Theranostics 2019, Vol.

9, Issue 20 5090

Ivyspring
International Publisher
Journal of Cancer
2019; 10(21): 5090-5098. doi: 10.7150/jca.30528
Research Paper

Diagnostic Value of Plasma MicroRNAs for Lung Cancer

Using Support Vector Machine Model
Wei Wang1,2, Mingcui Ding1, Xiaoran Duan3, Xiaolei Feng1, Pengpeng Wang1, Qingfeng Jiang4, Zhe
Cheng5, Wenjuan Zhang6, Songcheng Yu7, Wu Yao1, Liuxin Cui3, Yongjun Wu7, Feifei Feng8, Yongli
Yang9
1. Department of Occupational Health and Occupational Disease, College of Public Health, Zhengzhou University, Zhengzhou, China
2. The Key Laboratory of Nanomedicine and Health Inspection of Zhengzhou, Zhengzhou, China
3. Department of Environmental Health, College of Public Health, Zhengzhou University, Zhengzhou, China
4. Department of Thoracic Surgery, the Affiliated Cancer Hospital of Zhengzhou University (Henan Cancer Hospital), Zhengzhou, China
5. Department of Respiratory Medicine, The First Affiliated Hospital of Zhengzhou University, Zhengzhou, China
6. Department of Public Health and Preventive Medicine, School of Medicine, Jinan University, Guangzhou, China
7. Department of Sanitary Chemistry, College of Public Health, Zhengzhou University, Zhengzhou, China
8. Department of Health Toxicology, College of Public Health, Zhengzhou University, Zhengzhou, China
9. Department of Epidemiology and Biostatistics, College of Public Health, Zhengzhou University, Zhengzhou, China

 Corresponding author: Yongli Yang, PhD. Department of Epidemiology and Biostatistics, College of Public Health, Zhengzhou University, Zhengzhou,
China. Tel: +86 371 67781466; Fax: +86 371 67781868; E-mail: ylyang377@126.com

© The author(s). This is an open access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/).
See http://ivyspring.com/terms for full terms and conditions.

Received: 2018.10.09; Accepted: 2019.06.25; Published: 2019.08.28

Abstract
Aim: Small single-stranded non-coding RNAs (miRNAs) play an important role in carcinogenesis through
degrading target mRNAs. However, the diagnostic value of miRNAs was not explored in lung cancers. In
this study, a support-vector-machine (SVM) model for diagnosis of lung cancer was established based on
plasma miRNAs biomarkers, clinical symptoms and epidemiology material.
Methods: The expressions of plasma miRNA were examined with SYBR Green-based quantitative
real-time PCR.
Results: We identified that the expressions of 10 plasma miRNAs (miR-21, miR-20a, miR-210, miR-145,
miR-126, miR-223, miR-197, miR-30a, miR-30d, miR-25), smoking status, fever, cough, chest pain or
tightness, bloody phlegm, haemoptysis, were significantly different between lung cancer and control
groups (P<0.05). The accuracies of the combined SVM, miRNAs SVM, symptom SVM, combined Fisher,
miRNAs Fisher and symptom Fisher were 96.34%, 80.49%, 84.15%, 84.15%, 75.61%, and 80.49%,
respectively; AUC of these six model were 0.976, 0.841, 0.838, 0.865, 0.750, and 0.801, respectively. The
accuracy and AUC of combined SVM were higher than the other 5 models (P<0.05).
Conclusions: Our findings indicate that SVM model based on plasma miRNAs biomarkers may serve as
a novel, accurate, noninvasive method for auxiliary diagnosis of lung cancer.
Key words: Lung cancer, Plasma miRNAs, Support vector machine, Diagnosis

Introduction
Lung cancer is currently the number one cause of three major histological subtypes [3]. Yet the NSCLC
morbidity and mortality worldwide [1], which has patients at an early stage with no obvious clinical
been classified into small cell lung cancer (SCLC) and symptoms, and lack of sensitive biomarkers and
non-small-cell lung cancer (NSCLC). NSCLC could effective tools for early diagnosis, therefore more than
account for 85% of all lung cancers to become the 75% of NSCLC patients are still diagnosed at
main subgroup of lung cancer [2]. NSCLCs can be advanced stages with distant metastases [4]. Although
further divided into adenocarcinoma (AC), squamous novel therapies are improving the survival of lung
cell carcinoma (SCC) and large cell carcinoma (LCLC) cancer patients, 5-year survival rate was still less than

http://www.thno.org
Journal of Cancer 2019, Vol. 10 5091

15% for advanced NSCLCs. However, the 5-year Classification is part of the important functions of
survival rate was up to 80% for the initial stage data mining, which is often closely related to disease
NSCLCs [5]. Therefore, early diagnosis and early diagnosis. At present, data mining is used primarily
screening of lung cancer are particularly important. in the field of auxiliary diagnosis of diseases [18]. DM
Lung cancer is diagnosed by means of techniques include SVM, artificial neural networks
histological examination, diagnostic imaging, (ANN), decision tree (DT), genetic algorithms and so
low-dose spiral computed tomography (LDSCT) and on. SVM is a pattern recognition method based on
positron emission tomography (PET). Although these statistical learning theory (SLT) and structural risk
techniques have been improved, they still have some minimization, which has several advantages such as
limitations and the five-year death rate of lung cancer prominent generalization ability and non-linear
remains low [6]. For example, histological examination processing capacity and high-dimensional processing
is the golden standard for diagnosis of lung cancer, capacity in many areas [19].
but it is not suitable for early screening of lung cancer Based on the previous research [20], this study
because of its traumatic and highly technical explored the significance of the SVM model by using
requirements. Diagnostic imaging such as Chest X data of plasma miRNAs biomarkers for the auxiliary
Ray (CxR) and Computed Tomography (CT) have diagnosis of lung cancer.
been used for diagnosing NSCLC at an early stage,
however, there is a certain radiation hazard and Materials and methods
limited role in reducing lung cancer mortality [7].
Study population
While lung cancer mortality is reduced by 20% in
high-risk lung cancer patients through LDSCT The lung cancer patient group consisted of 148
method, the false-positive are as high as 90% [8]. cases (age rank 29-87 years) with primary lung cancer
Although the sensitivity and specificity of PET from the First Affiliated Hospital of Zhengzhou
method are up to 90%, there is still a 10% University, Henan Cancer Hospital and Henan
false-positive rate and the cost is expensive [9]. Provincial Chest Hospital, from Jun. 2016 to Feb. 2017.
Therefore, new biomarkers and therapeutic strategies Patients were selected on the basis of the following
urgently need to be developed for better management inclusion criteria: (1) patients had a pathological
of lung cancer. diagnostic primary lung cancer that met histological
miRNAs are small single-stranded non-coding or cytological criteria; (2) without undergoing surgical
RNAs that play vital regulatory roles by targeting resection, chemotherapy, or radiotherapy; (3) without
mRNAs for degradation or translational repression. It previous other organ tumors; (4) good compliance
acts as key regulators of cell proliferation, and availability of outcome data. Patients were
differentiation, apoptosis and other biological excluded with major organ function failure, pregnant,
processes [10]. A line of studies suggest that miRNAs or lactating. Pathologic diagnosis was based on WHO
are involved in human diseases and cancers. miRNAs criteria. Lung cancer staging for each patient was
expression is associated with lung cancer has been performed according to the AJCC Cancer Staging
identified in varieties of normal and cancer tissues [11, Manual, 7th edition.
12]. Moreover, it has been demonstrated that plasma Controls come from a company who take
miRNAs regulate numerous target genes and play a physical examinations in Qixian Center for Disease
critical role in lung carcinogenesis, which indicates Control and Prevention. The controls were excluded
that miRNAs might be a potential diagnostic tool for according to the following criteria: (1) without
lung cancer [13]. Published studies [14-16] have shown malignant tumors of the lung or other organs; (2)
that 11 plasma miRNAs (miR-16, miR-21, miR-20a, without major organ function failure; (3) without
miR-210, miR-145, miR-126, miR-223, miR-197, pregnant or lactating; (4) good compliance and
miR-30a, miR-30d, miR-25) in lung cancer patients are availability of outcome data. A total of 148 gender-
abnormal expressions. These results suggest that and age-frequency matched (±3 years) were enrolled
combined with several miRNAs can improve the in this study. The permission was got from each
sensitivity and specificity for the early diagnosis of participant. A questionnaire that included the
lung cancer. information of epidemiology was completed for each
Data mining(DM), also called Knowledge participant by trained interviewers. Smokers are
Discovery in Database(KDD), is extracting potentially defined as people who have smoked for six months or
useful information and knowledge of the process more in their lifetime according to the criteria of
from abundant, incomplete, noisy, fuzzy and random WHO. The alcohol-drinkers are defined as drinking
practical application data [17]. DM has a unique alcohol at least once a week and the consumption of
advantage in solving multi-parameter problems. pure alcohol is above 20 g.

http://www.jcancer.org
Journal of Cancer 2019, Vol. 10 5092

Main instruments and reagents The Fisher and SVM models were developed
The instruments and reagents used in the study using the training set, then the samples of the
included a Labcycler PCR amplifier (SensoQuest validation set were used to validate the quality of the
Company, China), a 7500 Fast Real-time PCR system models. The Combined model (16 items) with 16
(ABI, America), primers (Sangon Biotech), miRcute input variables of smoking status, fever, cough, chest
miRNA extraction and separation kit(Tiangen, pain or tightness, bloody phlegm, hemoptysis and
Beijing), MiRcute enhanced miRNA fluorescence expressions of 10 plasma miRNAs (miR-21, miR-20a,
quantitative detection kit(Tiangen, Beijing) and miR-210, miR-145, miR-126, miR-223, miR-197,
ChemiDoc MP gel imaging analyzer(Bio-RAD, miR-30a, miR-30d, miR-25); the miRNAs model (10
America). items) with 10 input variables of plasma miRNAs
(miR-21, miR-20a, miR-210, miR-145, miR-126,
Statistical analysis and model evaluation miR-223, miR-197, miR-30a, miR-30d, miR-25); the
The Ct values of the samples were calculated symptom model (6 items) with 6 input variables of
with the software for real-time PCR instrument. The fever, cough, chest pain or tightness, bloody phlegm,
comparison of multiple of the expression of miRNA in hemoptysis.
the lung cancer patients to the normal controls was
Fisher discrimination model
calculated using the formula of 2−ΔΔCt (ΔCt= CtmiR −
Ctexternal reference; ΔΔCt=ΔCtmiR –ΔCtaverage normal controls). Fisher discrimination is a widely used
The data was analyzed using SPSS 21.0 software. classification model in traditional statistical methods.
SPSS Clementine 21.0 software was used for data The basic idea: Projection before discriminant
mining. The analysis of the quantitative data was analysis, Projection is the core of the Fisher
analyzed with independent sample t-test or discrimination analysis. After repeating training, the
Mann-Whitney U. Each contingency table was tested Fisher discrimination parameter settings were: Use
by Chi-Square test. Binary logistic regression was partitioned data: no; method: Enter; Mode: Expert;
conducted to analyse the influencing factors of lung Prior probabilities: All groups equal; Use covariance
cancer. The significance level was set at 0.05. matrix: Within-groups.
This study assessed sensitivity, specificity, SVM model
accuracy positive predictive value (PPV), negative
The basic principle is to transform the input
predictive value (NPV), and area under the ROC
space into a high dimensional space by using the
curve (AUC) to estimate the models.
nonlinear transformation defined by the inner
Establishment of models product function, and to find the optimal linear
classification surface.
Data preprocessing After repeating training, the SVM parameter
Data transformation: The relative expression of settings were: Use partitioned data: no; Mode: Expert;
11 miRNAs did not follow a normal distribution, so Kernel type: Polynomial; Gamma: 1; Stopping criteria:
normal transformation was needed. The expression of 1.0E-3.
11 miRNAs was normalized based on 10 common
logarithm transformations. Results
Groups of training set and validation set: Based
on the random sampling function of the partition Demographic characteristics of lung cancer
node, according to a ratio of 3:1, the normalized data patients and controls
of each group were separated randomly into a The 148 lung cancer patients (mean age 60.97 ±
training set (114 controls, 100 cancer cases) and a 10.83 years) and 148 controls (mean age 60.14 ± 9.66
validation set (34 controls, 48 cancer cases). The years) were enrolled. The age distribution of subjects
training set was utilized to develop the model, while was in normal distribution, so the age group was
the validation set was used to verify the model. divided into two groups according to mean age (60
years). All the subjects were divided into four groups
Model derivation (Never smoking; Light smoking: <10 cigarettes/day;
The Data node is the source of data for the study; Moderate smoking: 10~20 cigarettes/day; Heavy
the variables are documented using Type node; the smoke>20 cigarettes/day) according to the smoking
samples were randomly divided into the training set status. As shown in Table 1, the average age, sex and
and validation set according to the proportion of 3:1 alcohol were no significant differences between the
using Partition node; Random number seed is two groups (P>0.05). However, the frequency of
1111111. smoking, fever, cough, chest pain or tightness, bloody

http://www.jcancer.org
 Journal of Cancer 2019, Vol. 10 5093

phlegm and hemoptysis were significantly higher in 87.38%, 74.77%, 79.91%, 98.13%, 85.51%, and 83.64%,
the cancer group than that in control group (P<0.001). respectively. The accuracies in the validation set were
84.15%, 75.61%, 80.49%, 96.34%, 80.49%, and 84.15%,
Table 1. Demographic characteristics of lung cancer patients and respectively.
controls
Variable Lung cancer Controls χ2/t P Table 2. Clinical and pathological characteristics of lung cancer
(n=148) (n=148) patients and controls
Age* 60.97±10.83 60.14±9.66 0.691 0.490
Age-grouped ≤60 67 68 0.014 0.907 Clinical and pathological characteristics n Percentage (%)
>60 81 80 Histological type SCLC 18 12.16
Gender Male 98 99 0.821 0.365 SCC 36 24.32
Female 50 49 AC 66 44.59
Fever No 131 146 12.654 <0.001 LCLC 2 11.35
Yes 17 2 Others 26 17.57
Cough No 57 130 77.387 <0.001 TNM stage* Ⅰ+Ⅱ 33 24.63
Yes 91 18 Ⅲ+Ⅳ 101 75.37
Chest pain or No 86 135 42.878 <0.001 lymphatic No 21 17.80
tightness Yes 62 13 metastasis * Yes 97 82.20
Bloody phlegm No 109 148 44.918 <0.001 distant metastases* No 83 70.34
Yes 39 0 Yes 35 29.67
Hemoptysis No 134 148 14.695 <0.001 Note: The * indicates data is missing.
Yes 14 0
Weak No 146 145 0.000 1.000
Yes 2 3 Table 3. The relative expression of 11 plasma miRNAs in lung
Alcohol Never 134 129 0.853 0.356 cancer and controls
Yes 14 19
Smoking status Never 76 89 16.989 <0.001 Lung cancer (n=148) Control (n=148)
Light 10 26 miRNAs M (P25, P75) M (P25, P75) Z P
Moderate 28 15 miR-16 1.60(0.70,2.93) 1.39(0.66,2.51) -1.184 0.236
Heavy 34 18 miR-21 1.05(0.77,2.09) 0.68(0.53,0.90) -6.017 <0.001
Note: The * indicates age according with normal distribution. miR-20a 1.93(0.81,4.40) 0.80(0.42,1.51) -6.264 <0.001
miR-210 1.10(0.53,3.09) 0.68(0.39,1.24) -4.267 <0.001
miR-145 1.11(0.56,2.93) 0.70(0.44,1.07) -4.242 <0.001
Clinical pathologic characteristics of lung miR-126 1.64(0.71,2.83) 0.77(0.32,1.58) -5.096 <0.001
miR-223 2.26(1.26,5.55) 0.76(0.41,1.36) -8.952 <0.001
cancer patients miR-197 1.13(0.59,2.29) 0.59(0.41,1.25) -5.008 <0.001
The clinical and pathological characteristics of miR-30a 0.82(0.51,2.81) 0.66(0.37,1.75) -2.908 <0.001
miR-30d 1.37(0.78,3.55) 0.69(0.48,1.24) -6.409 <0.001
lung cancer patients collected in this study are shown miR-25 1.36(0.77,3.27) 0.80(0.34,1.73) -4.925 <0.001
in Table 2. The lung cancer group was consisted of 36
SCC cases, 18 SCLC cases, 66 AC cases, 2 LCLC cases,
Table 4. Effect of data mining on distinguish lung cancer
and 26 other histological type cases; 33 cases of clinical
stage Ⅰ and Ⅱ, 101 cases of clinical stage Ⅲ and Ⅳ. Model Training set(n=214) Validation set(n=82)
Cancer Controls Cancer Controls
cases cases
Comparison of the expressions of 11 plasma Combined Fisher Cancer cases 79 21 41 7
miRNAs between the two groups model
Controls 6 108 6 28
As seen in Table 3, expressions of 10 plasma Total 85 129 47 35
miRNAs (miR-21, miR-20a, miR-210, miR-145, Accuracy 87.38% 84.15%
miR-126, miR-223, miR-197, miR-30a, miR-30d, miRNAs Fisher model Cancer cases 70 30 38 10
Controls 24 90 10 24
miR-25) were all significantly up-regulated in lung Total 94 120 48 34
cancer patients than controls (P<0.05). However, the Accuracy 74.77% 75.61%
expression of miR-16 was no significant difference Symptom Fisher model Cancer cases 71 29 38 10
Controls 14 100 6 28
between lung cancer patients and controls (P>0.05). Total 85 129 44 38
Accuracy 79.91% 80.49%
Data mining Combined SVM model Cancer cases 99 1 47 1
Controls 3 111 2 32
The data of the Fisher and SVM model based on Total 102 112 49 33
the smoking status, fever, cough, chest pain or Accuracy 98.13% 96.34%
tightness, bloody phlegm, haemoptysis and miRNAs SVM model Cancer cases 83 17 38 10
Controls 14 100 6 28
expressions of 10 plasma miRNAs (miR-21, miR-20a, Total 97 117 44 38
miR-210, miR-145, miR-126, miR-223, miR-197, Accuracy 85.51% 80.49%
miR-30a, miR-30d, and miR-25) are presented in Table Symptom SVM model Cancer cases 81 19 42 6
Controls 16 98 7 27
4. In the training set, the accuracies of combined Total 97 117 49 33
Fisher, miRNAs Fisher, symptom Fisher, combined Accuracy 83.64% 84.15%
SVM, miRNAs SVM, and symptom SVM model were

http://www.jcancer.org
Journal of Cancer 2019, Vol. 10 5094

The evaluation of models plasma miRNAs (miR-21, miR-126, miR-210, and

The results of the evaluation indexes of the 6 miR-486) could differentiate NSCLC from controls
models were presented in Table 5. Sensitivity of with 86.22% sensitivity and 96.55% specificity, which
combined SVM model reached 97.90%, and the also could to distinguish NSCLC with 73.33%
specificity was 94.10%. PPV and NPV were likewise sensitivity and 96.55% specificity in phase Ⅰ [23]. In the
highest. Meanwhile, AUC was greater than 0.9. On plasma of NSCLC patients, one study identified 15
the other hand, AUC of the miRNAs Fisher and types of miRNAs associated with lung cancer tissues
symptom Fisher models were slightly smaller than the from the literature and found that the expression of
other models. miR-155, miR-197, and miR-182 were significant
increase in phase Ⅰ [16]. The sensitivity and specificity
of diagnosis NSCLC patients were 81.33% and
Table 5. Comparison results in the validation set by SVM and
86.76%, respectively.
Fisher models
In this study, we compared the expression of 11
Model Sensitivity
(%)
Specificity
(%)
Accuracy
(%)
PPV
(%)
NPV
(%)
AUC (95% CI) plasma miRNAs in lung cancer patient to that in the
Fisher Combined 0.854 0.824 0.842 0.872 0.800 0.865(0.821,0.902) controls. Single-factor analysis showed that the
miRNAs 0.792 0.706 0.756 0.792 0.706 0.750(0.697,0.798) expressions of 10 plasma miRNAs (miR-21, miR-20a,
Symptom 0.792 0.824 0.805 0.864 0.737 0.801(0.751,0.845)
SVM Combined 0.979 0.941 0.963 0.959 0.970 0.976(0.952,0.990)
miR-210, miR-145, miR-126, miR-223, miR-197,
miRNAs 0.792 0.824 0.805 0.864 0.737 0.841(0.795,0.881) miR-30a, miR-30d, miR-25) in lung cancer group were
Symptom 0.875 0.794 0.842 0.857 0.818 0.838(0.791,0.878) statistically significant higher than the controls;
Multiple factor analysis revealed that elevated plasma
The results of the AUC of the 6 models were miR-20a levels and miR-223 were risk factors for lung
shown in Table 6. The AUC of combined SVM model cancer.
was superior to the other 5 models, and the difference The ten miRNAs explored in the present study
was statistically significant (P<0.05); The AUC of have been evaluated as lung cancer markers in other
combined Fisher model was higher than miRNAs studies. Potential mechanisms of the miRNAs on lung
Fisher model and symptom Fisher model (P<0.05). cancer were explored in previous study (Seen in Table
There were no statistical differences in AUC among 7), the main signaling pathways including
the other 3 models (P>0.05). PI3K/Akt/NF-Κb (miR-21, miR-223, miR-145,
miR-126, and miR-30a) [24-28], STAT3 (miR-126, and
Table 6. Comparison of results in validation set by SVM and miR-197) [27, 29, 30], estrogen (miR-21 and miR-210) [31], et
Fisher discriminant analysis al.
Comparison of models Z P
Combined SVM model vs Combined Fisher model 5.474 <0.0001
Combined SVM model vs miRNAs SVM model 6.445 <0.0001
Table 7. Potential mechanisms of the miRNAs on lung cancer
Combined SVM model vs Symptom SVM model 6.363 <0.0001
miRNAs Pathways
miRNAs SVM model vs Symptom SVM model 0.105 0.9168
miR-20a angiogenesis, TGF-β pathway, platelet-derived growth factor pathway,
miRNAs SVM model vs miRNAs Fisher model 4.032 0.0001 and oxidative stress response[32]
Symptom SVM model vs Symptom Fisher model 2.256 0.0241 miR-21 PI3K/AKT/NF-κB signaling pathway, and estrogen signaling
Combined Fisher model vs miRNAs Fisher model 4.179 <0.0001 pathway[24]; autophagy-related AMPK/ULK1 signaling pathway[56]
Combined Fisher model vs Symptom Fisher model 3.167 0.0015 miR-210 estrogen signaling pathway[31]
miRNAs Fisher model vs Symptom Fisher model 1.454 0.1459 miR-223 Notch/miR-223/FBXW7 pathway[57]; NF-κB signaling pathway[58];
IGF-1R/Akt/S6 signaling pathway[25]; IGF-1R/PI3K/AKT signaling
pathway[59]

Discussion miR-25 ERK signaling pathway[60]; cell cycle regulation[61]

miR-145 ERβ/MALAT1/miR145-5p/NEDD9 signaling pathway[62];
EGFR/PI3K/AKT signaling pathway[26]; JNK signaling pathway[63];
Early diagnosis and effective treatment of lung mTOR signaling pathway[64]
cancer is the key to improve the survival rate of miR-126 STAT3 signal pathway[30]; PI3K/AKT/Snail signal pathway[27]
patients. Therefore, early and non-invasive miR-30a PI3K/AKT signaling pathway[28]
miR-197 miR-197/CKS1B/STAT3-mediated PD-L1 network[29]
biomarkers for lung cancer diagnosis have been the
most popular research areas. It has been shown that Transcription factors such as HMBOX1, DDX5,
circulating miRNAs are stable under the actual and ZBTB5, which were identified to be co-regulated
experimental conditions and that abnormal by miR-20a and miR-15b, have implication on cancer
expression of cancer-related miRNAs may be earlier progression [32]. E2F1 and E2F8 belong to E2F
than the clinical symptoms, therefore, circulating transcription factor family that is essential for the
miRNAs may be used as tumor biomarkers [21]. A regulation of cell cycle progression [33]. miR-20a has
large body of studies have suggested that a series of been shown to directly inhibit E2F1 transcription
circulating miRNAs have the potential as diagnostic factor and highly express in NSCLC tissues [34].
tool in malignancies [22]. It has been shown that four miR-223 suppressed the proliferation, migration and

http://www.jcancer.org
Journal of Cancer 2019, Vol. 10 5095

invasion of NSCLC cells through directly inhibition of relative expressions of miR-126, miR-30a and miR-30d
E2F8 expression [35]. Increased expression of miR-20a in plasma lung cancer patients were greater than
and miR-223 were found in lung cancer in our study, controls, and the data differed from the studies above.
which could be used as a molecular biomarker in Various data mining algorithms have been
auxiliary diagnosis of lung cancer. improved in recent years, such as cluster analysis,
miR-25, acting as an oncogene or anti-oncogene decision tree and rough set, ANN and genetic
located on chromosome 7p22.1, is involved in the algorithm, SVM and fuzzy processing technology [47].
development of multiple malignant tumors at the Each method has advantages and limits as well as the
post-transcriptional level [36]. The expression of miR-25 applicable scope. Fisher discriminant analysis is one
in NSCLC tissues was significantly higher than that in of the most widely used method in multivariate
adjacent non-cancerous tissues, which plays a statistical pattern recognition, which requires the
carcinogenic effect by regulating cell cycle element E2 independent input variables without interaction effect
and is associated with cancer cell for resistance, and normal distribution and so on [48]. Therefore, the
proliferation, metastasis, invasion, and so on [37]. analysis of the nonlinear system has a couple of
miR-21 is over expressed in a variety of diseases and limitations. In order to get the best generalization
negatively responsible for the regulation of tumor ability, based on the statistical learning theory of VC
suppressor genes by participating in the proliferation, (Vapnik-Cher-Vonenkis) and structural risk
invasion, metastasis and vascular infiltration of tumor minimization principle, SVM finds the best
cells, thus promoting the development of tumor [38]. compromise between the complexity of the model and
miR-210 and hypoxia-inducible factor 1-alpha the ability to learn [49]. SVM is a classical method in
(HIF-1α) play a synergistic role in the proliferation, data mining. There are several advantages of SVM
differentiation, apoptosis, angiogenesis, DNA method. For example, structural risk minimization
damage repair and energy metabolism of the hypoxic and good generalization ability, what is based on
cell [39]. Study also demonstrated the high expression statistical learning theory [50]. The second, SVM can
of miR-210 in advanced lung cancer [40]. miR-197 is achieve similar results with different kernel functions
related to infiltration and metastasis of tumor cells like ANN, which depends on the selected model [51]. In
and located on chromosome 1p13.3. Recent research general, SVM is the optimal solution in the existing
bears out the miR-197/CKS1B/STAT3-mediated information situation, which makes up for the
PD-L1 network in chemoresistant NSCLC, deficiency of ANN in determining the reasonable
independent of immunosuppression signals [29]. structure and local optimal problem, and has a
Another research found decreased expression of significant improvement in learning methods. This
miR-197 induces p53-dependent lung cancer cell study deeply analyzed with more mature SVM
apoptosis, which may be oncogene [41]. The results of algorithms employed in the medical field.
the study are basically in line with these studies in At present, some studies have mostly focused on
which the increased expression of miR-25, miR-21, one or several biomarkers using traditional analysis
miR-197 and miR-210 in the plasma of lung cancer methods. One study explored serum miR-22,
compared with controls. Taken together, these results miR-125b, and miR-15b diagnosis compared with the
reflect the reliability and stability of miR-25, miR-21, current commonly used tumor marker CEA, which
miR-197 and miR-210 as the lung cancer biomarkers. indicates that the diagnostic significance of these three
The data of the expressions of miR-145 are not serum miRNAs(AUC=0.725, 0.704, and 0.619) for
consistent in different studies. It has been shown that NSCLC was higher than that of serum CEA
miR-145 is down-regulated in various malignancies (AUC=0.594) [52]. Meanwhile, some studies focused on
including lung adenocarcinoma, which inhibited cell gene and other biomarkers using ANN or decision
proliferation through targeting epidermal growth tree model and so on. The ANN and decision tree
factor receptor (EGFR) and nucleoside diphosphate 1 model of lung cancer based on the genetic
(NUDT1) [42]. However, Study found that the polymorphism of CYP1A1, GSTM1, mEH, XRCC1, the
increased expression of miR-145 in the plasma lung length of telomere, and the methylations of p16 and
cancer, which is consistent with our study [43]. miR-126 RASSF1A gene, the results showed that the accuracy
may inhibit the proliferation of lung cancer cells and for ANN and decision tree model validation sets was
the expression of miR-126 was lower than normal 89.62% and 93.00% [53]. The accuracy and sensitivity
tissue [44]. miR-30a can inhibit the invasion and were also improved by the above methods. In this
migration of lung cancer cells by directly inhibiting study, the SVM model and Fisher model were
the expression of the snail [45]. miR-30d could inhibit established based on miRNAs tumor biomarkers and
the cell proliferation and activity of NSCLC by clinical symptom characteristics for the first time.
directly regulating CCNE2 [46]. In this study, the

http://www.jcancer.org
Journal of Cancer 2019, Vol. 10 5096

We established the Fisher model with 10 miRNA diagnosis of lung cancer, which also has excellent
and 6 symptom for lung cancer diagnostic, and, the predictive power, such as all patients with stage Ⅰ and
AUCs of three models are combined Fisher model (16 Ⅱ lung cancer in validation set were correctly
items) (0.865, 95%CI=0.821-0.902), miRNAs Fisher predicted to be lung cancer.
model (10 items) (0.750, 95%CI 0.697-0.798), and This study showed that 10 plasma miRNAs
symptom Fisher model (6 items) (0.801, 95%CI expression levels were associated with lung cancer,
0.751-0.845), respectively. The accuracy for three which provides a theoretical possibility for further
model validation sets was 84.15%, 75.61%, and prospective studies or large-scale clinical trials. More
80.49%, respectively. The combined Fisher model importantly, the expression of the plasma miRNAs is
showed good ability to detect lung cancer, which is very stable under different harsh conditions, which
superior to the lung cancer diagnosis Fisher model indicating that the plasma miRNAs has the potential
(0.670, 95%CI 0.569-0.761) established with FHIT, to serve as biomarker for auxiliary diagnosis of lung
RASSF1A, p16 promoter methylation, and relative cancer. Our findings indicate that SVM model based
telomere length in our prophase research [20]. This may on plasma miRNAs biomarkers may serve as a novel,
be due to the miRNAs biomarkers has better accurate, noninvasive method for auxiliary diagnosis
specificity compared with gene or other biomarkers. of lung cancer. However, there are some limitations in
Our findings indicate that the changed expression this study. Firstly, the selection of 10 plasma miRNAs
levels could be used as potential biomarkers for were based on published studies rather than miRNA
diagnosis of lung cancer. Besides, probably because of array or bioinformatics method. More plasma
the data pretreatment before model established. After miRNAs need to be analyzed to for using as specific
the normal transformation, the expression levels of biomarkers. Secondly, compare to single study, large
miRNAs are approximately normal distribution and sample and multicenter clinical trial studies will yield
without missing values. more reliable results. Moreover, there are still things
miRNAs play a critical role in lung cancer for the further validation study need to be thought,
carcinogenesis, which were studied widely as cancer including health policy, ethics, cost, et al.
biomarkers. Zhang et al [54] established screening
method for early-stage NSCLC using four miRNAs Conclusions
(miR-145, miR-20a, miR-21, miR-223), and the AUC of In summary, this study suggests that the 10
the model was 0.897. To the best of our knowledge, plasma miRNAs are associated with lung cancer, and
there is no data mining model for lung cancer the changed expression levels could be used as
diagnosis based on miRNAs. SVM model were potential biomarkers for diagnosis of lung cancer.
established for lung cancer diagnostic in our study, SVM model has the superior diagnostic value for
which combined 10 miRNAs and 6 symptoms, had a auxiliary diagnosis of lung cancer based on miRNAs
higher accuracy. The combined SVM model with tumor biomarkers and clinical symptom
miRNAs was superior in lung cancer diagnosis in this characteristics.
study compared to models with methylation and
telomere biomarkers in our prophase research [20]. The Abbreviations
accuracy and AUC of combined SVM model in our Adenocarcinoma: AC; Artificial neural
study were also better than the results of other studies networks: ANN; Area under the ROC curve: AUC;
on gene and other biomarkers using ANN or SVM Data mining: DM; decision tree: DT; Knowledge
and so on. For example, one study explored eighteen Discovery in Database: KDD; Large cell carcinoma:
genes (including TTN, RHOH, RPS20, TRBC2) for six LCLC; Negative predictive value: NPV; Positron
cancer (including lung cancer) using SVM with emission tomography: PET; Positive predictive value:
accuracy of 75.10% [55]. PPV; Received operating characteristic: ROC; Small
As to the three models we established, the single-stranded non-coding RNAs: miRNAs; Support
accuracy of models (10 miRNAs SVM, 6 symptom vector machine: SVM; Small cell lung cancer: SCLC;
SVM model, and combined SVM) were 80.45%, Non-small-cell lung cancer: NSCLC; Squamous cell
84.15%, and 96.34%, respectively; the AUC of models carcinoma: SCC.
(10 miRNAs SVM, 6 symptom SVM model, and
combined SVM) were 0.841, 0.818, and 0.976, Acknowledgements
respectively. The AUC and accuracy of combined This study was supported by the Programs for
SVM model were better than the miRNAs SVM and National Nature Science Foundation of China
symptom SVM model. Overall, the SVM model based (81001239, 81872597, 81473014), Science and
on miRNAs and clinical symptom characteristics has a Technology Development of Henan Province
higher accuracy rate and might be useful for early

http://www.jcancer.org
 Journal of Cancer 2019, Vol. 10 5097

(142102310116) and the Outstanding Youth Grant of 22. Zhou X, Wen W, Zhu J, Huang ZB, Zhang L, Zhang H, et al. A six-microRNA
signature in plasma was identified as a potential biomarker in diagnosis of
Zhengzhou University (1521329035). esophageal squamous cell carcinoma. Oncotarget. 2017; 8: 34468-80.
23. Shen J, Todd NW, Zhang H, Yu L, Lingxiao X, Mei YP, et al. Plasma
Ethical approval microRNAs as potential biomarkers for non-small-cell lung cancer. Lab Invest.
2011; 91: 579-87.
All procedures performed in studies involving 24. Zhou B, Wang D, Sun G, Mei F, Cui Y, Xu H. Effect of miR-21 on Apoptosis in
Lung Cancer Cell Through Inhibiting the PI3K/ Akt/NF-kappaB Signaling
human partcipants were in accordance to the ethical Pathway in Vitro and in Vivo. Cell Physiol Biochem. 2018; 46: 999-1008.
standards of the insttutonal and/or natonal research 25. Zhao FY, Han J, Chen XW, Wang J, Wang XD, Sun JG, et al. miR-223 enhances
the sensitivity of non-small cell lung cancer cells to erlotinib by targeting the
commitee and with the 1964 Helsinki Declaraton and insulin-like growth factor-1 receptor. Int J Mol Med. 2016; 38: 183-91.
26. Li B, Ding CM, Li YX, Peng JC, Geng N, Qin WW. MicroRNA145 inhibits
its later amendments or comparable ethical standards. migration and induces apoptosis in human nonsmall cell lung cancer cells
through regulation of the EGFR/PI3K/AKT signaling pathway. Oncol Rep.
Competing Interests 2018; 40: 2944-54.
27. Jia Z, Zhang Y, Xu Q, Guo W, Guo A. miR-126 suppresses
The authors have declared that no competing epithelial-to-mesenchymal transition and metastasis by targeting
PI3K/AKT/Snail signaling of lung cancer cells. Oncol Lett. 2018; 15: 7369-75.
interest exists. 28. Meng F, Wang F, Wang L, Wong SC, Cho WC, Chan LW. MiR-30a-5p
Overexpression May Overcome EGFR-Inhibitor Resistance through
References Regulating PI3K/AKT Signaling Pathway in Non-small Cell Lung Cancer Cell
Lines. Front Genet. 2016; 7: 197.
1. Miller KD, Siegel RL, Lin CC, Mariotto AB, Kramer JL, Rowland JH, et al. 29. Fujita Y, Yagishita S, Hagiwara K, Yoshioka Y, Kosaka N, Takeshita F, et al.
Cancer treatment and survivorship statistics, 2016. CA Cancer J Clin. 2016; 66: The clinical relevance of the miR-197/CKS1B/STAT3-mediated PD-L1
271-89. network in chemoresistant non-small-cell lung cancer. Mol Ther. 2015; 23:
2. Zhu JF, Zhang YL, Yang X, Jin LT. Clinical Significance and 717-27.
Tumor-Suppressive Function of miR-516b in Nonsmall Cell Lung Cancer. 30. Zhang Z, Wang J, Cheng J, Yu X. Effects of miR-126 on the STAT3 signaling
Cancer Biother and Radiopharm. 2017; 32: 115-23. pathway and the regulation of malignant behavior in lung cancer cells. Oncol
3. Molina JR, Yang P, Cassivi SD, Schild SE, Adjei AA. Non-small cell lung lett. 2018; 15: 8412-6.
cancer: epidemiology, risk factors, treatment, and survivorship. Mayo Clin 31. Peng Z, Pan L, Niu Z, Li W, Dang X, Wan L, et al. Identification of microRNAs
Proc. 2008; 83: 584-94. as potential biomarkers for lung adenocarcinoma using integrating genomics
4. Foss KM, Sima C, Ugolini D, Neri M, Allen KE, Weiss GJ. miR-1254 and analysis. Oncotarget. 2017; 8: 64143-56.
miR-574-5p: serum-based microRNA biomarkers for early-stage non-small cell 32. Alhoshani A, Alrashdi A, Alhosaini K, Alanazi FE, Alajez NM, Altaf M, et al.
lung cancer. J Thorac Oncol: official publication of the International Gold-containing compound BDG-I inhibits the growth of A549 lung cancer
Association for the Study of Lung Cancer. 2011; 6: 482-8. cells through the deregulation of miRNA expression. Saudi Pharm J. 2018; 26:
5. Jemal A, Bray F, Center MM, Ferlay J, Ward E, Forman D. Global cancer 1035-43.
statistics. CA Cancer J Clin. 2011; 61: 69-90. 33. Attwooll C, Lazzerini Denchi E, Helin K. The E2F family: specific functions
6. Yang J, Zhu J, Zhang YH, Chen YS, Ding LL, Kensler TW, et al. Lung Cancer in and overlapping interests. EMBO J. 2004; 23: 4709-16.
a Rural Area of China: Rapid Rise in Incidence and Poor Improvement in 34. Babu KR, Muckenthaler MU. miR-20a regulates expression of the iron
Survival. Asian Pac J Cancer Prev. 2015; 16: 7295-302. exporter ferroportin in lung cancer. J Mol Med (Berl). 2016; 94: 347-59.
7. Swensen SJ, Jett JR, Hartman TE, Midthun DE, Mandrekar SJ, Hillman SL, et 35. Dou L, Han K, Xiao M, Lv F. miR-223-5p Suppresses Tumor Growth and
al. CT screening for lung cancer: five-year prospective experience. Radiology. Metastasis in Non-Small Cell Lung Cancer by Targeting E2F8. Oncol Res. 2019;
2005; 235: 259-65. 27: 261-8.
8. Aberle DR, Adams AM, Berg CD, Black WC, Clapp JD, et al. Reduced 36. Deng T, Yuan YZ, Zhang CN, Zhang CY, Yao WY, Wang C, et al. Identification
lung-cancer mortality with low-dose computed tomographic screening. N of Circulating MiR-25 as a Potential Biomarker for Pancreatic Cancer
Engl J Med. 2011; 365: 395-409. Diagnosis. Cell Physiol Biochem. 2016; 39: 1716-22.
9. Cao JQ, Rodrigues GB, Louie AV, Zaric GS. Systematic review of the 37. Zhao ZY, Liu JT, Wang CL, Wang Y, Jiang YG, Guo M. MicroRNA-25
cost-effectiveness of positron-emission tomography in staging of regulates small cell lung cancer cell development and cell cycle through cyclin
non--small-cell lung cancer and management of solitary pulmonary nodules. E2. Int J Clin Exp Pathol. 2014; 7: 7726-34.
Clin Lung Cancer. 2012; 13: 161-70. 38. Pfeffer SR, Yang CH, Pfeffer LM. The Role of miR-21 in Cancer. Drug Dev Res.
10. Sharma S, Kelly TK, Jones PA. Epigenetics in cancer. Carcinogenesis. 2010; 31: 2015; 76: 270-7.
27-36. 39. Xu TX, Zhao SZ, Dong M, Yu XR. Hypoxia responsive miR-210 promotes cell
11. Yanaihara N, Harris CC. MicroRNA Involvement in Human Cancers. Clin survival and autophagy of endometriotic cells in hypoxia. Eur Rev Med
Chem. 2013; 59: 1811-2. Pharmacol Sci. 2016; 20: 399-406.
12. Switlik W, Karbownik MS, Suwalski M, Kozak J, Szemraj J. miR-30a-5p 40. Puissegur MP, Mazure NM, Bertero T, Pradelli L, Grosso S, Robbe-Sermesant
together with miR-210-3p as a promising biomarker for non-small cell lung K, et al. miR-210 is overexpressed in late stages of lung cancer and mediates
cancer: A preliminary study. Cancer Biomarker. 2018; 21: 479-88. mitochondrial alterations associated with modulation of HIF-1 activity. Cell
13. Kunz M, Wolf B, Schulze H, Atlan D, Walles T, Walles H, et al. Non-Coding Death Differ. 2011; 18: 465-78.
RNAs in Lung Cancer: Contribution of Bioinformatics Analysis to the 41. Fiori ME, Barbini C, Haas TL, Marroncelli N, Patrizii M, Biffoni M, et al.
Development of Non-Invasive Diagnostic Tools. Genes. 2016; 8: 8. Antitumor effect of miR-197 targeting in p53 wild-type lung cancer. Cell Death
14. Geng Q, Fan T, Zhang B, Wang W, Xu Y, Hu H. Five microRNAs in plasma as Differ. 2014; 21: 774-82.
novel biomarkers for screening of early-stage non-small cell lung cancer. 42. Cho WC, Chow AS, Au JS. MiR-145 inhibits cell proliferation of human lung
Respir Res. 2014; 15: 149. adenocarcinoma by targeting EGFR and NUDT1. RNA Biol. 2011; 8: 125-31.
15. Inamura K, Ishikawa Y. MicroRNA In Lung Cancer: Novel Biomarkers and 43. Zhang H, Mao F, Shen TY, Luo QQ, Ding ZP, Qian LQ, et al. Plasma miR-145,
Potential Tools for Treatment. J Clin Med. 2016; 5: 36. miR-20a, miR-21 and miR-223 as novel biomarkers for screening early-stage
16. Zheng DL, Haddadin S, Wang Y, Gu LQ, Perry MC, Freter CE, et al. Plasma non-small cell lung cancer. Oncol lett. 2017; 13: 669-76.
microRNAs as novel biomarkers for early detection of lung cancer. Int J Clin 44. Miko E, Margitai Z, Czimmerer Z, Varkonyi I, Dezso B, Lanyi A, et al. miR-126
Exp Pathol. 2011; 4: 575-86. inhibits proliferation of small cell lung cancer cells by targeting SLC7A5. FEBS
17. Fang H, Lu B, Wang X, Zheng L, Sun K, Cai W. Application of data mining lett. 2011; 585: 1191-6.
techniques to explore predictors of upper urinary tract damage in patients 45. Kumar R, Kumari B, Kumar M. Prediction of endoplasmic reticulum resident
with neurogenic bladder. Braz J Med Biol Res. 2017; 50: e6638. proteins using fragmented amino acid composition and support vector
18. Chen HL, Yang B, Liu J, Liu DY. A support vector machine classifier with machine. PeerJ. 2017; 5: e3561.
rough set-based feature selection for breast cancer diagnosis. Expert Syst 46. Chen D, Guo W, Qiu Z, Wang Q, Li Y, Liang L, et al. MicroRNA-30d-5p
Appl. 2011; 38: 9014-22. inhibits tumour cell proliferation and motility by directly targeting CCNE2 in
19. Jayasurya K, Fung G, Yu S, Dehing-Oberije C, De Ruysscher D, Hope A, et al. non-small cell lung cancer. Cancer Lett. 2015; 362: 208-17.
Comparison of Bayesian network and support vector machine models for 47. Amani M, Amani P, Kasaeian A, Mahian O, Pop I, Wongwises S. Modeling
two-year survival prediction in lung cancer patients treated with and optimization of thermal conductivity and viscosity of MnFe2O4 nanofluid
radiotherapy. Med Phys. 2010; 37: 1401-7. under magnetic field using an ANN. Sci Rep. 2017; 7: 17369.
20. Wang W, Feng XL, Duan XR, Tan SJ, Wang SH, Wang TW, et al. Establishment 48. Hung CI, Lee PL, Wu YT, Chen LF, Yeh TC, Hsieh JC. Recognition of motor
of two data mining models of lung cancer screening based on three gene imagery electroencephalography using independent component analysis and
promoter methylations combined with telomere damage. Int J Biol Markers. machine classifiers. Ann Biomed Eng. 2005; 33: 1053-70.
2017; 32: E141-E6. 49. Sattari MT, Farkhondeh A, Patrick Abraham J. Estimation of sodium
21. Wang K, Yuan Y, Cho JH, McClarty S, Baxter D, Galas DJ. Comparing the adsorption ratio indicator using data mining methods: a case study in Urmia
MicroRNA Spectrum between Serum and Plasma. PloS one. 2012; 7: e41561. Lake basin, Iran. Environ Sci Pollut Res Int. 2018; 25: 4776-86.

http://www.jcancer.org
 Journal of Cancer 2019, Vol. 10 5098
50. Guo L, Liu X, Wu Y, Yan W, Shen X. Research on the segmentation of MRI
image based on multi-classification support vector machine. Conf Proc IEEE
Eng Med Bio Soc. 2007; 2007: 6020-3.
51. Nedaie A, Najafi AA. Support vector machine with Dirichlet feature mapping.
Neural Netw. 2017; 98: 87-101.
52. Shi GL, Chen Y, Sun Y, Yin YJ, Song CX. Significance of Serum MicroRNAs in
the Auxiliary Diagnosis of Non-Small Cell Lung Cancer. Clin Lab. 2017; 63:
133-40.
53. Wang Z, Feng F, Zhou X, Duan L, Wang J, Wu Y, et al. Development of
diagnostic model of lung cancer based on multiple tumor markers and data
mining. Oncotarget. 2017; 8: 94793-804.
54. Zhang H, Mao F, Shen T, Luo Q, Ding Z, Qian L, et al. Plasma miR-145,
miR-20a, miR-21 and miR-223 as novel biomarkers for screening early-stage
non-small cell lung cancer. Oncol lett. 2017; 13: 669-76.
55. Zhang YH, Huang T, Chen L, Xu Y, Hu Y, Hu LD, et al. Identifying and
analyzing different cancer subtypes using RNA-seq data of blood platelets.
Oncotarget. 2017; 8: 87494-511.
56. Li S, Zeng X, Ma R, Wang L. MicroRNA-21 promotes the proliferation,
migration and invasion of non-small cell lung cancer A549 cells by regulating
autophagy activity via AMPK/ULK1 signaling pathway. Exp Ther Med. 2018;
16: 2038-45.
57. Zhang H, Chen F, He Y, Yi L, Ge C, Shi X, et al. Sensitivity of non-small cell
lung cancer to erlotinib is regulated by the Notch/miR-223/FBXW7 pathway.
Biosci Rep. 2017; 37: BSR20160478.
58. Huang L, Li F, Deng P, Hu C. MicroRNA-223 Promotes Tumor Progression in
Lung Cancer A549 Cells via Activation of the NF-kappaB Signaling Pathway.
Oncol Res. 2016; 24: 405-13.
59. Han J, Zhao F, Zhang J, Zhu H, Ma H, Li X, et al. miR-223 reverses the
resistance of EGFR-TKIs through IGF1R/PI3K/Akt signaling pathway. Int J
Oncol. 2016; 48: 1855-67.
60. Ding X, Zhong T, Jiang L, Huang J, Xia Y, Hu R. miR-25 enhances cell
migration and invasion in non-small-cell lung cancer cells via ERK signaling
pathway by inhibiting KLF4. Mol Med Rep. 2018; 17: 7005-16.
61. Zhao Z, Liu J, Wang C, Wang Y, Jiang Y, Guo M. MicroRNA-25 regulates small
cell lung cancer cell development and cell cycle through cyclin E2. Int J Clin
Exp Pathol. 2014; 7: 7726-34.
62. Yu W, Ding J, He M, Chen Y, Wang R, Han Z, et al. Estrogen receptor beta
promotes the vasculogenic mimicry (VM) and cell invasion via altering the
lncRNA-MALAT1/miR-145-5p/NEDD9 signals in lung cancer. Oncogene.
2019; 38: 1225-38.
63. Chang Y, Yan W, Sun C, Liu Q, Wang J, Wang M. miR-145-5p inhibits
epithelial-mesenchymal transition via the JNK signaling pathway by targeting
MAP3K1 in non-small cell lung cancer cells. Oncol lett. 2017; 14: 6923-8.
64. Li JC, Zheng JQ. Effect of microRNA-145 on proliferation and apoptosis of
human non-small cell lung cancer A549 cells by regulating mTOR signaling
pathway. J Cell Biochem. 2017.DOI: 10.1002/jcb.26629.

http://www.jcancer.org