lly appeared
mundane and failed to
immediately
disclose novel, unifying mechanisms

of sepsis response
.
Ninety
-
one

and 276 plasma proteins
differed
between

sepsis survivors and

non
-
infected controls

at
t
0

and
t
24
, respecti
vely (stratified ANOVA, FDR
7


5%; Supplementary Tables 11 and 12).
Plasma protein changes were not stochastic, however, as evidenced by

concordance of
80%

of 133 proteins
that
were

detected at both times and
significant
ly different

at one or
more

(F
igure 2; Supplementary Figures 20 and

21
; Supplementary Table 12
)
.

As expected,
acute phase
reactants

/

inflammatory response markers

were elevated in sepsis

(c
omplement C8α,
C
-
reactive protein
,
interleukin 4
,
lipopolysaccharide binding protein
,

prothymosin α,

α1
-
antitrypsin (
SERPINA3
)
,
serum

amyloid

A
1, A2

and A3
,

leukocy
te immunoglobulin
-
like receptor A
3

and
tumor necrosis

f
actor receptor 12A
. CRP and
LBP
changes
are well documented and
have been used
as

adjunct
s

in
sepsis diagnosis
24, 25
.
Potential
ly

expla
ining
some
metabolome

perturbations in sepsis

were increased mitochondrial a
cyl
-
coenzyme A
synthetase 6

and
glyceraldehyde
-
3
-
phosphate dehydrogenase

and decreased glycogen phosphorylase and
phosphorylase B kinase
, as discussed below
.

Considerably greater than differences

between sepsis survivors and non
-
infected controls

were

plasma
proteome differences differen
tiating

s
epsis survivors and deaths

(
626 and 458

significant

plasma
protein
changes

at t
0

and t
24
, respectively
;
F
igure 2; Supplementary Figures 20 and 21
; Supp
lementary Table 12). Forty
seven
proteins

were significant
and concordant
at both time points
.
Furthermore, 67%

of 373 proteins were
significantly different at one time point
,
detected at the second

and

concordant
. Striking
in death
were
increased

complement
and decreased

thrombolytic proteins (
Figure 2; Supplemental Figure 22
), providing a
potent
ia
l molecular mechanism for
mitigation of

sepsis mortality

by

administration of
activated protein C
17
.
Consistent with
lipolysis and impaired fatty acid β
-
oxidation

in sepsis death were
altered

levels of
fatty acid
carrier proteins
(
including
group
-
specific component

(GC)
, serum amyloid
A
4
, apolipoprotein
s

A

-
I
,
-
II,
-
IV, L1,
C
-
IV, transthyretin and sterol carrier protein X
)

and lipid modification proteins
. Decreased GC has previously
shown association

with
death

in sepsis
30
.
M
ost elevated
in sepsis death

was

plasma

3
-
hydroxy
-
3
-
methyl
-
glutaryl
-
CoA reductase

(HMGCR,
EC 1.1.1.88
),
the rate
-
controlling enzyme of the mevalonate pathway
, which
helps

explain the

striking
changes

observed

in steroid levels.

Interestingly,
HMGCR

inhibitors

have
shown

association
with

sepsis

survival

31
.

8


In order to integrate the r
esults of t
he
plasma proteome and
metabolome

surveys
,

332
metabolites

and

924 proteins

that were

measured

at both
at t
0

and t
24

were correlated
in 150 patients
.

Gratifyingly, many
related analytes
were juxtaposed by u
nsupervised h
ierarchical clustering
of correlations

(Supp. Figure 23
)
. For
example, seven

acyl
-
c
arnitines were nearest neighbors
,

as were five

androgenic steroids
,

nine

acyl
-
GPCs

and
acyl
-
GPEs, 19

fatty acids

and 12 amino acids or their metabolites
.
Likewise, many
functionally or structurally
related proteins co
-
clustered
, such as
acute phase reactants LPS
-
binding protein, C
-
reactive protein and SAA1
and
SAA3
;
regenerating islet
-
derived 1α

and 3
α
;

complement c
omponent 1q A and B chains
; and

S100 calcium
binding protein
s

A8

and A9.

These data

suggest that
, as expected, many closely related analytes co
-
cluster

on
the basis of
similar

global correlations with plasma proteins or metabolites.

40,026 of ~2 milli
on
correlations were replicated and/or significant

(
concordant with
p
-
value

<
0.1

at
one time point and
<0.05

at the other
,
or Bonferroni correction

of
<10
7.03
, respectively
; Supplementary Table
13
)
.

Many high
-
confidence

correlations

re
capitulated known enzyme

substrate,

product,


allosteric
regulator

or biochemical

complex
-
component

associations
:

steroidogenic acute regulatory protein
,

for
example,

which regulates

steroid synthesis by enhancing
cholesterol

conversion to pregnenolone
,
correlated

negative
ly

with substrates cholesterol and 7α
-
hyroxy
-
3
-
oxo
-
4
-
cholesterol and positive
ly

with products
androsterone
-
1
-
sulfat
e and epiandrosterone
-
1 sulfate
;
ribonuclease A1

correlated closely (coefficients

of
>
0.7
)

with products
N6
-
carbamoylthreo
nyladenosine, pseudo
uridine, arabitol and N2,
N2
,

dimethylguanosine
;
hemoglobin subunits α, β, γ
, δ, ζ and
ε
correlated
with
component
heme
,

and
an
allosteric effector
,

adenosine
-
5
-
monophosphate
;
subunit D of the
succinate dehydrogenase

complex

(SDHD)

correlated
with

product

malate,
and downstream products
oxaloacetate

and citrate
.

Several

examples of

novel regulation of primary
metabolism

were suggested
:

SDHD
, for example,

correlated with pyruvate, lactate and acetyl
-
carnitine,
suggesting
regulation

of the

citric acid cycle

to
be more complex than previously appreciated
.
Particularly
compelling

were

clusters of
correlation
s of related analytes

that were
substantiated by
established

interactions and

included

novel
correlations with

proteins or metabolites, providing
putative
new

insights
into
9


pathway regulation
.

For example, five bile acid substrates
correlate
d

with glucosidase β (GBA2),

which
catalyzes the hydrolysis of bile acid 3
-
O
-
glucosides
.

Correlations of all plasma proteins
and metabolites

we
re
anticipat
ed
to provide insights into

human

biochemistry rather than
seps
is, since sepsis was utilized merely

as

a

means of

perturb
ation of

biochemical

homeostasis.

However,
it was hypothesized that
additional

insights into the pathophysiology of sepsis
outcomes would be garnered by
limiting correlations
to
the 70
annotated
proteins and 44 metabolites

that
exhibited

significant, concordant differences in either sepsis diagnosis or sepsis outcomes at both time p
oints.

Hierarchical clustering of
these
c
orrelations
revealed remarkable constancy

at t
0

and t
24

(
Figure 3
)
.
S
everal
correlations lent additional support
for

a dichotomic
host
metabolic
response to sepsis that differ
entiates

survivors and deaths.
For example, nine acyl
-
carnitines (a primary source of fatty acids in the mitochondrion)
correlated with
acyl
-
CoA synthet
ase
,

medium
-
chain family member 6

(ACSM6,

which
attaches fatty acids

(FA)

to Coenzyme A (CoASH)

in preparation for β oxidation),
lysoph
osphatidylcholine acyltransferase 2

(
LPCAT2
,
which

incorporates

FA into

the membrane components phosphatidyl

-
ethanolamine and

choline
) and
pantothenate kinase 4

(
which controls a
key step in
CoASH

biosynthesis)
.
Similarly, five FA substrates
correlated
inversely with ACSM6 and five acyl
-
GPC substrates correlated with LPCAT2
.
Thus correlation
s

of

analytes showing significant differences

unveiled close linkage between

metabolic and proteomic alterations in

plasma in sepsis that support
s

the existence of

a
dichotomic
host response that differentiates survivors and

death
s
.

Finally, in light of

an apparent dichotomy in host

response to sepsis, b
iomarker classifiers were sought
that could serve as
objective
adjuncts for

sepsis diagnosis

and prognostic determina
tion

in EDs
.
Sparse
p
redictive models were developed with

t
0

data

and
accuracies were
assessed in t
24

and

replication cohort data

using frequentist and Bayesian methods
.
In general, metabolite classifiers had superior performance than
those developed with plasma
proteins or clinical variables

(Table 1)
.

Among frequentist methods, logistic
regression was

consistently

superior.

S
epsis diagnosis

panels
had excellent performance (98.9% positive
predictive value (PPV) and 93.1% negative predictive value (NPV)
for classification of non
-
infected SIRS
-
positive
10


patients and sepsis survivors
at t
0

for BPFA
-
GPR). However,
classifiers for the more authentic diagn
ostic
comparison, between non
-
infected SIRS
-
positive and

sepsis patients

(irrespective of outcome), had

low NPV
,
and
would
be of

limited

use in excluding non
-
infected patients.
In contrast, m
etabolite
-
based sepsis prognosis
classifiers p
erform
ed
considerab
ly
better than
well
-
established prognostic tools (APACHE II, S
OFA and
capillary

lactate) at t
0

and t
24

and potentially could provide

novel,

daily treatment intensity guidance for hospitalized
patients with sepsis
14
-
16, 32
.

In summary, a systems survey
of

sepsis,
a complex
, heterogeneous and highly dynamic

pathologic
state
,

yielded new insights into molecular mechanisms of survival and death that appear to be sufficiently
accurate to serve as adjuncts
for predictive differentiation

and individualized treatment of patients. Additional
investigation of the dichotomous host
meta
bolic
response to sepsis is
needed

to address

more fully the
temporal dynamics and

breadth of relevance in community
-
acquired and nosocomial sepsis among diverse
patient populations

and infectious agents,

and pot
ential

extensibility to

outcomes in

related conditions, such
as trauma
,
hyperthermia

and drug
-
induced mitochondrial damage
. The proximity with which metabolic
dichotomy is detected during sepsis suggests it to be pre
-
programmed, potentially indicating Mendelian
causality. An exciting future

direction will be to assess reversibility of the death phenotype by

targeted

interventions such as early goal
-
directed therapy, steroid
or
HMGCR inhibitor

administration or
enhancement
of mitochondrial biogenesis
.

Of general import is the apparent utility

of global
and temporal
correlations of
metabolome

and
proteome

data from
relevant biological fluids and
well
-
phenotyped patient
groups

to expand

understanding of intermediary metabolism, particularly with respect to poorly annotated proteins and
metabolit
es, and for characterization of homogeneous subgroups in complex traits.

In light of studies showing
similar synergistic integration of genotype and functional genomic data, there exists the imminent possibility
of establishment of multi
-
dimensional molecu
lar models of disease that may be used to examine likely
network responses to perturbation.




11


Funding/Support: This work was supported by NIH Grant 1U01AI066569 and NIH Contract
HHSN266200400064C from the National Institute of Allergy and Infectious Diseases and NIH
Grant P20 RR
016480 from the National Center for Research Resources,
and grants fro
m Pfizer Inc. and Roche Diagnostics
Inc.

A deo lumen, ab amicis auxiliu
m
.



12


METHODS SUMMARY

A total of
1152 patients presenting at

EDs (Henry Ford Hospital, Duke University Hospital, and Durham
Veterans Administration Medical Center) with suspected

sepsis (≥2 SIRS criteria and infection) were enrolled.
Approval was obtained by the local ethical committees and written informed consent was given by each
patient or legal guardian.
Physical examination was performed and venous plasma was

collected at enr
ollment
(t
0
) and 24 hrs later (t
24
); patients were followed
for 28 days
. Demographic and clinical data was
anonymized
and
digitally stored
in com
pliance with HIPAA regulations (ProSanos

Inc., Harrisburg, PA
). Following
independent audit of
infection status

and outcomes
, 150 subjects were chosen for
discovery studies
. Patients
were classified
as

n
on
-
infected SIRS
-
positive
,

severe s
epsis
; septic s
hock
; uncomplicated s
epsis

or sepsis d
eath
.
t
0

and t
24

samples from a

replication
set of 52 sepsis survivors and
deaths w
ere

also

used.

Pl
asma m
etabolite
s
were prepared and analyzed by high performance liquid chromatography

(HPLC)

and

Thermo
-
Finnigan LTQ MS
with electrospray ionization and linear ion
-
trap mass analyzer and by gas chromatography and
Thermo
-
Finnigan
Trac
e DSQ fast
-
scanning single
-
quadrupole MS with electron impact ionization

(Metabolon Inc.,
Durham, NC)
.

Plasma proteins were
immunodepleted

and analyzed by
LTQ ion trap
MS

(
Thermo
-
Finnigan)

in
triple play mode

(Monarch LifeSciences Inc., Indianapolis, I
N)
.
Statistical analysis employed

JMP Genomics v4,
SAS
.
BPFA
-
GPR

employed a

mixture model
, with each metabolite employing fa
ctors from one cluster,
inference of t
he number of
clusters

needed t
o represent all metabolites
via the Dirichlet
process,
and
devel
opment of

a sparse probit
-
regression

classifier
20
.



13


Methods

CAPSOD Study Sites and Patients

The Community Acquired Pneumonia and Sepsis Outcome
Diagnostics (CAPSOD) study was approved by the
Institutional Review Boards of the National Center for Genome Resources (Santa Fe, NM), Duke University
Medical Center (Durham, NC), Durham Veteran Affairs Medical Center (Durham, NC) and Henry Ford Hospital
(
Detroit, MI) and filed at ClinicalTrials.gov (NCT00258869). Inclusion criteria were presentation at the
emergency department with known or suspected acute infection and presence of at least two of the systemic
inflammatory response syndrome (SIRS) criteri
a
33
. Exclusion crit
eria
were
as previously described
13
. Patients
were enrolled from 2005 through 2009 in emergency departments at each institution and written, informed
consent was obtained by all study participants or their legal designates.


Clinical Data

Collection

Patient demographics, exposure, symptoms, past medical history, results of physical examination, APACHE II
score, SOFA score, DIC score, MELD score, development of ALI and ARDS and treatment were recorded at
enrollment (t
0
) and at 24 hours (t
24
) by a nurse practitioner or physician using online electronic data capture
(Prosanos,
Harrisburg, PA
), as previously described
13, 18, 33
. Microbiologic evaluation was as indicated clinically
together with urinary pneumococcal and legionella antigen tests. Finger
-
stick lactate values were obtained.
Other laboratory, microbiology and radiographic tests were ordered by the treating physicia
n according to
standards of care. Following patient discharge or death, charts were reviewed and largest deviations of clinical
and laboratory parameters from normal were recorded, together with occurrence of outcome measures,
microbiologic results, treat
ment and time
-
to
-
events. Blood was collected in bar
-
coded EDTA
-
plasma tubes
at
enrollment (t
0
) and the following day (t
24
), incubated on ice until centrifuged (within 4 hours), plasma was
separated and aliquots were stored at
-
80°C.


Clinical Data Audit an
d Discovery Cohort Selection

14


Records from all enrollees were adjudicated independently during 2009 by a board

certified study physician at
least 30 days after discharge to determine whether presenting symptoms and signs were due to infection, and,
if so, e
tiologic agent, site of infection, patient outcomes and times
-
to
-
outcomes. Patients were assigned to
categories, as described
13
: Definite infection, causative organism identified; definite infection, causative
organism uncertain; Indeterm
inate, infection possible; No evidence of infection; and
n
o evidence of infection
and diagnosis of a non
-
infectious process that accounted for SIRS. 150 patients were selected from the
definite infection and non
-
infection categories for plasma metabolome
and proteome analyses
4, 34, 35
, as
follows: Non
-
infected patients with SIRS (NIS, n=29); Uncomplicated sepsis (UCS
4, 34, 35
, n=27); Severe sepsis (SS;
patients with definite infections who progressed
to severe sepsis by day 3, but did not develop septic shock or
death by day 28
4, 34, 35
, n=25,
supplementary table 1
); Septic shock (
patients with definite infections who
progressed to

septic shock by day 3 but did not die by day 28
4, 34, 35
, n=3
8, Table 1); Sepsis deaths (
patients with
definite infections who died by day 28, n=31). Patients with definite infection
s were further selected to
maximize the number caused by
Escherichia coli, Staphylococcus aureus, and Streptococcus pneumoniae
, and
to provide a wide range of APACHE II scores. Within these constraints, groups were matched for age, race, sex
and enrollmen
t site
(
Supplementary table 2
).


Metabolite Sample Preparation and Gas Chromatography/Mass
-
Spectrometry (GC/MS) and Liquid
Chromatography/Mass
-
Spectrometry (LC/MS/MS) analysis

Plasma samples were thawed on ice at Metabolon Inc. (Durham, NC)
, and 100
μL was
extracted using an
automated MicroLab STAR® system (Hamilton Company, Salt Lake City, UT), as described
29
. A well
characterized human plasma pool (“Matrix”, MTRX) was also included as a technical replicate, to
assess
variability and sensitivity in the measurement of all consistently detected chemicals. A single solvent
extraction was performed with 400μl of methanol, containing the recovery standards tridecanoic acid,
fluorophenylglycine, chlorophenylalanine, a
nd d6
-
cholesterol, by shaking for two minutes using a
Geno/Grinder 2000 (Glen Mills Inc., Clifton, NJ). After extraction, the sample was centrifuged, the supernatant
15


removed and split into four equal aliquots: two for LC/MS, one for GC/MS and a reserve al
iquot. Aliquots were
dried under vacuum overnight on a TurboVap® (Zymark). Samples were maintained at 4
o
C throughout the
extraction process. For LC/MS analysis, aliquots were reconstituted in either 0.1% formic acid for positive ion
LC/MS, or 6.5 mM ammoni
um bicarbonate pH 8.0 for negative ion LC/MS. For GC/MS analysis, aliquots were
derivatized using equal parts N,O
-
bistrimethylsilyl
-
trifluoroacetamide and a mixture of
acetonitrile:dichloromethane:cyclohexane (5:4:1) with 5% triethylamine at 60
o
C for 1 hou
r. The derivatization
mixture also contained a series of alkyl benzenes that served as retention time markers.


LC/MS was carried out using an Acquity UPLC (Waters Corporation, Milford, MA) coupled to a linear trap
quadrupole (LTQ) mass spectrometer (Ther
mo Fisher Scientific Inc., Waltham, MA) equipped with an
electrospray ionization source. Two separate LC/MS injections were performed on each sample: the first
optimized for positive ions and a second for negative ions. The mobile phase for positive ion a
nalysis consisted
of 0.1% formic acid in H
2
O (solvent A) and 0.1% formic acid in methanol (solvent B), while that for negative ion
analysis consisted of 6.5 mM ammonium bicarbonate, pH 8.0 (solvent A) and 6.5 mM ammonium bicarbonate
in 95% methanol (solven
t B). The acidic and basic extracts were monitored for positive and negative ions,
respectively, using separate acid/base dedicated 2.1 x 100 mm Waters BEH C18 1.7 µm particle columns
heated to 40°C. The extracts were loaded via a Waters Acquity autosample
r and gradient eluted (0% B to 98%
B, with an 11 minute runtime) directly into the mass spectrometer at a flow rate of 350 µl/min. The LTQ
alternated between full scan mass spectra (99
-
1000 m/z) and data dependent MS/MS scans, which used
dynamic exclusion.



Derivatized samples were analyzed on a Thermo
-
Finnigan Trace DSQ fast
-
scanning single
-
quadrupole MS

set at
unit mass resolving power. The GC column was 20 m x 0.18 mm with 0.18 µm film phase consisting of 5%
phenyldimethyl silicone. The temperature program ramped from 60°C to 340°C, with helium as the carrier gas.
The MS was operated using electron impac
t ionization with a 50
-
750 amu scan range, tuned and calibrated
16


daily for mass resolution and mass accuracy. Samples were randomized to avoid group block effects and were
analyzed over five days. Six MTRX aliquots, an internal standard sample (see below)

and control samples
(without plasma extract) were included in each run.


Metabolites were identified by automated comparison to a reference library of purified external standards
using Metabolon software developed for creating library entries from known c
hemical entities with automatic
fitting of reference to experimental spectra. Peaks that eluted from the LC or GC methods were compared to
the library at a particular retention time and associated spectra for that metabolite. An internal standard
composed

of 30 organic molecules was used in the GC and LC methods to calibrate retention times of
metabolites across all samples. Platform variability was determined by calculating the median relative
standard deviation for the internal standard compounds. Over
all variability (including sample preparation)
was determined by the median RSD for 261 metabolites present in all MTRX samples. Peptides were identified
using standard tandem mass spectrometry sequencing.


Raw area counts for each metabolite in each sa
mple were normalized to correct for variation resulting from
instrument inter
-
day tuning differences. For each metabolite, the raw area counts were divided by the median
value for each run
-
day, therefore setting the medians to 1
.0

for each run. This preser
ved variation between
samples, but allowed metabolites of widely different raw peak areas to be compared on a similar graphical
scale. Missing values were imputed with the observed minimum after normalization. However, metabolites
with missing values in >

50% of the samples were excluded from analysis.

Proteome

Sample Prep and Mass Spectrometer Analysis

Plasma samples were thawed on ice at Monarch Life Sciences Inc. (Indianapolis, IN) and the most abundant
proteins (albumin, IgG, fibrinogen, transferrin, IgA, IgM, haptoglobin, α2
-
macroglobulin, α1
-
acid glycoprotein,
α1
-
antitrypsin and apolipoprotein A
-
I a
nd A
-
II) were removed using Seppro IgY12 Columns (GenWay Biotech
Inc., San Diego, CA). Column flow
-
throughs were denatured by 8M urea, reduced by triethylphosphine,
17


alkylated by iodoethanol and digested by trypsin, as described
36
. Tryptic digests (~20 μg) were analyzed using
a Thermo
-
Fisher Scientific linear

ion
-
trap mass spectrometer (LTQ) coupled with a Surveyor HPLC system.
Peptides were separated on a C18 reverse phase column (i.d. = 2.1 mm, length = 50 mm) with a flow rate of
200μl/min and eluted with a gradient from 5 to 45% acetonitrile developed over

120 min. All injections were
randomized and the instrument was operated by the same operator for the study. Data were collected in the
triple
-
play mode (MS scan, zoom scan and MS/MS scan). Data were filtered and analyzed by a proprietary
algorithm
37, 38
. Databa
se searches against the IPI (International Protein Index) human database and the non
-
Redundant
-
Homo Sapiens database were carried out using both the X!Tandem and SEQUEST algorithms
39, 40
.
Observed peptide MS/MS spectrum and theoretically derived spectra were used to assign quality scores (Xcorr
in SEQUEST and e
-
Score in X!Tandem). Protein identities were assigned priority scores (fr
om one to four): 1,
high peptide confidence and multiple sequences; 2, high peptide confidence; 3, moderate peptide confidence
and multiple sequences; 4, moderate peptide confidence. High peptide confidence corresponded to a highest
confidence between 90
-
100%, and moderate confidence to 75
-
89%.


Protein quantification was carried out using a proprietary protein quantification algorithm
38
. Briefly, raw files
were acquired from the LTQ and all extracted ion chromatograms (XIC) were aligned by retention time. For
protein quantification, each aligned peak must match precurs
or ion, charge state, fragment ions (MS/MS data)
and retention time (within a one minute window). After alignment, area
-
under
-
the
-
curve (AUC) for each
individually aligned peak from each sample was measured and compared for relative abundance. Peak
intensi
ties were log2 transformed before quantile normalization
41

to ensure that every sample has a peptide
intensity histogram of the s
ame scale, location and shape. Normalization removed trends introduced by
sample handling, sample preparation, total protein differences and changes in instrument sensitivity while
running multiple samples (data not shown). If multiple peptides had the sam
e protein identification, then their
quantile normalized log2 intensities were averaged to obtain log2 protein intensities.


18


Statistical Analyses

Overlaid kernel density estimates, univariate distribution results, correlation coefficients of pair wise sa
mple
comparisons, unsupervised principal components analysis (by Pearson product
-
moment correlation) and Ward
hierarchal clustering of Pearson product
-
moment correlations were performed using log2 transformed data as
described
42

using JMP Genomics 4.0 (SAS Institute Inc., Cary, NC). Decomposition of principal components of
variance, including patient demographics, past medical history, laboratory and clinical values, was performed
to maximize sepsis
-
group
-
related components of va
riance and minimize residual variance (Table 2). Guided by
these analyses, ANOVA was performed between sepsis groups with five or 10% false discovery rate (FDR)
correction and inclusion of substantive non
-
hypothesis components of variance as fixed effects
. These
included renal function, as determined by the estimated glomerular filtration rate (eGFR) using the four
variable modification of diet in renal disease calculation, hemodialysis (HD), cirrhosis and liver disease,
hepatitis, neoplastic disease, con
genital disease, administration of exogenous immunosuppressants, drug
abuse, metabolic dysfunction, respiratory dysfunction, serum glucose levels and mean arterial pressure (MAP)
in analyses as non
-
hypothesis components of certain metabolome and/or proteom
e datasets. Box plots
(means, medians, upper and lower quartiles and limits of the distribution) were displayed for metabolites and
proteins identified by statistical tests using “R” from the Free Software Foundation, Inc. Predictive modeling
was perform
ed with JMP Genomics 4.0 using logistic regression. Cross validation was performed using 50
iterations and 10% sample omission.


Bayesian clinical factor analysis and
BPFA
-
GPR was performed jointly for all metabolites and samples
20
.
To distinguish the effects of SIRS outcomes (non
-
infected SIRS+, sepsis survivors, and sepsis death) and
relevant clinical
factors on the metabolome we used the Bayesian clinica
l factor analysis

[c
j

= By
j

+ A(s
j



z
j
) + ε
j
].
The analysis correlates relevant metabolite patient changes to the clinical features to define the relevance of
the clinical parameter. The formula defines B as the relationship between data and the clinical feature, while A
defines random or undefined
effects and ε
accounts for random noise. The clinical features were further
normalized to normal distribution with zero
-
mean and standard deviation.
The significant features were then
19


plotted on B
-
matrix as well as plotted as normalized energy of each clin
ical feature.

In the BPFA
-
GPR analysis,
for
each metabolite there was an associated vector of responses, manifested by the response of each sample
for that metabolite. Each such metabolite
-
dependent vector was represented as a linear combination of
factors
. A mixture model was employed, with each metabolite employing factors from one of the mixture
components (clusters). The number of clusters needed to represent all metabolites

was

inferred via the
Dirichlet
process, metabolites were grouped into clusters
and the latter were
employed to b
uild a sparse
probit
-
regression

classifier. Spec
ifically, the classifier imposed that only a small subset of clusters be

used (all
me
tabolites from unused clusters we
re not employed in the classifier). However, when a clust
er
was used,
multiple metabolites from the cluster we
re employed in the classifie
r. The use of a sparse set of metabolites
mitigated

over
-
training. Furth
er, since multiple metabolites were used from the
same cluster when that cluster
is employ
ed, correlate
d metabolites of

relevance

were

retained.


Cross Correlation Matrix


Pairwise c
ross

correlation
s

was performed using JMP Genomics 4.0 software to compare protein and
metabolite

values in 150 individuals

using Pearson moment
-
correlation. Briefly, all significant proteins
determined in the univariate analysis were used in the
proteome

analysis. All redundant entries that had the
same gene name and very best AA
-
sequence were removed from the analysis for a
total of 700 significant
proteins. All metabolites were included in the analysis, except for unannotated GC/MS determined com
pounds
as these are likely

to be techninical artifacts or compounds that would be difficult to identify the parent
metabolite. Meta
bolite and protein
log2 values were transposed int
o a wide format and the correlations were
merged based on patient identification.
Protein
-
metabolite correlations
were considered significant if
observed at t
0

and t
24
with

log10p
-
value
s ≥ 1.0 and >1.30
,
o
r at a single time point with a Bonferroni corrected
p value (

log10p
-
value
s

≥ 7
.0
3)
.


Biomarker Development

20



Objective biomarkers that supplement clinical criteria are needed for rapid, point
-
of
-
care diagnosis of
sepsis and for stratification of patients
for intensive care. A negative diagnostic biomarker result
, for example,

would avoid empiric antibiotic therapy and
prioritize the

pursuit of differential diagnose
s.
A
n adverse
prognosis phenotype is anticipated to be useful for patient selection for earl
y goal directed therapy

(EGDT),

hospital admission, intensive c
are.

Prognostic biomarkers
may

also provide metabolic surrogate end
-
points by
which to assess efficacy of interventions and guide treatment duration and intensity. Further, the metabolic
pathw
ay alterations that distinguish sepsis survivors and deaths provide a target profile for therapeutic
interventions in patients at high risk of sepsis death.
The current gold standards for prognostic assessment in
sepsis are the Sequential Organ Failure Ass
essment (SOFA) and the Acute Physiology and Chronic Health
Evaluation (APACHE II) tests
18, 33
. However, neither is used routinely since they ar
e time consuming and contain
subjective components with variable interpretation.

Arterial or capillary

lactate values
are

also
used for
severity assessment, with high values (≥ 4.0 mg/dl) indicating tissue

hypoxia
32
.
Conventional

prognostic
methods lack accuracy,

specificity or

sensitivity

(T
able 1).


To develop biomarker

panel
s
, two approaches were employed: Classical predictive modeling, and a
Bayesian elastic net approach using
beta
process factor analysis and group probit regression (BPFA
-
GPR)
43
. In
the
classical modeling approach
, four modeling methods

were

employed

(
JMP Genomics 4.0
)
: 1) discriminant
analysis; 2) general linear model selection; 3) logistic regression; and 4) partition trees. The results for all but
logistic regression with penalized par
ameter re
duction were poor, or exhibited poor cross validation or
replication (data not shown).
To distinguish

among sepsis outcomes,
t
0

data
,
including

clinical
parameters, was
used in a

training set
of
sepsis survivors versus sepsis death.
P
iperine, palm
itoycarnitine, 3
-
methoxytyrosine,
ocatanoylcarnitine, clinical blood lactate, and unannotated analytes X
-
12775, and the single sulfated
unannotated steroid X
-
11302

were selected

(
T
able 1). Accuracy,
receiver operating characteristic

area under
the curve (A
UC) were high (0.923, 0.957 respectively), while root mean square error (RMSE) was low (0.254).
Positive predictive value (PPV) was 96.7% and
negative predictive va
lue (NPV) was 80.7%
. Validity of
the model
was challenged in

three other

data sets. Accuracy

a
nd AUC remained high in these

(t
24
, Rt
0
, and Rt
24
; accuracy,
21


0.804, 0.784, 0.816 respectively; AUC, 0.838, 0.789, 0.784 respectively), while RMSE was low (0.379, 0.425,
0.435
,

respectively).
PPVs were excellent (
91.0%, 94.1%, 96.7% respectively), while
N
PVs were reasonable
(
51.7%, 47.1%, 50.0%
,

respectively).

A model was also developed for the more clinically relevant comparison of all sepsis patients versus
non
-
infected SIRS
-
positive controls (T
able 1): Galactonate, uridine, maltose, glutamate, creatine
and
unannotated metabolite X
-
12644 were selected and
exhibited high accuracy
in the t
0
, and t
24

datasets (0.880,
0.788
,

respe
ctively; AUC, 0.927, 0.742;

RMSE
,
0.2841,

0.424;
PPV, 95.9%, 87.9%
;
NPV, 55.2%, 40.0%). Results
improved when
confounding
sepsis
deaths were omitted from the analysis (sepsis diagnosis): Citrulline,
laurylcarnitine, androsterone sulfate, isoleucine, X
-
11838, X
-
12644, and X
-
11302 were selected

(
accuracy
,
0.924, 0.806
,

respectively
;

AUC
,
0.963, 0.852
, respectively;

RMSE
,
0.247, 0.424
,

respectively;
PPV, 94.4%,
84.6%
;
NPV, 84.6%, 68.0%).


For the Bayesian approach,
BPFA
-
GPR

was performed
as described in the statistical analysis
. In sepsis
outcomes, two tests were employed using BPFA
-
GPR. The first limited the number of feature groups, o
r
clusters
,

to a total of 11, of which eight were employed in the probit regression. The top 50 features
(metabolites) were determined to be pr
edictive (S
upplementary T
able 8
;
t
0

and t
24

accuracy 98.4%, 90.7%
,
respectively;

PPV
,
100.0%, 97.4%
, respectivel
y;
NPV
,
93.5% and 72.4%). However, t
he replication study results
were

less impressive
(
Rt
0

and Rt
24

accuracy
, 76.5%, 77.65 respectively
; PPV
, 100.0%, 100.0%
; NPV
,
29.4%,
31.3%
,

respectively). A second analysis was performed that
increased

sparseness in the model

(8

features:
creatinine, 4
-
vinylphenol sulfate, c
-
glycosyltrypto
phan, X
-
11261, X
-
12095, X
-
12100 and

X
-
13553
)
. Again t
0

and
t
24

performed well (PPV = 94.4%; 94.9% respectively; NPV = 58.1%; 69.0% respectively). However, While PPV
re
mained high in the replication dataset (Rt
0

= 94.1%; Rt
24
= 100.0%), NPV
was

less impressive (Rt
0

= 29.4%;
Rt
24
= 37.5%). For
diagnosis in
all
sepsis patients, factor analysis revealed

67 feature groups
,

of which most of
the top 50 featur
es were found within

group 67 (S
uppleme
ntary T
able 8).
Classification r
esults
, however,

were
not as impressive

in these groups, since

the model had difficulty distinguishing
non
-
infected SIRS
-
positive

patients at both t
0
and t
24

(accuracy, 88.0%, 83.3% respectively
; PPV
, 100
.0%, 99.1% respectively
; NPV
,
38.0%,
22


16.0%).
S
epsis diagnosis

in comparisons of sepsis survivors and controls was more accurate:

74 clusters

were

found in the factor analysis,
with
most of the significant features taken from the final cluster, and the top
50
wer
e determined t
o be predictive (Supplementary T
able 8; a
ccuracy, PPV and NPV at t
0

of 97.5%, 98.9% and
93.1%
,

respectively,

and
80.6%
,

89.7%

and
52.0%

at t
24
.


In summary, t
he logistic regression analyses provide
d

accurate classifiers for
sepsis
diagnosis and
sepsis outcomes
.

However, the Bayesian analyses provide
d

additional predictive features that
may serve
to
“tune” the model for improved accuracy, sensitivity and specificity.







1.

Williams, D.C.
God's wilds: John Muir's
vision of nature
. (Texas A&M University Press, College Station;
2002).

2.

Alberti, C.

et al.

Epidemiology of sepsis and infection in ICU patients from an international multicentre
cohort study.
Intensive care medicine

28, 108
-
121 (2002).

3.

Angus, D.C.

et
al.

Epidemiology of severe sepsis in the United States: analysis of incidence, outcome,
and associated costs of care.
Critical care medicine

29, 1303
-
1310 (2001).

4.

Bone, R.C.

et al.

Definitions for sepsis and organ failure and guidelines for the use of i
nnovative
therapies in sepsis. The ACCP/SCCM Consensus Conference Committee. American College of Chest
Physicians/Society of Critical Care Medicine.
Chest

101, 1644
-
1655 (1992).

5.

Dellinger, R.P.

et al.

Surviving Sepsis Campaign: international guidelines
for management of severe
sepsis and septic shock: 2008.
Critical care medicine

36, 296
-
327 (2008).

6.

Jaimes, F.

et al.

The systemic inflammatory response syndrome (SIRS) to identify infected patients in
the emergency room.
Intensive care medicine

29, 1368
-
1371 (2003).

7.

McCaig, L. & Nawar, E. (ed. U.D.o.H.a.H. Services)2006).

23


8.

Rangel
-
Frausto, M.S.

et al.

The natural history of the systemic inflammatory response syndrome (SIRS).
A prospective study.
Jama

273, 117
-
123 (1995).

9.

Woods, C.W., Feezor, R.J.
& Kingsmore, S.F. Sepsis and the Genomic Revolution: Risk, Diagnosis,
Prognosis, and Therapeutics, in
Genomic and Personalized Medicine
. (eds. H. Willard & G. Ginsburg)
(Elseiver, 2008).

10.

Sreekumar, A.

et al.

Metabolomic profiles delineate potential rol
e for sarcosine in prostate cancer
progression.
Nature

457, 910
-
914 (2009).

11.

Tabb, D.L.

et al.

Repeatability and reproducibility in proteomic identifications by liquid
chromatography
-
tandem mass spectrometry.
Journal of proteome research

9, 761
-
776 (201
0).

12.

Tsalik, E.L.

et al.

Multiplex PCR to diagnose bloodstream infections in patients admitted from the
emergency department with sepsis.
Journal of clinical microbiology

48, 26
-
33 (2010).

13.

Glickman, S.W.

et al.

Disease progression in hemodynamically

stable patients presenting to the
emergency department with sepsis.
Acad Emerg Med

17, 383
-
390 (2010).

14.

Rivers, E.P. &

Ahrens, T. Improving outcomes for severe sepsis and septic shock: tools for early
identification of at
-
risk patients and treatment protocol implementation.
Critical care clinics

24, S1
-
47
(2008).

15.

Rivers, E.P., Coba, V., Visbal, A., Whitmill, M. & Ampo
nsah, D. Management of sepsis: early
resuscitation.
Clinics in chest medicine

29, 689
-
704, ix
-
x (2008).

16.

Rivers, E.P., McIntyre, L., Morro, D.C. &

Rivers, K.K. Early and innovative interventions for severe sepsis
and septic shock: taking advantage of a window of opportunity.
CMAJ

173, 1054
-
1065 (2005).

17.

Bernard, G.R.

et al.

Efficacy and safety of recombinant human activated protein C for severe s
epsis.
The New England journal of medicine

344, 699
-
709 (2001).

18.

Knaus, W.A., Draper, E.A., Wagner, D.P. & Zimmerman, J.E. APACHE II: a severity of disease
classification system.
Critical care medicine

13, 818
-
829 (1985).

24


19.

Wishart, D.S.

et al.

HMDB:
a knowledgebase for the human metabolome.
Nucleic acids research

37,
D603
-
610 (2009).

20.

Chen, M.

et al.

Detection of Influenza via Statistical Gene
-
Expression Analysis
IEEE Trans. Biomedical
Engineering

(Submitted).

21.

Brealey, D.

et al.

Association bet
ween mitochondrial dysfunction and severity and outcome of septic
shock.
Lancet

360, 219
-
223 (2002).

22.

Fukumoto, K., Pierro, A., Spitz, L. & Eaton, S. Neonatal endotoxemia affects heart but not kidney
bioenergetics.
Journal of pediatric surgery

38, 690
-
6
93 (2003).

23.

Giovannini, I.

et al.

Respiratory quotient and patterns of substrate utilization in human sepsis and
trauma.
Jpen

7, 226
-
230 (1983).

24.

Herzum, I. & Renz, H. Inflammatory markers in SIRS, sepsis and septic shock.
Current medicinal
chemistry

15, 581
-
587 (2008).

25.

Pierrakos, C. & Vincent, J.L. Sepsis biomarkers: A review.
Critical care (London, England)

14, R15 (2010).

26.

Lissauer, E., Johnson, B., Shi, S., Gentle, T. & Scalea, M. 128 Decreased lysophosphatidylcholine levels
are associated
with sepsis compared to uninfected inflammation prior to onset of sepsis.
Journal of
Surgical Research

137, 206
-
206 ( 2007).

27.

Drobnik, W.

et al.

Plasma ceramide and lysophosphatidylcholine inversely correlate with mortality in
sepsis patients.
Journal o
f lipid research

44, 754
-
761 (2003).

28.

Omenn, G.S.

et al.

Overview of the HUPO Plasma Proteome Project: results from the pilot phase with
35 collaborating laboratories and multiple analytical groups, generating a core dataset of 3020
proteins and a publi
cly
-
available database.
Proteomics

5, 3226
-
3245 (2005).

29.

Lawton, K.A.

et al.

Analysis of the adult human plasma metabolome.
Pharmacogenomics

9, 383
-
397
(2008).

30.

Meier, U., Gressner, O., Lammert, F. & Gressner, A.M. Gc
-
globulin: roles in response to i
njury.
Clinical
chemistry

52, 1247
-
1253 (2006).

25


31.

Chua, D., Tsang, R.S. & Kuo, I.F. The role of statin therapy in sepsis.
The Annals of pharmacotherapy

41,
647
-
652 (2007).

32.

Rivers, E.

et al.

Early goal
-
directed therapy in the treatment of severe sepsis and septic shock.
The New
England journal of medicine

345, 1368
-
1377 (2001).

33.

Vincent, J.L.

et al.

The SOFA (Sepsis
-
related Organ Failure Assessment) score to describe organ
dysfunction/fail
ure. On behalf of the Working Group on Sepsis
-
Related Problems of the European
Society of Intensive Care Medicine.
Intensive care medicine

22, 707
-
710 (1996).

34.

Balk, R.A. Severe sepsis and septic shock. Definitions, epidemiology, and clinical manifestat
ions.
Critical
care clinics

16, 179
-
192 (2000).

35.

Bone, R.C.

et al.

Sepsis syndrome: a valid clinical entity. Methylprednisolone Severe Sepsis Study
Group.
Critical care medicine

17, 389
-
393 (1989).

36.

Hale, J.E., Butler, J.P., Gelfanova, V., You, J.S.
& Knierman, M.D. A simplified procedure for the
reduction and alkylation of cysteine residues in proteins prior to proteolytic digestion and mass
spectral analysis.
Analytical biochemistry

333, 174
-
181 (2004).

37.

Higgs, R.E.

et al.

Estimating the statisti
cal significance of peptide identifications from shotgun
proteomics experiments.
Journal of proteome research

6, 1758
-
1767 (2007).

38.

Higgs, R.E., Knierman, M.D., Gelfanova, V., Butler, J.P. & Hale, J.E. Comprehensive label
-
free method
for the relative qu
antification of proteins from biological samples.
Journal of proteome research

4,
1442
-
1450 (2005).

39.

Craig, R. & Beavis, R.C. TANDEM: matching proteins with tandem mass spectra.
Bioinformatics (Oxford,
England)

20, 1466
-
1467 (2004).

40.

Yates, J.R., 3rd
, Eng, J.K., McCormack, A.L. & Schieltz, D. Method to correlate tandem mass spectra of
modified peptides to amino acid sequences in the protein database.
Analytical chemistry

67, 1426
-
1436 (1995).

26


41.

Bolstad, B.M., Irizarry, R.A., Astrand, M. &

Speed, T.P. A comparison of normalization methods for high
density oligonucleotide array data based on variance and bias.
Bioinformatics (Oxford, England)

19,
185
-
193 (2003).

42.

Mudge, J.

et al.

Genomic convergence analysis of schizophrenia: mRNA sequenc
ing reveals altered
synaptic vesicular transport in post
-
mortem cerebellum.
PloS one

3, e3625 (2008).

43.

Zaas, A.K.

et al.

Gene expression signatures diagnose influenza and other symptomatic respiratory viral
infections in humans.
Cell host & microbe

6, 2
07
-
217 (2009).