Accepted: 20 April 2018

DOI: 10.1111/jre.12575

ORIGINAL ARTICLE

High-­throughput proteomic analysis of candidate biomarker

changes in gingival crevicular fluid after treatment of chronic
periodontitis

Y. A. Guzman1,2,3 | D. Sakellari4 | K. Papadimitriou4 | C. A. Floudas1,2,†

1
Artie McFerrin Department of Chemical
Engineering, Texas A&M University, College Background and Objective: Untargeted, high-­throughput proteomics methodologies
Station, USA have great potential to aid in identifying biomarkers for the diagnosis of periodontal
2
Texas A&M Energy Institute, Texas A&M disease. The application of such methods to the discovery of candidate biomarkers
University, College Station, USA
3 for the resolution of periodontal inflammation after periodontal therapy has been
Department of Chemical and Biological
Engineering, Princeton University, Princeton, investigated.
USA
Material and Methods: Gingival crevicular fluid samples were collected from 10 pa-
4
Department of Preventive Dentistry,
Periodontology and Implant
tients diagnosed with chronic periodontitis at baseline and 1, 5, 9 and 13 weeks after
Biology, Aristotle University of Thessaloniki, completion of mechanical periodontal treatment. Clinical indices of periodontal dis-
Thessaloniki, Greece
ease, including probing depth, recession, clinical attachment level and bleeding on
Correspondence probing, were recorded at baseline and 13 weeks. Samples were analyzed using an
Dimitra Sakellari, Department of Preventive
Dentistry, Periodontology and Implant
online liquid chromatography-­nanoelectrospray-­hybrid ion trap-­Orbitrap mass spec-
Biology, Aristotle University of Thessaloniki, trometer. Spectra were processed with the PILOT_PROTEIN proteomics software
Thessaloniki, Greece.
Email: dimisak@med.auth.gr
suite.
Results: Clinical parameters were significantly improved 13 weeks after treatment
Funding information
National Institutes of Health, Grant/Award
(Wilcoxon signed ranks test, P < .05). From the substantial number of identified pro-
Number: r01lm009338; National Science teins, a small subset was extracted by filter methods that included temporal pattern
Foundation, Grant/Award Number: CBET-
0941143
matching, logistic function fitting and mixed-­integer linear optimization. This subset
includes azurocidin, lysozyme C and myosin-­9 as candidate biomarkers prominent at
baseline and alpha-­smooth muscle actin as prominent 13 weeks after treatment.
Cross-­validation studies yielded average predictive accuracy and area under the
curve of 0.900 and 0.930, respectively.
Conclusions: High-­throughput proteomic analysis can contribute to identifying end-
points of periodontal therapy. These candidate biomarkers should be evaluated for
clinical efficacy.

KEYWORDS
biomarkers, gingival crevicular fluid, nonsurgical periodontal therapy

1 | I NTRO D U C TI O N received considerable interest in recent years. The complexity

The search for biomarkers, specifically molecular biomarkers,
for the diagnosis, staging or prognosis of periodontitis has
†Deceased, August 2016.
of periodontitis has led to the search for biomarkers that would
indicate early disease onset, aid in differential diagnoses, profile
disease progression and treatment responses, and indicate clear
clinical endpoints.1-4 The protein domain is a promising source
for biomarkers due to the relationship between the temporal

J Periodont Res. 2018;1–8. wileyonlinelibrary.com/journal/jre © 2018 John Wiley & Sons A/S. | 1
Published by John Wiley & Sons Ltd
2 |
expression of the static genome and biological activity. As an
alternative to hypothesis-­
d riven approaches to protein bio-
marker discovery, the past 5 years has seen the emergence of
untargeted studies using modern high-­throughput proteomics
technologies. The bottom-­u p proteomics methodology involves
enzymatic digestion of proteins, application of a separations
method on the resulting peptides, ionization and isolation of
peptides in a mass spectrometer (MS), and often fragmenta-
tion of the peptides (tandem MS/MS). To identify a peptide,
the mass of each peptide ion is measured before fragmentation,
with fragmentation occurring under known mechanisms and
yielding spectra that can elucidate the sequence of the peptide.
Identified peptides can then be attributed to proteins using pro-
tein sequence databases.
Proteomics studies of periodontitis have featured different
biofluids and experimental protocols. Gingival crevicular fluid
is most commonly used as a source for biomarkers due to the
ability to reflect serum composition and its role as an exudate in
periodontal disease. 5 Gingival crevicular fluid samples are often
pooled per-­p atient to produce a more robust protein source
that amplifies low-­a bundance proteins. 6 High-­t hroughput pro-
teomics platforms have been used to study gingival crevicu-
lar fluid samples from patients with chronic periodontitis,7-10
generalized aggressive periodontitis11 and maintenance-­p hase
chronic periodontitis.12,13 Saliva can act as a surrogate to gin-
gival crevicular fluid, though gingival crevicular fluid only ac-
counts for <10% of the total protein content.14 Saliva collection
is not as technically demanding as gingival crevicular fluid col-
lection. However, the utilization of saliva as a biomarker source
in carefully controlled clinical studies is more challenging due
to large abundance ranges, the presence of proteases and an
extremely complex protein composition that can be affected by
environmental or/and psychological stimuli and therefore in the
present study gingival crevicular fluid was chosen as the bio-
logical fluid for investigation. 4,15-17 One-­ and two-­d imensional
electrophoresis separations have been utilized with success, but
recent studies featured liquid chromatography (LC) separations,
typically paired with electrospray ionization, and typically pro-
duced an order of magnitude higher number of protein identifi-
cations. We refer to recent reviews for a more comprehensive
overview. 4,18
The aim of this study was to identify temporal proteomic profiles
of chronic periodontitis. Previous studies from our group, defined
TA B L E 1 Clinical parameters of participants
PD CAL
Time point n = 10 (mm) n = 10 (mm)
Baseline 4.13 ± 0.85* 3.87 ± 1.40*
13 weeks 3.09 ± 0.35* 3.06 ± 1.44*
BOP, bleeding on probing; CAL, clinical attachment level; PD, probing depth. Reported values are the mean ± SD.
*Statistically significant differences in clinical parameters between time points (Wilcoxon signed ranks test, P < .05).
GUZMAN et al.
proteomic profiles of patients with chronic periodontitis versus
healthy controls8 and developed a predictive model based on mixed-­
integer linear optimization (MILP) for identifying small subsets of di-
agnostic protein biomarker candidates.9
The current study observed dynamic changes in 10 patients di-
agnosed with chronic periodontitis as they undertook a 13-­week
treatment program. Two mathematical filter methods were applied
to the temporal profile of each protein, and mixed-­integer linear
programming was used to analyze baseline and endpoint data. The
clinical and biological relevance of the top candidates obtained
across all methods is discussed. The top candidates were evaluated
with cross-­validation studies, which involved every possible 10-­fold
cross-­validation partition from the experimental data, resulting in
100 total training-­testing evaluations. The top protein biomarker
candidates have the potential to act as indicators of a clinical end-
point for chronic periodontitis.
2 | M ATE R I A L A N D M E TH O DS
2.1 | Subject sample collection
Ten patients diagnosed with chronic periodontitis and had not
previously received treatment for the disease were selected
to participate in this study. All participants were patients
of the Department of Preventive Dentistry, Periodontology
and Implant Biology, Dental School, Aristotle University,
Thessaloniki, Greece. The participants included 5 male subjects
(mean age ± SD of 50.2 ± 3.5 years) and 5 age-­m atched female
subjects (mean age ± SD of 47.4 ± 3.4 years). Clinical parameters
for participants are presented in Table 1. All subjects were non-­
smokers who were systemically healthy and were not taking
medication known to affect periodontal tissues. Subjects who re-
ported antibiotic intake during the previous 6 months and those
who were pregnant or lactating women were excluded from the
present study. Subjects were diagnosed with chronic periodonti-
tis according to the analytical criteria of the American Academy
of Periodontology.19 Patients were required to meet the thresh-
olds for moderate or advanced chronic periodontitis (presence of
proximal attachment loss of ≥5 mm in ≥30% of teeth present 20 )
and were willing to perform oral hygiene as instructed. All sub-
jects signed an informed consent, and the study was conducted
according to the protocol outlined by the Research Committee,
Aristotle University of Thessaloniki, Greece, and approved by
Sampled sites, n = 40
BOP
n = 10 PD (mm) CAL (mm) BOP
0.53 ± 0.27* 6.15 ± 0.69* 6.60 ± 0.80* 1.00 ± 0.00*
0.20 ± 0.18* 3.65 ± 0.42* 3.77 ± 0.58* 0.05 ± 0.10*
GUZMAN et al.
the Ethical Committee of the School of Dentistry (protocol no.
237).
2.2 | Clinical recordings and treatment procedures
Clinical parameters were assessed at 6 sites of all teeth present
in the dentition, and included probing depth, recession, clinical
attachment level and bleeding on probing. Recordings were per-
formed by a calibrated examiner (DS) using an automated probe
(Florida probe®; Florida Probe Corporation, Gainesville, FL, USA).
The examiner has reproducible assessments (Pearson correlation
coefficient of 0.971) as determined in 10% of weekly registrations.
After clinical assessments and instructions for proper oral hygiene,
all patients were provided with identical nylon, soft, multi-­tufted
toothbrushes (Oral-­B Indicator manual toothbrush; Oral-­B , USA)
and were reassessed for compliance to the oral hygiene instruc-
tions after 2 weeks.
Mechanical treatment was provided by 1 therapist (KP), and
to ensure the same quality of root debridement, all patients were
checked by one of the authors (DS) upon completion of treatment.
Mechanical therapy included supragingival ultrasonic instrumenta-
tion (Piezon®, Instrument A; EMS, Nyon, Switzerland) and subgingi-
val scaling and root planing using hand instruments (Gracey curettes
3/4, 11/12, 13/14; Hu-­Friedy, Chicago, IL, USA) under local anes-
thesia in 2 sessions, and were completed without time restrictions
and within 1 week of each other. Clinical recordings were repeated
13 weeks after completion of mechanical treatment.
2.3 | Gingival crevicular fluid samples
One pooled gingival crevicular fluid sample was collected from each
participant at each time point. Samples were taken from 4 pre-­
selected sites that displayed a baseline probing depth >6 mm and
<8 mm as previously described.8,9 After isolation of a site with cotton
rolls to prevent contamination with saliva, supragingival plaque was
removed and the tooth was air-­dried. A 30-­second gingival crevicular
fluid sample was collected on filter strips (Periopaper®; Interstate
Drug Exchange, Amityville, NY, USA). Periopaper strips were gen-
tly inserted into the orifice of the periodontal pocket (1-­2 mm sub-
gingivally) for 30 seconds. The samples were immediately placed in
Eppendorf tubes containing 100 μL of 100 mmol/L ammonium bi-
carbonate, frozen in liquid nitrogen and stored at −80°C. Gingival
crevicular fluid samples were collected before the clinical measure-
ments and were discarded when visibly contaminated with blood.
Samples were lyophilized for 5 hours at −55°C and 0.03 mbar in an
ALPHA 1-­4 lyophilizer (Martin Christ, Gefriertrocknungsan-­langen
GmbH). Before lyophilization, they were vortexed for 20 minutes
and centrifuged for 10 minutes at 8161.4 g. Gingival crevicular
fluid samples were collected at the treatment baseline and 1, 5, 9
and 13 weeks after treatment. One sample was mistransported
and thus was not included in the study. Samples were processed
at the Princeton Proteomics & Mass Spectrometry Core facility at
Princeton University.
| 3
2.4 | Statistical analysis for clinical parameters
The statistical analysis for clinical parameters was carried out with
the statistical package SPSS Statistics (v. 19; IBM Corp., Armonk,
NY, USA). Differences in clinical parameters between baseline and
13 weeks after treatment were quantified by applying the Wilcoxon
signed ranks test with a significance level of .05 (Table 1).
2.5 | Sample preparation and mass spectrometry
data analysis
Lyophilized samples were reconstituted in 100 μL of water and pro-
tein concentrations were estimated with a bicinchoninic acid assay
(Thermo Scientific, Waltham, MA, USA). Samples were reduced,
alkylated, extracted and loaded to the mass spectrometer as previ-
ously described.8 Nanoflow LC-­MS/MS was performed on an LTQ
Orbitrap XL (Thermo Scientific) coupled to an Agilent 1200 Series
binary HPLC pump (Agilent Technologies, Santa Clara, CA, USA) and
an Eksigent AS2 autosampler (Eksigent Technologies, Dublin, CA,
USA). The Orbitrap obtained full MS spectra and operated in a data-­
dependent mode by selecting 7 peaks per MS spectrum for MS/MS
fragmentation. Raw MS data were centroided and converted to the
mzXML format using the msconvert tool from ProteoWizard tool-
box 21 within the Trans-­Proteomic Pipeline. 22 All MS/MS spectra
were processed using an improved version of the PILOT_PROTEIN
protein identification suite23-26 using a subset of the Swissprot
database derived from the Homo sapiens taxonomy. Search toler-
ances included a value of 0.1 Da for the precursor ion and 0.5 Da
for the fragment ion. Searches were performed using a maximum of
2 missed cleavages and a static cysteine modification of 57 Da due
to the iodoacetimide treatment. The false discovery rate for protein
identification utilized in this study was 2% and was calculated using
a reverse-­sequence decoy database. 25 Positive identification of a
protein was allowed for one peptide if that particular amino acid se-
quence could not be associated with another protein in the database
and meets a scoring threshold. 25 All other positive protein identifica-
tions required at least 2 annotated peptides.
2.6 | Biomarker selection by temporal profiling
Temporal profiling of periodontitis was performed by analyzing
per-­
p atient trends and population-­
b ased trends. For each pa-
tient, any proteins that were detected at baseline, not detected
at treatment conclusion and transitioned once between presence
and absence over time in the patient’s samples, or vice versa (ie,
not detected and then detected), were considered to be candi-
date biomarkers for that patient. For example, a protein that is
detected in patient 1 at baseline and at weeks 1 and 5, and not
detected in patient 1 at weeks 9 and 13, would be considered a
candidate biomarker for patient 1. This will be referred to as the
pattern matching test for the detection of patterns consistent with
upregulation or downregulation of the protein as the patient un-
dergoes treatment.
4 | GUZMAN et al.

TA B L E 2 Candidate protein biomarkers with strong temporal ∑

zk = m
profiles k

Uniprot accession Protein DP HP |fp(13) − fp(0)| ξi ,μi ≥ 0, ∀i zk ∈ {0,1}, ∀k

P20160 Azurocidin 6 0 −0.707
P35579 Myosin-­9 6 0 −0.545 Briefly, indices i and k correspond to instances (ie, samples) and
P61626 Lysozyme C 5 0 −0.699 features (ie, protein biomarker candidates), respectively. Samples from
P62736|P63267 α-­sm actin | 0 5 0.554 the baseline are assigned yi = 1, while samples at week 13 are assigned
γ-­sm actin yi = −1. Parameter xik represents the (z-­score normalized) presence of
DP, disease patterns, protein downregulates as treatment course pro- candidate biomarker k in sample i. Binary variables zk determine if can-
gresses; HP, health patterns, protein upregulates as treatment course didate biomarker k is selected in the biomarker subset of size m. The
progresses. model thus selects the best m features by minimizing slack variables
ξi (which quantify prediction error) and maximizing slack variables μi
(which quantify correct prediction margin) by adjusting variables wk
TA B L E 3 Summary of cross-­validation studies with the
candidate biomarkers and b; only those wk whose zk = 1 are allowed to be non-­zero. With the
assigned values of yi, it can be seen that if wk > 0 in an optimal solution,
Fraction of
then the model has evaluated biomarker k as an indicator of disease,
Average accuracy Average AUC perfect partitions
and vice versa for wk < 0. The MILP model was solved to global opti-
0.900 0.930 0.800 mality to select the top 5 candidate biomarkers (m = 5); objective func-
AUC, area under the curve. Averages are over all 100 possible training-­ tion parameter C was varied from 8 to 32 with no change in the results.
testing partitions of the endpoint data. Perfect partitions are those with
perfect predictive accuracy and AUC.
2.8 | Cross-­validation studies
The 2 temporal biomarker selection filters yielded 4 candidate biomark-
Across all patients, the fractions of patients for which a protein
ers (see Table 2), which may be used to assess clinical endpoints for
p was detected as a function of time, fp (t) was fit to the logistic
patients with chronic periodontitis. The proteomic profile data for the
function
4 candidate biomarkers were then evaluated in cross-­validation stud-
ies. Data from the 4 candidates across all 10 patients at baseline and at
( )−1
fp (t) = 1 + exp(β0 + β1 t)

by performing logistic regression to find parameters β 0 and β1. week 13 were treated as instances in a classification learning problem

The parameters determine the rate of change and direction of to determine disease state. Performing 1 round 10-­fold cross-­validation

the logistic function as it transitions from 0 to 1 or from 1 to 0. involved dividing the data into 10 balanced segments, ie, with equal

Good candidate biomarkers would be expected in this scenario to numbers of instances from baseline and week 13. Each segment is then

transition somewhere between t = 0 and t = 13, and for much of utilized in turn as the testing set for evaluating a classification model

the transition to occur within this time range. Thus, proteins for constructed on the remaining 9 segments, ie, the training set. The low

which |fp (13) − fp (0)| ≥ 0.5 were kept as candidate biomarkers. This number of samples allows for the enumeration of every possible 10-­

will be referred to as the logistic function test for the detection fold segment partitioning; thus, 10 rounds of 10-­fold cross-­validation

of population-­level up-­or downregulation of proteins across the were performed, covering all 100 possible balanced training-­testing

samples. partitions where 90% of the data are used in training. All classifiers
consisted of simple linear classifiers via C-­parameterized support vec-
tor machines (C-­SVM)28,29 with linear kernels. Software implementa-
2.7 | Biomarker selection by mixed-­integer linear tion LIBSVM30 with C = 0.125 was used for all training and testing. The
optimization predictive accuracy and area under the curve (AUC)31 was evaluated

Data from the baseline and the endpoint for each patient was used for each partition, and the averages are summarized in Table 3.

to identify the best features for distinguishing the 2 extremes via a

MILP model for feature selection27:
3 | R E S U LT S
∑(
Clinical parameters for participants are summarized in Table 1. All
)
min Cξi − μi
w,b,ξ,μ,z
i
clinical parameters of periodontal disease were significantly dif-
ferent between baseline and 13 weeks after treatment (Wilcoxon
( )
∑
s.t. yi wk xik + b = 1 − ξi + μi ∀i signed ranks test, P < .05).
k
The numbers of proteins or protein groups identified in each
B
−w zk ≤ wk ≤ w zk ∀k B sample are shown in Table 4. When a set of peptides identified in a
GUZMAN et al. | 5

TA B L E 4 Numbers of proteins or protein groups identified per Table 2. Interestingly, these 4 proteins were also included in the top
sample 5 selected by the MILP model. The endpoint data (baseline and week

Week 13) involving the 4 candidates of Table 2 were utilized in the cross-­
validation studies. Averaged over all 100 training-­testing partitions,
Patient 0 1 5 9 13
the testing accuracy was 0.900 and the AUC was 0.930. Owing to
1 185 123 163 156 169 the low numbers of testing instances and qualitative nature of the
2 179 201 184 166 64 data, the results were highly sensitive to individual samples. Eighty
3 142 152 172 179 131 percent of the partitions had a testing accuracy and AUC of 1; it is
4 124 189 194 192 147 predicted that larger datasets and more quantitative data would fur-

5 144 179 154 163 176 ther increase these metrics for the candidate biomarkers.

6 171 152 169 183 144

7 112 183 164 X 108
8 146 180 120 179 169
9 133 204 177 139 141
10 126 150 68 55 123
sample are able to be attributed to more than 1 protein, the proteins
are kept together as a protein group and only contribute as 1 identi-
fication count in Table 4. These protein groups remained unbroken
even after the per-­sample protein inference methods implemented
in the PILOT_PROTEIN suite. The specific proteins identified per
patient and per sample are given in Table S1 of the Supporting in-
formation. An alternative protein inference algorithm based on prin-
ciples of parsimony was attempted, which took into account protein
identifications across samples. Briefly, over each set of samples per
week, protein groups were reduced to only include proteins, which
were identified as single proteins (ie, not within a group) elsewhere
in the data. This simplified a number of protein groups but did not
alter the results, and thus for simplicity was not included. Thus, in
the view of the temporal profiling, all proteins belonging to a group
were first ungrouped and considered to be present individually, and
were then regrouped post-­analysis given that they generated the
same temporal profile and were always found grouped together in
the initial data.
The MILP model utilized only baseline and endpoint data, which
was grouped, and the top 5 candidate biomarkers are shown in
Table 5. The model also determined which status, disease or health,
each candidate biomarker indicated. The model was solved in less
than 2 CPU seconds using GAMS (version 24.3) and the ILOG CPLEX
solver.
Four proteins or protein groups displayed strong temporal pro-
files, and were independently selected by both the pattern matching
and logistic function tests. The results of the 2 filters are shown in
4 | D I S CU S S I O N
From the perspective of the current analyses, the strongest candidate
biomarker for periodontitis is azurocidin, which was upregulated at
baseline and downregulated as the treatment course proceeded. No
azurocidin was detected in any patient at week 13 of the treatment
course. Azurocidin is a neutrophil granule-­derived protein known for
its antimicrobial activity, particularly against gram-­negative bacte-
ria.32,33 It is present in the azurophilic granules of polymorphonu-
clear neutrophils along with myeloperoxidase, cathepsin-­G , elastase,
lysozyme and other antimicrobial compounds. Owing to its cationic
nature, azurocidin is able to disrupt the structure of the outer mem-
brane of gram-­negative bacteria via binding with lipopolysaccha-
rides.33 This is particularly relevant to periodontitis, as all members
of the “red complex” group of bacteria, identified by Socransky and
coworkers as highly associated with periodontal disease, are gram-­
negative.34 Azurocidin has been previously identified in gingival cre-
vicular fluid samples but not in saliva collected from patients with
chronic periodontitis, using both LC-­MS/MS and enzyme-­linked im-
munosorbent assay analyses.7,35
The profile of the bacteriolytic protein lysozyme C was similar
to that of azurocidin. No lysozyme C was detected in any patient
at week 13 of the treatment course. Lysozyme is a cationic protein
component of gingival crevicular fluid and saliva with antimicrobial
activity against both gram-­positive and gram-­negative bacteria via
depriving microorganisms of iron.36-38 It has been detected via high-­
throughput proteomics studies in gingival crevicular fluid samples
obtained from patients with chronic periodontitis and gingivitis.6,9
No changes in gingival crevicular fluid levels of lysozyme have
been reported 2 weeks following surgical periodontal treatment.39
Though the periodontal treatments are not comparable, in the
present study, detection levels of lysozyme C were still high (80%

TA B L E 5 Candidate protein biomarkers

Indicator of periodontal
selected by the mixed-­integer linear
Uniprot accession Protein disease or health
optimization model
P20160 Azurocidin Disease
P35579 Myosin-­9 Disease
P61626 Lysozyme C Disease
P62736|P63267 α-­sm actin|γ-­sm actin Health
P60174 Triosephosphate isomerase Health
6 |
incidence) after 5 weeks of treatment before dropping to 50% and
then 0% over the next 2 time points.
It should be noticed, that in the present study, other important
constituents of polymorphonuclear neutrophil granules such as my-
eloperoxidase or cathepsin G have not been identified as biomarkers
for evaluation of periodontal tissue status. A possible explanation
for this observation is that these important mediators of inflamma-
tion in the periodontal pockets are at low abundance compared to
highly abundant serum-­related proteins in gingival crevicular fluid
such as albumin or immunoglobulins and were therefore masked and
not easily identifiable by the current methodologies usually applied
for sample processing and analysis.18
Myosin 9 is a part of the myosin family of proteins, which are re-
sponsible for actin-­based motility. The myosin-­9 heavy chain is one
subunit of the myosin IIA protein. Myosin 9 has not been extensively
investigated in the literature and is considered a non-­muscle myosin.
It has been detected with high-­throughput proteomics in gingival
crevicular fluid in patients with chronic periodontitis and generalized
aggressive periodontitis,7,11,40 and patients with maintenance-­phase
chronic periodontitis12 but its possible role in periodontitis remains
unclear. An interesting possibility is that while some cells use more
than one type of myosin II, certain blood cells such as platelets and
white blood cells (leukocytes) use only myosin IIA and therefore the
constant detection of myosin-­9 from patients with periodontitis,
particularly at baseline, might be a marker of the well-­known abun-
dance of various types of leukocytes in the periodontal pocket.41
After periodontal treatment, and during the resolution of inflamma-
tion, numbers of inflammatory cells reduce, which might account for
the depletion of myosin-­9 at later time points.
The only strong temporal profile that was upregulated as treat-
ment progressed was the protein group composed of alpha-­smooth
muscle actin and gamma-­smooth muscle actin. These proteins have
an almost identical sequence (99.2% sequence similarity) and are
thus highly unlikely to be distinguished from each other by MS/
MS analysis. It is likely that the attributed peptides originate from
alpha-­smooth muscle actin, as it is known to be expressed by peri-
42
odontal ligament fibroblasts and was detected in patients with
maintenance-­phase chronic periodontitis. 12
Alpha-­smooth muscle
actin enables contraction in mesenchymal stem cells and in their con-
nective tissue progeny,43 and can upregulate fibroblast contractility
in wound healing,44 while collagen gel contraction is dependent on
the expression of this molecule.45 Most interestingly, alpha-­smooth
muscle actin has been shown to be a marker for a fibroblast subtype
that can rapidly remodel the extracellular matrix.45
The MILP model only utilized endpoint data and, with m = 5,
selected the 4 temporal biomarker candidates described above
as well as the glycolytic enzyme triosephosphate isomerase. The
model assigned triosephosphate isomerase as an indicator of
health, and the enzyme was observed in 80% of patients at week
13. Triosephosphate isomerase was eliminated by the other tests
due to the high detection rate (90%) at week 1. Triosephosphate
isomerase has been detected in saliva in healthy patients,46 and was
found to be significantly downregulated in diseased periodontal
GUZMAN et al.
pocket tissues from patients with chronic periodontitis.47 More re-
search must be done to determine the exact role of host or bacte-
rial triosephosphate isomerase in periodontitis, as existing studies
have focused on triosephosphate isomerase and other glycolytic
enzymes in oral bacteria, which are associated with adhesion.48,49
In conclusion, in this study, albeit the low number of partici-
pants and the absence of quantitative data, an untargeted high-­
throughput proteomics platform was used to generate temporal
proteomic profiles of patients undergoing treatment for chronic
periodontitis. Filtering techniques including logistic function
fitting and pattern matching were used to identify azurocidin,
lysozyme C, myosin 9 and smooth muscle actin as candidate bio-
markers, which can aide in the determination of clinical endpoints
for chronic periodontitis. These 4 biomarker candidates were also
independently selected by applying an MILP model for feature
selection and biomarker discovery. Cross-­validation studies using
the proteins and simple classifiers yielded high patient predictive
accuracy and AUC. These proteins have biological support in the
literature for their importance in chronic periodontitis progression
and recovery. The candidates can be the subjects of targeted stud-
ies of the mechanisms of periodontitis and the evaluation of their
clinical efficacy.
AC K N OW L E D G E M E N T S
CAF acknowledges financial support from the National Science
Foundation (CBET-­0941143) and the National Institutes of Health
(R01LM009338). The authors have stated explicitly that there are no
conflicts of interest in connection with this article.
ORCID
D. Sakellari http://orcid.org/0000-0002-4365-8406
REFERENCES
1. Zhang L, Henson BS, Camargo PM, Wong DT. The clinical value
of salivary biomarkers for periodontal disease. Periodontol 2000.
2009;51:25‐37.
2. Loos BG, Tjoa S. Host-­derived diagnostic markers for periodon-
titis: do they exist in gingival crevice fluid? Periodontol 2000.
2005;39:53‐72.
3. Buduneli N, Kinane DF. Host-­derived diagnostic markers related to
soft tissue destruction and bone degradation in periodontitis. J Clin
Periodontol. 2011;38:85‐105.
4. Guzman YA, Sakellari D, Arsenakis M, Floudas CA. Proteomics for
the discovery of biomarkers and diagnosis of periodontitis: a critical
review. Expert Rev Proteomics. 2014;11:31‐41.
5. Delima AJ, Van Dyke TE. Origin and function of the cellular compo-
nents in gingival crevice fluid. Periodontol 2000. 2003;31:55‐76.
6. Bostanci N, Ramberg P, Wahlander Å, et al. Label-­free quantitative
proteomics reveals differentially regulated proteins in experimental
gingivitis. J Proteome Res. 2013;12:657‐678.
7. Choi YJ, Heo SH, Lee JM, Cho JY. Identification of azurocidin as a
potential periodontitis biomarker by a proteomic analysis of gingival
crevicular fluid. Proteome Sci. 2011;9:42.
GUZMAN et al.
8. Baliban RC, Sakellari D, Li Z, DiMaggio PA, Garcia BA, Floudas CA.
Novel protein identification methods for biomarker discovery via a
proteomic analysis of periodontally healthy and diseased gingival
crevicular fluid samples. J Clin Periodontol. 2012;39:203‐212.
9. Baliban RC, Sakellari D, Li Z, Guzman YA, Garcia BA, Floudas CA.
Discovery of biomarker combinations that predict periodontal
health or disease with high accuracy from GCF samples based on
high-­throughput proteomic analysis and mixed-­integer linear opti-
mization. J Clin Periodontol. 2013;40:131‐139.
10. Tsuchida S, Satoh M, Kawashima Y, et al. Application of quantitative
proteomic analysis using tandem mass tags for discovery and iden-
tification of novel biomarkers in periodontal disease. Proteomics.
2013;13:2339‐2350.
11. Bostanci N, Heywood W, Mills K, Parkar M, Nibali L, Donos N.
Application of label-­free absolute quantitative proteomics in human
gingival crevicular fluid by LC/MSE (gingival exudatome). J Proteome
Res. 2010;9:2191‐2199.
12. Ngo LH, Veith PD, Chen YY, Chen D, Darby IB, Reynolds EC. Mass
spectrometric analyses of peptides and proteins in human gingival
crevicular fluid. J Proteome Res. 2010;9:1683‐1693.
13. Ngo LH, Darby IB, Veith PD, Locke AG, Reynolds EC. Mass
spectrometric analysis of gingival crevicular fluid biomarkers
can predict periodontal disease progression. J Periodontal Res.
2013;48:331‐341.
14. Castagnola M, Cabras T, Iavarone F, et al. The human salivary pro-
teome: a critical overview of the results obtained by different pro-
teomic platforms. Expert Rev Proteomics. 2012;9:33‐46.
15. Messana I, Inzitari R, Fanali C, Cabras T, Castagnola M. Facts and
artifacts in proteomics of body fluids. What proteomics of saliva is
telling us? J Sep Sci. 2008;31:1948‐1963.
16. Siqueira WL, Dawes C. The salivary proteome: challenges and per-
spectives. Proteomics Clin Appl. 2011;5:575‐579.
17. Cabras T, Pisano E, Boi R, et al. Age-­dependent modifications of
the human salivary secretory protein complex. J Proteome Res.
2009;8:4126‐4134.
18. Bostanci N, Bao K. Contribution of proteomics to our understand-
ing of periodontal inflammation. Proteomics. 2017;17:3‐4.
19. Armitage GC. Development of a classification system for periodon-
tal diseases and conditions. Ann Periodontol. 1999;4:1‐6.
20. Tonetti M, Claffey N. Advances in the progression of periodon-
titis and proposal of definitions of a periodontitis case and dis-
ease progression for use in risk factor research. J Clin Periodontol.
2005;32:210‐213.
21. Chambers MC, Maclean B, Burke R, et al. A cross-­platform tool-
kit for mass spectrometry and proteomics. Nat Biotechnol.
2012;30:918‐920.
22. Deutsch EW, Mendoza L, Shteynberg D. A guided tour of the Trans-­
Proteomic Pipeline. Proteomics. 2010;10:1150‐1159.
23. DiMaggio PA, Floudas CA. De novo peptide identification via tan-
dem mass spectrometry and integer linear optimization. Anal Chem.
2007;79:1433‐1446.
24. DiMaggio PA, Floudas CA, Lu B, Yates I. A hybrid method for pep-
tide identification using integer linear optimization, local database
search, and quadrupole time-­of-­flight or OrbiTrap tandem mass
spectrometry. J Proteome Res. 2008;7:1584‐1593.
25. Baliban RC, DiMaggio PA, Plazas-Mayorca MD, Garcia BA, Floudas
CA. PILOT_PROTEIN: identification of unmodified and modified
proteins via high-­resolution mass spectrometry and mixed-­integer
linear optimization. J Proteome Res. 2012;11:4615‐4629.
26. Baliban RC, DiMaggio PA, Plazas-Mayorca MD, Young NL,
Garcia BA, Floudas CA. A novel approach for untargeted post-­
translational modification identification using integer linear op-
timization and tandem mass spectrometry. Mol Cell Proteomics.
2010;9:764‐779.
| 7
27. Guzman YA. Theoretical advances in robust optimization, feature
selection, and biomarker discovery. 2016, Doctoral dissertation,
Princeton University.
28. Cortes C, Vapnik V. Support-­ vector networks. Mach Learn.
1995;20:273‐297.
29. Vapnik VN. The Nature of Statistical Learning Theory. New York:
Springer; 1995.
30. Chang CC, Lin CJ. LIBSVM: a library for support vector machines.
ACM Trans Intell Syst Technol. 2011;27:27.
31. Hanley JA, McNeil BJ. The meaning and use of the area under
a receiver operating characteristic (ROC) curve. Radiology.
1982;143:29‐36.
32. Campanelli D, Detmers PA, Nathan CF, Gabay JE. Azurocidin and a
homologous serine protease from neutrophils. Differential antimi-
crobial and proteolytic properties. J Clin Investig 1990;85:29‐36.
33. Almeida RP, Vanet A, Witko-Sarsat V, Melchior M, McCabe D,
Gabay JE. Azurocidin, a natural antibiotic from human neutrophils:
expression, antimicrobial activity, and secretion. Protein Expr Purif.
1996;7:355‐366.
34. Socransky S, Haffajee A, Cugini M, Smith C, Kent R. Microbial com-
plexes in subgingival plaque. J Clin Periodontol. 1998;25:134‐144.
35. Leppilahti JM, Hernández-Ríos PA, Gamonal JA, et al. Matrix metal-
loproteinases and myeloperoxidase in gingival crevicular fluid pro-
vide site-­specific diagnostic value for chronic periodontitis. J Clin
Periodontol. 2014;41:348‐356.
36. Friedman SA, Mandel ID, Herrera MS. Lysozyme and lactoferrin
quantitation in the crevicular fluid. J Periodontol. 1983;54:347‐350.
37. Perera S, Sabin E, Nelson P, Lowe D. Increases in salivary lysozyme
and IgA concentrations and secretory rates independent of salivary
flow rates following viewing of a humorous videotape. Int J Behav
Med. 1998;5:118‐128.
38. West NP, Pyne DB, Renshaw G, Cripps AW. Antimicrobial peptides
and proteins, exercise and innate mucosal immunity. FEMS Immunol
Med Microbiol. 2006;48:293‐304.
39. Jentsch H, Sievert Y, Göcke R. Lactoferrin and other markers from
gingival crevicular fluid and saliva before and after periodontal
treatment. J Clin Periodontol. 2004;31:511‐514.
40. Kido J, Bando M, Hiroshima Y, et al. Analysis of proteins in human
gingival crevicular fluid by mass spectrometry. J Periodontal Res.
2012;47:488‐499.
41. Berglundh T, Zitzmann NU, Donati M. Are peri-­ implantitis le-
sions different from periodontitis lesions? J Clin Periodontol.
2011;38:188‐202.
42. Giannopoulou C, Cimasoni G. Functional characteristics of gingival
and periodontal ligament fibroblasts. J Dent Res. 1996;75:895‐902.
43. Kinner B, Zaleskas J, Spector M. Regulation of smooth muscle actin
expression and contraction in adult human mesenchymal stem cells.
Exp Cell Res. 2002;278:72‐83.
44. Hinz B, Celetta G, Tomasek JJ, Gabbiani G, Chaponnier C. Alpha-­
smooth muscle actin expression upregulates fibroblast contractile
activity. Mol Biol Cell. 2001;12:2730‐2741.
45. Arora PD, McCulloch CA. Dependence of collagen remodelling on
α-­smooth muscle actin expression by fibroblasts. J Cellular Physiol.
1994;159:161‐175.
46. Jagtap P, McGowan T, Bandhakavi S, et al. Deep metaproteomic anal-
ysis of human salivary supernatant. Proteomics. 2012;12:992‐1001.
47. Bertoldi C, Bellei E, Pellacani C, et al. Non bacterial protein expres-
sion in periodontal pockets by proteome analysis. J Clin Periodontol.
2013;40:573‐582.
48. Fletcher JM, Nair SP, Ward JM, Henderson B, Wilson M. Analysis
of the effect of changing environmental conditions on the expres-
sion patterns of exported surface-­associated proteins of the oral
pathogen Actinobacillus actinomycetemcomitans. Microb Pathog.
2001;30:359‐368.
8 |
49. Kinnby B, Booth NA, Svensäter G. Plasminogen binding by
oral streptococci from dental plaque and inflammatory lesions.
Microbiology. 2008;2008:924‐931.
S U P P O R T I N G I N FO R M AT I O N
Additional supporting information may be found online in the
Supporting Information section at the end of the article.
GUZMAN et al.
How to cite this article: Guzman YA, Sakellari D,
Papadimitriou K, Floudas CA. High-­throughput proteomic
analysis of candidate biomarker changes in gingival crevicular
fluid after treatment of chronic periodontitis. J Periodont Res.
2018;00:1–8. https://doi.org/10.1111/jre.12575