Improving Detection Accuracy of Lung Cancer Serum Proteomic Profiling via Two-Stage Training Process.

Abstract

Surface-Enhanced Laser Desorption/Ionization Time-of-Flight Mass Spectrometry (SELDI-TOF-MS) is a frequently used technique for cancer biomarker research. The specificity of biomarkers detected by SELDI can be influenced by concomitant inflammation. This study aimed to increase detection accuracy using a two-stage analysis process.
Sera from 118 lung cancer patients, 72 healthy individuals, and 31 patients with inflammatory disease were randomly divided into training and testing groups by 3:2 ratio. In the training group, the traditional method of using SELDI profile analysis to directly distinguish lung cancer patients from sera was used. The two-stage analysis of distinguishing the healthy people and non-healthy patients (1st-stage) and then differentiating cancer patients from inflammatory disease patients (2nd-stage) to minimize the influence of inflammation was validated in the test group.
In the test group, the one-stage method had 87.2% sensitivity, 37.5% specificity, and 64.4% accuracy. The two-stage method had lower sensitivity (> 70.1%) but statistically higher specificity (80%) and accuracy (74.7%). The predominantly expressed protein peak at 11480 Da was the primary splitter regardless of one- or two-stage analysis. This peak was suspected to be SAA (Serum Amyloid A) due to the similar m/z countered around this area. This hypothesis was further tested using an SAA ELISA assay.
Inflammatory disease can severely interfere with the detection accuracy of SELDI profiles for lung cancer. Using a two-stage training process will improve the specificity and accuracy of detecting lung cancer.

Full text

Page 1

RESEARCH Open Access
Improving Detection Accuracy of Lung Cancer
Serum Proteomic Profiling via Two-Stage Training
Process
Pei-Sung Hsu1, Yu-Shan Wang2, Su-Chen Huang2, Yi-Hsien Lin3, Chih-Chia Chang2, Yuk-Wah Tsang2,
Jiunn-Song Jiang1, Shang-Jyh Kao1, Wu-Ching Uen4*and Kwan-Hwa Chi2,5*
Abstract
Background: Surface-Enhanced Laser Desorption/Ionization Time-of-Flight Mass Spectrometry (SELDI-TOF-MS) is a
frequently used technique for cancer biomarker research. The specificity of biomarkers detected by SELDI can be
influenced by concomitant inflammation. This study aimed to increase detection accuracy using a two-stage
analysis process.
Methods: Sera from 118 lung cancer patients, 72 healthy individuals, and 31 patients with inflammatory disease
were randomly divided into training and testing groups by 3:2 ratio. In the training group, the traditional method
of using SELDI profile analysis to directly distinguish lung cancer patients from sera was used. The two-stage
analysis of distinguishing the healthy people and non-healthy patients (1st-stage) and then differentiating cancer
patients from inflammatory disease patients (2nd-stage) to minimize the influence of inflammation was validated in
the test group.
Results: In the test group, the one-stage method had 87.2% sensitivity, 37.5% specificity, and 64.4% accuracy. The
two-stage method had lower sensitivity (> 70.1%) but statistically higher specificity (80%) and accuracy (74.7%). The
predominantly expressed protein peak at 11480 Da was the primary splitter regardless of one- or two-stage
analysis. This peak was suspected to be SAA (Serum Amyloid A) due to the similar m/z countered around this area.
This hypothesis was further tested using an SAA ELISA assay.
Conclusions: Inflammatory disease can severely interfere with the detection accuracy of SELDI profiles for lung
cancer. Using a two-stage training process will improve the specificity and accuracy of detecting lung cancer.
Keywords: SELDI SAA, lung cancer
Background
Lung cancer is the most common malignancy in the
world, with poor overall 5-year survival rate [1].
Although advances in non-invasive imaging have
improved its detection, > 75% of lung cancer patients
present with advanced stages, when therapeutic options
are limited [2]. Even patients who present with clinical
stage I have at best a 60% 5-year survival rate, signifying
that a large percentage of patients regardless of stage
have undetectable metastatic disease at the time of pre-
sentation [2,3]. Early detection is the most important
factor to improve survival. However, the current tools
used for early detection are non-satisfactory in accuracy
and timing, despite routine chest X-ray, computed
tomography, bronchoscopy, sputum cytology, and tumor
markers. Current serum biomarkers are convenient but
lack of adequate sensitivity and/or specificity [4-8]. Bet-
ter biomarkers are needed for screening, follow-up, and
differential diagnosis.
SELDI-TOF-MS is currently one of the techniques
used for cancer biomarker screening, including ovarian,
breast, prostate, and liver cancers [9-13]. SELDI profiling
reportedly detects lung cancer from healthy individuals
* Correspondence: M002047@ms.skh.org.tw; M006565@ms.skh.org.tw
2Division of Radiation Therapy and Oncology, Shin Kong Wu Ho-Su
Memorial Hospital, Taipei, Taiwan
4Division of Hematology and Oncology, Shin Kong Wu Ho-Su Memorial
Hospital, Taipei, Taiwan
Full list of author information is available at the end of the article
Hsu et al. Proteome Science 2011, 9:20
http://www.proteomesci.com/content/9/1/20
© 2011 Hsu et al; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in
any medium, provided the original work is properly cited.

Page 2

[14-20] and most of the SELDI profiling comes from
analyzing samples from cancer patients and healthy
individuals. However, inflammatory reaction is com-
monly associated with lung cancer diagnosis. Many
identified serum peaks turn out to be the fragments of
acute phase proteins. Serum amyloid A (SAA) is a sam-
ple of acute phase protein that has been detected by
SELDI study. It belonged to a family of apolipoproteins
and is reported as a serum biomarker of lung cancer
and other malignancies, as well as many inflammatory
diseases [14,21-27].
A positive association between previous lung disease,
such as chronic bronchitis or pneumonia, with lung can-
cer has been found [28-30]. Pneumonia frequently co-
exists when lung cancer is diagnosed. A study shows
that > 50% of newly-proved lung cancer patients have
been diagnosed within 3 months after hospitalization for
pneumonia [31]. A history of previous lung disease,
including asthma, chronic bronchitis, pneumonia, and
tuberculosis, is associated with significantly increased
risk of lung cancer [28,32].
Thus, differentiating lung cancer from inflammatory
disease is an important concern. The impact of inflam-
matory disease on the detection of lung cancer by
SELDI profile has not been highlighted. This study
aimed to improve the detection accuracy of lung cancer
by SELDI in mixed individuals, those who are healthy
and those with inflammatory disease.
Methods
Patients and samples
The serum samples from cancer patients were collected
from the Department of Chest Medicine of Shin Kong
Wu Ho-Su Memorial Hospital. Samples from patients
with other diseases and healthy volunteers were col-
lected from the Department of Health Management dur-
ing this same period. The ethics committee/institutional
review board approved the study protocol and all
patients provided informed consent.
Each sample was collected into a 10 ml serum separa-
tor vacutainer tube and laid up at 4°C for 3 h, and then
centrifuged for 10 min at 3000 rpm. The serum was dis-
tributed into 100 μl aliquots and stored frozen at -80°C
until analysis. A total of 221 serum specimens were col-
lected. In the 118 cancer patients (80 males and 38
females), histologic distribution were 27 squamous cell
carcinomas, 71 adenocarcinomas, 9 small cell cancers,
and 11 others. Patients with inflammation (23 males and
8 females) included 15 pneumonia, 6 tuberculosis, and
10 urinary tract infection patients. None of them had
any evidence of malignancy. The 72 healthy controls (39
males and 33 females) came from general survey of
health were used as normal control and had no cancer
history or acute/chronic inflammatory disease.
The serum specimens were randomly assigned to a
training group and testing group by a 3:2 ratio from the
three sets of patients. Thus, the training group had 71 can-
cer patients (40 adenocarcinoma, 19 squamous cell carci-
noma, 5 small cell carcinoma, and 7 other types), 46
healthy controls, and 17 inflammatory disease patients.
The testing group has 47 cancer patients (31 adenocarci-
noma, 8 squamous cell carcinoma, 4 small cell carcinoma,
and 4 other types). The mean age of the training and test-
ing groups was 60 ± 14 and 59 ± 15 years, respectively
(p < 0.39) and the ratio of male to female was 59%:41%
and 67%:33%, respectively (p < 0.1), Two groups were
matched for sex and age. The demographic information
for all collected samples was provided in Table 1.
SELDI protein profiling
For analysis, ProteinChip Arrays (CM10) with anionic
surface chemistry was used. The CM10 ProteinChip
Arrays incorporated a carboxylate group that acted as a
weak cation exchanger. The chips were put in a
Table 1 Patient characteristics
Histological classification Training group
No.
F
Testing group
No.
F
age age
MM
Lung cancerAdenocarcinoma (n = 1) 22 68(52-91)
62(41-86)
68(50-85)
69(56-80)
68(64-85)
59(28-74)
53(51-55)
1761(32-76)
61(33-76)
65(60-80)
50(50)
53(48-61)
68(50-81)
1814
Squamous cell carcinoma (n = 7)167
31
Small cell carcinoma (n = 9)
Other (n = 11)
5
5
4
4
2
Non-malignancyHealthy (n = 72)23 54(42-67)
51(32-73)
58(36-86)
43(20-74)
1652(34-81)
46(33-63)
58(27-84)
71(53-92)
23 10
Inflammatory disease (n = 31)1211
53
Hsu et al. Proteome Science 2011, 9:20
http://www.proteomesci.com/content/9/1/20
Page 2 of 10

Page 3

bioprocessor (Ciphergen Biosystems, Inc.), which
allowed larger volumes of serum to each chip array.
Within the bioprocessor, the chips were equilibrated for
15 min in 150 μl binding/washing buffer (100 mM
sodium acetate, pH 4.0) per well. Six-microliter serum
samples were denatured in 12 μl urea buffer (9 M urea,
2% CHAPS, and 50 mM Tris). Then, 220 μl of binding
buffer were added to the serum mixture, and 100 μl of
the denatured, diluted serum were randomly applied to
each chip spot. The bioprocessor was then sealed and
shaken at a speed of 400 rpm on a platform shaker for
30 min.
After the incubation, the bioprocessor was washed
with 150 μl of washing buffer in each well. This step
was repeated three times, and each time the binding
buffer was discarded by inverting the bioprocessor on a
paper towel. The bioprocessor was rinsed with 1 mmol/l
HEPES and the chips were removed from the bioproces-
sor. After the arrays were air-dried, 1 μl of 50% satu-
rated matrix solution (sinapinic acid in 50% acetonitrile
and 0.5% trifluoroacetic acid) was applied on the array
and allowed to air dry. Arrays were analyzed using a
ProteinChip Reader (ProteinChip PBS IIC, Ciphergen
Biosystems Inc.), with the following settings: laser inten-
sity 185, detector sensitivity 9, and molecular mass
range 2,000-20,000 Da. Analysis were performed in
duplicate on serum specimens from each patient in the
study group. Mass accuracy was calibrated externally
using the all-in-one peptide molecular mass standard
(Ciphergen Biosystems, Inc.).
Peak Detection, Data Analysis, and Decision Tree
Classification
Peak detection was performed using Ciphergen Pro-
teinChip Software, version 3.0.2 (Ciphergen Biosystems).
The spectra were normalized to total ion current inten-
sity of m/z between 2,000 and 20,000, and baselines
were subtracted. The part of the spectrum with m/z
values < 2,000 was eliminated for analysis, since the
energy absorbing matrix signal usually interfered with
peak detection in this range. ProteinChip Software auto-
detected the peaks to clusters between 2,000 and 20,000
m/z ratios with a signal-to-noise ratio of > 5. The clus-
ter mass window was 0.3% of mass. The peaks clustered
using second-pass peak selection with signal-to-noise
ratio > 2. The M/Z values that were within the 0.3%
mass accuracy window were regarded as equal between
replicates. In order to minimize data variation data
between multiple samples, the peak intensity values
were logarithmically transformed.
Based on the peak intensities of each signal cluster,
three decision trees were constructed from the training
groups. For each sample, the intensity values for each
peak within the 2,000-20,000 m/z range were inputted
into the Biomarker Patterns Software (Ciphergen Biosys-
tems) and classified according to the tree analysis
described. A data set was divided into two nodes by tree
analysis pattern, using one rule at a time in the form of
a question. The presence or absence, and the intensity
levels of one peak defined the splitting decision. This
splitting process continued until the terminal nodes or
leaves were produced or further splitting had no gain.
Classification of terminal nodes was determined by the
group (“class”) of samples (i.e., lung cancer, healthy, or
inflammatory) representing the majority of samples in
that node. Peaks selected by this process formed the
splitting rules to achieve the maximum reduction of
cost in the two descendant nodes.
Serum SAA determination
Serum levels of SAA were measured in human serum
samples using a commercial enzyme-linked immuno-
sorbent assay (ELISA) kit (Invitrogen, Camarillo, CA,
USA), according to the manufacturers’ instructions.
ELISA tests were performed in triplicate on serum spe-
cimens from each patient in the study group. SAA levels
were assayed in serum diluted by a factor of 200. If the
absorbance readings exceeded the linear range of the
standard curves, the ELISA assay was repeated after
serial dilution of the supernatants.
Statistical analysis
The SPSS software version 13.0 (SPSS INC.,Chicago, IL,
USA) or Ciphergen ProteinChip Software, version 3.0.2
was used for statistical analysis and results were
expressed as mean ± SD. Statistical significance was ana-
lyzed by using a two-tailed Student’s t test and p < 0.05
was considered statistically significant. The discrimina-
tory power for each putative marker was shown using
the receiver operating characteristics (ROC) area under
the curve (AUC).
Results
Reproducibility
To examine the reproducibility of the assay system,
pooled normal sera from 60 individuals were used as
control samples. Ten protein peaks randomly selected
over the course of the study were used to calculate the
coefficient of variance (CV) as described previously [33].
The reproducibility of the SELDI spectra was deter-
mined, both within and between arrays (intra-assay and
inter-assay, respectively). The inter-assay (chip to chip)
CV was 17.9% (14.9% to 20.3%) for peak intensity and
0.08% (0.056% to 0.102%) for mass accuracy (peak
range: 5902-17340). The minimal variation of day-to-day
instrumentation was shown in Figure 1.
Hsu et al. Proteome Science 2011, 9:20
http://www.proteomesci.com/content/9/1/20
Page 3 of 10

Page 4

Detection of lung cancers by one stage SELDI profile
Serum samples were first analyzed from the training
group using 134 random samples with or without lung
cancer, using the SELDI Protein-Chip system. Peaks
were detected automatically after baseline subtraction
using Ciphergen ProteinChip Software, version 3.0.2.
This analysis identified 51 signal peak protein clusters,
seen in the spectrum representations of the two groups
(cancer and non-cancer) within the 2,000 to 20,000 m/z
range. No single peak could adequately discriminate
lung cancer sera from healthy sera and inflammatory
disease patients.
Using the spectrum, a decision tree classification algo-
rithm was built and three protein peaks at 11480, 8802
and 3185 Da were automatically selected as splitters
(p < 0.05, for each). The 11480 Da peak was used as the
Figure 1 Day-to-day verification of SELDI peaks at 9418 and 15113Da in the pooled normal sera. The representative peaks from Days 1
to 5 are showed. Red lines, peaks used for calibration
Hsu et al. Proteome Science 2011, 9:20
http://www.proteomesci.com/content/9/1/20
Page 4 of 10

Page 5

root node in the classification tree to divide the 134
samples into two groups (Figure 2A). The left node
(node 2) included cases with peak intensity ≦ 0.53. The
right node (node 3) contained the remaining with peak
intensity ≦-0.061. Finally, all 134 cases in the training
set were classified in the 4 terminal nodes, and a classifi-
cation tree was obtained (Figure 2A). The sensitivity and
specificity of the training set were 81.7% (58 of 71) and
85.7% (54 of 63), respectively. When the validity of this
classification tree algorithm was challenged by the test
set (87 cases), there was still high sensitivity (87.2%, 41
of 47) but low specificity (37.5%, 15 of 40) and accuracy
(64.4%, 56 of 87) (Table 2).
The training set was further divide into two cohorts.
The first cohort was lung cancer patients and healthy
control. The decision tree was shown in Figure 2B. The
11480 Da peak was used as the root node in the classifi-
cation tree and followed three peaks to classify the
cancer and healthy cohorts. The sensitivity and specifi-
city of the training set were 88.7% (63 of 71) and 89.1%
(41 of 46), respectively, and 83% (39 of 47) and 84.6%
(22 of 26), respectively, as validated in the test group.
The second cohort was lung cancer and inflammation
patients. The decision tree was classified by three peaks
as shown in Figure 2C. The 12580 Da peak was used as
the root node in this classification tree. The sensitivity
and specificity from the training set were 94.4% (67 of
71) and 70.6% (12 of 17), respectively, and 85.1% (40 of
47) and 57.1% (8 of 14), respectively, in the test set.
Detection of lung cancers by two stage SELDI profile
To further improve the outcome of lung cancer detec-
tion rate, the two original cohorts shown in Figures 2B
and C were re-combined into a two-stage training set as
shown in Figure 3. The validity of this classification tree
algorithm was then challenged by the test set. The first
Figure 2 Decision tree algorithms from three different training protocols. (A) Serum samples of individuals with lung cancer and
inflammatory disease, and healthy controls. (B) Serum samples of individuals with lung cancer and healthy controls. (C) Serum samples of
individuals with lung cancer and inflammatory disease.
Hsu et al. Proteome Science 2011, 9:20
http://www.proteomesci.com/content/9/1/20
Page 5 of 10