1 Introduction

1.1 What is MIRit

MIRit (miRNA integration tool) is an open-source R package that aims to facilitate the comprehension of microRNA (miRNA) biology through the integrative analysis of gene and miRNA expression data deriving from different platforms, including microarrays, RNA-Seq, miRNA-Seq, proteomics and single-cell transcriptomics. Given their regulatory importance, a complete characterization of miRNA dysregulations results crucial to explore the molecular networks that may lead to the insurgence of complex diseases. Unfortunately, there are no currently available options for thoroughly interpreting the biological consequences of miRNA dysregulations, thus limiting the ability to identify the affected pathways and reconstruct the compromised molecular networks. To tackle this limitation, we developed MIRit, an all-in-one framework that provides flexible and powerful methods for performing integrative miRNA-mRNA multi-omic analyses from start to finish. In particular, MIRit includes multiple modules that allow to perform:

  1. Differential expression analysis, which identifies miRNAs and genes that vary across biological conditions for both RNA-Seq and microarray experiments (even though other technologies can also be used);
  2. Functional enrichment analysis, which allows to understand the consequences of gene dysregulations through different strategies, including over-representation analysis (ORA), gene-set enrichment analysis (GSEA) and Correlation Adjusted MEan RAnk gene set test (CAMERA);
  3. SNP association, that links miRNA dysregulations to disease-associated SNPs occurring at miRNA gene loci;
  4. Target identification, which retrieves predicted and validated miRNA-target interactions through innovative state-of-the-art approaches that limit false discoveries;
  5. Expression levels integration, which integrates the expression levels of miRNAs and mRNAs for both paired data, through correlation analyses, and unpaired data, through association tests and rotation gene-set tests;
  6. Topological analysis, which implements a novel approach called Topologically-Aware Integrative Pathway Analysis (TAIPA) for identifying the impaired molecular networks that affect biological pathways retrieved from KEGG, Reactome and WikiPathways.

1.2 How to cite MIRit

If you use MIRit in published research, please cite:

Ronchi J and Foti M. ‘MIRit: an integrative R framework for the identification of impaired miRNA-mRNA regulatory networks in complex diseases’. bioRxiv (2023). doi:10.1101/2023.11.24.568528

This package internally uses different R/Bioconductor packages, remember to cite the appropriate publication.

1.3 Installation

Before starting, MIRit must be installed on your system. You can do this through Bioconductor.

## install MIRit from Bioconductor
if (!requireNamespace("BiocManager", quietly = TRUE)) {
    install.packages("BiocManager")
}
BiocManager::install("MIRit")

If needed, you could also install the development version of MIRit directly from GitHub:

## install the development version from GitHub
if (!requireNamespace("BiocManager", quietly = TRUE)) {
    install.packages("BiocManager")
}
BiocManager::install("jacopo-ronchi/MIRit")

When MIRit is installed, we can load it through:

## load MIRit
library(MIRit)

2 Data preparation

2.1 Load example data

To demonstrate the capabilities of MIRit we will use RNA-Seq data from Riesco-Eizaguirre et al. (2015). This experiment collected samples from 8 papillary thyroid carcinoma tumors and contralateral normal thyroid tissue from the same patients. These samples were profiled for gene expression through RNA-Seq, and for miRNA expression through miRNA-Seq. To provide an easy access to the user, raw count matrices have been retrieved from GEO and included in this package.

To load the example data, we can simply use the data() function for both gene and miRNA count matrices.

## load count matrix for genes
data(geneCounts, package = "MIRit")

## load count matrix for miRNAs
data(mirnaCounts, package = "MIRit")

2.2 Paired vs unpaired data

When using MIRit, we must specify whether miRNA and gene expression values derive from the same individuals or not. As already mentioned, paired data are those where individuals used to measure gene expression are the same subjects used to measure miRNA expression. On the other hand, unpaired data are those where gene expression and miRNA expression derive from different cohorts of donors. Importantly, MIRit considers as paired samples those data sets where paired measurements are available for at least some samples.

In our case, miRNA and gene expression data originate from the same subjects, and therefore we will conduct a paired samples analysis.

2.3 Set up expression matrices

As input data, MIRit requires miRNA and gene expression measurements as matrices with samples as columns, and genes/miRNAs as rows. Further, the row names of miRNA expression matrix should contain miRNA names according to miRBase nomenclature (e.g. hsa-miR-151a-5p, hsa-miR-21-5p), whereas for gene expression matrix, row names must contain gene symbols according to HGNC (e.g. TYK2, BDNF, NTRK2).

These matrices may handle different types of values deriving from multiple technologies, including microarrays, RNA-Seq and proteomics. The only requirement is that, for microarray studies, expression matrices must be normalized and log2 transformed. This can be achieved by applying the RMA algorithm implemented in the oligo (Carvalho and Irizarry 2010) package, or by applying other quantile normalization strategies. On the contrary, for RNA-Seq and miRNA-Seq experiments, the simple count matrix must be supplied.

Eventually, expression matrices required by MIRit should appear as those in mirnaCounts and geneCounts, which are displayed in Tables 1 and 2.

Table 1: Gene expression matrix
An expression matrix containing normalized values, with row names according to hgnc.
PTC 1 PTC 2 PTC 3 PTC 4 PTC 5
A1BG 7 6 9 12 7
A1BG-AS1 20 8 22 6 16
A1CF 0 0 0 1 1
A2M 9 11 9 37 18
A2M-AS1 1486 722 801 968 1787
Table 2: MiRNA expression matrix
An expression matrix containing normalized values, with row names according to miRBase.
PTC 1 PTC 2 PTC 3 PTC 4 PTC 5
hsa-let-7a-2-3p 3 0 9 1 4
hsa-let-7a-3p 472 82 228 122 313
hsa-let-7a-5p 141101 45543 105598 45503 159598
hsa-let-7b-3p 412 81 120 147 164
hsa-let-7b-5p 16337 6586 8121 7993 16516

2.4 Define sample metadata

Once we have expression matrices in the proper format, we need to inform MIRit about the samples in study and the biological conditions of interest. To do so, we need to create a data.frame that must contain:

  • a column named primary, specifying a unique identifier for each different subject;
  • a column named mirnaCol, matching the column name used for each sample in the miRNA expression matrix;
  • a column named geneCol, matching the column name used for each sample in the gene expression matrix;
  • a column that defines the biological condition of interest;
  • other optional columns that store specific sample metadata, such as age, sex and so on…

Firstly, let’s take a look at the column names used for miRNA and gene expression matrices.

## print sample names in geneCounts
colnames(geneCounts)
#>  [1] "PTC 1" "PTC 2" "PTC 3" "PTC 4" "PTC 5" "PTC 6" "PTC 7" "PTC 8" "NTH 1"
#> [10] "NTH 2" "NTH 3" "NTH 4" "NTH 5" "NTH 6" "NTH 7" "NTH 8"

## print sample names in mirnaCounts
colnames(mirnaCounts)
#>  [1] "PTC 1" "PTC 2" "PTC 3" "PTC 4" "PTC 5" "PTC 6" "PTC 7" "PTC 8" "NTH 1"
#> [10] "NTH 2" "NTH 3" "NTH 4" "NTH 5" "NTH 6" "NTH 7" "NTH 8"

## check if samples in geneCounts are equal to those in mirnaCounts
identical(colnames(geneCounts), colnames(mirnaCounts))
#> [1] TRUE

In our case, we see that both expression matrices have the same column names, and therefore mirnaCol and geneCol will contain the same values. However, note that is not always the case, especially for unpaired data, where miRNA and gene expression values derive from different subjects. In these cases, mirnaCol and geneCol must map each column of miRNA and gene expression matrices to the relative subjects indicated in the primary column. Notably, for unpaired data, NAs can be used for missing entries in mirnaCol/geneCol.

That said, we can proceed to create the data.frame with sample metadata as follows.

## create a data.frame containing sample metadata
meta <- data.frame(
    primary = colnames(mirnaCounts),
    mirnaCol = colnames(mirnaCounts),
    geneCol = colnames(geneCounts),
    disease = c(rep("PTC", 8), rep("NTH", 8)),
    patient = c(rep(paste("Sample_", seq(8), sep = ""), 2))
)

2.5 Create a MirnaExperiment object

At this point, after setting up expression matrices, and after defining sample metadata, we need to create an object of class MirnaExperiment, which is the main class used in MIRit to store all the data that are necessary for the integrative miRNA-mRNA analysis. In particular, this class extends the MultiAssayExperiment class from the homonym package (Ramos et al. 2017) to store expression levels of both miRNAs and genes, differential expression results, miRNA-target pairs and integrative miRNA-gene analysis.

The easiest way to create a valid MirnaExperiment object is to use the MirnaExperiment() function, which automatically handles the formatting of input data and the creation of the object.

## create the MirnaExperiment object
experiment <- MirnaExperiment(
    mirnaExpr = mirnaCounts,
    geneExpr = geneCounts,
    samplesMetadata = meta,
    pairedSamples = TRUE
)

3 Differential expression analysis

Now that the MirnaExperiment object has been created, we can move to differential expression analysis, which aims to define differentially expressed features across biological conditions.

3.1 Visualize expression variability

Firstly, before doing anything else, it is useful to explore miRNA and gene expression variability through dimensionality reduction techniques. This is useful because it allows us to visualize the main drivers of expression variability. In this regard, MIRit offers the plotDimensions() function, which enables to visualize both miRNA and gene expression in the multidimensional space (MDS plots). Moreover, it is possible to color samples based on specific variables, hence allowing to explore miRNA/gene expression variation between distinct biological groups.

In our example, let’s examine expression variability for both miRNAs and genes, and let’s color the samples based on “disease”, a variable included in the previously defined metadata.

geneMDS <- plotDimensions(experiment,
    assay = "genes",
    condition = "disease",
    title = "MDS plot for genes"
)
mirnaMDS <- plotDimensions(experiment,
    assay = "microRNA",
    condition = "disease",
    title = "MDS plot for miRNAs"
)

ggpubr::ggarrange(geneMDS, mirnaMDS,
    nrow = 1, labels = "AUTO",
    common.legend = TRUE
)
MDS plots for miRNAs and genes. Both plots show that the main source of variability is given by the disease condition.

Figure 1: MDS plots for miRNAs and genes
Both plots show that the main source of variability is given by the disease condition.

As we can see from Figures 1A and 1B, the samples are very well separated on the basis of disease condition, thus suggesting that this is the major factor that influences expression variability. This is exactly what we want, since we aim to evaluate the differences between cancer and normal tissue. If this wasn’t the case, we should identify the confounding variables, and include them in the model used for differential expression analysis (see Section 3.2.2).

3.2 Perform miRNA and gene differential expression

3.2.1 Available methods for RNA-Seq and microarrays

Now, we are ready to perform differential expression analysis. In this concern, MIRit provides different options based on the technology used for generating expression data. Indeed, when expression measurements derive from microarrays, MIRit calculates differentially expressed features through the pipeline implemented in the limma package. On the other hand, when expression values derive from RNA-Seq experiments, MIRit allows to choose between different approaches, including:

  • the approach defined in the edgeR package;
  • the approach defined in the DESeq2 package;
  • the limma-voom approach defined in the limma package.

Moreover, MIRit gives the possibility of fully customizing the parameters used for differential expression analysis, thus allowing a finer control that makes it easy to adopt strategies that differ from the standard pipelines proposed in these packages. For additional information, see Section 3.2.4.

3.2.2 Model design

After identifying the variable of interest and the confounding factors, we must indicate the experimental model used for fitting expression values. Notably, MIRit will automatically take care of model fitting, so that we only need to indicate a formula with the appropriate variables.

In our case, we want to evaluate the differences between cancer and normal thyroid. Therefore, disease condition is our variable of interest. However, in this experimental design, each individual has been assayed twice, one time for cancer tissue, and one time for healthy contralateral thyroid. Thus, we also need to include the patient ID as a covariate in order to prevent the individual differences between subjects from confounding the differences due to the disease.

## design the linear model for both genes and miRNAs
model <- ~ disease + patient

If other variables affecting miRNA and gene expression are observed, they should be included in this formula.

3.2.3 The performMirnaDE() and performGeneDE() functions

Once the linear model has been defined, we can perform the differential expression analysis through the performMirnaDE() and performGeneDE() functions. Indeed, these two functions take as input a MirnaExperiment object, and compute differential expression for miRNAs and genes.

Additionally, when we run these functions, we must define different arguments, namely:

  • group, which corresponds to the name of the variable of interest as specified in Section 2.4. In our case, we are interested in studying the differences between cancer tissue and normal tissue, and therefore our variable of interest is “disease”.
  • contrast, which indicates the levels of the variable of interest to be compared. In particular, this parameter takes as input a string where the levels are separated by a dash, and where the second level corresponds to the reference group. In our example, we want to compare samples from papillary thyroid cancer (PTC) against normal thyroid tissue (NTH), thus we set contrast to “PTC-NTH”.
  • design, which specifies the linear model with the variable of interest and eventual covariates. To do so, we pass to this parameter the R formula that we defined in Section 3.2.2.
  • method, which tells MIRit which pipeline we want to use for computing differentially expressed features. As stated in Section 3.2.1, for microarray studies the only option available is limma, while for RNA-Seq experiments, the user can choose between edgeR, DESeq2, and voom (for limma-voom). In our case we are going to perform differential expression analysis through the pipeline implemented in the edgeR package.

Following our example, let’s calculate differentially expressed genes and differentially expressed miRNAs in thyroid cancer.

## perform differential expression for genes
experiment <- performGeneDE(experiment, group = "disease", contrast = "PTC-NTH", design = model)
#> Performing differential expression analysis with edgeR...
#> Differential expression analysis reported 267 significant genes with p < 0.05 (correction: fdr). You can use the 'geneDE()' function to access results.

## perform differential expression for miRNAs
experiment <- performMirnaDE(experiment, group = "disease", contrast = "PTC-NTH", design = model)
#> Performing differential expression analysis with edgeR...
#> Differential expression analysis reported 40 significant miRNAs with p < 0.05 (correction: fdr). You can use the 'mirnaDE()' function to access results.

In the end, the analysis reported 265 differentially expressed genes, and 40 differentially expressed miRNAs. If not specified, the performMirnaDE() and performGeneDE() functions will define differentially expressed genes/miRNAs as those having an adjusted p-value lower than 0.05, and an absolute log2 fold change higher than 1 (FC > 2). However, this behavior can be changed by tweaking the pCutoff parameter, that specifies the statistical significance threshold; the logFC parameter, which indicates the minimum log2 fold change that features must display for being considered as differentially expressed; and the pAdjustment parameter, which specifies the approach used for multiple testing correction (default is fdr to use the Benjamini-Hochberg method).

3.2.4 Advanced parameters

In addition to the above mentioned settings, other parameters can be passed to the performMirnaDE() and performGeneDE() functions. Specifically, depending on the method adopted for differential expression analysis, the user can finely control the arguments passed to each function involved in the pipeline. In particular, these two functions include the following advanced parameters:

  • filterByExpr.args,
  • calcNormFactors.args,
  • estimateDisp.args,
  • glmQLFit.args,
  • glmQLFTest.args,
  • DESeq.args,
  • useVoomWithQualityWeights,
  • voom.args,
  • lmFit.args,
  • eBayes.args,
  • useArrayWeights,
  • useWsva,
  • arrayWeights.args,
  • wsva.args.
  • useDuplicateCorrelation
  • correlationBlockVariable
  • duplicateCorrelation.args

In this regard, when using limma-voom strategy, the useVoomWithQualityWeights parameter tells MIRit whether to use voomWithWualityWeights() instead of the standard voom() function. In the same way, for microarray studies, the useArrayWeights specifies whether to consider array quality weights during the limma pipeline. Similarly, useWsva can be set to TRUE to include a weighted surrogate variable analysis for batch effect correction. Moreover, useDuplicateCorrelation can be set to TRUE if you want to consider the effect of correlated samples through the duplicateCorrelation() function in limma. In this concern, the correlationBlockVariable specifies the blocking variable to use. All the other parameters ending with “.args”, accept a list object with additional parameters to be passed to the relative functions. In this way, the user has full control over the strategy used for differential expression analysis.

For a complete reference on the usage of these parameters, check the help page of these functions. Instead, for further instructions on how to use these tools, please refer to their original manuals, which represent exceptional resources for learning the basics of differential expression analysis:

  • limma User’s Guide,
  • edgeR User’s Guide,
  • Analysing RNA-Seq data with DESeq2.

3.2.5 Add differential expression results from other technologies

Even though MIRit implements all the most commonly used strategies for differential expression analyses, these methods may not be suitable for all kind of experiments. For instance, expression data deriving from technologies different from microarrays and RNA-Seq can’t be processed through performGeneDE() and performMirnaDE() functions. Therefore, MIRit grants the possibility to perform differential expression analysis with every method of choice, and then add the results to an existing MirnaExperiment object. This is particularly valuable for proteomic studies, where different normalization strategies are used in differential expression pipelines. In this way, MIRit fully supports the use of proteomic data for conducting miRNA integrative analyses.

To do so, we can make use of the addDifferentialExpression() function, which allows to manually add the results of the analysis. This function takes as input a MirnaExperiment object, and a table containing the differential expression results for all miRNAs/genes analyzed, not just for statistically significant species. If we want to manually set differential expression results for both miRNAs and genes, two different tables must be supplied. These tables must include:

  • One column containing miRNA/gene names (according to miRBase/HGNC nomenclature). Accepted column names are: ID, Symbol, Gene_Symbol, Mirna, mir, Gene, gene.symbol, Gene.symbol.
  • One column with log2 fold changes. Accepted column names are: logFC, log2FoldChange, FC, lFC.
  • One column with average expression. Accepted column names are: AveExpr, baseMean, logCPM.
  • One column with the p-values resulting from the differential expression analysis. Accepted column names are: P.Value, pvalue, PValue, Pvalue.
  • One column containing p-values adjusted for multiple testing. Accepted column names are: adj.P.Val, padj, FDR, fdr, adj, adj.p, adjp.

Further, we must specify the cutoff levels used to consider miRNAs/genes as significantly differentially expressed. This can be done through the mirna.logFC, mirna.pCutoff, mirna.pAdjustment, gene.logFC, gene.pCutoff, gene.pAdjustment parameters.

3.3 Visualize differentially expressed features

3.3.1 Access differential expression tables

Once differential expression analysis has been performed, we can use the mirnaDE() and geneDE() functions to access a table with differentially expressed features. Therefore, let’s access the results of differential expression analysis in thyroid cancer for both miRNAs and genes.

## access DE results for genes
deGenes <- geneDE(experiment)

## access DE results for miRNAs
deMirnas <- mirnaDE(experiment)

3.3.2 Create a volcano plot for miRNAs and genes

In addition to differential expression tables, we can also generate a graphical overview of differential expression through volcano plots, which are extremely useful for visualizing features changing across biological conditions. To produce volcano plots, MIRit offers the plotVolcano() function.

In our example, let’s create volcano plots for both miRNA and gene differential expression.

## create a volcano plot for genes
geneVolcano <- plotVolcano(experiment,
    assay = "genes",
    title = "Gene differential expression"
)

## create a volcano plot for miRNAs
mirnaVolcano <- plotVolcano(experiment,
    assay = "microRNA",
    title = "miRNA differential expression"
)

## plot graphs side by side
ggpubr::ggarrange(geneVolcano, mirnaVolcano,
    nrow = 1, labels = "AUTO",
    common.legend = TRUE
)
Volcano plots of gene and miRNA differential expression.

Figure 2: Volcano plots of gene and miRNA differential expression

3.3.3 Produce differential expression bar plots

Finally, if we are interested in specific genes/miRNAs, MIRit implements the plotDE() function that allows to represent expression changes of specific features as box plots, bar plots, or violin plots. In our example, we can use this function to visualize expression changes of different genes involved in the normal functioning of thyroid gland. Note that we use the linear = FALSE option to plot data in log scale (useful when multiple genes have very different expression levels).

## create a bar plot for all thyroid features
thyrBar <- plotDE(experiment,
    features = c("TG", "TPO", "DIO2", "PAX8"),
    graph = "barplot", linear = FALSE
)

## show the resulting plot
thyrBar
#> Warning: Computation failed in `stat_summary()`
#> Caused by error in `get()`:
#> ! object 'mean_sd' of mode 'function' was not found
Differential expression bar plots for different thyroid genes. Differential expression analysis demonstrated how TG, TPO, DIO2 and PAX8 result downregulated in thyroid cancer.

Figure 3: Differential expression bar plots for different thyroid genes
Differential expression analysis demonstrated how TG, TPO, DIO2 and PAX8 result downregulated in thyroid cancer.

4 Functional enrichment analysis

After finding differentially expressed genes, we usually end up having long lists of features whose expression changes between biological conditions. However, this is usually not very informative, and we seek to understand which biological functions result impaired in our experiments. In this regard, different methods exist for determining which cellular processes are dysregulated in our conditions.

4.1 Available approaches: ORA, GSEA and CAMERA

In particular, MIRit supports different strategies for functional enrichment of genes, including over-representation analysis (ORA), gene-set enrichment analysis (GSEA), and Correlation Adjusted MEan RAnk gene set test (CAMERA). In this way, the user can infer compromised biological functions according to the approach of choice.

Among these methods, the first one that has been developed is named over-representation analysis (Boyle et al. 2004), often abbreviated as ORA. This analysis aims to define whether genes annotated to specific gene sets (such as ontological terms or biological pathways) are differentially expressed more than would be expected by chance. To do this, a p-value is calculated by the hypergeometric distribution for each gene set as in Equation (1).

\[\begin{equation} p = 1 - \sum_{i = 0}^{k - 1}{\frac{\binom{M}{i}\binom{N - M} {n - 1}}{\binom{N}{n}}} \tag{1} \end{equation}\]

Here, \(N\) is the total number of genes tested, \(M\) is the number of genes that are annotated to a particular gene set, \(n\) is the number of differentially expressed genes, and \(k\) is the number of differentially expressed genes that are annotated to the gene set.

Additionally, another available approach is the gene set enrichment analysis (Subramanian et al. 2005), often refereed to with the acronym GSEA, which is suitable to find categories whose genes change in a small but coordinated way. The GSEA algorithm takes as input a list of genes ranked with a particular criterion, and then walks down the list to evaluate whether members of a given gene set are normally distributed or are mainly present at the top or at the bottom of the list. To check this out, the algorithm uses a running-sum that increases when finding a gene belonging to a given category, and decreases when a gene not contained in that specific set is found. The maximum distance from zero occurred in the running score is defined as the enrichment score (ES). To estimate the statistical significance of enrichment scores, a permutation test is performed by swapping gene labels annotated to a gene set.

Even though GSEA is arguably the most commonly used approach for functional enrichment, Wu and Smyth (2012) demonstrated that inter-gene correlations might affect the reliability of functional enrichment analyses. To overcome this issue, they developed Correlation Adjusted MEan RAnk gene set test (CAMERA), which is another competitive test used for functional enrichment analysis of genes. The main advantage of this method is that it adjusts the gene set test statistic according to inter-gene correlations.

4.2 Available databases and categories

As described above, functional enrichment analysis relies on gene sets, which consist in collections of genes that are annotated to specific functions or terms. Independently from the strategy used for the analysis, functional enrichment methods need access to these properly curated collections of genes. In the effort of providing access to a vast number of these resources, MIRit uses the geneset package to support multiple databases, including Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), MsigDB, WikiPathways, Reactome, Enrichr, Disease Ontology (DO), Network of Cancer Genes (NCG), DisGeNET, and COVID-19. However, the majority of these databases contain multiple categories. To see the complete list of available gene sets for each database refer to the documentation of the enrichGenes() function.

4.3 Supported species

The above-mentioned collections have their own lists of supported species. To check the available species for a given database, MIRit provides a practical helper function named supportedOrganisms(). For example, to retrieve the species supported by Reactome database, we can simply run the following piece of code.

## list available species for Reactome database
supportedOrganisms("Reactome")
#>  [1] "Bos taurus"               "Caenorhabditis elegans"  
#>  [3] "Danio rerio"              "Drosophila melanogaster" 
#>  [5] "Gallus gallus"            "Homo sapiens"            
#>  [7] "Mus musculus"             "Rattus norvegicus"       
#>  [9] "Saccharomyces cerevisiae" "Sus scrofa"              
#> [11] "Xenopus tropicalis"

4.4 Perform functional enrichment with the enrichGenes() function

The main function in MIRit for the functional enrichment analysis of genes is enrichGenes(), which requires as input:

  • the MirnaExperiment object that we get after running differential expression analysis;
  • method, which specifies the desired approach among ORA, GSEA, and CAMERA;
  • database and category, which define the gene set that you want to use for the enrichment (see Section 4.2 for a complete reference of available databases and categories);
  • organism, which indicates the specie under investigation (defaults to “Homo sapiens”);
  • pCutoff and pAdjustment, which specify the cutoff for statistical significance and the multiple testing correction, respectively.

In our example, we are going to perform the enrichment analysis by using ORA with GO database (biological processes).

## perform over-representation analysis with GO
ora <- enrichGenes(experiment, method = "ORA", database = "GO", organism = "Homo sapiens")
#> Since not specified, 'category' for GO database is set to bp (default).
#> Reading GO gene-sets from cache...
#> Performing the enrichment of upregulated genes...
#> Performing the enrichment of downregulated genes...
#> The enrichment of genes reported 12 significantly enriched terms for downregulated genes and 0 for upregulated genes.

MIRit performs ORA separately for upregulated and downregulated genes, as it has been shown that this is more powerful compared to enriching all DE genes (Hong et al. 2014). Therefore, when we use ORA, enrichGenes() returns a list containing two objects of class FunctionalEnrichment, one storing enrichment results for upregulated genes, and one for downregulated genes.

Before exploring the results of the analysis, we will also demonstrate the capabilities of MIRit by performing GSEA with the gene sets provided by the KEGG pathway database.

## perform gene-set enrichment analysis with KEGG
gse <- enrichGenes(experiment, method = "GSEA", database = "KEGG", organism = "Homo sapiens")
#> Since not specified, 'category' for KEGG database is set to pathway (default).
#> Reading KEGG gene-sets from cache...
#> Ranking genes based on signed.pval...
#> Performing gene-set enrichment analysis (GSEA)...
#> GSEA reported 2 significantly enriched terms.

In this case, the enrichGenes() function returns a single object of class FunctionalEnrichment, containing GSEA results.

4.5 Visualize enriched sets

4.5.1 Access results table

After running the enrichGenes() function, we get FunctionalEnrichment objects holding the results of enrichment analyses. To access the full table containing significantly affected biological functions, we can use the enrichmentResults() function. In our case, we can check the results of our GSEA analysis to investigate the human pathways that result affected in thyroid cancer. In Table 3 we can see the output of enrichmentResults(gse).

Table 3: GSEA results
A table listing the affected KEGG pathways in thyroid cancer.
pathway pval padj log2err ES NES size leadingEdge
Protein processing in endoplasmic reticulum 0 0.01 0.54 -0.55 -1.87 55 BCL2, HY….
Thyroid hormone synthesis 0 0.01 0.52 -0.67 -2.02 27 SLC26A4,….

4.5.2 Enrichment dot plots and bar plots

Further, in addition to exploring results table, MIRit offers several options for the visualization of enrichment analyses, including dot plots and bar plots. These plots are available for every FunctionalEnrichment object independently from the method used.

Following our example, we can visualize the results of the ORA that we performed in Section 4.4 through a simple dot plot.

## create a dot plot for ORA
enrichmentDotplot(ora$downregulated, title = "Depleted functions")
ORA results for downregulated genes. The enrichment of downregulated genes through the gene sets provided by GO database.

Figure 4: ORA results for downregulated genes
The enrichment of downregulated genes through the gene sets provided by GO database.

4.5.3 Other plots for GSEA

Additionally, MIRit provides specific visualization methods that are exclusive for GSEA, including ridge plots and GSEA plots. For instance, after running GSEA through enrichGenes(), we can produce a GSEA-style plot through the gseaPlot() function. In our case, we are going to use this plotting method for the “Thyroid hormone synthesis” pathway.

## create a GSEA plot
gseaPlot(gse, "Thyroid hormone synthesis", rankingMetric = TRUE)
GSEA-style plot for Thyroid hormone synthesis. This type of plot shows the running sum that GSEA uses to determinate the enrichment score for each pathway.

Figure 5: GSEA-style plot for Thyroid hormone synthesis
This type of plot shows the running sum that GSEA uses to determinate the enrichment score for each pathway.

5 Associate miRNAs with disease-SNPs

Interestingly, MIRit enables to explore the presence of disease-associated SNPs occurring at loci of differentially expressed miRNAs. In this concern, SNPs occurring within miRNAs may have important effects on the biological function of these transcripts, as they might alter its expression or the spectrum of miRNA targets. To verify the presence of disease-SNPs within miRNA loci, MIRit directly queries the NHGRI-EBI Catalog of published genome-wide association studies through the gwasrapidd package, and then retains only SNPs that affect DE-miRNA genes or their relative host genes (for intragenic miRNAs).

In our case, there are no SNPs associated with thyroid cancer that occur witin DE-miRNA loci. Therefore, we will explain how this function works with reference to the analysis on Alzheimer’s disease reported in the MIRit paper.

5.2 Identify miRNA genes overlapping with disease-SNPs

Now, we can use the findMirnaSNPs() function to identify the disease-related SNPs that affect miRNA loci.

## detect disease-SNPs occuring at DE-miRNAs loci
association <- findMirnaSNPs(experiment, "Alzheimer disease")

5.3 Build a track plot to display miRNA-SNP associations

Finally, if any disease-related SNPs is present within DE-miRNA loci, we can use the mirVariantPlot() function to graphically build a track plot displaying the polymorphism along with the relevant genomic context, including the corresponding miRNA locus.

## create a track plot to represent disease-SNPs at DE-miRNA loci
mirVariantPlot("rs2632516", association, showContext = TRUE)

6 Retrieve miRNA targets

Before performing integrative miRNA-mRNA analyses, we need to identify the targets of differentially expressed miRNAs, so that we can test whether they really affect the levels of their targets or not.

6.1 Databases with miRNA-mRNA interactions

Different resources have been developed over the years to predict and collect miRNA-target interactions, and we can categorize them in two main types:

  • Prediction databases, that contain information about computationally determined miRNA-target interactions;
  • Validated databases, which only contain interactions that have been proven through biomolecular experiments.

The choice of which type of resources to use for identifying miRNA targets drastically influences the outcome of the analysis. In this regard, some researchers tend to give the priority to validated interactions, even though they are usually fewer than predicted ones. On the other hand, predicted pairs are much more numerous, but they exhibit a high number of false positive hits.

6.2 The mirDIP approach

The downside of miRNA target prediction algorithms is also the scarce extend of overlap existing between different tools. To address this issue, several ensemble methods have been developed, trying to aggregate the predictions obtained by different algorithms. Initially, several researchers determined as significant miRNA-target pairs those predicted by more than one tool (intersection method). However, this method is not able to capture an important number of meaningful interactions. Alternatively, other strategies used to merge predictions from several algorithms (union method). Despite identifying more true relationships, the union method leads to a higher proportion of false discoveries. Therefore, other ensemble methods started using other statistics to rank miRNA-target predictions obtained by multiple algorithms. Among these newly developed ensemble methods, one of the best performing one is the microRNA Data Integration Portal (mirDIP) database, which aggregates miRNA target predictions from 24 different resources by using an integrated score inferred from different prediction metrics. In this way, mirDIP reports more accurate predictions compared to those of individual tools. For additional information on mirDIP database and its ranking metric refer to Tokar et al. (2018) and Hauschild et al. (2023).

6.3 Download predicted and validated interactions with getTargets()

Given the above, MIRit allows the prediction of miRNA-target interactions via the mirDIP database (version 5.2). In addition, in order to raise the number of true interactions, MIRit combines the miRNA-target pairs returned by mirDIP with the experimentally validated interactions contained in the miRTarBase database (version 9) (Huang et al. 2022). In practice, to identify miRNA targets MIRit implements the getTargets() function, which allows to download both type of interactions. Specifically, this function also includes a parameter called score that determines the degree of confidence required for the targets predicted by mirDIP. The value of this parameter must be one of Very High, High (default), Medium, and Low, which correspond to ranks among top 1%, top 5% (excluding top 1%), top 1/3 (excluding top 5%) and remaining predictions, respectively. Moreover, the includeValidated parameter tells MIRit whether to include experimentally validate interactions deriving from miRTarBase (default is TRUE). Please note that mirDIP database is only available for human miRNAs; thus, for species other than Homo sapiens, only validated interactions contained in miRTarBase are used.

In our example, we are going to retrieve both predicted and validated interactions by using default settings.

## retrieve miRNA target genes
experiment <- getTargets(experiment)

After running this function, we obtain a MirnaExperiment object containing miRNA-target interactions in its targets slot. The user can access a data.frame detailing these interactions through the mirnaTargets() function.

7 Investigate the effects of miRNA expression changes on target genes

Now that we have defined the targets of differentially expressed miRNAs, we can continue with the integrative analysis of miRNA and gene expression levels. This analysis is useful as it allows to only consider miNA-target pairs where an inverse relationship is observed. As already mentioned, MIRit can work with both paired and unpaired data by using different statistical approaches, including:

  • Correlation analysis, which is the recommended method when samples are paired;
  • Association tests, like Fisher’s exact test and Boschloo’s exact test;
  • Rotation gene-set tests, as implemented in the fry function from limma package.

For unpaired data, only association tests and rotation gene-set tests are available, whereas correlation analysis is the best performing strategy for paired data. The integrative analysis, either performed through correlation, association tests, or rotation gene-set tests, is implemented in the mirnaIntegration() function. When using the default option test = "auto", MIRit automatically performs the appropriate test for paired and unpaired samples. If only some samples of the data set have paired measurements, a correlation analysis will be carried out only for those subjects.

7.1 Correlation analysis for paired data

When both miRNA and gene expression measurements are available for the same samples, a correlation analysis is the recommended procedure. In statistics, correlation is a measure that expresses the extent to which two random variables are dependent. In our case, we want to assess whether a statistical relationship is present between the expression of a miRNA and the expression of its targets.

7.1.1 Statistical correlation coefficients

Several statistical coefficients can be used to weigh the degree of a correlation. Among them, the most commonly used are Pearson’s correlation coefficient \(r\), Spearman’s correlation coefficient \(\rho\), and Kendall’s Tau-b correlation coefficient \(\tau_b\). Pearson’s \(r\) is probably the most diffused for determining the association between miRNA and gene expression. However, it assumes that the relationships between miRNA and gene expression values is linear. This is typically not true for miRNAs, whose interactions with their targets are characterized by imperfect complementarity. Additionally, miRNAs can target multiple genes with different binding sites, and this implies that a simple linear relationship may not be sufficient to properly describe the complexity of these interactions. In contrast, Spearman’s and Kendall’s Tau-b correlation coefficients result more suitable for representing the interplay between miRNAs and target genes, because they are robust to non-linear relationships and outliers. However, Kendall’s correlation just relies on the number of concordant and discordant pairs, and is less sensitive then Spearman’s correlation; so, when many ties are present or when the sample size is small, it may have a lower detection power. This is the reason why Spearman’s correlation coefficient is the default coefficient used in the mirnaIntegration() function to measure the correlation between miRNA and gene expression. Moreover, since miRNAs mainly act as negative regulators, only negatively correlated miRNA-target pairs are considered, and statistical significance is estimated through a one-tailed t-test.

7.1.2 Perform a correlation analysis in MIRit

To sum up the steps that MIRit follows when evaluating the correlation between miRNAs and genes, what the mirnaIntegration() function does during a correlation analysis is to:

  1. consider the miRNA-target interactions retrieved with the getTargets() function;
  2. calculate the correlation coefficient for each miRNA-target pair based on their expression values;
  3. compute the statistical significance of all miRNA-target pairs;
  4. adjust p-values for multiple testing before reporting significant results.

In our thyroid cancer example, we want to find the miRNA-target pairs that exhibit a negative correlation with a Spearman’s coefficient lower than -0.5 and with an adjusted p-value lower than 0.05.

## perform a correlation analysis
experiment <- mirnaIntegration(experiment, test = "correlation")
#> As specified by the user, a correlation will be used.
#> Performing Spearman's correlation analysis...
#> A statistically significant correlation between 215 miRNA-target pairs was found!

Please note that all the parameters used for the correlation analysis are customizable. For instance, the user can change the significance threshold and the multiple testing correction method by setting the pCutoff and pAdjustment parameters, respectively. Further, it is also possible to change the correlation coefficient used, by editing the corMethod option, and the minimum required value of the correlation coefficient, by changing the corCutoff setting.

7.1.3 Account for batch effects prior to correlation analysis

Sometimes, when exploring expression variability through MDS plots, as we do with the plotDimensions() function, we notice the presence of batch effects that prevent a clear separation of our biological groups. Indeed, batch effects consist in unwanted sources of technical variation that confound expression variability and limit downstream analyses. Since the reliability of biological conclusions of integrative miRNA-mRNA analyses depends on the correlation between miRNA and gene expression levels, it is pivotal to ensure that expression measurements are not affected by technical variation. In this regard, if batch effects are noticed in the data, MIRit provides the batchCorrection() function, which removes batch effects from expression data before moving to correlation analysis. Please note that this procedure can only be used prior to correlation analysis, because for differential expression analysis it is more appropriate to include batch variables in the linear model, as specified in Section 3.2.2. For additional information, please refer to the manual of the batchCorrection() function.

7.1.4 Explore the succesfully integrated miRNA-target pairs

Before moving on to the identification of the altered miRNAs regulatory networks, we can explore correlated miRNA-target pairs thanks to the integration() function, which returns a data.frame object with comprehensive details about the computed correlations.

## extract correlation results
integrationResults <- integration(experiment)

## take a look at correlation results
head(integrationResults)
#>                     microRNA Target microRNA.Direction Corr.Coefficient
#> hsa.miR.1179.7  hsa-miR-1179  ITGA6      downregulated       -0.8617647
#> hsa.miR.1179.10 hsa-miR-1179  RBMS2      downregulated       -0.7764706
#> hsa.miR.1179.12 hsa-miR-1179 SPIRE1      downregulated       -0.7117647
#> hsa.miR.1179.14 hsa-miR-1179   TAF5      downregulated       -0.5970588
#> hsa.miR.1275.1  hsa-miR-1275  CCND2      downregulated       -0.7794118
#> hsa.miR.1275.2  hsa-miR-1275   CD44      downregulated       -0.8382353
#>                 Corr.P.Value Corr.Adjusted.P.Val
#> hsa.miR.1179.7  8.902592e-06        0.0006579303
#> hsa.miR.1179.10 2.021754e-04        0.0034840545
#> hsa.miR.1179.12 9.920216e-04        0.0072485216
#> hsa.miR.1179.14 7.304866e-03        0.0266876638
#> hsa.miR.1275.1  1.858401e-04        0.0034840545
#> hsa.miR.1275.2  2.505171e-05        0.0013423542

7.1.5 Visualize the correlation between miRNAs and genes

Additionally, for correlation analyses, MIRit allows to graphically represent inverse correlations through a scatter plot. Indeed, we can make use of the plotCorrelation() function to display the correlation between specific miRNA-target pairs. For example, the correlation analysis performed in Section 7.1.2 revealed how miR-146b-5p, the most upregulated miRNA, is inversely correlated with the expression of DIO2, which is crucial for thyroid hormone functioning. Furthermore, it has also emerged that miR-146b-3p results negatively correlated with PAX8, which directly induces TG transcription.

## plot the correlation between miR-146b-5p and DIO2
cor1 <- plotCorrelation(experiment,
    mirna = "hsa-miR-146b-5p",
    gene = "DIO2", condition = "disease"
)

## plot the correlation between miR-146b-3p and PAX8
cor2 <- plotCorrelation(experiment,
    mirna = "hsa-miR-146b-3p",
    gene = "PAX8", condition = "disease"
)

## plot graphs side by side
ggpubr::ggarrange(cor1, cor2,
    nrow = 1, labels = "AUTO",
    common.legend = TRUE
)
Correlation between miRNAs and key thyroid genes. These plots suggest that the upregulation of miR-146b-5p and miR-146b-3p may be responsible for the downregulation of DIO2 and PAX8, respectively.

Figure 6: Correlation between miRNAs and key thyroid genes
These plots suggest that the upregulation of miR-146b-5p and miR-146b-3p may be responsible for the downregulation of DIO2 and PAX8, respectively.

7.2 Association tests for unpaired data

For unpaired data, we cannot directly quantify the influence of miRNA expression on the levels of their targets, because we do not have any sample correspondence between miRNA and gene measurements. However, one-sided association tests can be applied in these cases to evaluate if targets of downregulated miRNAs are statistically enriched in upregulated genes, and, conversely, if targets of upregulated miRNAs are statistically enriched in downregulated genes. In this regard, to estimate the effects of differentially expressed miRNAs on their target genes, MIRit can use two different one-sided association tests, namely:

  • Fisher’s exact test,
  • Boschloo’s exact test (default).

Both these tests consist in a statistical procedure that estimates the association between two dichotomous categorical variables. In our case, for each miRNA, we want to evaluate whether the proportion of targets within the differentially expressed genes significantly differs from the proportion of targets in non-differentially expressed genes. To do this, a 2x2 contingency table is built as shown in Table 4.

Table 4: The 2x2 contingency table that MIRit uses for one-sided association tests
This is the table that the mirnaIntegration() function creates to determine if differentially expressed genes are enriched in miRNA targets.
Target genes Non target genes Row total
Differentially expressed \(a\) \(b\) \(a + b\)
Non differentially expressed \(c\) \(d\) \(c + d\)
Column total \(a + c\) \(b + d\) \(a + b + c + d = n\)

7.2.1 Fisher’s exact test

After that the contingency table has been defined, Fisher’s exact test p-value can be calculated through Equation (2).

\[\begin{equation} p = \frac{(a + b)!\ (c + d)!\ (a + c)!\ (b + d)!}{a!\ b!\ c!\ d!\ n!} \tag{2} \end{equation}\]

Additionally, it is also possible to compute Fisher’s p-values with Lancaster’s mid-p adjustment, since it has been proven that it increases statistical power while retaining Type I error rates.

7.2.2 Boschloo’s exact test

In contrast to Fisher’s exact test, a more appropriate method for the integrative analysis between miRNAs and genes is Boschloo’s exact test. Indeed, the major drawback of the Fisher’s exact test is that it consists in a conditional test that requires the sum of both rows and columns of a contingency table to be fixed. Notably, this is not true for genomic data because it is likely that different data sets may lead to a different number of DEGs. Therefore, the default behavior in MIRit is to use a variant of Barnard’s exact test, named Boschloo’s exact test, that is suitable when group sizes of contingency tables are variable. Moreover, it is possible to demonstrate that Boschloo’s test is uniformly more powerful compared to Fisher’s exact test. However, keep in mind that Boschloo’s test is much more computational intensive compared to Fisher’s exact test, and it may require some time.

7.2.3 Perform one-sideded association tests in MIRit

In MIRit, the mirnaIntegration() function automatically performs association tests for unpaired data when test = "auto". Moreover, the type of association test to use can be specified through the associationMethod parameter, which can be set to:

  • fisher, to perform a simple one-sided Fisher’s exact test;
  • fisher-midp, to perform a one-sided Fisher’s exact test with Lancaster’s mid-p correction;
  • boschloo, to perform a one-sided Boschloo’s exact test (default option).

For example, we could use the Boschloo’s exact to evaluate the inverse association between miRNA and gene expression values through a simple call to mirnaIntegration() function.

## perform a one-sided inverse association
exp.association <- mirnaIntegration(experiment,
    test = "association",
    associationMethod = "boschloo",
    pAdjustment = "none"
)

Finally, after performing the association analysis, results can be accessed through the integration() function in the same way as we can do for correlation analyses.

7.3 Rotation gene-set tests for unpaired data

Lastly, for unpaired data, the effect of DE-miRNAs on the expression of target genes can be estimated through rotation gene-set tests. In this approach, we want to evaluate for each miRNA whether its target genes tend to be differentially expressed in the opposite direction. In particular, a fast approximation to rotation gene-set testing called fry, implemented in the limma package, can be used to statistically quantify the influence of miRNAs on expression changes of their target genes.

To perform the integrative analysis through rotation gene-set tests, we must simply set test = "fry" when calling mirnaIntegration() function.

## perform the integrative analysis through 'fry' method
exp.fry <- mirnaIntegration(experiment,
    test = "fry",
    pCutoff = 0.1
)

7.4 Functional enrichment of integrated target genes

Additionally, after finding miRNA-target pairs that appear to have an inverse association, we can try to identify the impaired biological functions as a result of miRNA dysregulations through ORA. To do this, MIRit provides the enrichTargets() function, which automatically performs ORA for the functional enrichment of target genes that result associated with differentially expressed miRNAs.

In our example, we are going to enrich the significantly anti-correlated targets that we have found in Section 7.1.2 through the Disease Ontology database.

## enrichment of integrated targets
oraTarg <- enrichTargets(experiment, database = "DO")
#> Reading DO gene-sets from cache...
#> Performing the enrichment of upregulated genes...
#> Performing the enrichment of downregulated genes...
#> The enrichment of genes reported 113 significantly enriched terms for downregulated genes and 0 for upregulated genes.

## show a dot plot for the enrichment of downregulated genes
enrichmentDotplot(oraTarg$downregulated,
    showTerms = 7,
    showTermsParam = "padj", title = "Depleted diseases"
)
Functional enrichment of integrated targets. This dot plot shows the enriched diseases for downregulated genes.

Figure 7: Functional enrichment of integrated targets
This dot plot shows the enriched diseases for downregulated genes.

In Figure 7, we appreciate the depletion of diseases where thyroid gland is overly active, such as goiter and hyperthyroidism, therefore suggesting the involvement of miRNAs in thyroid malfunctioning.

8 Identify the impaired miRNA-mRNA regulatory networks

Once the dysregulated miRNA-mRNA regulatory networks have been identified, the typical goal is to infer altered cellular processes and functions. To do so, MIRit introduces a novel approach named Topology-Aware Integrative Pathway Analysis (TAIPA), which specifically focuses on detecting altered molecular networks in miRNA-mRNA multi-omic analyses by considering the topology of biological pathways and miRNA-mRNA interactions.

8.1 Topologically-Aware Integrative Pathway Analysis (TAIPA)

This analysis aims to identify the biological pathways that result affected by miRNA and mRNA dysregulations. In this analysis, biological pathways are retrieved from a pathway database such as KEGG, and the interplay between miRNAs and genes is then added to the networks. Each network is defined as a graph \(G(V, E)\), where \(V\) represents nodes, and \(E\) represents the relationships between nodes. Then, nodes that are not significantly differentially expressed are assigned a weight \(w_i = 1\), whereas differentially expressed nodes are assigned a weight \(w_i = \left| \Delta E_i \right|\), where \(\Delta E_i\) is the linear fold change of the node. Moreover, to consider the biological interaction between two nodes, namely \(i\) and \(j\), we define an interaction parameter \(\beta_{i \rightarrow j} = 1\) for activation interactions and \(\beta_{i \rightarrow j} = -1\) for repression interactions. Subsequently, the concordance coefficient \(\gamma_{i \rightarrow j}\) is defined as in Equation (3):

\[\begin{equation} \gamma_{i \rightarrow j} = \begin{cases} \beta_{i \rightarrow j} &\text{if } sign(\Delta E_i) = sign(\Delta E_j) \\ - \beta_{i \rightarrow j} &\text{if } sign(\Delta E_i) \not= sign(\Delta E_j) \end{cases}\,. \tag{3} \end{equation}\]

Later in the process, a breadth-first search (BFS) algorithm is applied to topologically sort pathway nodes so that each individual node occurs after all its upstream nodes. Nodes within cycles are considered leaf nodes. At this point, a node score \(\phi\) is calculated for each pathway node \(i\) as in Equation (4):

\[\begin{equation} \phi_i = w_i + \sum_{j=1}^{U} \gamma_{i \rightarrow j} \cdot k_j\,, \tag{4} \end{equation}\]

where \(U\) represents the number of upstream nodes, \(\gamma_{i \rightarrow j}\) denotes the concordance coefficient, and \(k_j\) is a propagation factor defined as in Equation (5):

\[\begin{equation} k_j = \begin{cases} w_j &\text{if } \phi_j = 0 \\ \phi_j &\text{if } \phi_j \not = 0 \end{cases}\,. \tag{5} \end{equation}\]

Finally, the pathway score \(\Psi\) is calculated as in Equation (6):

\[\begin{equation} \Psi = \frac{1 - M}{N} \cdot \sum_{i=1}^{N} \phi_i\,, \tag{6} \end{equation}\]

where \(M\) represents the proportion of miRNAs in the pathway, and \(N\) represents the total number of nodes in the network. Then, to compute the statistical significance of each pathway score, a permutation procedure is applied. Later, both observed pathway scores and permuted scores are standardized by subtracting the mean score of the permuted sets \(\mu_{\Psi_P}\) and then dividing by the standard deviation of the permuted scores \(\sigma_{\Psi_P}\).

Finally, the p-value is defined based on the fraction of permutations that reported a higher normalized pathway score than the observed one. However, to prevent p-values equal to zero, we define p-values as in Equation (7):

\[\begin{equation} p = \frac{\sum_{n=1}^{N_p} \left[ \Psi_{P_N} \ge \Psi_N \right] + 1} {N_p + 1}\,. \tag{7} \end{equation}\]

In the end, either p-values are corrected for multiple testing through the max-T procedure (default option) which is particularly suited for permutation tests, or through the standard multiple testing approaches.

8.2 Perform TAIPA in MIRit

Before performing TAIPA, we need to create miRNA-augmented networks. To do so, MIRit implements the preparePathways() function, which automatically uses the graphite R package to download biological networks from multiple pathway databases, namely KEGG, WikiPathways and Reactome. Then, each pathway is converted to a graph object and significant miRNA-mRNA pairs are added to the network. Further, edge weights are included according to interaction type. After running this function, we obtain a list containing all the miRNA-augmented nwtworks as graph objects.

In our example, we want to use the significant miRNA-target pairs that we identified in Section 7.1.2 to augment biological pathways retrieved from the KEGG database.

## create miRNA-augmented networks using KEGG pathways
networks <- preparePathways(experiment,
    database = "KEGG",
    organism = "Homo sapiens"
)

After running this function, pathways with less than 10% of nodes with expression measurements are removed. This option can be changed by specifying the minPc parameter.

Now, we are ready to perform TAIPA through the topologicalAnalysis() function, which is used to calculate pathway scores for all the augmented networks and to evaluate their statistical significance through permutation tests.

## set seed for reproducible results
set.seed(1234)

## perform TAIPA
taipa <- topologicalAnalysis(experiment,
    pathways = networks,
    nPerm = 10000
)

As a result of the analysis, an object of class IntegrativePathwayAnalysis storing the results of TAIPA is returned. Notably, the user can change the behavior of the topologicalAnalysis() in several ways. For example, the pCutoff and pAdjustment parameters can be used to change the significance threshold and the multiple testing correction method, respectively. Moreover, the nPerm parameter can be tweaked to change the number of permutations to use for evaluating statistical significance. In this regard, we recommend using at least 10000 permutations, with no less than 1000.

8.3 C++ code and parallel computing with BiocParallel

For computational efficiency, pathway score computation has been implemented in C++ language. Furthermore, since computing pathway score for 10000 networks for each pathway is computationally intensive, parallel computing has been employed to reduce running time. The user can modify the parallel computing behavior by specifying the BPPARAM parameter. See the help page of the BiocParallel package for further details. Both the preparePathways() and the topologicalAnalysis() functions accept the BPPARAM option.

8.4 Visualize the significantly affected pathways

After running the topologicalAnalysis() function, we can inspect the significantly perturbed pathways contained in the IntegrativePathwayAnalysis object by using the integratedPathways() function, which returns a data.frame reporting the results of TAIPA.

## extract the results of TAIPA
perturbedNetworks <- integratedPathways(taipa)

8.5 Visualize the impaired networks within biological pathways

As with functional enrichment analyses, we can plot perturbed miRNA-mRNA networks as dot plots. To do so, the integrationDotplot() function can be used. In particular, we will graphically represent the most perturbed pathways in thyroid cancer.

## produce a dotplot that shows the most affected networks
intDot <- integrationDotplot(taipa)
intDot
The perturbation of miRNA-mRNA networks in thyroid cancer. This dot plot display the impairment of the biological processes involved in the production of thyroid hormone, further highlighting the disruption of this mechanism in this disease.

Figure 8: The perturbation of miRNA-mRNA networks in thyroid cancer
This dot plot display the impairment of the biological processes involved in the production of thyroid hormone, further highlighting the disruption of this mechanism in this disease.

Finally, after identifying the impaired molecular networks, MIRit provides the possibility of exploring the molecular perturbations. In this concern, the visualizeNetworks() function can be used to visually represent the compromised pathways along with expression changes of both miRNAs and genes, so that users can easily interpret the functional consequences of miRNA and gene dysregulations. For example, we can explore the perturbed molecular events that are responsible for diminished production of thyroid hormone in thyroid cancer.

## plot the impaired network responsible for reduced TG synthesis
visualizeNetwork(taipa, "Thyroid hormone synthesis")
Impaired miRNA-mRNA regulatory network prevents thyroid hormone synthesis in thyroid cancer. The network created by MIRit suggests that the upregulation of miR-146b-5p and miR-146b-3p may be responsible for diminished expression of PAX8, which in turn causes reduced transcription of thyroid hormone.

Figure 9: Impaired miRNA-mRNA regulatory network prevents thyroid hormone synthesis in thyroid cancer
The network created by MIRit suggests that the upregulation of miR-146b-5p and miR-146b-3p may be responsible for diminished expression of PAX8, which in turn causes reduced transcription of thyroid hormone.

Session info

#> R Under development (unstable) (2024-01-16 r85808)
#> Platform: x86_64-pc-linux-gnu
#> Running under: Ubuntu 22.04.3 LTS
#> 
#> Matrix products: default
#> BLAS:   /home/biocbuild/bbs-3.19-bioc/R/lib/libRblas.so 
#> LAPACK: /usr/lib/x86_64-linux-gnu/lapack/liblapack.so.3.10.0
#> 
#> locale:
#>  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
#>  [3] LC_TIME=en_GB              LC_COLLATE=C              
#>  [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
#>  [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                 
#>  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
#> [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       
#> 
#> time zone: America/New_York
#> tzcode source: system (glibc)
#> 
#> attached base packages:
#> [1] stats     graphics  grDevices utils     datasets  methods   base     
#> 
#> other attached packages:
#> [1] MIRit_0.99.10    BiocStyle_2.31.0
#> 
#> loaded via a namespace (and not attached):
#>   [1] DBI_1.2.1                   bitops_1.0-7               
#>   [3] gridExtra_2.3               rlang_1.1.3                
#>   [5] magrittr_2.0.3              MonoPoly_0.3-10            
#>   [7] matrixStats_1.2.0           compiler_4.4.0             
#>   [9] RSQLite_2.3.5               vctrs_0.6.5                
#>  [11] quadprog_1.5-8              pkgconfig_2.0.3            
#>  [13] crayon_1.5.2                fastmap_1.1.1              
#>  [15] backports_1.4.1             dbplyr_2.4.0               
#>  [17] XVector_0.43.1              labeling_0.4.3             
#>  [19] utf8_1.2.4                  rmarkdown_2.25             
#>  [21] graph_1.81.0                purrr_1.0.2                
#>  [23] bit_4.0.5                   xfun_0.41                  
#>  [25] MultiAssayExperiment_1.29.0 zlibbioc_1.49.0            
#>  [27] cachem_1.0.8                GenomeInfoDb_1.39.5        
#>  [29] jsonlite_1.8.8              blob_1.2.4                 
#>  [31] highr_0.10                  DelayedArray_0.29.1        
#>  [33] BiocParallel_1.37.0         broom_1.0.5                
#>  [35] parallel_4.4.0              R6_2.5.1                   
#>  [37] bslib_0.6.1                 limma_3.59.1               
#>  [39] car_3.1-2                   GenomicRanges_1.55.2       
#>  [41] jquerylib_0.1.4             Rcpp_1.0.12                
#>  [43] bookdown_0.37               SummarizedExperiment_1.33.3
#>  [45] knitr_1.45                  IRanges_2.37.1             
#>  [47] BiocBaseUtils_1.5.0         Matrix_1.6-5               
#>  [49] splines_4.4.0               tidyselect_1.2.0           
#>  [51] abind_1.4-5                 yaml_2.3.8                 
#>  [53] codetools_0.2-19            curl_5.2.0                 
#>  [55] geneset_0.2.7               lattice_0.22-5             
#>  [57] tibble_3.2.1                Biobase_2.63.0             
#>  [59] withr_3.0.0                 evaluate_0.23              
#>  [61] BiocFileCache_2.11.1        pillar_1.9.0               
#>  [63] BiocManager_1.30.22         ggpubr_0.6.0               
#>  [65] filelock_1.0.3              MatrixGenerics_1.15.0      
#>  [67] carData_3.0-5               stats4_4.4.0               
#>  [69] generics_0.1.3              RCurl_1.98-1.14            
#>  [71] S4Vectors_0.41.3            ggplot2_3.4.4              
#>  [73] munsell_0.5.0               scales_1.3.0               
#>  [75] glue_1.7.0                  tools_4.4.0                
#>  [77] data.table_1.15.0           fgsea_1.29.0               
#>  [79] locfit_1.5-9.8              ggsignif_0.6.4             
#>  [81] fastmatch_1.1-4             cowplot_1.1.3              
#>  [83] grid_4.4.0                  tidyr_1.3.1                
#>  [85] edgeR_4.1.15                colorspace_2.1-0           
#>  [87] GenomeInfoDbData_1.2.11     cli_3.6.2                  
#>  [89] fansi_1.0.6                 S4Arrays_1.3.3             
#>  [91] dplyr_1.1.4                 Rgraphviz_2.47.0           
#>  [93] gtable_0.3.4                rstatix_0.7.2              
#>  [95] sass_0.4.8                  digest_0.6.34              
#>  [97] BiocGenerics_0.49.1         SparseArray_1.3.3          
#>  [99] farver_2.1.1                memoise_2.0.1              
#> [101] htmltools_0.5.7             lifecycle_1.0.4            
#> [103] httr_1.4.7                  statmod_1.5.0              
#> [105] bit64_4.0.5

References

Boyle, Elizabeth I., Shuai Weng, Jeremy Gollub, Heng Jin, David Botstein, J. Michael Cherry, and Gavin Sherlock. 2004. “GO::TermFinder—Open Source Software for Accessing Gene Ontology Information and Finding Significantly Enriched Gene Ontology Terms Associated with a List of Genes.” Bioinformatics 20 (18): 3710–5. https://doi.org/10.1093/bioinformatics/bth456.

Carvalho, Benilton S., and Rafael A. Irizarry. 2010. “A Framework for Oligonucleotide Microarray Preprocessing.” Bioinformatics 26 (19): 2363–7. https://doi.org/10.1093/bioinformatics/btq431.

Hauschild, Anne-Christin, Chiara Pastrello, Gitta Kirana Anindya Ekaputeri, Dylan Bethune-Waddell, Mark Abovsky, Zuhaib Ahmed, Max Kotlyar, Richard Lu, and Igor Jurisica. 2023. “MirDIP 5.2: Tissue Context Annotation and Novel microRNA Curation.” Nucleic Acids Research 51 (D1): D217–D225. https://doi.org/10.1093/nar/gkac1070.

Hong, Guini, Wenjing Zhang, Hongdong Li, Xiaopei Shen, and Zheng Guo. 2014. “Separate Enrichment Analysis of Pathways for up- and Downregulated Genes.” Journal of the Royal Society Interface 11 (92): 20130950. https://doi.org/10.1098/rsif.2013.0950.

Huang, Hsi-Yuan, Yang-Chi-Dung Lin, Shidong Cui, Yixian Huang, Yun Tang, Jiatong Xu, Jiayang Bao, et al. 2022. “miRTarBase Update 2022: An Informative Resource for Experimentally Validated miRNA–Target Interactions.” Nucleic Acids Research 50 (D1): D222–D230. https://doi.org/10.1093/nar/gkab1079.

Ramos, Marcel, Lucas Schiffer, Angela Re, Rimsha Azhar, Azfar Basunia, Carmen Rodriguez, Tiffany Chan, et al. 2017. “Software for the Integration of Multiomics Experiments in Bioconductor.” Cancer Research 77 (21): e39–e42. https://doi.org/10.1158/0008-5472.CAN-17-0344.

Riesco-Eizaguirre, Garcilaso, León Wert-Lamas, Javier Perales-Patón, Ana Sastre-Perona, Lara P. Fernández, and Pilar Santisteban. 2015. “The miR-146b-3p/PAX8/NIS Regulatory Circuit Modulates the Differentiation Phenotype and Function of Thyroid Cells During Carcinogenesis.” Cancer Research 75 (19): 4119–30. https://doi.org/10.1158/0008-5472.CAN-14-3547.

Subramanian, Aravind, Pablo Tamayo, Vamsi K. Mootha, Sayan Mukherjee, Benjamin L. Ebert, Michael A. Gillette, Amanda Paulovich, et al. 2005. “Gene Set Enrichment Analysis: A Knowledge-Based Approach for Interpreting Genome-Wide Expression Profiles.” Proceedings of the National Academy of Sciences 102 (43): 15545–50. https://doi.org/10.1073/pnas.0506580102.

Tokar, Tomas, Chiara Pastrello, Andrea E M Rossos, Mark Abovsky, Anne-Christin Hauschild, Mike Tsay, Richard Lu, and Igor Jurisica. 2018. “mirDIP 4.1—Integrative Database of Human microRNA Target Predictions.” Nucleic Acids Research 46 (D1): D360–D370. https://doi.org/10.1093/nar/gkx1144.

Wu, Di, and Gordon K. Smyth. 2012. “Camera: A Competitive Gene Set Test Accounting for Inter-Gene Correlation.” Nucleic Acids Research 40 (17): e133. https://doi.org/10.1093/nar/gks461.