Computational drug repositioning for peripheral arterial disease: prediction of anti-inflammatory and pro-angiogenic therapeutics

Peripheral arterial disease (PAD) results from atherosclerosis that leads to blocked arteries and reduced blood flow, most commonly in the arteries of the legs. PAD clinical trials to induce angiogenesis to improve blood flow conducted in the last decade have not succeeded. We have recently constructed PADPIN, protein-protein interaction network (PIN) of PAD, and here we combine it with the drug-target relations to identify potential drug targets for PAD. Specifically, the proteins in the PADPIN were classified as belonging to the angiome, immunome, and arteriome, characterizing the processes of angiogenesis, immune response/inflammation, and arteriogenesis, respectively. Using the network-based approach we predict the candidate drugs for repositioning that have potential applications to PAD. By compiling the drug information in two drug databases DrugBank and PharmGKB, we predict FDA-approved drugs whose targets are the proteins annotated as anti-angiogenic and pro-inflammatory, respectively. Examples of pro-angiogenic drugs are carvedilol and urokinase. Examples of anti-inflammatory drugs are ACE inhibitors and maraviroc. This is the first computational drug repositioning study for PAD.


Introduction
Recent pharmaceutical research and development (R&D) reports show that the probability of success for a new pharmaceutical compound to get to the market has declined in the last 10 years (Pammolli et al., 2011). The average time of drug development has increased from 9.7 years during the 1990s to 13.9 years from 2000 onwards. The average probability of success of total numbers of R&D projects in the cardiovascular system is only 4.86%. Drug repositioning, new use of old drugs, can shorten the development time and provide solutions for the high cost and declined number of new successful drugs of the pharmaceutical companies (Dudley et al., 2011). Computational repositioning strategies can predict new therapeutic indications for FDA-approved drugs, which then have to undergo clinical trials for the new indication (Belch et al., 2003;Ostchega et al., 2007;Shameer et al., 2015). In this study, we primarily use the network-based approach in computational drug repositioning.
Peripheral arterial disease (PAD) results from atherosclerosis, the plaque built-up inside the arteries, which blocks the blood flow in the peripheral arteries and most commonly in the arteries that perfuse the legs (Belch et al., 2003;Annex, 2013). Age, diabetes, and cigarette smoking are the major risk factors for the development of PAD (Belch et al., 2003;Ostchega et al., 2007;Annex, 2013). There are 8-12 million people with PAD in the United States (Writing Group et al., 2010). The clinical manifestations of PAD range from patients who do not report leg pain but have a lower functional capacity (approximately 50% of all PAD subjects) to patients who have intermittent claudication manifested as leg pain with walking/exercise that is relieved with rest (approximately 33-40% of all PAD subjects) (Hirsch et al., 2006;Norgren et al., 2007). With the goal to increase blood flow around blockages, clinical trials using drugs and gene delivery for therapeutic angiogenesis such as VEGF (vascular endothelial growth factor) gene delivery have been performed for the last two decades but have not been successful. Hoier et al. showed that there was no difference in basal skeletal muscle VEGF mRNA content before and after passive or active exercise between PAD patients and control (Hoier et al., 2013). However, the basal level of anti-angiogenic protein thrombospondin-1 (TSP1) was remarkably higher in the PAD patients than control groups. They conclude that the antiangiogenic factors dominate the pro-angiogenic factors in PAD patients. The up-regulation of TSP1 has been shown in various gene expression microarray studies of mouse (Chu et al., 2015) and human samples of PAD (Fu et al., 2008;Masud et al., 2012). Currently there are no FDA-approved drugs targeting TSP1. Therefore, the computational drug repositioning approach to predict the drugs targeting other endogenous anti-angiogenic proteins should be helpful for designing clinical trials for therapeutic angiogenesis in PAD.
Inflammation plays an important role in initiation and progression of PAD, and many circulating biomarkers such as matrix metalloproteinases (MMPs) and interleukin are considered as the clinical manifestation of PAD (Signorelli et al., 2014). Atherosclerosis is the dominant cause of many cardiovascular diseases, including myocardial infarction, heart failure, coronary artery disease (CAD), and stroke (Frostegård, 2013). Atherosclerosis is a chronic inflammatory condition. Potential anti-inflammatory treatments in atherosclerosis are reviewed in Frostegård (2013). The interplay between inflammation and endothelial progenitor cells is critical in cardiovascular diseases (Grisar et al., 2011). Combination of anti-inflammatory and pro-angiogenic treatments for PAD was suggested and validated in vivo by Zachman et al. (2014). However, a systematic bioinformatics approach to identify the potential drug repositioning for inhibition of anti-angiogenic and pro-inflammatory proteins for PAD is still lacking.
We previously constructed the PADPIN, protein-protein interaction network (PIN) in PAD that includes angiome, immunome, and arteriome, characterizing the processes of angiogenesis, immune response/inflammation and arteriogenesis, respectively (Chu et al., 2015). We have analyzed several available microarray gene expression datasets from ischemic and non-ischemic muscles in two mouse models of PAD (in C57BL/6 and BALB/c mouse species) from Hazarika et al. (2013) to identify important genes/proteins in PAD, such as THBS1 (thrombospondin-1), TLR4 (toll-like receptor 4), EphA4 (EPH receptor A4), and TSPAN7 (tetraspanin 7). However, none of the four genes (THBS1, TLR4, EphA4, and TSPAN7) have FDA-approved drugs to target them. Considering the time (>10 years) and cost (>$1 billion) for developing a new drug agent, drug repositioning in PAD offers promise of providing effective therapeutics in shorter time and at lower cost compared to conventional de-novo drug discovery and development. In addition, drug repurposing is an approach of taking agents in development that have achieved adequate safety for one indication but are tested for efficacy in another when safety is already evident.

Resources for Drugs and Drug-target Interactions
We rely on two major resources for drug information and drug-target, DrugBank 3.0 http://www.drugbank.ca/ (Knox et al., 2011) and Pharmacogenomics Knowledge Base (PharmGKB) http://www.pharmgkb.org/ (Whirl-Carrillo et al., 2012). DrugBank contains extensive omics data, such as pharmacogenomic, pharmacoproteomic, and pharmacometabolomic data. We use DTome (Drug-Target interactome tool) (Sun et al., 2012) to compile all the drugs included in DrugBank 3.0 (Knox et al., 2011), including the approved, experimental, nutraceutical, illicit, and withdrawn drugs. We compile three binary relations in DrugBank from DTome: drug-drug, drug-gene, and drug-target interactions. This compilation provides the rich resources for the potential repositioning or repurposing. By considering the drug safety and development time, we focus on FDA-approved drugs in this study. We compiled the three binary relations from PharmGKB: gene-disease, gene-drug, and gene-gene interactions. The drugtarget interactions were compiled from both DrugBank (Knox et al., 2011) andPharmGKB (Whirl-Carrillo et al., 2012).

Proteins in PADPIN and Therapeutic Angiogenesis in PAD
Details of the construction of PADPIN, protein-protein interaction (PIN) of PAD in angiogenesis, immune response and arteriogenesis, are described in Chu et al. (2015). The methodology is similar to that used for constructing the global PIN of angiogenesis (angiome) that comprises 1233 proteins and 5726 interactions (Chu et al., 2012). The PIN of immune response (immunome) comprises 3490 proteins and 21,164 interactions. The PIN of arteriogenesis (arteriome) comprises 289 proteins and 803 interactions. The degree of node represents the number of links to a node in the network. The network parameter was calculated by NetworkAnalyzer (Assenov et al., 2008) in Cytoscape (Smoot et al., 2011). We start with the genes listed in the three PINs, to find the interactive drugs from the DrugBank and PharmGKB. Note that in bioinformatics publications, and specifically in protein-protein networks publications, the terms "gene" and "protein" are sometimes used interchangeably; while we mostly use "protein" term in this context, we sometime use "gene" to be consistent with previous publications.

List of Anti-angiogenic and Pro-inflammatory Genes
The activation of a specific biological process can be implemented using two strategies. One is direct activation of the genes involved in positive regulation of that biological process; the other is inhibition of the genes involved in negative regulation of that biological process. Specifically for PAD, to stimulate vascular growth and remodeling and increase the blood flow, we propose inhibition of genes annotated as negative regulation of angiogenesis as a therapeutic approach to stimulating angiogenesis. The rationale for this approach is that numerous clinical trials aimed at stimulating angiogenesis by growth factors such as VEGF-A and FGF-2 have not been successful. We identified 39 anti-angiogenic genes, chosen by Gene Ontology (GO: 0016525) and literature (Chu et al., 2014). The endothelial dysfunction in patients with PAD is characterized by impaired nitric oxide signaling, excessive inflammation and diminished response to angiogenic factors (Annex, 2013). To inhibit the inflammation, we propose inhibition of pro-inflammatory responses as a therapeutic approach for anti-inflammatory treatment of PAD. There are 89 genes classified in positive regulation of inflammatory response (GO:0050729). We list these genes in Table 1.

Drug-targets Relations in Angiome, Immunome and Arteriome of PADPIN
We collected 11,043 binary relations between the drug and drug targets from DrugBank 3.0 (Knox et al., 2011) and 3138 binary relations between the drug and associated genes of that drug, which may not be the direct targets, from PharmGKB (Whirl-Carrillo et al., 2012). By matching the genes in angiome, immunome, and arteriome with the drug targets listed in the drug-gene binary relations from DrugBank and PharmGKB, we build the complete tables of genes and repositioning drugs (Tables S1-S3). Table S1 shows 409 and 174 drug targets listed in angiome for the drugs from DrugBank and PharmGKB, respectively. We select the genes with at least one drug targeting that gene in angiome, and skip the genes without any drug-gene relations. There might be multiple drugs targeting the same drug target; we list the multiple drugs in the same row of the table. Table S2 shows 865 and 382 drug targets in immunome for the drugs from DrugBank and PharmGKB, respectively. Table S3 shows 82 and 46 drug targets in arteriome for the drugs from DrugBank and PharmGKB, respectively.
We rank the genes in angiome, immunome, and arteriome by the degree of nodes, i.e., number of links of the nodes in the network, in Tables S1-S3, respectively. Tables S1-S3 provide the complete list of drugs and drug targets which are annotated in angiogenesis, immune response/inflammation, and arteriogenesis. Tables S1-S3 provide the complete list of drugs in DrugBank and PharmGKB, including approved, experimental, nutraceutical, illicit, and withdrawn drugs. Considering the drug safety and efficacy issues, we mostly consider the FDA-approved drugs in the predictions of repositioning drugs (Table S4).

Inhibition of Pro-inflammatory Genes
We match the 89 pro-inflammatory genes with drug targets and drugs listed in Table S2, and only list the FDA-approved drugs from DrugBank in Table 3 (see the list of pro-inflammatory genes in Methods). The corresponding FDA-approved drugs include maraviroc (an antiretroviral drug, a CCR5 inhibitor), bosentan (a dual endothelin receptor antagonist that affects both endothelin A and B receptors, used in the treatment of pulmonary artery hypertension), sitaxentan (endothelin A receptor antagonist, used in the treatment of pulmonary artery hypertension), cetuximab (EGFR antagonist, used in several types of cancer) and imiquimod (an immune response modulator, used for skin diseases including skin cancer).
To find the physiological relevance of these pro-inflammatory genes in PAD, we continue to use PubMed to find the relevant references. References in Table 3 support our hypothesis that anti-inflammatory drugs have high potential for repositioning for PAD. Some drugs cannot improve ABI (ankle-pressure index) of PAD patients but can improve the walking ability in patients with critical limb ischemia (CLI), such as ACE inhibitors (Hunter et al., 2013;Shahin et al., 2013). Some genes are indicated as related with PAD, such as C3 (complement component 3) (Fehervari et al., 2014), PTGS2 (prostaglandinendoperoxide synthase 2) (Flórez et al., 2009), SERPINE1 (Björck et al., 2013), S100A12 (Shiotsu et al., 2011), and TNF (Botti et al., 2012;Wozniak et al., 2012;Gardner et al., 2014). Some genes are potential biomarkers or associated with other cardiovascular diseases, such as AGTR1 (angiotensin II receptor, type 1) in coronary occlusive disease (Baños et al., 2011), CCR5 in pulmonary arterial hypertension (Amsellem et al., 2014), LTA (lymphotoxin alpha) in CAD (Topol et al., 2006), and PRKCA (protein kinase C, alpha) in atherosclerosis (Konopatskaya and Poole, 2010). Many of the anti-inflammatory genes in Table 3 are not directly associated with PAD or CAD based on PubMed search, such as ADORA2B (adenosine A2b receptor), EDNRA (endothelin receptor type A), FCER1G (Fc receptor, IgE, high affinity I, gamma polypeptide), STAT5B (signal transducer and activator of transcription 5B), and TLR9 (toll-like receptor 9). In general, the physiological evidence of these anti-inflammatory genes listed in Table 3 strongly supports our hypothesis that inhibition of pro-inflammatory genes is a viable drug repositioning strategy in PAD.

Visualization of Drug-target Network
Graph representation is used to visualize pro-angiogenic and anti-inflammatory repositioning drugs for PAD in Figures 1, 2, respectively. We plot the drug-target networks of the antiangiogenic and pro-inflammatory proteins for the drugs in Tables 2, 3, respectively. We represent the drug target by pink circle and the drug by blue square. Figure 1 shows several compounds targeting the proteins which are annotated as negative regulation of angiogenesis. Figure 2 shows the drugtarget networks of the anti-inflammatory drugs and targets from Table 3. The number of inflammation targets and drugs in Figure 2 is much larger than anti-angiogenic targets and drugs in Figure 1. This gives the insight for the development of clinical trials of anti-inflammatory drugs in PAD in the future. We will discuss the potential clinical trials in Discussion.

Discussion
The clinical trials aimed at stimulating VEGF in PAD and CAD have been unsuccessful (Annex, 2013). The exercise therapy has been demonstrated as the beneficial treatment Frontiers in Pharmacology | www.frontiersin.org

FIGURE 2 | Anti-inflammatory drug-target interaction networks.
for PAD, including walking tolerance, modified inflammatory markers, and adaptation of the limb (e.g., angiogenesis and arteriogenesis) (Haas et al., 2017). Clinical trials with agents targeting angiogenesis and inflammation, other than stimulation of VEGF, should be considered in the future. Below we provide insights for the potential repositioning drugs in PAD identified in this study, including the mechanism of action of these drugs, case studies for several selected drugs in clinical trials, and future experimental validations.

Mechanism of Action of Repositioning Drugs for PAD
Tables 1, 2 provide the anti-angiogenic and pro-inflammatory genes/proteins, drugs targeting these molecules, and physiological evidence for the involvement of these molecules in PAD. However, even though these drug-targets have been identified by our bioinformatics approaches, the mechanism of action of these drugs in PAD and the feasibility of the clinical trials need to be elucidated. Specifically, the effect of some of these drugs to promote angiogenesis in PAD by targeting anti-angiogenic proteins is unknown. Therefore, we search PubMed for the drugs listed in Table 2 using the keywords "(drug name) AND angiogenesis" to understand the mechanism and original use of these putative proangiogenic drugs. We list the drugs with at least one supporting reference found in PubMed in Table 4. These drugs include beta-1 adrenergic receptors blocker (carvedilol, targeting NPPB), vasodilator (isosorbide dinitrate, targeting  Le et al., 2013;Stati et al., 2014 NPR1), and plasminogen activator (alteplase, targeting SERPINE1). We further search PubMed for the anti-inflammatory drugs in Table 3 using the keywords "(drug name) AND inflammation" to elucidate the mechanism and original use of these antiinflammatory drugs ( Table 5). These drugs include antiplatelet (abciximab targeting FCGR1A, acetylsalicylic acid targeting PTGS2), monoclonal antibody (adalimumab targeintg TNFalpha), immune suppressant (alefacept targeting FCGR1A and FCGR2A), ACE inhibitor (benazepril, captopril and enalapril), non-steroidal anti-inflammatory drug (NSAID, e.g., bromfenac, celecoxib, diclofenac, ketorolac, nepafenac, sulindac), and PDE5 inhibitor (tadalafil, vardenafil).

Case Studies of Potential Drug Targets and Drug Repositioning in PAD
We choose three candidate drugs for repositioning in PAD as case studies of our predictions. We selected several drugs that are anti-inflammatory or pro-angiogenic and had no effects on each other. These drugs include bosentan, carvedilol, and maraviroc. We compared the drug targets with the up-regulated genes in the microarray dataset of PAD, including the mouse data of Hazarika et al. (2013), and human microarray studies of Masud et al. (2012), Fu et al. (2008), and Croner et al. (2012.

Case I: Bosentan Targeting EDNRA
The endothelin receptor antagonists (bosentan and ambrisentan) have been approved for use in pulmonary arterial hypertension (PAH) and have been assigned orphan drug status. Details of hepatotoxicity of bosentan, ambrisentan, and sitaxentan are reviewed in de Haro Miralles et al. (2010). Endothelin-1 is a powerful endogenous vasoconstrictor (Frumkin, 2012) and thus blocking endothelin could improve perfusion to the lower extremities in patients with PAD. In a pre-clinical PAD model, Luyt et al. (2000) demonstrated that endothelin, antagonists, bosentan, and darusentan (LU13525) increased tissue blood flow measured by laser Doppler perfusion imaging. de Haro Miralles et al. (2010) examined plasma levels of endothelin and showed that endothelin levels were increased in patients with intermittent claudication compared to non-PAD controls. Just as importantly patients with the most severe form of PAD, CLI, did not demonstrate elevated levels of endothelin, which suggests that an elevation of endothelin is specific to the pathophysiology of intermittent claudication and not all forms of PAD. The original indication of zibotentan was in oncology and pulmonary artery hypertension. The reuse of an endothelin receptor antagonists in PAD patients with intermittent claudication is now in a Phase II clinical trial; the details of the clinical trial of zibotentan are provided in ClinicalTrials.gov https://clinicaltrials.gov/ct2/show/ NCT01890135?term=NCT01890135&rank=1.

Case II: Carvedilol Targeting NPPB
Carvedilol has anti-inflammatory and pro-angiogenic effects in chronic ischemic cardiomyopathy (Le et al., 2013). Carvedilol showed improvement of myocardial flow and reduction of inflammation in the canine model of multivessel cardiomyopathy. The anti-inflammatory cytokine IL-10, which inhibits inflammatory cytokines such as TNFα, IL-1, IL-6, IL-8, and IL-12, was up-regulated in the carvedilol-treated animals. In the PAD microarray data, the inflammatory cytokine IL-8 was up-regulated as found in Masud et al. (2012) and Croner et al. (2012). Though beta-blockers are commonly used in patients with PAD, currently there are no specific clinical trials for carvedilol being compared to placebo or other beta-blockers in PAD patients.

Case III: Maraviroc Targeting CCR5
Maraviroc is an HIV drug targeting CCR5, which is involved in the inflammation pathway (Francisci et al., 2014). Therefore, maraviroc could have anti-inflammatory and anti-atherosclerosis effects, and become a potential repositioning drug in PAD. Croner et al. (2012) show the up-regulation of CCR5 in microarrays from the human femoral artery in PAD. CCR5 inhibitor maraviroc also blocks cell migration and metastasis, but not directly affects the angiogenesis pathway in triple negative breast cancer cell lines . Currently there are no clinical trials for maraviroc in PAD patients.

Limitations of Computational Drug Repositioning Approaches
There are several limitations by the computational approaches to predict the repositioning drugs in PAD. First, PAD is a Treats high blood pressure and heart failure Bai et al., 2012 Frontiers in Pharmacology | www.frontiersin.org complex disease caused by many risk factors and classified by different stages of diseases. Our methods cannot predict the repositioning drugs based on various conditions in PAD patients. Second, the current available clinical trials based on these predicted repositioning drugs in PAD patients are very limited. The available gene expression dataset in human PAD and mouse PAD model is limited. It is difficult to validate our predictions by current clinical trials and available microarray data. Third, the pro-angiogenic and anti-inflammatory drugtarget networks cannot directly link the drugs to PAD based on the current physiological evidence in PAD. The value of the computational drug repositioning might be limited for clinical trial design.

Conclusions
Our study provides comprehensive predictions of potential proangiogenic and anti-inflammatory drugs and drug targets for PAD patients. Based on the protein-protein interaction network PADPIN, we collected the binary relations between FDAapproved drugs and genes annotated in PADPIN. By gathering FDA-approved drugs, these predictions form a basis for further validation and future translational research in PAD.