We developed PS4DR, an automated workflow that enables the integration of multimodal datasets together with pathway information from different canonical pathway databases to predict drug repositioning candidates in different diseases (Figure 1). We showcase PS4DR using real-world gene expression signatures (i.e., Open Targets (Koscielny et. al., 2016) and LINCS) and GWAS data (i.e., GWASdb (Li et al., 2011), GWAS Catalog (Welter et al., 2013), GRASP (Leslie et al., 2014), PheWAS (Denny et al., 2010)). First, the workflow filters disease and drug transcriptomics (i.e., gene expression signatures) with the help of genome-wide association study data. The next step involves calculating pathway signatures for diseases and drugs via pathway enrichment analysis with the filtered dataset. Finally, PS4DR performs an anti-correlation analysis by calculating correlation scores between the pathway signatures of drugs and diseases to prioritize drugs for each disease. Below, we show the utility of the workflow with three applications on how this tool can serve to i) identify drug repositioning candidates, ii) prioritize drug combinations, and iii) propose drugs that simulate disease dysregulations.
Figure 1: An overview of the PS4DR workflow. The workflow requires three different datasets as inputs, (i) disease perturbed gene expression signatures, (ii) genome-wide association study (GWAS) data, and (iii) drug perturbed gene expression signatures. The first and optional part of the workflow involves different filtering steps based on gene set intersection operations that enable the identification of genes in the gene expression signatures that have also been identified in a GWAS of the studied disease. To retain the maximum flexibility in the workflow, users can decide which of the filtering steps want to apply, if any. The next step uses the transcriptomics datasets, filtered or not, to conduct pathway enrichment analysis and evaluate the direction of perturbation for each affected pathway in a particular disease context. While the dotted lines in the figure represent all possible combinations of the filtering steps that can be applied and lead to the pathway enrichment step, solid lines show the option we chose to demonstrate the workflow. Finally, the last step uses the correlation of the pathway scores calculated by the previous step to prioritize drugs that are predicted to invert the pathway signatures observed in a given disease context.
1.1. Identifying Drug Repositioning Candidates
As a first application, we explored the list of 27 diseases for which our workflow predicted drug repositioning candidates. While our workflow predicted plenty of drug candidates, we considered two criteria to prioritize predicted drugs. First, we prioritized all drugs in each disease based on their negative correlation scores. However, a drug could have a negative correlation score by only reverting a minority of the pathways dysregulated in the disease. Therefore, we also consider the relative number of the dysregulated pathways reverted by a drug for the prioritization process. While this prioritization approach facilitated us narrowing down our candidate lists, we are aware that each of the drugs exhibiting negative correlation scores might have the potential to revert the disease condition even if they alter very few dysregulated pathways.
Figure 2 shows the normalized correlation scores of all drugs in all tested diseases. The distribution plots demonstrate correlation scores for all drugs in each disease follows a normal distribution. Next, a threshold of the correlation score smaller than -0.4 and affected pathway (number of disease pathways affected by drugs/total disease pathways) greater than 50% was applied to prioritize the proposed candidate drugs in each disease. This filtering step gives a list of predicted drug candidates for 17 diseases (Additional file 1: Table S1). We further investigated the drugs in some of the conditions to see whether PS4DR was able to identify FDA-approved drugs for their known indications and predict new indications for existing drugs in the prioritized list. We first focused on investigating different types of cancer, since they are the most predominant disease group in the filtered list.
Figure 2: Distribution plots of the correlation scores of the drugs in all the diseases. Plots show the distribution of the Z-score normalized correlation scores of all drugs in each disease. Correlation scores of drugs from 27 diseases are plotted along with the x-axis and density distribution on the y-axis. Drugs with higher negative correlation scores in each disease are of potential interest for repositioning. On the other hand, drugs with positive correlation could be used as a candidate for creating disease cell lines.
First, we looked into the predicted drug list in melanoma. We searched DrugBank (Wishart et. al., 2017) and scientific literature for collecting the evidence for these drugs and summarized it in Table 1. Seven of nine predicted drugs are either already being used as cancer drugs or currently being studied in different clinical trials. Collected evidence suggests that these drugs worth further investigation as the repositioning candidates for the treatment of melanoma. The remaining two drugs, olmesartan and fluspirilene, are used for the treatment of hypertension and schizophrenia respectively and could reveal novel insights into the pathways regulating this condition.
Table 1: Drug repositioning candidates for Melanoma. Drugs showing a negative correlation score less than or equal to -0.40 and affecting more than 50% of the dysregulated pathways in melanoma. The last column outlines the current uses of the given drug in other conditions according to DrugBank and scientific literature.
In breast carcinoma, the only shortlisted drug, AT7519, a selective inhibitor of specific Cyclin-Dependent Kinases (CDKs), is under investigation for the treatment of leukemia, lymphoma, myelodysplastic syndrome, and solid tumors (Wishart et al., 2017) (Additional file 1: Table S1). This is in concordance with the study by Yu et al. (2006) describing how a subgroup of breast cancer patients benefited from the treatment of CDK4 kinase inhibitors. Similarly, we found that three out of five drugs proposed for pancreatic carcinoma are either already being used in different cancer or suggested in the literature, as we discuss below (Additional file 1: Table S1). First drug fenofibrate, an antilipemic agent, was reported to inhibit pancreatic cancer cells proliferation via activation of p53 mediated by upregulation of MEG3 (Hu et al., 2016). Another drug AZD-6482, a selective PI3Kβ inhibitor, could be useful in pancreatic cancer treatment because of its apoptotic effect in cancer cell lines (Xu et al., 2019). Finally, the last drug praziquantel was reported to inhibit cancer cell growth when used synergistically with paclitaxel via downregulating the expression of X-linked inhibitor of apoptosis protein (XIAP) (Wu et al., 2012).
While our workflow showed very promising results in cancer, we wanted to explore the results in complex disorders with no available treatments such as Alzheimer’s disease (AD) and Parkinson’s disease. In the case of AD, the workflow provided a dozen of shortlisted candidates (Table 2). The top drug on the list is rapamycin, an immunosuppressant, already proposed for the treatment of AD by different studies (Spilman et al., 2010; Bové et al., 2011; Cai and Yan 2013). It has been suggested that the therapeutic effect of this drug is due to the reduction of amyloid-beta levels caused by its inhibition of the mTOR signaling pathway (Spilman et al., 2010). Another compound primozide, an antipsychotic agent, was recently suggested as a potential AD therapeutic which was reported to reduce of toxic forms of tau protein by enhanced autophagy activity via AMPK-ULK1 axis stimulation (Kim et al., 2017).
Animal studies have demonstrated that blockade of muscarinic receptors results in increased levels of acetylcholine and improve cognition (Clader and Wang, 2005). Therefore, another proposed drug, terfenadine which is a muscarinic receptor antagonist and has not yet been linked to AD, could be a potential repositioning candidate. Similarly, several 5-HT6R antagonists have advanced to different phases of clinical trials (Benhamú et al., 2014; NCT02258152; NCT02580305) as treatments for AD. The results also suggest another drug in the list, ritanserin, that has not been directly indicated for AD. The high score proposed by our workflow to this serotonin receptor antagonist may be explained by its regulation of the neuronal cholinergic and glutamatergic pathways, both dysregulated in AD. Furthermore, there is increasing evidence showing that neuroinflammation significantly contributes to AD pathogenesis (Lee et al., 2010; Rubio-Perez and Morillas-Ruiz, 2012). Hence, it is not surprising to find two anti-inflammatory agents in our list (i.e., betamethasone and halcinonide) that could be worth investigating as potential repositioning drugs. Finally, doxylamine, a neurotransmitter agent and histamine antagonist, is also a promising candidate since the beneficial effects of histamine antagonists in AD have been reported in multiple studies (Nuutinen and Panula, 2010; Passani and Blandina, 2011; Vohora and Bhowmik, 2012).
Table 2: Drug repositioning candidates for Alzheimer’s disease (AD). Drugs showing a negative correlation score less than or equal to -0.40 and affecting more than 50% of the dysregulated pathways in AD.
1.2. Prioritizing Drug Combinations
Although we have illustrated that our workflow is able to identify candidate compounds for drug repositioning, combining multiple drugs can provide more advantages since the number of affected pathways can be increased by taking advantage of their synergistic effect. Therefore, we applied our workflow to all drug pair combinations in all diseases in order to identify therapies that could have a greater effect than single-drug treatments. For this showcase, we exclusively considered combinations of two drugs for two reasons: i) application of multiple drugs is usually counterproductive since it increases the number of side effects and ii) calculation time increases exponentially as the number of drugs increases.
We investigated our workflow predictions in breast cancer to verify if we have more drugs with a good negative correlation score and affected pathways (%). While we had only 1 drug from our single-drug prediction approach, we were able to retrieve 150 drug pairs from the combination approach with the same thresholds. We increased our threshold of correlation score to less than or equal to -0.50 and affected pathways (%) more than or equal to 80 and still able to manage to retrieve 14 drug pairs. Interestingly, all these drug pairs contain AT-7519 (DB08142), a selective cyclin-dependent kinases (CDKs) inhibitor and under investigation for the treatment of different cancer (Wishart et al., 2017). Further, we found that six drugs partnering with AT-7519 are already being investigated as a potential drug in various cancer. First, selumetinib, a tyrosine kinase inhibitor, is currently being tested in several clinical trials (i.e., NCT03162627; NCT03742102; NCT02503358) for different cancer including breast cancer. Second, AZD-1775, a drug that inhibits the G2–M cell-cycle checkpoint gatekeeper WEE1 kinase, has been used in multiple trials studying the treatment of lymphoma, ovarian cancer, and adult glioblastoma, etc. (Matheson et al., 2016; Wishart et. al., 2017). Third, axitinib, a selective vascular endothelial growth factor receptor (VEGFR) inhibitor, is under investigation in different clinical trials for various cancer (i.e., NCT02129647; NCT03494816; NCT03472560). The fourth drug, TAK-715 is a p38 MAP kinase inhibitor that cross-reacts with casein kinase ɛ (CKIɛ). Since CKIɛ mutations have been linked with the proliferation of different breast cancer cell lines, this drug could be explored to repurpose it for breast cancer treatment (Verkaar et al., 2011). Fifth, raltitrexed, a thymidylate synthase inhibitor, is in several clinical trials for colorectal cancer (Viéitez et al., 2011; NCT03053167; NCT03344614). Finally, megestrol acetate, a progesterone receptor agonist, is under various clinical trials either alone or in combination with other cancer drugs for breast cancer treatment (i.e., NCT03306472, NCT03024580).
Table 3: Drug repositioning candidates from combined drugs strategy for breast cancer. This table lists all drugs that partner with the investigational cancer drug AT-7519 (DB08142) in breast cancer using the drug repositioning combination strategy.
1.3. Proposing Drugs that Simulate Disease Pathway Signatures
While we have initially focused on the drugs with the most negative correlation scores, we also anticipated a potential utility for drugs showing positive correlations. Well-characterized drugs with high positive correlation scores can provide information about how pathways or targets could be implicated in the molecular basis of the disease. Hence, as an extended application, the workflow may be used additionally as a prioritization tool to identify drugs that could be potentially employed to generate in-vitro or in-vivo models. By investigating the correlation scores (Figure 3), researchers can quickly identify drugs that could be used for this purpose. Our workflow predicted induction of disease pathway signatures for pevonedistat in diabetes mellitus, alvocidib in Crohn's disease; entinostat and panobinostat in systemic lupus erythematosus (SLE) through very high positive correlation scores in addition to their broad coverage of affecting disease pathways. We see the need for further investigations of all the drugs with both high positive correlation scores and a high percentage of affected pathways for their use in potential disease model development.
Figure 3: Scatter Plots of the drug’s correlation scores against affected pathways (%) in each disease. The relative number of target pathways by the drug in the disease context (%) is plotted along with the x-axis and correlation scores on the y-axis. Drugs in the top-right corner of the plot are of high interest due to their high positive correlation scores and overlap of pathways for developing potential disease models.