Machine learning algorithms reveal potential miRNAs biomarkers in gastric cancer

demographic information

Clinicopathologic information for 348 (64.9%) men and 188 (35.1%) women included in this study was downloaded from TCGA and summarized in Table 1. The average age was 65.3 years, and approximately 250 (46.6%) had advanced gastric cancer.

Correlation analysis

Of the clinicopathological data, only stage was significantly associated with cancer. An important criterion for measuring relevance is p– < 0.05 value (Figure 2A).

Data collection

As mentioned in the Materials and Methods section, the source of raw data from clinical information and sequencing was the TCGA database. Based on the criteria mentioned, 536 samples were selected for further study, of which approximately 465 were related to GC patients and 72 to age- and sex-matched controls.

Data preprocessing and identification of differentially expressed miRNAs (DeMiRs)

The dataset contained 1882 miRNAs, which was reduced to 220 miRNAs after normalization using the Limma package, R software. In a processing step using heat maps, the most significant features were selected (Figure 2B) and classified using machine learning algorithms. Then 5 algorithms (SVM, dts, rf, logistic regression, and knn) were tested with 4 different metrics (accuracy, f1score, ROC_curve, and confusion matrix), and finally obtained from these 4 metrics According to the scores, the SVM algorithm was chosen as the most accurate algorithm. (DTS, Accuracy: 88%, AUC = 47%; random forest, Accuracy: 93%, AUC = 39.5%; SVMs, Accuracy: 93%, AUC = 88.5%; KNN, Accuracy: 93%, AUC = 41.7%; logistics, Accuracy: 93%, AUC = 88%). The confusion matrix can also be seen in Figure 2C. As a result, a list of 29 miRNAs with 5 significant up and 24 significant down expression in gastric cancer was selected for further analysis (Table 2) Figure 3.

ROC curve analysis for identification of diagnostic biomarkers

The results of ROC curve analysis were hsa-miR-93 (combined AUC was 0.76, sensitivity was 0.69, specificity was 0.73, cutoff was 0.86) (Figure 2D).

Survival analysis of demiR

Survival analysis of demiR was performed using SPSS version 20, p– Values ​​were considered < 0.05. As a result, 13 miRNAs (Hsa-miR-21, Hsa-miR-146b, Hsa-miR-185, Hsa-miR-1.1, Hsa-miR-1.2, Hsa-miR-143, Hsa-miR-4652, Hsa-miR-1911, Hsa-miR-29c, Hsa-miR-3170, Hsa-miR-139, Hsa-miR-5683, Hsa-miR-133a.2) have prognostic function (Figure 4).

Validation of candidate microRNAs in the dataset

Of the 29 candidate microRNAs obtained from the machine learning algorithm using the online web server described in the Materials and Methods section, hsa-miR-21, hsa-miR-133a, hsa-miR-146b, hsa-miR-146b, hsa- miR-29c, and HSA-MIR-204 can be found in （gse54397）、exp004526（gse54397）で高度に検証されていました。 GSE106817)、EXP00405、EXP00118(GSE28700)、EXP00406、EXP00666、EXP00444 (GSE78775)、EXP00476 (GSE99415)、EXP00316 (GSE77380)、および EXP00175(GSE33743 )) was performed using the supplemental file (httpsR miPath heatmap server 1, run online/mpd.bioinf.uni-sb.de/, Figure 5A).

miRNA target prediction

Prediction of miRNA targets was achieved using several databases including miRwalk, miRdb and Targetscan. Using the Venn diagram online database, a list containing 407 common genes was identified (Figure 5B).

Protein-protein interaction network analysis

The candidate genes predicted in the previous step were submitted to the STRING database to build a PPI network based on the criteria described in Materials and methods. To obtain hub genes with important roles, PPI networks were imported and visualized by Cytoscape software. We used the Cytohubba tool to select 100 high-scoring genes based on degree (Fig. 5C). Finally, we imported the hub genes into the string database and re-analyzed the PPI network (Figure 5D).

Functional analysis

To reveal the role of selected hub genes, we performed enrichment analysis using R software. Results demonstrate hub gene transcription factor binding, enzyme binding, RNA polymerase II cis-regulatory region sequence-specific DNA binding, protein binding, double-stranded DNA binding, arrestin family protein binding, sequence-specific DNA binding, and chromatin binding. bottom. In terms of molecular function, most genes are involved in miRNA-mediated inhibition of translation, positive regulation of viral transcription by the host, regulation of gene expression by gene imprinting, production of miRNAs involved in gene silencing by miRNAs, Gene silencing, enriched Wnt signaling pathways, calcium regulatory pathways, regulation of cellular senescence, negative regulation of gene expression, and epigenetic gene silencing on biological processes. Chromatin, euchromatin, nucleoplasm, non-membrane-bound organelles, and cytosol were the most enriched cellular components. Identification of key signaling pathways using the KEGG database indicates that candidate hub genes are primarily associated with glioma, melanoma, prostate cancer, non-small cell lung cancer, renal cell carcinoma, GnRH secretion, aldosterone-regulated sodium reabsorption, and pancreatic It has been shown to be involved in cancer (Fig. 6).