Abstract Network-based computational method, with the emphasis on biomolecular interactions and biological data integration, has succeeded in drug development and created new directions, such as drug repositioning and drug combination. Drug repositioning, that is finding new uses for existing drugs to treat more patients, offers time, cost and efficiency benefits in drug development, especially when in silico techniques are used. MicroRNAs (miRNAs) play important roles in multiple biological processes and have attracted much scientific attention recently. Moreover, cumulative studies demonstrate that the mature miRNAs as well as their precursors can be targeted by small molecular drugs. At the same time, human diseases result from the disordered interplay of tissue- and cell lineage-specific processes. However, few computational researches predict drug-disease potential relationships based on miRNA data and tissue specificity. Therefore, based on miRNA data and the tissue specificity of diseases, we propose a new method named as miTS to predict the potential treatments for diseases. Firstly, based on miRNAs data, target genes and information of FDA (Food and Drug Administration) approved drugs, we evaluate the relationships between miRNAs and drugs in the tissue-specific PPI (protein-protein) network. Then, we construct a tripartite network: drug-miRNA-disease Finally, we obtain the potential drug-disease associations based on the tripartite network. In this paper, we take breast cancer as case study and focus on the top-30 predicted drugs. 25 of them (83.3%) are found having known connections with breast cancer in CTD (Comparative Toxicogenomics Database) benchmark and the other 5 drugs are potential drugs for breast cancer. We further evaluate the 5 newly predicted drugs from clinical records, literature mining, KEGG pathways enrichment analysis and overlapping genes between enriched pathways. For each of the 5 new drugs, strongly supported evidences can be found in three or more aspects. In particular, Regorafenib (DB08896) has 15 overlapping KEGG pathways with breast cancer and their p-values are all very small. In addition, whether in the literature curation or clinical validation, Regorafenib has a strong correlation with breast cancer. All the facts show that Regorafenib is likely to be a truly effective drug, worthy of our further study. It further follows that our method miTS is effective and practical for predicting new drug indications, which will provide potential values for treatments of complex diseases. Keywords: drug repositioning, miRNAs, tissue specificity, module distance Introduction The identification of therapeutic approaches for the treatment of cancer is an arduous, costly, and often inefficient process. By conservative estimates, it now takes over 15 years and $800 million to $1 billion to bring a new drug to market [29]^1. Drug repositioning, which is the discovery of new indications for existing drugs, is an increasingly attractive mode of therapeutic discovery. A repositioned drug does not need the initial six to nine years required for the development of new drugs, but instead goes directly to preclinical testing and clinical trials, thus reducing risk and costs [30]^2. Repositioning drugs has been implemented in several ways. One of the well-known examples is sildenafil citrate, which was repositioned from a hypertension drug to a therapy for erectile dysfunction [31]^3. Drugs treat diseases by targeting the proteins related to the phenotypes arising from the disease. However, drug development does not accurately follow the “one gene, one drug, one disease” principle, which has been challenged in many cases [32]^4 and the traditional drug repositioning methods by accident makes it hard to satisfy medical needs by successfully repositioning a large number of existing drugs. Computational methods are able to solve this question by high-level integration of available biological data and elucidation of unknown mechanisms. In recent years, systems biology continues to make important progress to solve fundamental problems in biology and leading to practical applications in medicine and drug discovery [33]^5. Network-based computational systems biology emphasizes the interactions among biomolecules and highlights the network concept. Typically, a network comprises a set of nodes and edges, and is described by graph theory in a mathematical manner [34]^6. A node can be a biological molecule, for example, gene, RNA, protein, metabolite, and pathway. A node can also be at the phenotype level such as disease and drug. An edge can represent the complex interaction between two nodes such as protein-protein interaction, drug-disease therapeutic relationship, drug-protein target relationship, and so on. The accumulation of different high-throughput biology data, such as gene expression data, miRNA expression data and drug-target data, has made the reconstruction of biomolecular and cellular networks possible. There are many network-based methods to predict the new indications of drugs [35]^7^,[36]^8^,[37]^9^,[38]^10^,[39]^11. Cheng et al. built a bipartite graph composed of the approved drugs and proteins linked by drug target binary associations and relied on a supervised network-based inference method to predict drug-target interactions [40]^12. Chen et al. constructed a general heterogeneous network which comprised drugs and proteins linked by protein-protein sequence similarity, drug-drug chemical similarity, and the known drug-target interaction [41]^13. Yeh et al. [42]^14 developed a network flow approach for identifying potential target proteins, which have a strong influence on disease genes in the context of biomolecular networks. The biomolecular networks are weighted by degree of co-expression of interacting protein pair. More recently, many studies have demonstrated that drugs can regulate microRNA (miRNAs) expression and mature miRNAs as well as their precursors can be targeted by small molecular drugs [43]^15^,[44]^16^,[45]^17^,[46]^18. For example, Miravirsen (SPC3649) is the first miRNA-targeted drug in clinical trials, which can successfully inhibit miR-122 expression that is required by hepatitis C virus replication [47]^19. The expression levels of 32 miRNAs (significant up-regulation of 22 miRNAs and down-regulation of 10 miRNAs) were changed after the treatment of trichostatin A in human breast cancer cell lines [48]^20. miRNAs are non-coding small RNAs (∼23 nucleotides) that downregulate gene expression at the post transcriptional level by inhibiting translation or initiating mRNA degradation and are dysregulated in most of human cancers [49]^21. Increasingly evidences have demonstrated that miRNAs play significant roles in many important biological processes, such as cell growth [50]^22^,[51]^23^,[52]^24, cellular signaling [53]^25, tissue development [54]^26 and disease process [55]^27^,[56]^28^,[57]^29^, [58]^30^,[59]^31^,[60]^32. Although only approximately 2000 miRNAs exist in humans, they regulate 30% of all genes [61]^33. miRNAs have been identified to play a crucial role in various human disease, especially in cancers. Therefore, targeting miRNAs with drugs will provide a new type of therapy for complex diseases [62]^34 and a new direction for drug repositioning. However, few computational researches predict drug-disease relationships based on miRNA data. Moreover, many genes with tissue-specific expression and function are expected to underlie many human diseases [63]^35^,[64]^36^,[65]^37. Therefore, in this study, we propose a new method based on miRNA data and tissue specificity of diseases, named as miTS, to predict potential drugs for diseases. The framework of miTS is shown in Figure [66]1. Firstly, we download miRNA expression data of diseases from TCGA [67]^38, miRNA-target gene relationship data from three experimentally validated databases: miRecords [68]^39, miRTarbase [69]^40 and TarBase [70]^41, and the drug-target gene data from Drugbank [71]^42 and KEGG [72]^43. Secondly, we select differentially expressed miRNAs of diseases based on a threshold and preprocess the target information of FDA approved drugs. Thirdly, we evaluate the relationships between miRNAs and drugs in the tissue-specific PPI network. And then, we construct a tripartite network: drug-miRNA-disease. Finally, we obtain the potential drug-disease associations based on the tripartite network. In this paper, we take breast cancer as case study and evaluate the results from CTD (Comparative Toxicogenomics Database) benchmark, clinical records, literature mining, KEGG pathways enrichment analysis and overlapping genes between enriched pathways. In the top-30 drugs, we find 5 new drugs for breast cancer. In particular, Regorafenib (DB08896) has 15 overlapping KEGG pathways with breast cancer and their p-values are all very small. In addition, whether in the literature curation or clinical validation, Regorafenib has a strong correlation with breast cancer. All the facts show that Regorafenib is likely to be a truly effective drug, worthy of our further study. Figure 1. [73]Figure 1 [74]Open in a new tab The framework of our method miTS. (A) Data preparation: miRNA expression data of breast cancer got from TCGA, miRNA-target gene data got from miRecords, miRTarbase and TarBase, and drug-target gene data got from Drugbank and KEGG. (B) Data preprocessing: we use Z-score to obtain the differentially expressed miRNAs for diseases and preprocess the target information of drugs. (C) In the tissue-specific PPI network, the targets of drug and miRNA are mapped to the PPI network. Orange nodes represent the target genes of miRNAs. Purple nodes represent the target genes of drugs. Green nodes represent the background genes. (D) Based on the module distance algorithm, we construct a drug-miRNA-disease tripartite network, and then based on the tripartite network, we get potential drugs for diseases. d[A,B]represents the association score between a drug and a disease. Methods Datasets Drug-target data FDA-approved drugs of human and their corresponding targets are downloaded from KEGG database and Drugbank. We merge the two datasets and get 1,732 drugs, 1,714 targets and 12,361 drug-target pairs. miRNA-target data The target genes of miRNAs are downloaded from miRecords, miRTarbase, and TarBase databases. We merge the three datasets and get 340 miRNAs, 2,028 targets and 3,652 miRNA-target pairs. miRNA-disease data The miRNA-disease curated relationships are downloaded from HMDD (the Human MicroRNA Disease Database) [75]^44. HMDD presents more detailed and comprehensive annotations to the human miRNA-disease association data, including miRNA-disease association data from the evidence of genetics, epigenetics, circulating miRNAs, and miRNA-target interactions. Finally, we get 578 miRNAs, 383 diseases and 6,448 miRNA-disease relationships. miRNA expression data Taking breast cancer as case study, we download the miRNAs expression data related with breast cancer from TCGA and get a matrix of 503 rows and 1,189 columns, row representing miRNA, column representing cancer sample, and the values in the matrix representing the RPKM (Reads Per Kilobase per Million mapped reads) for the miRNAs. We take the mean value of the RPKM values for 1,189 samples as the final value. Disease-gene data The genes related with breast cancer are downloaded from OMIM [76]^45 database. Tissue-specific PPI Interaction network We download the mammary tissue-specific PPI network marked as “Top Edges” from GIANT (Genome-scale Integrated Analysis of gene Networks in Tissues) database [77]^46 ([78]http://giant.princeton.edu/) (2017 version). GIANT proposes a tissue-specific benchmark to automatically up-weight datasets relevant to a tissue from a large data of different tissues and cell-types. Finally, we get 15,269 proteins and 883,071 protein-protein interactions. The weights on the edges are proportional to the relationships between nodes. In order to apply module distance algorithm [79]^47 (details shown in Figure [80]2) to calculate the relationships between drugs and miRNAs, we use the Gaussian kernel Inline graphic to transfer protein-protein closeness Inline graphic to protein-protein distance Inline graphic , as shown in formula (1). Figure 2. [81]Figure 2 [82]Open in a new tab An example for calculating the distance between target set of miRNA A and target set of drug B. Orange and purple nodes represent genes related to miRNA A and drug B, respectively. Node c is a shared node, so it is marked by two colors. graphic file with name ijbsv14p0971i004.jpg (1) Screening differentially expressed miRNAs In order to obtain the differentially expressed miRNAs of breast cancer, we first filter the miRNAs expression data downloaded from TCGA. For a miRNA Inline graphic , we use formula (2) to calculate its Inline graphic . graphic file with name ijbsv14p0971i007.jpg (2) Where Inline graphic is the RPKM value of miRNA Inline graphic ; Inline graphic and Inline graphic represent mean value and standard deviation of r, respectively. Then we choose Inline graphic =1.645 (p-value = 0.05) as threshold to screen differentially expressed miRNAs. Finally, we get a total of 40 differentially expressed miRNAs of breast cancer (see Table [83]1). In Table [84]1, the miRNAs marked by “*” represent they have connections with breast cancer in HMDD. We find 34 of 40 (85%) differentially expressed miRNAs are related with breast cancer, which indicates that miRNAs associated with breast cancer tend to be highly expressed in breast cancer patients. Then, we choose the 34 miRNAs marked by “*” in Table [85]1 for further study. Table 1. Differentially expressed miRNAs of breast cancer miRNA name Z-score miRNA name Z-score miRNA name Z-score hsa-mir-21* 3.32 hsa-mir-375* 2.32 hsa-mir-23a* 1.91 hsa-mir-22* 2.94 hsa-mir-101-1* 2.29 hsa-mir-199a-2* 1.90 hsa-mir-10b* 2.93 hsa-mir-200c* 2.28 hsa-mir-126* 1.90 hsa-mir-30a* 2.85 hsa-mir-25* 2.27 hsa-mir-100* 1.86 hsa-mir-148a* 2.77 hsa-let-7a-3* 2.21 hsa-let-7c* 1.79 hsa-mir-99b 2.73 hsa-let-7a-1* 2.21 hsa-mir-151 1.78 hsa-mir-143* 2.73 hsa-mir-30d* 2.19 hsa-mir-199a-1* 1.73 hsa-mir-182* 2.72 hsa-mir-92a-2* 2.18 hsa-mir-26a-2* 1.72 hsa-let-7b* 2.61 hsa-let-7f-2* 2.12 hsa-mir-142 1.72 hsa-mir-10a* 2.56 hsa-mir-93* 2.03 hsa-mir-29c* 1.70 hsa-mir-103-1 2.50 hsa-mir-29a* 2.03 hsa-mir-181a-1* 1.69 hsa-let-7a-2* 2.44 hsa-mir-28 2.00 hsa-mir-141* 1.66 hsa-mir-30e 2.38 hsa-mir-199b* 1.98 hsa-mir-183* 2.37 hsa-mir-203* 1.94 [86]Open in a new tab The miRNAs marked by “*” represent they have relationship with breast cancer in HMDD. Construct drug-miRNA-disease tripartite network The relationship between a miRNA and a drug is derived by measuring the correlation between their target sets. Because miRNA target genes, drug target genes and protein-protein interaction (PPI) networks remain largely incomplete, we calculate the distance between two modules based on the shortest path in incomplete networks [87]^44. Figure [88]2 gives an example to calculate the distance between miRNA A and drug B in a weighted tissue-specific PPI network. As shown in Figure [89]2, miRNA A has three target genes, marked as a, b, c and drug B has four targets, marked as c, d, e, f. For the node a, its distance to targets {c, d, e, f} of drug B are 0.8, 1.0, 1.1 and 1.9 respectively, so its shortest distance to drug B is 0.8. In this way, we can obtain the distances between each node in gene set {a, b, c} and drug B, and the distances between each node in target set {c, d, e, f} and miRNA A, shown in Figure [90]2. Finally, the distance between miRNA A and drug B, Inline graphic , is equals to the sum of all the distances divided by the total number of nodes related to miRNA A and drug B. Here, the total number is 7. Through the above calculation process, we get 1,017 drugs, 25 miRNAs and 25,425 drug-miRNA relationships. Combining the drug-miRNA relations with the miRNA-breast cancer information, we construct a drug-miRNA-breast cancer tripartite network. Predicting potential drugs for breast cancer Based on the drug-miRNA-breast cancer tripartite network, we predicting potential drugs for breast cancer. If a drug and breast cancer have common miRNA neighbors, there will be a connection between them. Finally, there are 25 common miRNAs between drugs and breast cancer. We use formula (3) to calculate the average distance between the 25 miRNAs related to breast cancer and drugs as the drug-breast cancer relationship distance score, Inline graphic . graphic file with name ijbsv14p0971i015.jpg (3) Where Inline graphic represents the distance between the i-th miRNA of disease C and the drug B (the calculation of Inline graphic as shown in Figure [91]2); n represents the number of miRNAs corresponding to disease Inline graphic . Here, C= breast cancer and n= 25. In order to make the drug-disease distances be proportional to their direct correlations, we use formula (4) to normalize Inline graphic as Inline graphic : graphic file with name ijbsv14p0971i021.jpg (4) Where Max[d] and Min[d] represent the maximum and the minimum of all the drug-disease distances, respectively; Inline graphic represents the distance between disease C and drug B; Inline graphic represents the direct association between disease C and drug B. Results CTD benchmark verification In our study, we choose breast cancer as case, the drug-breast cancer associations are ranked in descending order according to their scores. In order to verify the accuracy of our results, we use the drug-breast cancer relationships data in Comparative Toxicogenomics Database (CTD) [92]^48 as benchmark. As shown in Figure [93]3, we give the precision curves of predicted drug-breast cancer relationship results. For each given threshold, the precision of our method is calculated by formula (5). Figure 3. [94]Figure 3 [95]Open in a new tab The precision of our predictions at different top-x% drug-breast cancer pairs. graphic file with name ijbsv14p0971i024.jpg (5) Where P represents the number of predicted drug-disease pairs; P[CTD] represents the number of drug-disease pairs, which can be found in CTD database. In Figure [96]3, we give the precision curves of predicted drug-breast cancer pairs at different top-x%. From the figure, we find the higher the associations ranking, the higher the accuracy. Hence, for the breast cancer, we choose top 30 drugs for further analysis. The top 30 drugs related to breast cancer are shown in Table [97]2. We validate the 30 drugs by CTD database and find 11 (36.7%) of them are marked as “therapeutic (T)”, which means that they have a highly correlation with breast cancer. In addition, we find 14 of the rest 19 drugs also have connections with breast cancer in CTD database with inference score [98]^49 over 0 and they are marked as “Ref” in Table [99]2. The inference score [100]^46 reflects the degree of similarity between CTD chemical-gene-disease networks and a similar scale-free random network, which is computed as shown below: Table 2. The top 30 drugs related to breast cancer Rank Drugbank ID Drug name Marker Inference Score Similarity Score 1 DB08818 Hyaluronic acid Ref 61.1 1.00000 2 DB00570 Vinblastine T 40.6 0.97977 3 DB00642 Pemetrexed T 12.38 0.97666 4 DB01169 Arsenic trioxide T 212.32 0.96998 5 DB00242 Cladribine Ref 14.83 0.96343 6 DB04967 Lucanthone Ref 45.07 0.96120 7 DB09073 Palbociclib T 53.75 0.96083 8 DB02701 Nicotinamide Ref 63.26 0.96013 9 DB01005 Hydroxyurea Ref 32.57 0.95909 10 DB01204 Mitoxantrone T 25.1 0.95847 11 DB00309 Vindesine sulfate T 2.54 0.95388 12 DB00361 Vinorelbine T 4.36 0.95388 13 DB08871 Eribulin mesylate None None 0.95388 14 DB01394 Colchicine Ref 50.29 0.95213 15 DB01229 Paclitaxel T 111.41 0.95050 16 DB01248 Docetaxel T 72.35 0.95050 17 DB00440 Trimethoprim Ref 6.82 0.94971 18 DB01179 Podofilox Ref 2.87 0.94918 19 DB05260 Gallium nitrate Ref 19.88 0.94671 20 DB00441 Gemcitabine T 112.67 0.94582 21 DB00031 Tenecteplase None None 0.94357 22 DB08813 Nadroparin None None 0.94325 23 DB00432 Trifluridine Ref 12.49 0.94205 24 DB01073 Fludarabine Ref 59.43 0.94184 25 DB00694 Daunorubicin Ref 85.05 0.94044 26 DB00970 Dactinomycin Ref 98.94 0.93988 27 DB08896 Regorafenib None None 0.93853 28 DB06813 Pralatrexate None None 0.93853 29 DB00563 Methotrexate T 123.36 0.93799 30 DB00615 Rifabutin Ref 2.89 0.93433 [101]Open in a new tab Ranked by drug-breast cancer similarity score. Marker has three values: T(therapeutic), Ref (inferred by genes) and None (no record in CTD database). Inference Score represents the score for the inference based on the topology of the network consisting of the chemical, disease, and one or more genes used to make the inference. graphic file with name ijbsv14p0971i025.jpg graphic file with name ijbsv14p0971i026.jpg (6) where Y represents the inference score; P represents the probability that a vertex in a large network interacts with another vertex decays according to a power law [102]^50; G, C, and D represent a gene, chemical, and disease respectively; k represents the number of connection between G, C, or D; n[G] represents a gene set. The higher the Y score, the more likely the inference network has atypical connectivity [103]^51. That is to say, there are 83.3% (25/30) drugs can be found in the CTD database and we predict five potential drugs for breast cancer (DB08871, DB00031, DB08813, DB08896, and DB06813, marked as boldface in Table [104]2). Clinical evaluation For the five predicted drugs, we further analyze them based on the ClinicalTrials.gov ([105]https://clinicaltrials.gov/). ClinicalTrials.gov is a registry and results database of publicly and privately supported clinical studies of human participants conducted around the world. Currently, it lists 242,537 studies with locations in all 50 states and in 198 countries (April 25, 2017). From the ClinicalTrials.gov, we can find 84 records for drug Eribulin mesylate (DB08871) treat breast cancer. For example, “Eribulin Mesylate Phase IV Clinical Trial in Korean Patients with Metastatic or Locally Advanced Breast Cancer (ESKIMO) ([106]NCT01961544)”, the purpose is assessing the safety of Eribulin which is approved for the treatment of the patients in Korea with locally advanced or metastatic breast cancer; “Eribulin with Trastuzumab as First-line Therapy for Locally Recurrent or Metastatic HER2 Positive Breast Cancer ([107]NCT01269346)”, the purpose is evaluating the safety and efficacy of Eribulin mesylate in combination with trastuzumab as first line treatment in female subjects with locally recurrent or metastatic human epidermal growth factor receptor (HER2) positive breast cancer; “Eribulin Mesylate in Treating Patients with Previously Treated Metastatic Breast Cancer ([108]NCT01908101)”, and so on. For drug Nadroparin (DB08813), we find one record: “Prevention of Venous and Arterial Thromboembolism, in Cancer Patients Undergoing Chemotherapy, With a Low Molecular Weight Heparin (Nadroparin Calcium) ([109]NCT00951574)”, 1200 patients with lung, breast, gastrointestinal (stomach, colon-rectum, pancreas), ovarian or head and neck cancer undergoing chemotherapy will be randomly assigned in a 2:1 ratio and in double-blind conditions to a treatment with subcutaneous low molecular weight heparin (nadroparin calcium, one injection/day) or placebo for the overall duration of chemotherapy or up to a maximum of 4 months. For drug Regorafenib (DB08896), we find three records related with breast cancer, “Refametinib in Combination with Regorafenib in Patients with Advanced or Metastatic Cancer ([110]NCT02168777)”, “Effect of Regorafenib on Digoxin and Rosuvastatin in Patients with Advanced Solid Malignant Tumors ([111]NCT02106845)”, and so on. For drug Pralatrexate (DB06813), we find a clinical study of Pralatrexate in 22 female patients with previously-treated breast cancer ([112]NCT01118624). Only one drug, Tenecteplase (DB00031) was not found in ClinicalTrials.gov. Literature curation In the above section, the top 30 drugs related with breast cancer are validated by CTD database and Clinical database. After our analysis, we obtain five potential drugs (Eribulin mesylate, Tenecteplase, Nadroparin, Regorafenib, Pralatrexate) for breast cancer. In ClinicalTrials.gov database, only one drug, Tenecteplase, cannot be found its corresponding record. In this section, we will analyze the five potential drugs for breast cancer by literature mining. The summarized results are shown in Table [113]3. Table 3. Literature verification for the five potential drugs Drug names Drugbank ID Original indications Supported references for