This article is mentioned in:

Introduction

Gastric cancer (GC) is one of the most threatening worldwide diseases. An estimated 951,600 new stomach cancer cases and 723,100 deaths occurred in 2012 (1). GC is still the second most frequent malignancy globally despite great improvement in the diagnosis and treatment of GC. In China, which belongs to an area with the highest incidence rate, a total of 679,000 new cases of GC and 498,000 deaths occurred in 2015 (2). The high mortality of GC is attributed to the low rate of early diagnosis. Thus, the majority of patients are diagnosed at an advanced stage with a poor patient prognosis. Therefore, it is critical to improve the sensitivity and specificity of diagnostic tools for the prevention and detection of GC. Although the gold standard diagnostic methods for GC, endoscopy and random biopsy endoscopy or image examination, could facilitate the early diagnosis of GC, the invasive nature, potential sampling errors and high expenditure impact their use only for patients at a high-risk of GC.

MicroRNAs (miRNAs) are small non-coding RNAs comprised of approximately 21 nucleotides in length that crucially participate in regulating the translation and degradation of mRNAs (3). miRNAs preferentially bind to complementary sites at the 3′UTR of their target mRNAs, which contribute to their pivotal regulation in a wide variety of biological processes, including cell growth, development, differentiation and apoptosis (4). In the development of GC, aberrant expression of miRNAs has been found to be correlated with the clinical features of GC, such as occurrence, development and metastasis (5).

Recent studies have aimed to evaluate the microRNAs present in serum/plasma as potential molecular biomarkers on account of their high stability and convenience in biological samples (e.g. miR-21, miR-148a and miR-124) (6,7). MicroRNA expression profiling has been used to achieve this goal in an effective way (8). An innovative screening strategy which is called machine learning was carried out in our study to select miRNAs from profiles and verify their efficacy more efficiently. Machine learning is the programming of computers to optimize a performance criterion using example data or past experience. A mathematical model is built by the theory of statistics, and learning is the execution of a computer program to optimize the parameters of the model using the training data with high speed and efficiency.

Herein, we excavated a profile of three combined serum miRNAs using machine learning and confirmed its excellent accuracy and reliability in the detection of GC.

Materials and methods

Ethics statements

The study was approved by the Human Research Review Committee of Huashan Hospital, Fudan University and written informed consents were obtained from all of the patients. The study conformed to The Code of Ethics of the World Medical Association (Declaration of Helsinki).

Study and microRNA selection

GEO (Gene Expression Omnibus) dataset search engine was used for the GC microRNA expression profiling studies. The following keywords: ‘miRNA’ OR ‘microRNA’ OR ‘miR’, ‘gastric’ OR ‘stomach’, ‘profiling’ OR ‘microarray’ were used to search for potential studies. We selected a total of three GEO profiles (GSE23739, GSE26595, GSE28700) as their sample sizes were sufficiently large. The GSE23739 profile represented 723 human and 76 human viral miRNAs in 40 normal and 40 cancerous gastric tissues. MicroRNAs identified from 60 primary GC tissues and 8 surrounding non-cancer tissues were used for microarray analysis in GSE26595. A total of 22 paired GC and normal tissues were processed in GSE28700. We eradicated any repetitive miRNAs and normalized data of each profile.

Machine learning

Initially, we treated GSE23739 as a training set and prepared to use GSE26595 and GSE28700 as validation sets. In order to analyze the data from different profiles, total data were standardized using a series of classifiers. These six classifiers (9) were performed to screen specific microRNAs to distinguish GC tissues from normal tissues: Compound covariate classifier (10), Diagonal linear discriminant analysis (DLDA) classifier (11), Bayesian CCP classifier (12,13), 1/3-Nearest Neighbor classifier (14), Nearest centroid classifier (15), and Support vector machines classifier (16). 1-Nearest Neighbor classifier and 3-Nearest Neighbor classifier are different types of this classifier to assess the stability of data. Leave one out cross validation was run to ensure stability and accuracy of the output.

Serum samples

All serum samples were collected from GC patients and healthy individuals treated at Huashan Hospital (Shanghai, China), affiliated to the Fudan University between 2014 and 2016. All of the 24 patients were clinically and pathologically diagnosed with GC. We evaluated the clinical and pathological features of the patients and these data are summarized in Table I.

Table I. Clinical characteristics of the GC patients vs. normal controls.

Table I.

Clinical characteristics of the GC patients vs. normal controls.

Characteristics	GC n=24	Controls n=20	P-value
Age (years), mean ± SD	57.6±14.0	53.2±12.5	P=0.2823
Sex			P=0.2929
Male	13	11
Female	11	9
Tumor location
Cardia	10
Body	9
Antrum	3
Other	2
Histology
Adenocarcinoma	24
Tumor size (cm)
≥5	13
<5	11
TNM stage
I+II	5
III+IV	19
Metastatic status
Yes	3
No	21

[i] GC, gastric cancer; TNM, Tumor, Node, Metastasis.

Serum preparation and microRNA extraction

Venous blood was collected in EDTA anticoagulation vacuum tubes and was centrifuged at 1000 × g for 15 min at 4ºC and then the separated serum was transferred into 1.5 ml RNase-free tubes stored at −80°C until RNA extraction. Small RNAs were extracted from 200 µl of serum using the miRcute miRNA Isolation kit (Tiangen Biotech Co., Beijing, China) according to the manufacturer's instructions.

Quantification of miRNA expression in serum by qRT-PCR

Reverse transcription (RT) reactions were performed using miRcute Plus miRNA First-Strand cDNA Synthesis kit (Tiangen Biotech Co.). Singleplex reactions were conducted in a volume of 20 µl which consisted of 10 µl 2X miRNA RT reaction buffer, 2 µl miRNA RT Enzyme Mix, and 8 µl miRNA template. Then, the RT reaction was carried out in a thermocycler under the following conditions: 42°C for 60 min, 95°C for 3 min, followed by a hold at 4°C. All the operations were performed with caution to exclude RNase contamination, and the total end products were preserved at −20°C for further analysis.

Quantitative PCR (qPCR) was performed using 7500 Fast Real-Time PCR System (Applied Biosystems, Foster City, CA, USA) with miRcute miRNA qPCR Detection kit (Tiangen Biotech Co.). All forward primers were obtained from Tiangen Biotech Co. Their catalog numbers are CD201-0092 (hsa-miR-21-5p), CD201-0404 (hsa-miR-29c-3p) and CD201-0305 (hsa-miR-22-3p). Then they were diluted to 10 µM before being adding to the PCR reaction mixture. qRT-PCR was performed in 3 duplicate reactions comprising 10 µl 2X miRcute miRNA Premix (with SYBR and ROX), forward and reverse primer, each with 0.4 µl, 2 µl miRNA First-Strand cDNA template, and 7.2 µl RNase-free ddH2O. Mixtures were denatured at 94°C for 2 min and then run for 40 cycles (94°C for 20 sec, 60°C for 34 sec). Melting curve analysis was run after all these procedures were completed. The expression levels of miRNAs in serum were normalized to U6 for the next quantification which was calculated using the 2−∆∆CT method.

Target gene analysis

The union of predicted target genes was searched using starBase v2.0 (http://starbase.sysu.edu.cn/index.php). The Gene Ontology and Genome Pathway were processed and produced by OmicsBean (http://www.omicsbean.cn/). Then we created biological networks employing Cytoscape v3.2 open-source software with CyTargetLinker App (17) and we treated miRTarBase (http://mirtarbase.mbc.nctu.edu.tw/) as the tool for selecting targets intersected by the results of three miRNAs in this software. Gene-disease association data were retrieved from the DisGeNET database (http://www.disgenet.org/). The term ‘gastric adenocarcinoma’ (umls: C0278701) was used to identify GC-associated genes. STRING (http://www.string-db.org/) was used to analysis the interaction between different proteins.

Statistical analysis

The clinical characteristics among groups were compared using the χ2 test and Fisher's exact test for qualitative data, and t-test for quantitative data. A receiver operating characteristic (ROC) curve was generated for the specificity and sensitivity value calculated by classifiers, which are represented by the area under the curve (AUC) value and 95% confidence intervals (CI). Experimental data are presented as means ± SD. The results were considered to be statistically significant at *P<0.05, **P<0.01, ***P<0.001, ****P<0.001.

Results

Training set marker selection

Firstly, treating GSE23739 as a training set after normalization, six types of classifiers (see Machine learning in Materials and method) were used to select markers. Since Compound covariate classifier, Diagonal linear discriminant analysis (DLDA) classifier and Support vector machine classifier are linear, we achieved the linear discriminant and calculated the gene weight value of diverse classifiers using maximum likelihood estimate (MLE). Using the same method, the threshold values of the Compound covariate classifier, DLDA classifier, and Support vector machine classifier were determined as 1.469, 2.724 and 0.747, respectively. We set each gene weight value as ωi and the expression of gene as xi. If a sample's ∑iωixi > threshold value, then it will be classified as cancerous. Following this principle, we calculated the accuracy of single or a small cluster of miRNAs to discriminate GC from normal tissue.

As elaborated in Fig. 1A, we found that the combination (miR-21-5p, miR-22-3p and miR-29c-3p) of these three markers had the greatest accuracy of the total 175 markers following three steps. The weight values of three miRNAs in the linear classifiers are listed in Table II. The accuracy predicted using 7 markers is documented in Table III. The heat map in Fig. 1B represents the expression data clustering analysis of the three markers in all 80 samples of GSE23739. By employing leave-one-out cross validation, we found the results for sensitivity and specificity in diverse classifiers. Then, we drew 3 ROC curves corresponding to Compound covariate classifier (Fig. 1C), DLDA classifier (Fig. 1D) and Support vector machine classifier (Fig. 1E). Thus, we determined the AUC values for the three curves, 0.9437, 0.9456 and 0.9563, which were high and reliable confirming these markers as having potential diagnostic criteria.

Figure 1. The selection of three miRNAs (miR-21, miR-22 and miR-29c) as markers in the training set. (A) The diagram illustrating the screening of microRNAs in GSE23739 following three steps. Step 1: Selection of 175 overlapping miRNAs (red color). Step 2: Using seven types of classifiers combined with Leave-one-out cross validation to screen a panel of miRNAs (number: 2–5). Step 3: Discovering the most highly accurate panel of three-miRNA group (red color), when setting the mis-classification rate as 0.2. (B) Heat-map diagram consisting of three miRNAs was drawn according to the miRNA profile (GSE23739). The 1–3 lines represent miR-21, miR-22 and miR-29c, respectively. The green color represents normal patients; the yellow color indicates cancer patients. The darkness of the heat map color represents the gene expression level; the color which is closer to red (the number 2) has the higher level of gene expression. (C-E) Receiver operating characteristic (ROC) curves of three linear classifiers are displayed in the following sequence: Compound covariate classifier, Diagonal linear discriminant analysis (DLDA) classifier and Support vector machine classifier.

Table II. Weight values of three miRNAs.

Table II.

Weight values of three miRNAs.

Genes	Compound covariate	Diagonal linear discriminant analysis	Support vector machines
hsa-miR-21-5p	−6.6501	−0.5599	−0.469
hsa-miR-22-3p	4.8583	0.815	0.2505
hsa-miR-29c-3p	6.9725	0.9744	0.9594

Table III. The accuracy of miR-21-5p, miR-22-3p and miR-29c-3p in GSE23739 using 7 classifiers (shown as percentages).

Table III.

The accuracy of miR-21-5p, miR-22-3p and miR-29c-3p in GSE23739 using 7 classifiers (shown as percentages).

Classifier	Compound covariate	Diagonal linear discriminant analysis	1-Nearest neighbor	3-Nearest neighbor	Nearest centroid	Support vector machines	Bayesian CCP
Accuracy	84	88	86	85	80	88	88

Marker validation

We performed 2 validation sets to validate the three markers. As mentioned above, we investigated the accuracy of prediction in GSE26595 and GES28700 by 7 classifiers (Table IV). The heat map of GSE26595 is shown in Fig. 2A. After calculating the sensitivity and specificity of the data in GSE26595, we determined the AUC of three curves by three linear classifiers separately (0.9563, 0.9625, 0.9688), which confirmed the feasibility of the diagnostic criteria (Fig. 2B-D). The three markers were validated again in another GEO microRNA profile (GSE28700). The heat map is presented in Fig. 2E and the AUC of ROC curves are 0.7645, 0.75 and 0.7789 (Fig. 2F-H).

Figure 2. The validation of two prediction sets. Heat map of GSE26595 (A) and GSE28700 (E) was created according to the relative expression of three miRNAs. The lines represent miR-21, miR-22 and miR-29c from the first to the last row. The discriminative ability of the three-miRNA panel is shown in GSE26595 (B-D) and GSE28700 (F-H) using ROC analysis following the sequence mentioned above. FPR, false-positive rate; TPR, true-positive rate; AUC, area under the receiver operating characteristic (ROC) curve.

Table IV. The accuracy of three miRNAs in validation sets (GSE26595 and GSE28700) using 7 classifiers (shown as percentages).

Table IV.

The accuracy of three miRNAs in validation sets (GSE26595 and GSE28700) using 7 classifiers (shown as percentages).

Classifier	Compound covariate	Diagonal linear discriminant analysis	1-Nearest neighbor	3-Nearest neighbor	Nearest centroid	Support vector machines	Bayesian CCP
GSE26595	88	88	88	88	87	93	93
GSE28700	64	56	56	60	73	69	71

Confirmation of the selected miRNAs in serum of GC patients

There were 20 samples from healthy volunteers regarded as control subjects and 24 serum samples from GC patients in this study. No significant differences in sex or age (Table I)were noted between the GC patients and the healthy volunteers (P=0.2823, P=0.2929, Student's t-test, respectively).

The expression of three candidate miRNAs (miR-21-5p, miR-22-3p and miR-29c-3p) was assessed by qRT-PCR in individual serum samples. The level of miR-21 was significantly upregulated in GC (P<0.0001) (Fig. 3A). Reversely, in Fig. 3B and C, levels of miR-22 and miR-29c were downregulated in the tumor group (both P<0.0001). These changes were consistent with the results in the training and validation set. Furthermore, we explored the relationship between the expression of these miRNAs with the clinical and pathological features of GC (Table V). There was a higher expression of miR-21 in GC patients with larger tumor sizes (≥5 cm) as previous reported (18).

Figure 3. The expression of candidate miRNAs in serum samples and analysis of target genes of the miRNAs. (A-C) The expression of miR-21 was significantly upregulated, while the expression of miR-22 and miR-29c were downregulated in 24 GC tissues compared to 20 control subjects. ****P<0.0001 compared to the healthy control group. Each value is the mean ± SD of three experiments. (D) Venn diagram illustrating the target genes targeted by three miRNAs, which were predicted by Starbase website. (E) Gene ontology (GO) analysis was conducted and the highest percentage of genes are involved in the nucleic acid binding activity of enriched processes of level 4. GC, gastric cancer; NC, normal controls.

Table V. Association between the three selected microRNAs and clinicopathological features of the GC patients.

Table V.

Association between the three selected microRNAs and clinicopathological features of the GC patients.

	miR-21		miR-22		miR-29c
Clinicopathological features	∆Ct	P-value	∆Ct	P-value	∆Ct	P-value
Tumor location		0.07		0.5		0.35
Cardia	3.07±0.93		4.87±0.83		0.19±0.86
Not in cardia	3.78±0.86		6.02±1.07		0.83±0.96
Tumor size (cm)		0.05		0.25		0.99
≥5	3.14±1.05		5.19±1.20		0.23±0.91
<5	3.89±0.62		5.95±0.89		0.96±0.89
TNM stage		0.38		0.29		0.27
Early (I+II)	3.14±0.49		4.77±0.69		−0.26±0.52
Later (III+IV)	3.58±1.03		5.74±1.14		0.78±0.95
Metastatic status		0.38		0.85		0.74
Yes	3.04±1.25		4.80±1.02		−0.14±0.81
No	3.55±0.89		5.65±1.11		0.66±0.95

a P<0.05 compared to the control group. GC, gastric cancer; TNM, Tumor, Node, Metastasis.

Target prediction and analysis

Initially, StarBase v2.0 was used to search for the targets of miR-21, miR-22, and miR-29c. All predicted targets of the deregulated miRNAs are illustrated in Fig. 3D. An overview of the Gene Ontology (GO) analysis indicated that the binding attribute of molecular function was high by OmicsBean website (data not shown). Fig. 3E further shows that the highest percentage of genes are involved in the nucleic acid binding activity of enriched processes of level 4. The top 20 highly enriched KEGG pathways are listed in Table VI. Noteworthy, the tumor-suppressor gene, PTEN, was discovered in the class of enriched KEGG pathways, such as: focal adhesion and PI3K-Akt signaling pathway.

Table VI. KEGG pathway analysis of shared target genes of the three miRNAs.

Table VI.

KEGG pathway analysis of shared target genes of the three miRNAs.

Pathway name	Pathway ID	P-value	Genes
Protein digestion and absorption	hsa04974	1.30E-10	COL7A1; COL3A1
ECM-receptor interaction	hsa04512	1.38E-06	COL4A5; COL4A4
Focal adhesion	hsa04510	6.21E-06	PTEN; COL4A5; COL4A4
PI3K-Akt signaling pathway	hsa04151	3.19E-05	PTEN; COL4A5; COL4A4
Small cell lung cancer	hsa05222	8.28E-05	PTEN; COL4A5; COL4A4
Amoebiasis	hsa05146	2.94E-04	COL3A1; COL4A5; COL4A4
Insulin resistance	hsa04931	5.89E-04	PTEN; PPARA; RPS6KA3
Proteoglycans in cancer	hsa05205	1.04E-03	ESR1; CBL; FLNA; HGF
Phosphatidylinositol signaling system	hsa04070	4.31E-03	PTEN; PIKFYVE; CALM1
MAPK signaling pathway	hsa04010	7.61E-03	RASGRP1; RPS6KA3; FLNA
Pathways in cancer	hsa05200	8.69E-03	PTEN; RASGRP1; COL4A5
N-Glycan biosynthesis	hsa00510	9.19E-03	ALG9; ALG1; MAN1A2
Inositol phosphate metabolism	hsa00562	1.10E-02	PTEN; PIKFYVE; MTMR2
Melanoma	hsa05218	1.10E-02	PTEN; HGF; PTEN; CDK6
Neurotrophin signaling pathway	hsa04722	1.41E-02	CAMK4; CALM1; RPS6KA3
AGE-RAGE signaling pathway in diabetic complications	hsa04933	1.74E-02	COL3A1; COL4A5; COL4A4
Glioma	hsa05214	2.84E-02	PTEN; CALM1; PTEN; CDK6
Long-term potentiation	hsa04720	3.01E-02	CAMK4; GRM5; CALM1
Prostate cancer	hsa05215	3.03E-02	PTEN; CREB1; PTEN; CREB5

Furthermore, Cytoscape v3.2 software was applied to focus on the shared genes by both miR-22 and miR-29c. Their decreased level may contribute to GC by upregulating various oncogenes. By applying miRTarBase database in CyTargetLinker App to search for targets, we created a biological network with numerous nodes which stand for target genes and gave attention to the shared targets between two miRNAs (Fig. 4A). However, there may have been several missing targets as this database has not been included. Sp1, which is a transcription factor and may be associated with poor prognosis of GC patients, was selected from the targets of miR-22 and its node is indicated in orange (19). Recent research also reveals that miR-29c could target Sp1 in lung cancer (20). Therefore, SP1 has been confirmed to be a shared target between two downregulated miRNAs (miR-22 and miR-29c).

Figure 4. The network of candidate miRNAs and the possible mechanisms. (A) The network includes the individual miRNAs (red circles) and their predicted mRNA target genes (hexagons), obtained from miRTarBase databases. The orange and yellow colors highlight the predicted genes (SP1 and PTEN). (B) The protein-protein-interaction was drawn using STRING website. (C) The diagram illustrates the possible mechanism involving the three miRNAs. GC, gastric cancer.

STRING was used to screen genes between Sp1 and 476 targets of miR-21 predicted by miRTarBase. A total of 38 relevant genes were selected out. Next, we search for the gene list related to gastric adenocarcinoma in DisGeNet database (umls: C0149826). Then we took the intersection between 38 genes related to Sp1 and 284 GC-related genes. Twelve genes were selected out and STRING was used to predict the interaction between them (Fig. 4B). Finally, we gave attention to PTEN, which has been reported to be targeted by miR-21 in GC and was transcriptionally inhibited by Sp1 (21–24). Sp1 and PTEN were consistent with the results we found in GO analysis and KEGG pathway. In conclusion, Fig. 4C illustrates that the higher level of SP1 and lower level of PTEN may contribute to the progression of GC.

Discussion

It is crucial to identify practical biomarkers for the detection of gastric cancer (GC) in order to improve patient outcomes. In comparison to tissue biopsy, using serum miRNAs as biomarkers is simple, has a lower cost and is non-invasive, which has benefit for the screening and monitoring of tumors (25). A combination of miR-21, miR-22 and miR-29c was identified using machine learning. Their ROC analyses in the training set revealed marked AUC (0.9437, 0.9456 and 0.9563) with more than 80% positive predictive value (PPV)and negative predictive value (NPV) in three linear classifiers. Then we validated them in two training sets and the serum of patients for further confirmation. It should be noted that the number of samples was few and we required further proof to validate these three biomarkers in GC. In the final step, we used tools to identify their targets and elucidate the possible mechanisms.

Application of machine learning to large databases is also called data excavation which means that a large volume of raw data are processed into a small amount of precious material using classifier models. During recent years, machine learning has been widely utilized as a method to predict the progression, susceptibility and recurrence of cancerous conditions (26). For example, machine learning models, including Support vector machine (SVM) classifier, was used to predict childhood acute lymphoblastic leukemia (ALL) relapse based on medical data (27). In the field of diagnostics, Chen et al (28) applied four classical machine learning-based classifications to estimate the stage of hepatic fibrosis. Radiomic machine-learning classifiers were applied for prognostic biomarkers of advanced nasopharyngeal carcinoma (29). In this study, we fully utilized the predominance of machine learning and used it for the screening of biological biomarkers in GC. The panel of miR-21, miR-22, and miR-29c was found to have the highest accuracy in predicting GC tissues (Fig. 1A). Machine learning is a novel method that is worthy to be popularized in identifying biomarkers in different types of disease.

Among the three miRNAs identified in this study, miR-21-5p was upregulated in the serum of GC patients, which was consistent with previous research (6,30). As early as 2008, Zhang et al (31) found that miR-21 could regulate GC cell invasion and migration. Concurrently, Chan et al (32) verified that miR-21 was overexpressed in 92% (34/37) of GC samples and PTEN may be a target gene of miR-21 (21). H. pylori infection was found to induce miR-21 and the level of miR-21 was upregulated in gastric juice of GC patients (33). Accumulating evidence indicates that miR-21 can serve as a diagnostic candidate for GC. In contrast, the levels of miR-22 and miR-29c were decreased in our serum samples, which was in accordance with results in other research (34). miR-29 family plays a vital role in tumor-related changes including cell proliferation, cell cycle, cell differentiation, apoptosis and metastasis (35). Han et al (36) showed that miR-29c suppressed the initiation of gastric carcinogenesis in transgenic mouse models. Sufficient evidence revealed that the level of miR-22 was downregulated in GC, which was related to lymph node metastasis, poor prognosis in patients (37), and acted as a metastasis suppressor by directly targeting Sp1 (38). Therefore, the combination of these three miRNAs would achieve more specificity than separate miRNAs in the prediction of GC, achieving high accuracy.

Next, we aimed to explain how these three miRNAs work together. Sp1, which functions as a transcription factor, is a ubiquitously expressed, zinc finger-containing DNA binding protein that can activate or repress transcription in a variety of diseases (39). It is overexpressed in GC and is closely correlated with poor outcome (40). miR-22 targets Sp1 and represses GC (38), while miR-29c may function in the same way (41). PTEN is one of the well-known tumor suppressor gene that plays a crucial role in various types of tumors including GC (42) and was validated to be targeted by miR-21 (21–24). Sp1 can inhibit PTEN promoter activity through a specific Sp1-binding site at the PTEN core promote (43). The mechanism is summarized in Fig. 4C. miR-22 and miR-29c both suppress the level of Sp1 and miR-21 suppresses the expression of PTEN inhibited by Sp1, which contributes to the development of GC.

In summary, our study revealed that miRNAs or other biomarkers could be excavated effectively by machine learning. Three miRNAs were screened: miR-21, miR-22 and miR-29c. Their diagnostic potential was evaluated by various classifiers and AUC curves. We then verified their differential expression in the serum of patients and explained this phenomenon by predicting their targets. Further studies will aid in confirming this serum miRNA panel for the diagnosis of GC.

Acknowledgements

This research was supported by a grant from the Shanghai Municipal Commission of Health and Family Planning (no. 20134132).

References