Identification of hub genes and diagnostic efficacy for triple-negative breast cancer through WGCNA and Mendelian randomization

All measured genes expression and grouping information from this dataset can be gained from the Gene Expression Omnibus (GEO) database. This dataset (GSE38959) was counted with transcriptome microarray assays in the mammary ductal cells obtained from TNBC tissues by means of immunohistochemical staining (N = 30) and normal tissues (N = 13) [17].

Initially, the dataset GSE38959 was read through R software version 4.2.1. The dataset underwent preprocessing for correction and data normalization. Following this, DEG analysis searching was conducted by means of the "limma" package between TNBC and normal samples, and the adjusted p-value and |log fold changes (FC)| were calculated for each gene. Genes that met the criteria, adjusted p-value < 0.05 and |logFC|≥ 1.0, were considered as DEGs. Expression levels were analyzed, and volcano diagram and DEGs expression heatmap were generated using the R packages "pheatmap" and "ggplot2".

The study employed a methodical procedure of WGCNA to construct a gene co-expression network specific to triple-negative breast cancer. The WGCNA approach is frequently used to identify highly synergistic genomes and possible markers through an assessment of the interrelationship between such genomes and their relationship to phenomena [9]. By evaluating the interaction between each module and the molecular mechanism of triple-negative breast cancer, the most prominent module was selected as the central gene chosen by WGCNA.

To gain insight into triple-negative breast pathogenesis, we generated intersections and Venn plots for the candidate hub genes of WGCNA and DEG. To comprehend the potential mechanisms underlying progression and pathogenesis, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis were conducted using the "clusterProfiler" R package [18].

Trough the STRING (https://​string-db.​org/​) platform and Cytoscape (https://​cytoscape.​org/​) software, molecular interaction as well as PPI networks was predicted and visualized. The first step involved the use of the Degree algorithm in Cytoscape (https://​cytoscape.​org/​) to rank and score the significant genes in the PPI networks. In the next step, we focused on the top 50 proteins arranged by degree and drew the protein–protein interaction for further analysis. Furthermore, top 5 proteins arranged by degree were selected was hub genes.

The "rms" package was utilized to construct a nomogram model for assessing the risk of TNBC [19]. The predictive power of the nomogram model was evaluated by Harrell's concordance index [20]. Additionally, the "DynNom" R package was employed to explore the dynamics of TNBC risks. In order to determine the diagnostic efficacy of the candidate genes, the "ROC" package was employed to construct the receiver operating characteristic (ROC) curve. The accuracy of the ROC curve was indicated by the area under the curve (AUC) classified as high (AUC ≥ 0.9) [21].

The involvement of immune cells in TNBC was investigated by evaluating the degree of infiltration of 22 immune cells using cibersort analysis based on DEGs [22]. Stacked diagram was used to investigate the proportion in different samples. Next, we filtered the samples that met the conditions by p-value (p < 0.05). We plotted heatmaps to explore the infiltration of 22 immunocytes in each sample and violin maps to explore differential immune cells between TNBC and normal breast tissue cells. We utilized the 'corrplot' package to create heatmaps visualizing the correlation between 22 types infiltration of immunocytes. Furthermore, we used the 'ggplot2' package to analyze the correlation between immune cells and CCNB1 gene expression to investigate its role in the development of immune cells in TNBC tissues.

We further validated our results by applying a consistent DEG selection method (|logFC|≥ 1.0, adjusted p-value < 0.05) to two additional independent external datasets (GSE45827 and GSE65194), included 41 TNBC specimens and 11 normal breast specimens respectively, and dataset used in this study. Venn diagrams were generated to compare the DEGs identified from the three datasets. Interestingly, we found that the CCNB1 gene was located at the intersection of the three datasets, indicating the robustness of our findings.

To assess the clinical outcome, the CCNB1 gene was subjected to the Kaplan–Meier (KM) plotter (https://​kmplot.​com/​analysis/​) [23]. The KM plotter mRNA breast cancer database was applied to evaluate the prognostic values of CCNB1 in TNBC patients. In this study, TNBC patients were selected based on ER-negative, PR-negative assessed by IHC and HER2 -negative assessed by array. Probes of genes were selected based on the “only JetSet best probe set”. We plotted KM survival curves for the three main survival outcomes, including recurrence-free survival (RFS), overall survival (OS), and distant metastasis-free survival (DMFS).

All statistics of the study was utilized in the open database. The Genome-Wide Association Study (GWAS) data source of CCNB1 was attained from ieu open GWAS project. The GWAS of the phenotype “G2/mitotic-specific cyclin-B1” was obtained in this study, including 3,301 samples and 10,534,735 SNPs. GWAS summary statistics of TNBC were obtained from the Breast Association (BRAC) and Discovery, Biology and Risk of Inherited Variants in Breast Cancer Consortium (DRIVE) [24]. In this study, inverse variance weighted (IVW) estimates was applied for the main analysis, which combined the Wald ratio of each SNP on the outcome and obtained an overall causal estimate. The assumption that the genetic variant influences the outcome only through the exposure was assessed for potential violation due to horizontal pleiotropy. If such pleiotropy exists, it would lead to bias in the causal estimates. To address this, analytical approaches were employed. Heterogeneity of the analyses was estimated by means of Cochran's Q test and its corresponding p-value. Furthermore, several statistical tests were performed to detect potential bias and pleiotropy. These tests included MR-Egger, weighted median, MR-PRESSO, single SNP analysis, and leave-one-out analysis. The MR-Egger method was used to correct for potential pleiotropy and obtain consistent causal inference in the presence of invalid instrumental variables. On the other hand, the weighted median approach was employed if invalid instrumental variables contributed to at least half of the weight in the analyses [25, 26]. The MR-PRSSO method was used to identify the Outlier SNP and correct the results to avoid potential horizontal pleiotropy [27]. In order to visualize our results, plots, including forest plot, funnel plot, scatter plot, and leave-one-out plot, were made to describe the robustness of the causal estimates of the MR analyses.

The TNBC dataset (GSE38959) was obtained from the GEO database and analyzed. By comparing the TNBC group with the normal group, 1850 DEGs were identified, consisting of 1004 upregulated genes and 846 downregulated genes (Fig. 1A, B; Supplementary Table S1).

To explore the relationship between the potential gene modules and TNBC, we conducted WGCNA analysis on all candidate genes from the TNBC dataset (GSE38959) (Fig. 2A). Through this analysis, we identified 16 distinct modules (Fig. 2B). Subsequently, by analyzing the positive correlation coefficients, we were able to isolate the module turquoise from the GSE38959 dataset (Fig. 2C; Supplementary Table S2).

By employing Venn diagrams, we identified 1585 overlapping genes as candidate hub genes, demonstrating potential significance in the progression of TNBC (Fig. 3A). To explore the co-expression of genes between the candidate hub genes derived from WGCNA and the DEGs, we conducted GO and KEGG analyses. The GO enrichment analysis revealed that these overlapping genes primarily impacted biological functions such as organelle fission, nuclear division, chromosome segregation, and mitotic cell cycle phase transition (Fig. 3B, C, D). Moreover, the KEGG enrichment analysis demonstrated the influence of these overlapping genes on cellular functions such as the cell cycle, amyotrophic lateral sclerosis, motor proteins, and cellular senescence (Fig. 3E, F).

In order to create a protein–protein interaction network of hub genes, we utilized the STRING online tool. (Fig. 4A). The Degree algorithm in “Cytoscape” was used to rank and score the significant genes in the PPI networks. Protein interaction networks were mapped for the top 50 proteins to investigate potential mechanisms of TNBC development (Fig. 4B). Mainly, top5 proteins (CDC2, CCNB1, CCNA2, TOP2A, CCNB2) were selected was hub genes. The darker color of the circle stands for the higher score.

A nomogram model was created to estimate TNBC risk (Fig. 5A). The "DynNom" R package was used to achieve the predict risk of TNBC with dynamic data. We then computed ROC curves of the five genes to investigate their diagnostic efficacy. The AUC of our nomogram model was also calculated to differentiate between TNBC and controls (Fig. 5B), demonstrating its effectiveness. The under areas of CDC2, CCNB1, CCNA2, TOP2A, and CCNB2 were 0.967, 0.974, 0.938, 0.910, and 0.867. Thus, our nomogram model accurately predicted the risk of TNBC, as demonstrated by the AUC values providing an accurate assessment of the diagnostic effect.

For confirming the relationship between CCNB1 level and immune cells, the proportion of 22 immunocytes was analyzed using the ‘cibersort’ R package. A stacked diagram was drawn to show immune cell infiltration proportion in different samples (Fig. 6A). A heatmap of infiltration of 22 immunocytes in each sample (Fig. 6B) was plotted and violin plots indicated significant differences in 11 immune cells between TNBC samples and normal tissue (Fig. 6C). We also calculated the correlation between immune cells by correlation analysis (Fig. 6D) were identified. The correlation analysis between CCNB1 expression and immune cells revealed significant correlations with three immune cell types (Fig. 7A). Memory B cells (Fig. 7B) and follicular helper T cells (Fig. 7C) exhibited a negative correlation with the expression of CCNB1, whereas there was a favorable link between activated CD4 memory T cells and CCNB1 (Fig. 7D). This study provides further evidence supporting the hypothesis that immune cell activity and infiltration may be influenced by the level of the hub gene CCNB1.

We obtained 4406 genes from GSE45827 and 3627 genes from GSE65194, respectively. Venn diagrams were generated to compare the DEGs identified from the three datasets (Fig. 8A). The expression of CCNB1 shows significant differences in each dataset (Fig. 8B–D). Interestingly, we found that the CCNB1 gene was located at the intersection of the three datasets, indicating the robustness of our findings.

To investigate the prognostic values of the CCNB1, the KM plotter bioinformatics analysis platform was used. We found that high expression of CCNB1 was associated with unfavorable overall survival of TNBC patients but it did not reach statistical significance (HR = 1.61; P = 0.23; n = 144) (Fig. 8E). While, overexpression of CCNB1 was an unfavorable prognostic factor of recurrence-free survival (HR = 1.90; 95% CI 1.13–3.19; P = 0.014; n = 220) (Fig. 8F) and distant metastasis-free survival in TNBC patients (HR = 1.97; 95% CI 1.10–3.52; P = 0.02; n = 190) (Fig. 8G).

Supplementary Table S3 displayed the SNP characteristics of CCNB1 (P < 5*10^–5). None of SNPs were considered weak instrumental variables. According to the three main assumptions of Mendelian randomization, the removal of five SNPs (rs215086, rs34383011, rs12198798, rs622354, rs117318310). The causal relationships of each genetic variation on TNBC were illustrated in Fig. 9A and B. Using the IVW method, we examined the causal relationship between CCNB1 and TNBC. The results revealed that each one-unit increase in log odds of CCNB1 led to a 2.8% higher risk of TNBC (OR: 1.028, 95% CI 1.002–1.055, p = 0.032). Additionally, significant statistical significance was observed with the MR–Egger method (OR = 1.092, 95% CI 1.024–1.166, p = 0.009) and the weighted median method (OR = 1.035, 95% CI 1.000–1.071, p = 0.049). As demonstrated in Fig. 9C, the funnel plot exhibited an approximate symmetrical causal effect. Moreover, there was no indication of heterogeneity, according to the MR Egger regression intercept (p = 0.369), suggesting that pleiotropy did not appear (p = 0.052). In Fig. 9D, we conducted a systematic MR analysis on the remaining SNPs after excluding every SNP individually, and the results remained significant. This demonstrates that all SNPs contributed significantly to the causality. Therefore, we can conclude that there was no dominant SNP in the relationship between CCNB1 levels and TNBC, validating the previous MR findings.