Abstract Background MicroRNAs always function cooperatively in their regulation of gene expression. Dysfunctions of these co-functional microRNAs can play significant roles in disease development. We are interested in those multi-disease associated co-functional microRNAs that regulate their common dysfunctional target genes cooperatively in the development of multiple diseases. The research is potentially useful for human disease studies at the transcriptional level and for the study of multi-purpose microRNA therapeutics. Methods and results We designed a computational method to detect multi-disease associated co-functional microRNA pairs and conducted cross disease analysis on a reconstructed disease-gene-microRNA (DGR) tripartite network. The construction of the DGR tripartite network is by the integration of newly predicted disease-microRNA associations with those relationships of diseases, microRNAs and genes maintained by existing databases. The prediction method uses a set of reliable negative samples of disease-microRNA association and a pre-computed kernel matrix instead of kernel functions. From this reconstructed DGR tripartite network, multi-disease associated co-functional microRNA pairs are detected together with their common dysfunctional target genes and ranked by a novel scoring method. We also conducted proof-of-concept case studies on cancer-related co-functional microRNA pairs as well as on non-cancer disease-related microRNA pairs. Conclusions With the prioritization of the co-functional microRNAs that relate to a series of diseases, we found that the co-function phenomenon is not unusual. We also confirmed that the regulation of the microRNAs for the development of cancers is more complex and have more unique properties than those of non-cancer diseases. Electronic supplementary material The online version of this article (doi:10.1186/s12859-017-1605-0) contains supplementary material, which is available to authorized users. Keywords: Cross-disease analysis, Disease-microRNA associations prediction, Co-functional microRNA pair Background MicroRNAs (miRNAs), a class of small non-coding RNA of ∼22 nucleotides, are significant regulation molecules for diverse cellular processes such as cell development, proliferation and differentiation [[35]1–[36]7]. Pairs of miRNAs can work cooperatively to regulate an individual gene or a cohort of genes that participate in similar processes [[37]8, [38]9]. This cooperativity (or co-function) is a frequent regulation mechanism of miRNAs for an enhanced target repression which has exhibited distinctive and fine-tuned target gene expression patterns [[39]10]. Investigation on miRNA cooperativity can systematically understand miRNA functions [[40]11] to detect their potential disease links [[41]12]. Using miRNAs as diagnostic and therapeutic targets, miRNA therapeutics is a promising research area that designs sophisticated strategies to restore or inhibit miRNA expression for the treatment of cancer and other diseases. For example, a therapy with the vector-encoded pair miR-15a and miR-16-1 has been proposed for the treatment of chronic lymphocytic leukaemia (CLL) [[42]13]; The microRNA cluster miR-216a/217 was reported to target genes PTEN and SMAD7 to induce the epithelial-mesenchymal transition, which can promote the drug resistance and recurrence of liver cancer [[43]14]. Such co-functional miRNA pairs are more suitable as drug targets instead of using individual ones. Large scale detection of novel co-functional miRNA pairs is an important pre-step to identify proper miRNA pairs as more effective drug targets. Currently, abundant disease-gene association information are stored in Online Mendelian Inheritance In Man (OMIM) [[44]15] and Comparative Toxicogenomics Database(CTD) [[45]16]; disease-miRNA associations are recorded in miR2Disease [[46]17] and HMDD [[47]18]; and miRNA-target regulations are recorded in miRecord [[48]19] and miRTarBase [[49]20]. Linking and integrating these databases, it can be inferred which diseases are correlated with the same genes or with the same miRNAs, and which miRNAs have the same target disease genes. Our hypothesis is that some of the miRNAs can regulate their common targets cooperatively and have roles in the development of a series of diseases. The focus of this work is on the detection and prioritization of multi-disease associated co-functional miRNA pairs. A multi-disease associated co-functional miRNA pair is a pair of miRNAs whose common target genes are associated with a series of diseases. Here, the definition of co-function for the miRNA pairs is broader than the definition of cooperativity as proposed in [[50]21, [51]22]. Figure [52]1 shows an example of multi-disease associated co-functional miRNA pairs detected from a disease-gene-miRNA (DGR) tripartite network. From this example, we can see that multi-disease associated co-functional miRNA pairs may hold a vast mechanism underlying multiple disease development, similarly like the basic cellular functions maintained by housekeeping genes. More importantly, these miRNAs can be considered as the common drug targets of these diseases for the design and development of multi-purpose drugs. Fig. 1. Fig. 1 [53]Open in a new tab An example: From a DGR tripartite network to a co-functional miRNA pair. The network in panel a contains known associations between the genes g1, g2, g3, g4, and g5, the diseases d1, d2, d3, and d4, and the miRNAs R1, R2, R3, and R4. In this example, miRNAs R2 and R3 are both associated with all the four diseases. However, the other three miRNAs are each associated with only one of these diseases. All these four diseases are associated with two common genes g4 and g5. Meanwhile, both of g4 and g5 are the targets of miRNAs R2 and R3. It is believed that R2-R3-g4-g5 in panel b may form a functional module that associated with the development of all the four diseases MiRNA co-function mechanisms have attracted intensive research recently [[54]9, [55]11, [56]12, [57]23, [58]24], with the focus on the analysis of miRNA-target networks or on the analysis of disease-miRNA associations for a specific disease. Our work advances the current research with two steps: (i) We reconstruct a DGR tripartite network through the integration of existing databases with our newly predicted disease-miRNA associations, and (ii) we propose a novel scoring method to prioritize the potential multi-disease associated co-functional miRNA pairs. Since the relationships between the exact miRNAs and diseases are largely unknown, computational methods are required to make prediction of disease-related miRNAs for constructing the disease-miRNA network in the DGR tripartite. For example, network-based or semi-supervised prediction methods [[59]25–[60]27], or the methods via support vector machines [[61]28, [62]29] can be used among some other prediction methods [[63]30–[64]32]. The key idea in the similarity assessment adopted by most of these methods is that: similar RNAs (functionally similar) are always associated with similar diseases (phenotypically similar, genotypically similar or semantically similar). During the training of the existing prediction methods, the disease-miRNA pairs without known relationships are thought to be ranked at bad positions or are regarded as negative samples directly. As some (probably many) of the unknown disease-miRNA pairs in the training data are true in fact, the false positive rates by the literature methods are high in the prediction of disease related miRNAs. On the other hand, the use of negative samples by the literature methods is straightforward without consideration of gene expression properties of miRNAs. To improve the prediction performance, we propose a new method to make predictions of disease-related miRNAs. Two new ideas are explored. One is the construction of a set of reliable negative samples of disease-miRNA association through miRNA expression comparison between control and diseased subjects. The second idea is the use of precomputed kernel matrix for support vector machines, which can avoid the step to tune the parameters of the kernel functions. The area under the ROC curve(AUC) performance of our method is much superior to the literature methods on bench-marking data sets. Our case studies have demonstrated that our prediction method can also work well even when a disease has no currently known disease-related miRNAs. Combining our predicted disease-miRNA associations with those literature-maintained associations between diseases, miRNAs and genes, we construct a more complete DGR tripartite network to detect and prioritize multi-disease associated co-functional miRNA pairs. Given a miRNA pair, our scoring method cfscore considers the function relationship between the two miRNAs, the co-dysexpression of the two miRNAs in the disease tissues and the relationship between the common target genes and the associated diseases of these miRNAs. We are also interested in finding the exact targets dysregulated by the co-functional miRNA pair during the diseases’ development. We call them the co-functional targets of the co-functional miRNA pair. The flowchart of our work is described in Fig. [65]2. Fig. 2. Fig. 2 [66]Open in a new tab The flowchart of our prediction and scoring method. Our work includes the parts such as material collection, similarity computing, association prediction, network reconstruction, scoring and prioritization of the co-function miRNA pairs and result output This method was tested on the cancer and non-cancer disease related DGR tripartite networks. The top 50 multi-disease associated co-functional miRNA pairs were concentrated for deep analysis. We found that most of them were from the same miRNA families or miRNA clusters. The comparison of the co-functional pairs from the two DGR networks suggests that the dysregulation mechanisms of miRNAs in the cancers are more complex. It has also been shown that the analysis of multi-disease associated co-functional miRNAs can help understand the regulation mechanisms of miRNAs in the development of different diseases and thus can provide new knowledge for the diagnosis or treatment of the diseases. Results Multi-disease associated co-functional miRNA pairs and their common dysfunctional target genes Two cancer-gene-miRNA tripartite networks were constructed to investigate the performance of our method for detecting and ranking multi-cancer associated co-functional miRNA pairs. As a pre-processing step, we merged the miRCancer database [[67]33] with miR2Disease [[68]17] and HMDD [[69]18], and collected 3655 cancer-miRNA associations between 83 cancers and 503 miRNAs. Connecting these miRNAs and diseases to their associated genes, the first cancer-gene-miRNA tripartite network was constructed. Then, all the 3655 cancer-miRNA associations (as positive samples) and a balanced set of 3655 negative samples of cancer-miRNA association in this tripartite network were used together to train our prediction model for inferring new cancer-miRNA associations. The prediction model was applied to all the un-connected disease-miRNA pairs between the 83 cancers and 503 miRNAs to predict whether some of them have associations or not. When a pair was predicted to have an association between a cancer and a miRNA, a probability was also estimated. A total of 3000 top-ranked associations were added to the first cancer-gene-miRNA tripartite network to form the second cancer-gene-miRNA tripartite network (i.e., a reconstructed network by adding the predicted cancer-miRNA associations). Those associations can be found in the Additional file [70]1. On average, the 503 miRNAs are associated with 7 or 13 cancers for the first and the reconstructed network respectively; and there are 2532 and 5634 miRNA pairs in these two networks that have a cfScore larger than 0 and that are associated with at least 10 cancers. There are very few literature proving the miRNA pairs can co-function in the development of more than 10 different diseases. To understand whether these miRNA pairs co-function in the development of some of the diseases, we manually searched and examined relevant literature to confirm that the individual miRNAs in the pairs can function cooperatively to regulate the same targets. Of the top-ranked 50 miRNA pairs from our reconstructed network, 40 pairs can be validated to be co-functional pairs by the literature, in comparison with 35 of the top 50 pairs from the first tripartite network. This implies that the addition of the predicted disease-miRNA associations into the tripartite network is useful and effective for the study of co-functional miRNA pairs. Here, we can just confirm these pairs of miRNAs are co-functional miRNA pairs but not multi-disease associated co-functional ones. We could not find any literature that discusses the relationship between miRNAs and a series of diseases. Details of the 50 miRNA pairs are shown in Fig. [71]3, where on the label of every edge, the first number represents the ranking position of the miRNA pair. If the rank number is followed by one or more gene names, it represents that the miRNA pair is a co-functional pair and has validated common targets. The number at the end of the label is the number of diseases that may associate with this co-functional pair. These multi-cancer associated co-functional miRNA pairs are mostly from the same clusters or families such as from the let-7 family (let-7a ∼7e and miR-98) and the miR-17 ∼92 cluster (miR-17-3p, miR-17-5p, miR-18a, miR-19a, miR-19b, miR-20a and miR-92). It has been known that clustered miRNAs or those miRNAs from the same family are evolved from a common ancestor and can target functionally related genes [[72]34]. Thus, it can be easily understood that miRNAs from the same cluster or family have similar functions and can always function cooperatively. However, not all those miRNAs in the same families or clusters can co-function with each other as their target genes are not completely overlapped. Moreover, some miRNAs that belong to different families or clusters can be co-functional miRNAs. For example, the 17th-ranked pair miR-497-5p-miR-424-5p is a co-functional miRNA pair. However, as recorded by miRBase, miR-424-5p is a member of mir-322 gene family while miR-497-5p stems from the mir-497 family. The pair is also not clustered. Fig. 3. Fig. 3 [73]Open in a new tab The 50 top-ranked co-functional miRNA pairs from the reconstructed cancer-miRNA-gene network. The labels along the edges illustrate the co-function information of the miRNAs. The first number of each label is the rank of the corresponding pair according to our prioritization method. The following gene symbols are the validated common targets during the co-functioning of the pair of miRNAs. The last number shows the potential diseases that related to this co-function pair The 5th-ranked pair, miR-15b and miR-195, both belong to the miR-15 family, and both of them can target gene BCL2, an important apoptosis inhibitor. This pair of miRNAs can also work together with another miRNA (miR-16) in regulation [[74]35]. We hypothesize that this co-functional pair may dysregulate their targets cooperatively, leading to the development of 38 different cancers such as prostate cancer (DOID:10283), prostate carcinoma (DOID:10286), stomach cancer (DOID:10534), and breast cancer (DOID:1612). The top three potential common targets of this miRNA pair are genes BCL2 (entrez id:596), CDKN1A (entrez id:1026), and CCND1(entrez id:595). We have verified that these three genes are individually related to most of (68%, 68% or 66%) the 38 cancers. Furthermore, these three genes are all involved in four KEGG [[75]36] pathways: hsa05215: Prostate cancer (p-value=1.5E-4), hsa05206: MicroRNAs in cancer (p-value=1.7E-3), hsa04151: PI3K-Akt signaling pathway (p-value=2.5E-3) and hsa05200: Pathways in cancer (p-value=3.2E-3) as revealed by the DAVID functional annotation tool [[76]37, [77]38]. Moreover, the three genes all have the functions of the cellular response to DNA damage stimulus (GO:0006974, p-value=1.4E-4) and response to drug (GO:0042493, p-value=4.0E-4), which are important functions for the normal cells. Based on these analysis and evidences, it is suggested that the pair of miR-15b and miR-195 may contribute to the development of all the 38 different types of cancers via a similar regulation mechanism. More details of the discovered miRNA pairs and references are listed in