miR2Trait: an integrated resource for investigating miRNA-disease associations

miRNA: disease mapping

Two expert curated databases were used as the source databases in the creation of miR2Trait: (1) miRTarbase, a database of experimentally validated miRNA–target interactions (MTIs), with >13,400 entries (Chou et al., 2018); and (2) DisGeNET, a widely used and standardized knowledge management platform with >1,000,000 gene–disease associations (Piñero et al., 2015). Both databases involved the mapping of genetic information, in one case to miRNAs, and in the other, to diseases. This provided a clue to relating miRNAs and disease via the bridge of genetic information (Fig. 2). In-house Python scripts were used to extract the gene:miRNA and the gene:disease mappings from the respective databases as dictionary data structures. The two dictionaries were merged on ‘gene’ to establish a dictionary of miRNA:disease mappings. Similarly, an inverse dictionary of disease:miRNA mappings was created.

Database creation

The dictionaries obtained above were used to populate relational tables in the construction of two databases: (i) miRNAs and their associated diseases, and (ii) diseases and their associated miRNAs. MySQL was used for encoding the databases.

Spectrum width calculation

The number of diseases with which an miRNA is associated is indicative of the breadth of its regulatory impact, which could flag master regulator miRNAs (or regulatory hubs). The disease spectrum width (DSW) of the ith miRNA is calculated after Qiu, Chen & Cui (2012): (1)${\displaystyle \text{DSW}\left(i\right)=d_{\mathrm{i}}/D_{\mathrm{N}}}$ where d_i isthe number of diseases associated with miRNA i and D_N is the total number of diseases in the database. Conversely the number of miRNAs associated with a given disease could be indicative of a complex multifactorial pathology, and the miRNA spectrum width (MSW) of the jth disease is given by: (2)${\displaystyle \text{MSW}\left(j\right)=m_{\mathrm{j}}/m_{\mathrm{N}}}$ where m_j is the number of miRNAs associated with the jth disease and M_N isthe total number of miRNAs in the database.

miRNA list enrichment analysis

Given a list of miRNAs that are collectively dys-regulated, it would be of interest to identify which diseases would be enriched (or more likely to occur). Statistical techniques to quantify such enrichment for genes as well as miRNAs exist (Li et al., 2018; Backes et al., 2016; Çorapçıoğlu & Oğul, 2015; Rivals et al., 2007). Here we implement an open-source tool to identify diseases in the miR2Trait database that are enriched for an input set of miRNAs. The method uses a hypergeometric test to quantify the enrichment of each disease in miR2Trait for the given set of miRNAs (in this case identical to the Fisher’s exact test) (Fisher, 1922). The test follows the construction of the Fisher contingency table for each disease, as illustrated in Table 1.

The statistical significance (p-value) of enrichment is computed using Python scipy routines (http://www.scipy.org/), and diseases that pass a user-specified adjusted p-value threshold (with an option to choose among three methods for multiple hypothesis correction) are returned sorted by significance. In addition to significance, both the web server and standalone tool calculate the effect size (log-fold) of each enrichment.

Diseasome analysis

Given an input set of diseases related in some way (called a diseasome), it would be of interest to find shared dysregulated miRNAs. We approached this problem with an abundance analysis of miRNAs in the diseasome, and identifying miRNAs that passed a user-specified count. Such a set of miRNAs could constitute master regulators acting on common pathways in the specified set of diseases. If we are interested in the miRNA spectra of umbrella disease terms such as ’carcinoma’, instead of a specific set of diseases, this is also allowed. In this case, diseases containing the general keyword are first identified, and then the miRNA abundance for this diseasome is computed in a manner similar to above. A CLI script to compute the miRNA profile of a specified diseasome is provided with miR2Trait (https://github.com/miR2Trait).

miR2Trait web server and standalone tools

A web interface using PHP to connect the user HTML front-end with the MySQL backend was created to offer access to all the services developed, including the two databases and the assorted tools. MiRNA list enrichment analysis (miLEA) was implemented using the python DBI module, and receives the user’s miRNA list from the HTML form via PHP intermediate. The command-line version of the enrichment analysis provides adjusted p-value calculation using Bonferroni, Holm-Bonferroni or Benjamini–Hochberg (default) correction. Other CLI scripts developed include:

(i) Query standalone miR2Trait by miRNA or disease

(ii) Retrieve miRNA-specific information from miRBase by user-given miRNA name(s) or sequence(s).

(iii) Network creation: creating an miRNA-miRNA/disease-disease adjacency network from miR2Trait dictionaries

(iv) Postprocessing constructed networks for consensus central nodes (miRNAs or diseases)

The user is referred to the wiki (https://github.com/miR2Trait/miR2Trait/wiki) for reproducible examples and use-cases of all the scripts developed.

Benchmarking

Dedicated miRNA omics databases have been developed to aid researchers unravel the role of miRNAs in biological processes (for e.g., see Ruepp et al., 2010; Keller et al., 2011). Computational methods have also been advanced for inferring miRNA-disease connections (Chen, Liu & Yan, 2012; Xuan et al., 2013; Qu et al., 2018). HMDDv3.2 is a manually curated database of miRNA-disease associations based on text-mining. While miR2Trait contains the full mature miRNA information, HMDD does not provide subtype information such as -5p or -3p; 1a or 1b. The miRNA arm designation (5p or 3p) is a major ingredient of disease etiology, and key to naming miRNAs (for e.g., the ones with the top DSWs in Table 3). To compare the databases, it is therefore necessary to ’normalize’ the miRNA nomenclature between the databases (by suppressing the subtype information in miR2Trait). This contracted the unique miRNAs in miR2Trait to 1,759 in number. Following this, the databases were compared, and the results are shown in Fig. 4. Even after contraction, miR2Trait appears more comprehensive in the catalog of miRNA-disease associations. miR2Trait-unique miRNAs are 2.5-fold more in number than HMDDv3.2 - unique miRNAs (884 vs 334). The unique miRNA complements suggest directions for research. Some of the diseases missed by miR2Trait could be those that involve deleterious mutations in the miRNA-binding region of the gene itself. Finally, the significant size of the overlap between the two databases increases confidence in the methods advanced here, which are also easily extensible with updates to the underlying primary databases. These observations further suggest the use of miR2Trait to complement HMDD.

Network analysis

The availability of two-way mappings between miRNAs and diseases with various analysis tools paves the way to varied investigations. One direction would be to identify pairs of miRNAs that lead to clusters of similar disease phenotypes. We addressed this question by using the miRNA enrichment analyzer discussed herein to compute the disease enrichment scores (2-tuples of BH-adjusted p-value and odds-ratio) of each miRNA-miRNA pair using Fisher’s exact test. This process is iterated over the entire set of miRNA-miRNA pairs (2595*2595) to obtain an adjacency matrix of ∼3 million interactions. This matrix was used to construct the corresponding network made of significant edges weighted by the odds-ratio. The interactions were sorted by -log₁₀ p-value and the top 10,000 interactions were used to construct the miRNA-miRNA adjacency network, yielding 2,442 nodes (and 10,000 edges). The network was read into Cytoscape (Shannon et al., 2003), and centrality analysis was performed using CytoNCA (Tang et al., 2015). The weighted versions of the following seven centrality measures were used: betweenness, closeness, eigenvector, degree, network, information, and local average centralities. The top 100 miRNAs from each centrality measure were used to obtain the consensus (five or more measures) top miRNAs (Table 7; algorithm details and Python script given here: https://github.com/miR2Trait/miR2Trait/wiki/Top-Nodes-from-Adjacency-Networks). Edge-percolation centrality of the network was also computed using cytoHubba to rank the miRNAs (Chin et al., 2014). The top 23 miRNAs contained 14 consensus miRNAs and the top 75 contained all but one of the consensus miRNAs. Such a consensus list could identify the critical regulatory miRNAs in human for further investigation. A disease-disease adjacency network was similarly constructed, and both the above adjacency networks and the scripts to perform the network construction and analysis are available from miR2Trait (https://github.com/miR2Trait).

Deployment

The miR2Trait web server and standalone CLI could be queried by either disease or miRNA. The standard use-case would be to identify miRNAs associated with some disease. For instance, a search by disease for ‘Myocardial infarction’ returned 1,445 associated miRNAs. The top hits included hsa-miR-302a-3p, hsa-miR-302b-3p, hsa-miR-302c-3p, hsa-miR-302d-3p, hsa-miR-367-3p and hsa-miR-367-5p, all of which are well-documented in the literature (Sun et al., 2017). The top hits for a search by disease of ‘Alzheimer’s disease’ included hsa-miR-9, hsa-miR-29a, hsa-miR34a, hsa-miR-106b, hsa-miR-125b, hsa-miR-146a, and hsa-miR-155, again all of which are well-documented in the literature (Leidinger et al., 2013). As a last example, the top hits for a search by disease of ‘Diabetes mellitus’ included hsa-miR-577, hsa-miR-37a, hsa-miR-375, hsa-miR-181a, hsa-miR-17, and hsa-miR-24, all of which are reported in the literature (Kim & Zhang, 2019; Guay et al., 2011). The search term is case-insensitive, matches anywhere, and interpreted according to regex grammar (https://dev.mysql.com/doc/refman/8.0/en/regexp.html#operator_regexp). Such searches might yield miRNAs lacking documented disease connections, and offer hypotheses for future investigations. In this regard, miR2Trait provides a script to retrieve known information about miRNAs from miRBase upon querying with the miRNA name or sequence (https://github.com/miR2Trait/miR2Trait/wiki/Querying-miRNA-information-from-miRBase).

In summary, we have developed miR2Trait, a comprehensive resource for investigating miRNA-disease associations, based on a simple yet surprisingly effective technique for fusing two curated primary databases. One outcome of the approach has been the enhanced interpretability and transparency of the resource constituents, aspects that are lost in more complex methods to arrive at reliable miRNA-disease associations.