IDMIR: identification of dysregulated miRNAs associated with disease based on a miRNA-miRNA interaction network constructed through gene expression data

Abstract

Micro ribonucleic acids (miRNAs) play a pivotal role in governing the human transcriptome in various biological phenomena. Hence, the accumulation of miRNA expression dysregulation frequently assumes a noteworthy role in the initiation and progression of complex diseases. However, accurate identification of dysregulated miRNAs still faces challenges at the current stage. Several bioinformatics tools have recently emerged for forecasting the associations between miRNAs and diseases. Nonetheless, the existing reference tools mainly identify the miRNA-disease associations in a general state and fall short of pinpointing dysregulated miRNAs within a specific disease state. Additionally, no studies adequately consider miRNA-miRNA interactions (MMIs) when analyzing the miRNA-disease associations. Here, we introduced a systematic approach, called IDMIR, which enabled the identification of expression dysregulated miRNAs through an MMI network under the gene expression context, where the network's architecture was designed to implicitly connect miRNAs based on their shared biological functions within a particular disease context. The advantage of IDMIR is that it uses gene expression data for the identification of dysregulated miRNAs by analyzing variations in MMIs. We illustrated the excellent predictive power for dysregulated miRNAs of the IDMIR approach through data analysis on breast cancer and bladder urothelial cancer. IDMIR could surpass several existing miRNA-disease association prediction approaches through comparison. We believe the approach complements the deficiencies in predicting miRNA-disease association and may provide new insights and possibilities for diagnosing and treating diseases. The IDMIR approach is now available as a free R package on CRAN (https://CRAN.R-project.org/package=IDMIR).

Keywords: immunotherapy; miRNA-disease association; miRNA–miRNA interaction; network analysis; specific disease context.

Figure 1

Schematic diagram of the IDMIR approach. The details of each step are shown in the diagram. In step 1, we perform data collection: G, GO terms from MSigDB. The circles represent the genes in the GO terms; miR, miRNAs integrated from four databases. The circles represent genes targeted by miRNAs; GEP, a disease gene expression profile of interest, which used to calculate the GDEscore. In step 2, a miRNA/GO bipartite graph in the context of a disease is constructed. In step 3, an MMI network in the context of a disease is constructed. In step 4, the dysregulated miRNAs are identified through a network diffusion algorithm.

Figure 2

Assessing the predictive outcomes of IDMIR in BRCA. (A) A miR/G bipartite network consisting of nine dysregulated miRNAs interacting with corresponding GO terms. All GO terms were categorized into six classes. (B) Comparison of IDMIR with three other approaches. We apply IDMIR to the TCGA-BRCA cohort to compare the performance with the three other approaches based on AUROC values. (C) Box plots display AUROC values for predicted results after different data removal scenarios. The dashed line denotes the initial AUROC value. (D) Venn diagram showing the overlapped number of the top 50 dysregulated miRNAs identified in the TCGA-BRCA, ICGC-BRCA and GSE42568 cohorts.

Figure 3

Training and validation of miR-21-5p-based IMRS model. (A) Heatmap of the expression values of miR-21-5p target genes in tumor and normal samples of the training set. The heatmap displays 38 target genes with significantly differential expression in the two sample classes (two-sided Wilcoxon rank-sum test; ^****, P < 0.0001; ^***, P < 0.001; ^**, 0.001 < P < 0.01; ^*, 0.01 < P < 0.05). (B) The forest plot shows HR, the 95% CI and beta coefficients of multivariable Cox regression analysis. (C–E) Kaplan–Meier survival curves were generated by using the miR-21-5p-based IMRS model to classify patients into IMRS-high and IMRS-low groups in both the training and validation sets.

Figure 4

Evaluation of the MMRS model in TCGA-BRCA cohort. (A) Kaplan–Meier survival curves of patients classified into MMRS-high and MMRS-low groups using the MMRS model. (B) Time-dependent ROC curves for prognosis prediction of the MMRS model for 1-, 3- and 5-year overall survival. (C) Comparison of the time-dependent AUROC values from 1–10 years of overall survival between MMRS and different IMRS models. (D) Multivariable Cox analysis of the MMRS and clinicopathological factors (age, T stage, N stage, M stage and tumor grade) for overall survival. (E) The line chart presents AUROC values for predicting 1–10 years of overall survival by the MMRS model and clinicopathological factors.

Figure 5

Biological mechanism analysis between MMRS-high and -low groups. (A) Ridgeline plot depicting the expression distribution of core genes for the significant pathways. The distribution located to the right of the dashed line indicates pathway enrichment in the MMRS-high group, while the opposite indicates pathway enrichment in the MMRS-low group. (B) Enrichment curves of GSEA for some important pathways. (C) Analyzing the immune microenvironment characteristics in relation to MMRS (two-sided Wilcoxon rank-sum test).

Figure 6

Analysis of MMRS in relation to clinical outcomes of BLCA patients. (A–C) Assessing the predictive outcomes of IDMIR in BLCA. (D) Kaplan–Meier survival curves of patients classified into MMRS-high and MMRS-low groups using the MMRS model in the TCGA-BLCA cohort. (E) Time-dependent ROC curves for prognosis prediction of the MMRS model for 1-, 3- and 5-year overall survival in the TCGA-BLCA cohort. (F) Validation of the prognosis prediction of the MMRS model in the IMvigor210 cohort. (G) Assessing the relationship between immunotherapy response and MMRS in the IMvigor210 cohort (two-sided chi-squared test). The stacked bar plot displays the sample count of responders (CR and PR) in the MMRS-high and MMRS-low groups. (H and I) performing the same analysis as (F) and (G) in the GSE176307 cohort.

Figure 7

Identification of dysregulated miRNAs across multi-cancer types. (A) Comparison of top ten dysregulated miRNAs identified in 14 cancer types. The miRNAs identified in at least two cancers were displayed. (B) Comparison of miR-21-5p expression levels between different tumor and normal samples (two-sided Wilcoxon rank-sum test; ^****, P < 0.0001). (C) Sankey diagram illustrates the relationship between miR-21-5p, its targets and pathways. (D) The lollipop chart displayed the P-value of univariate Cox analysis miR-21-5p-based IMRS in different cancers.