A novel computational approach for drug repurposing using systems biology

Abstract

Motivation

Identification of novel therapeutic effects for existing US Food and Drug Administration (FDA)-approved drugs, drug repurposing, is an approach aimed to dramatically shorten the drug discovery process, which is costly, slow and risky. Several computational approaches use transcriptional data to find potential repurposing candidates. The main hypothesis of such approaches is that if gene expression signature of a particular drug is opposite to the gene expression signature of a disease, that drug may have a potential therapeutic effect on the disease. However, this may not be optimal since it fails to consider the different roles of genes and their dependencies at the system level.

Results

We propose a systems biology approach to discover novel therapeutic roles for established drugs that addresses some of the issues in the current approaches. To do so, we use publicly available drug and disease data to build a drug-disease network by considering all interactions between drug targets and disease-related genes in the context of all known signaling pathways. This network is integrated with gene-expression measurements to identify drugs with new desired therapeutic effects based on a system-level analysis method. We compare the proposed approach with the drug repurposing approach proposed by Sirota et al. on four human diseases: idiopathic pulmonary fibrosis, non-small cell lung cancer, prostate cancer and breast cancer. We evaluate the proposed approach based on its ability to re-discover drugs that are already FDA-approved for a given disease.

Availability and implementation

The R package DrugDiseaseNet is under review for publication in Bioconductor and is available at https://github.com/azampvd/DrugDiseaseNet.

Supplementary information

Supplementary data are available at Bioinformatics online.

1 Introduction

Despite enormous investments in research and developments (R&D), it still takes approximately $800 million to $1 billion and 10–17 years to approve a new drug for clinical use (Adams and Brantner, 2006; Dickson and Gagnon, 2009; DiMasi et al., 2003). More than 90% of drugs fail to pass beyond the early stage of development and toxicity tests, and many of the drugs that go through early phases of the clinical trials fail because of adverse reactions, side effects, or lack of efficiency. Indeed, the rate of failure is still significantly higher than the rate of approval (Booth and Zemmel, 2004; Dickson and Gagnon, 2009; DiMasi et al., 2003). In order to overcome these challenges, drug repurposing, an approach aiming to find new indications for existing drugs (Chong and Sullivan, 2007), has emerged as an important strategy for drug discovery (Ashburn and Thor, 2004). This approach can also rescue drugs that are safe but fail to get to market due to the lack of efficacy against their initial clinical indication (Collins, 2011).

Repurposing approaches can be categorized as drug-based or disease-based. Disease-based approaches are developed to overcome the lack of knowledge about the pharmacology of a drug (Dudley et al., 2011). Drug-based approaches are preferred when drug data (e.g. transcriptomic data) are available. While each one of these approaches faces several challenges, successful repurposing approaches often take advantage of both drug and disease data. In this area, a number of approaches have been developed based on the analysis of transcriptomic data, such as gene expression signatures, defined as the changes in the expression of genes under a certain condition (e.g. administration of a drug, or a disease). Some of these approaches are based on the idea that if there is an anti-correlation between a drug-exposure gene expression signature and a disease gene expression signature, that drug may have a potential therapeutic effect on the disease (Lamb et al., 2006; Sirota et al., 2011). Drugs that are strongly anti-correlated with a disease are likely to be candidates for repurposing. Resources such as LINCS [new version of Connectivity Map (Lamb et al., 2006)] allow for systematic search of candidates for drug repurposing.

The Connectivity Map (CMap) project (Lamb et al., 2006) was the first systematic approach aimed at exploring functional connections between drugs, as well as between drugs and diseases. This project led to the first repository of genome-wide expression data from five human cancer cell lines exposed with 1309 compounds at different dosages, and integrated with other sources such as NCBI Gene Expression Omnibus (GEO). (Lamb et al., 2006) evaluate the similarity of a query signature, that can be a drug-exposure gene expression signature or a disease gene expression signature, to each drug signature in Connectivity Map database (reference data). In (Sirota et al., 2011) the authors developed a systematic approach based on the same idea originally proposed by (Lamb et al., 2006). In this work, they use drug-exposure gene expression signature from Connectivity Map as the reference data and query this reference data with every single disease gene expression signature by applying a pattern-matching method.

Systems biology can be used as an effective platform in drug discovery and development by leveraging the understanding of interactions between the different system components (Butcher et al., 2004; Kitano, 2002). In this paper, we propose a systems biology approach that takes advantage of prior knowledge of drug targets, disease-related genes and signaling pathways to construct a drug-disease network (DDN) composed of the genes that are most likely perturbed by a drug. By performing a system-level analysis on this network using disease gene expression signatures and drug-exposure gene expression signatures, our approach estimates the amount of perturbation caused by a drug on the genes that are associated to a disease of interest. Drugs are ranked based on the amount of perturbation they exercise on specific disease-related genes, and highest ranking drugs are proposed as candidates for repurposing.

We compare the results of our approach with the computational drug-repurposing approach proposed by (Sirota et al., 2011) using 19 datasets involving 4 diseases: idiopathic pulmonary fibrosis (IPF), non-small cell lung cancer (NSCLC), prostate cancer and breast cancer. We show that our approach provides a more accurate prediction based on its ability to identify drugs that are already approved for the disease of interest.

2 Materials and methods

2.1 Data sources

Disease and drug gene expression data. Large scale drug-exposure gene expression data are obtained from two databases: Connectivity Map and the Library of Integrated Network-Based Cellular Signatures (LINCS; Lamb et al., 2006; http://www.lincsproject.org/). Disease expression data are obtained from NCBI Gene Expression Omnibus (GEO; Edgar et al., 2002) and Lung Genomics Research Consortium (http://www.lung-genomics.org).

In Connectivity Map, drug expression data are measured from the exposure of five human cell lines to bioactive small molecules. Differentially expressed genes (DEGs) are identified using a moderated t-test (Smyth, 2005) by comparing treated samples and the corresponding control (untreated) samples. The resulting P-values are FDR adjusted (Benjamini and Yekutieli, 2001) to correct for multiple comparisons.

The LINCS program, the successor of Connectivity Map (Lamb et al., 2006), generated transcriptional gene expression data from cultured human cells exposed to small molecules and knock-down/overexpression of a single gene. The data is also available in GEO (GSE70138). This program provides DEGs in terms of z-score signatures by comparing two groups of samples (treatment versus control). In both Connectivity Map and LINCS, there are often more than one replicate for each drug. Replicates with at least (1%) DEGs (FDR-adjusted P-value < 0.025) are selected. Since measurements are carried out on different platforms, we standardize gene identifiers from chip specific probe identifiers to NCBI GeneID identifiers using the affy package (Gautier et al., 2004). We average across distinct probe expression values when multiple probes mapped to the same NCBI GeneID.

Drug-targets and disease-related genes. The proposed approach needs to construct a network that includes all the shortest paths between the drug targets and genes known to be associated to the disease of interest. Drug targets and disease-related genes (genes associated with the disease of interest) are retrieved from the Comparative Toxicogenomics Database (CTD; Mattingly et al., 2006) and Drugbank (Wishart et al., 2006). CTD is a database that provides curated data describing cross-species chemical-gene/protein interactions and gene-disease associations. Drugs with no known targets are removed from the study. Such drugs are mostly not FDA-approved.

Signaling pathways. We obtain signaling pathways from Kyoto Encyclopedia of Genes Genomics (KEGG; http://www.genome.ad.jp/kegg/). A signaling pathway in KEGG is modeled by a graph in which nodes represent genes or proteins, and directed edges between them represent signals between genes or proteins. The edges are weighted based on the various types of signals, such as activation, inhibition, etc.

2.2 Framework overview

The first part of the framework consists in building the drug-disease network (DDN) by integrating knowledge about the disease-related genes, drug targets and gene-gene interaction knowledge. Then, a repurposing score is computed for each drug-disease pair by integrating expression data into this network. Figure 1 represents the proposed framework.

Fig. 1.

Framework overview. (A) We construct a global network (GN) that is the union of all KEGG human signaling pathways. For each drug-disease pair, we extract a subgraph of GN, namely DDN, consisting of all shortest paths between two sets of disease-related genes and drug targets. (B) We then generate gene perturbation signatures of drug-disease pairs by applying a system-level analysis on their gene expression signatures in the drug-disease network (DDN). A comparative analysis is applied on drug and disease gene perturbation signatures. A repurposing score is assigned to each drug-disease pair. Finally, a ranked list of drugs with potential therapeutic effects for the given disease is generated based on repurposing scores

Drug-disease network (DDN) construction. As shown in Figure 1A, first, we construct a global network (GN) by performing the union of all nodes and edges of KEGG human signaling pathways. In a number of KEGG pathways, a gene ‘a’ interacts with gene ‘b’, through an intermediate pathway ‘A’. This is represented by a link that starts from gene ‘a’ to gene ‘b’ through pathway ‘A’. For example, in the Adherence Junction pathway, TGFβR activates Smad3 through the TGF-beta signaling pathway. Interactions between genes belonging to the pathway ‘A’ and genes ‘a’ and ‘b’ are not included in our model. There are some interactions between genes/pathways through DNA or small molecules in KEGG. For instance, there is a link between MAPK signaling pathway and Phosphatidylinositol signaling system through a small molecule (compound) IP3 in KEGG. Such interactions are not part of the scope of this analysis and we do not include them in constructing the global network. We used ROntoTools package (Voichiţa and Drăghici, 2013; version 1.2.0) to calculate the union all KEGG signaling pathways that are represented by the adjacency matrices and obtain a unified adjacency matrix. In this step, we included some implicit interactions between the genes by performing the union of adjacency matrices representing KEGG signaling pathways. For example, suppose gene ‘a’ activate gene ‘b’ in pathway ‘A’ and gene ‘b’ activates gene ‘c’ in pathway ‘B’. A path between gene ‘a’ and gene ‘c’ may be constructed by our analysis, while there was no path between them before this analysis.

Next, given the two sets of disease-related genes as $Diseas{e}_t=\{x_1,x_2,\dots ,x_n\}$, and drug targets as $Dru{g}_t=\{y_1,y_2,\dots ,y_n\}$, we extract a subgraph of GN that consists of all the shortest paths connecting genes belonging to these sets. It means that a gene from either Disease_t or Drug_t can be a source or destination of the shortest path extracted from GN. This subgraph called Drug-disease network (DDN) represents all the interactions between drug targets and genes related to the given disease, through all the interactions described in KEGG signaling pathways.

Drug-disease repurposing score computation. In this stage, we capture the impact caused by a drug exposure or a disease on the genes that are specific to the condition of interest. In order to integrate the drug and disease gene expressions signatures, we generate gene perturbation signatures by computing the amount of perturbation upon the genes belonging to the drug-disease network (DDN) for all drug-disease pairs, as shown in Figure 1B. The gene perturbation signatures are calculated using the impact analysis method Drǎghici et al. (2007) on the subgraph of global network we constructed in previous step. The impact analysis (IA) takes into account the structure and dynamics of a signaling pathway by considering a number of important aspects, including the measured gene expression changes, the direction and type of every gene signal and the position and role of every gene in a pathway. A perturbation factor for each gene, PF(g_i), is calculated using the impact analysis method Drǎghici et al. (2007), as follows:

A perturbation factor for each gene, PF(g_i), is calculated using the impact analysis method (Drǎghici et al., 2007), as follows:

$$\text{PF}\left(g_i\right)=\Delta E\left(g_i\right)+\sum _{j=1}^n{\beta }_{ij}\frac{\text{PF}\left(g_j\right)}{N_{ds}\left(g_j\right)},$$ (1)

where the term $\Delta E\left(g_i\right)$ denotes the signed normalized measured expression change of a gene g_i, added to the sum of all perturbation factors of the genes g_j that are direct upstream of the gene g_i, normalized by the number of downstream genes of g_j, $N_{ds}\left(g_j\right)$. The coefficient β_ij represents the type of the interaction, ${\beta }_{ij}=1$ for activation and induction and ${\beta }_{ij}=-1$ for inhibition and repression. The second term in Equation (1) involves the PF values of those genes that are upstream of the gene for which the perturbation factor is calculated. For a gene with no upstream genes, the PF will be the measured expression gene $\Delta E\left(g\right)$.

Next, we calculate the repurposing scores for drug-disease pairs by computing the Pearson correlation coefficient between their gene perturbation signatures. The result score is from –1 to 1, where a high positive score shows that the drug and the disease both cause similar perturbations in the system, and therefore, that drug may cause the same effect as the disease. Conversely, a high negative score shows that the drug and disease have opposite gene perturbation signatures. Our hypothesis is that if the perturbation caused by a particular drug in the system is the reverse of the perturbation caused by a disease, that drug may have the potential to treat the given disease. Thus, we rank drugs from the strongly anti-correlated to the strongly correlated, according to their repurposing pathway perturbation scores.

In order to estimate the statistical significance of drug candidate repurposing scores, we generate 1000 random drug gene expression signatures (by permuting gene labels) and then calculate random repurposing scores for all drug-disease pairs. We compute P-values as the percentage of the random scores higher than the observed score.

A systematic method to select repurposing candidates. We used a systematic method in order to rank repurposing candidates. To do this, given a ranked-list of drugs (drug instances) obtained by applying our approach on a disease dataset, we first compute a score for each drug that indicates how better or worse that drug is ranked in comparison to already FDA-approved drugs as follows:

$$\text{score}\left({\text{Drug}}_{\mathrm{x}}\right)=a\hspace{0.17em}–\hspace{0.17em}b,$$ (2)

where a denotes the number of already FDA-approved drugs (gold standards) that are ranked worse than Drug_x, and b denotes the number of FDA-approved drugs that are ranked better than Drug_x (Supplementary Fig. S3). For instance, if there were N FDA-approved drugs for a condition and an instance of a repurposing candidate were ranked higher than all N FDA approved drugs, the score of this candidate would be N. Conversely, if the candidate were ranked lower than all N FDA approved drugs, its score would be-N.

Using this objective measure, we then calculate an average score for each drug across different disease datasets (Supplementary Fig. S4B). Finally, we compute an average score for each distinct drug across different instances, if there are multiple instances for that drug (Supplementary Fig. S4C). We select the top 5% drug candidates from the ranked lists obtained by applying our approach on disease datasets and rank such drugs based on the scores computed by the this method, from highest to the lowest.

3 Results

To validate our approach, we analyzed 19 datasets from four different conditions: idiopathic pulmonary fibrosis (IPF; 6 datasets), non-small cell lung cancer (NSCLC; 4 datasets), prostate cancer (3 datasets) and breast cancer (6 datasets). The results of NSCLC, prostate cancer, and breast cancer are included in Supplementary Material. We compare the results of three computational drug repurposing approaches: our system-level approach, the most popular approach proposed by (Sirota et al., 2011; henceforth drug-disease) and a classical method based on disease and drug signature anti-correlation (henceforth anti-correlation).

Both the drug-disease and the anti-correlation approaches are based on the hypothesis that if gene expression signature is perturbed in one direction in a disease state, and in the opposite (reverse) direction upon a drug exposure, then that drug may have the potential therapeutic effect for the disease. The difference between the two approaches is related on the approach used to calculated the match between a disease and a drug. Given a disease gene expression signature (query signature) and a drug gene expression signatures (reference signature), the Sirota et al.’s drug-disease similarity approach calculate an enrichment score for the up-regulated and down-regulated disease genes [by applying a Kolmogorov–Smirnov (KS) test]. We use the R implementation of this approach available in the package DrugVsDisease (Pacini, 2013).

In contrast, the classical anti-correlation method calculates a similarity score for drug-disease pairs by computing the Pearson correlation coefficient between the drug gene expression signature and the given disease gene expression signature. Drugs are ranked from the highly anti-correlated to the highly correlated, according to their score.

In this study, we compare the various approaches based on their ability to identify drugs that have already been FDA-approved for that condition (gold standard), based exclusively on the molecular data. In essence, a good repurposing approach should place already approved drugs at the very top of the list of drugs proposed for that particular disease. We used the Wilcoxon rank sum test (Wilcoxon, 1945) to determine whether the proposed approach is significantly better than the existing approaches.

Supplementary Table S2 shows the proposed candidates and the preliminary evidence that support the usefulness of those candidates in treatment of four human diseases: IPF, NSCLC, prostate cancer and breast cancer.

Table 1.

Preliminary support by pre-clinical or clinical studies showing the therapeutic potential of the proposed candidates

Disease	Proposed candidate	Pre-clinical/clinical evidence	ClinicalTrials.gov ID
IPF	Sunitinib	(Grimminger et al., 2010; Knoerzer et al., 2013; Richeldi et al., 2011)
	Dabrafenib	(Madala et al., 2012; Olsen et al., 2014; Yoshida et al., 2002)
	Nilotinib	(Abdollahi et al., 2005; Allen and Spiteri, 2001; Antoniades et al., 1990; Cao et al., 2000; Chaudhary et al., 2007; Grimminger et al., 2015; Homma et al., 1995; Rhee et al., 2011)
NSCLC	Sunitinib	(Novello et al., 2011; Socinski et al., 2006)	NCT00372775, NCT00092001, NCT00864721, NCT00693992
	Sirolimus	(Boffa et al., 2004; Ekman et al., 2012; Gridelli et al., 2008; Sarkaria et al., 2007)	NCT00923273
	Everolimus	(Gridelli et al., 2008; Price et al., 2010; Soria et al., 2009)	NCT01061788
	Ponatinib	(Chao et al., 2012; Chase et al., 2013; Gainor and Shaw, 2013; Gautschi et al., 2017; Ren et al., 2013; Thomas et al., 2015)	NCT01813734
Prostate cancer	Podophyllotoxin	(Benzina et al., 2015; Chen et al., 2013; Choi et al., 2015; Gordaliza et al., 2000; Hu et al., 2016; Kim et al., 2013; Ma et al., 2016)
	Acetylsalicylic acid	(Bilani et al., 2017; Choe et al., 2012; Habel et al., 2002; Jacobs et al., 2005; Ma and Brusselaers, 2017; Nelson and Harris, 2000; Rothwell et al., 2011; Skriver et al., 2016)	NCT02757365, NCT03103152, NCT02804815
	Papaverine	(Goto et al., 1999; Huang et al., 2017; Savai et al., 2010)
	Mefloquine	(Geng et al., 2010; Sharma et al., 2012; Yan et al., 2013b)
	Vorinostat	(Bradley et al., 2009; Carter et al., 2016; Detchokul and Frauman, 2011; Kaushik et al., 2015)	NCT00330161, NCT00589472
	Sirolimus	(Amato et al., 2008; Chamie et al., 2008; Imrali et al., 2016; Rai et al., 2010)	NCT00311623, NCT02565901
Breast cancer	Captopril	(Jones et al., 2004; Kast and Halatsch, 2012; Krusche et al., 2013; Namazi et al., 2014; Rundle-Thiele et al., 2016; Rooprai et al., 2001)	NCT00086723
	Glibenclimiade	(Abdul and Hoosein, 2002; Núñez et al., 2013; Payen et al., 2001; Qian et al., 2008; Rong et al., 2013; Yasukagawa et al., 2012)
	Fluorometholone	(Karmakar et al., 2013; Keith, 2008; Lin and Wang, 2016)
	Etoposide	(Atienza et al., 1995; Yuan et al., 2012, 2015)	NCT01492556, NCT00026949, NCT01589159
	Colchicine	(Sun et al., 2016)
	Tretinoin	(Budd et al., 1998; Garattini et al., 2012; Lu et al., 2005; Paroni et al., 2012; Sutton et al., 1997)

Note: These candidates are currently FDA-approved but for other indications.

3.1 Drug repurposing using IPF data

The list of IPF datasets we used in our analysis is summarized in Supplementary Table S3. We compare the results of our approach with the existing approach proposed by (Sirota et al., 2011; drug-disease), as well as the classical method (anti-correlation). The lists of the top 10 drugs are summarized in Table 1.

Table 1.

A comparison between the results of three approaches: proposed, drug-disease, anti-correlation using IPF datasets (the top 10 drugs)

Proposed	Drug-disease	Anti-correlation	Proposed	Drug-disease	Anti-correlation
GSE24206-early			GSE24206-advanced
GSM1740570_saracatinib	GSM1746916_radicicol	GSM1746916_radicicol	GSM1740570_saracatinib	GSM1746916_radicicol	GSM1746916_radicicol
GSM1743214_nintedanib	GSM1746864_radicicol	GSM1746864_radicicol	GSM1743214_nintedanib	GSM1746864_radicicol	GSM1746864_radicicol
GSM1742836_celastrol	GSM1738326_mocetinostata	GSM1746893_radicicol	GSM1745714_buparlisib	GSM1738291_azacitidinea	GSM1746893_radicicol
GSM1745714_buparlisib	GSM1738291_azacitidinea	GSM1742836_celastrol	GSM1742836_celastrol	GSM1746893_radicicol	GSM1738290_azacitidinea
GSM1742552_linifanib	GSM1738794_garcinola	GSM1738290_azacitidinea	GSM1742552_linifanib	GSM1738772_ischemin	GSM1742836_celastrol
GSM1742850_nintedanib	GSM1745213_nilotiniba	GSM1738291_azacitidinea	GSM1740917_saracatinib	GSM1742836_celastrol	GSM1738291_azacitidinea
GSM1740917_saracatinib	GSM1737397_quizartiniba	GSM1745530_nilotiniba	GSM1739549_saracatinib	GSM1737397_quizartiniba	GSM1742552_linifanib
GSM1740731_CH5424802	GSM1743996_sirolimusa	GSM1745213_nilotiniba	GSM1742850_nintedanib	GSM1738326_mocetinostata	GSM1743996_sirolimusa
GSM1743268_linifanib	GSM1742836_celastrol	GSM1742552_linifanib	GSM1740731_CH5424802	GSM1742718_sorafeniba	GSM1742716_sorafeniba
GSM1739549_saracatinib	GSM1746893_radicicol	GSM1742716_sorafeniba	GSM1745213_nilotiniba	GSM1746811_ruxolitiniba	GSM1745530_nilotiniba
GSE44723			GSE21369
GSM1741104_sunitiniba	GSM1737411_NVP-BGT226	GSM1737409_NVP-BGT226	GSM1737700_rucapariba	GSM1737352_everolimusa	GSM1737700_rucapariba
GSM1740080_sunitiniba	GSM1743823_fostamatiniba	GSM1737411_NVP-BGT226	GSM1740570_saracatinib	GSM1737699_rucapariba	GSM1737699_rucapariba
GSM1742552_linifanib	GSM1745509_NVP-BEZ235	GSM1740923_BI-2536a	GSM1745714_buparlisib	GSM1737700_rucapariba	GSM1738767_decitabinea
GSM1744393_gefitiniba	GSM1737409_NVP-BGT226	GSM1740576_BI-2536a	GSM1740731_CH5424802	GSM1738767_decitabinea	GSM1740731_CH5424802
GSM1737353_everolimusa	GSM1738100_tranylcyprominea	GSM1745509_NVP-BEZ235	GSM1745213_nilotiniba	GSM1746916_radicicol	GSM1737385_motesaniba
GSM1742436_GDC-0941	GSM1741779_vorinostata	GSM1740570_saracatinib	GSM1742504_nintedanib	GSM1737385_motesaniba	GSM1737624_entinostat
GSM1743268_linifanib	GSM1742856_canertiniba	GSM1737412_NVP-BGT226	GSM1737448_idelalisiba	GSM1742800_palbocicliba	GSM1744048_imatiniba
GSM1743214_nintedanib	GSM1742795_palbocicliba	GSM1737353_everolimusa	GSM1745530_nilotiniba	GSM1737990_mocetinostata	GSM1737698_rucapariba
GSM1741743_sirolimusa	GSM1739241_olapariba	GSM1745194_NVP-BEZ235	GSM1742836_celastrol	GSM1737443_idelalisiba	GSM1738290_azacitidinea
GSM1740570_saracatinib	GSM1740387_CH5424802	GSM1738308_entinostat	GSM1739679_dabrafeniba	GSM1742836_celastrol	GSM1741754_sirolimusa
LGRC-ILD			GSE1724
GSM1740570_saracatinib	GSM1738326_mocetinostata	GSM1738326_mocetinostata	GSM1741104_sunitiniba	GSM1746800_nilotiniba	GSM1746916_radicicol
GSM1739549_saracatinib	GSM1742795_palbocicliba	GSM1737624_entinostat	GSM1740570_saracatinib	GSM1746916_radicicol	GSM1742552_linifanib
GSM1740917_saracatinib	GSM1737410_NVP-BGT226	GSM1739358_mocetinostata	GSM1739549_saracatinib	GSM1743953_linifanib	GSM1742836_celastrol
GSM1743214_nintedanib	GSM1741779_vorinostata	GSM1739435_belinostata	GSM1742552_linifanib	GSM1742836_celastrol	GSM1745213_nilotiniba
GSM1743268_linifanib	GSM1737624_entinostat	GSM1741779_vorinostata	GSM1745714_buparlisib	GSM1745958_dasatiniba	GSM1743268_linifanib
GSM1742552_linifanib	GSM1737411_NVP-BGT226	GSM1737410_NVP-BGT226	GSM1743268_linifanib	GSM1743268_linifanib	GSM1746864_radicicol
GSM1742504_nintedanib	GSM1739435_belinostata	GSM1746864_radicicol	GSM1740917_saracatinib	GSM1741215_velipariba	GSM1745626_vemurafenib
GSM1744170_GDC-0941	GSM1737409_NVP-BGT226	GSM1741767_vorinostata	GSM1740080_sunitiniba	GSM1741184_regorafeniba	GSM1743197_celastrol
GSM1745714_buparlisib	GSM1737455_vandetaniba	GSM1737409_NVP-BGT226	GSM1742706_alvocidib	GSM1747067_mitoxantronea	GSM1744371_selumetinib
GSM1741265_saracatinib	GSM1738350_pracinostat	GSM1737642_mocetinostata	GSM1742850_nintedanib	GSM1741566_velipariba	GSM1746881_mitoxantronea

Note: The P-values for Wilcoxon rank sum test comparing the results of the proposed approach and drug-disease approach using datasets GSE24206-early, GSE24206-advanced, GSE44723, GSE21369, LGRC- ILD and GSE1724 are 0.02, 0.02, 0.01, 0.02, 0.01 and 0.01, respectively. Highlighted drugs are FDA-approved for the treatment of IPF. The proposed approach was the only one that was able to rank the FDA-approved Nintedanib in the top 10. In contrast, none of the existing approaches was able to retrieve the FDA-approved drug in any of these six datasets.

^aDrugs that are currently FDA-approved but for other indications.

Gold standard: The gold standard for this disease is Nintedanib. This drug was approved for the treatment of IPF by FDA on October 2014. It is a small molecule inhibiting multiple tyrosine kinases (RTKs) and non-receptor tyrosine kinases (nRTKs). It is highlighted in Table 1.

As shown in Supplementary Table S4, We select the top 5% of drugs ranked lists obtained by applying our approach on 6 IPF datasets. We rank these drugs based on the scores computed by the systematic method from the highest to the lowest.

Proposed candidates: We propose Sunitinib (P = 0.0009), Dabrafenib (P = 0.0009) and Nilotinib (P = 0.0009) as repurposing candidates for treatment of IPF. Saracatinib, Linifanib, Buparlisib, GDC-0941 and Alvocidib are also highly ranked by our approach for treatment of IPF. Although these drugs are not approved by FDA yet, they can be considered for further experimental tests.

Sunitinib is a small molecule that inhibits multiple receptor tyrosine kinases (RTKs), including vascular endothelial growth factor receptors (VEGFR) and platelet-derived growth factor receptors (PDGFR). It is approved by FDA for the treatment of Gastrointestinal stromal tumor, advanced renal cell carcinoma and progressive well-differentiated pancreatic neuroendocrine tumors (Demetri et al., 2006; Motzer et al., 2007). It was investigated for its anti-fibrotic and anti-angiogenic properties. Its efficiency was experimentally proved in a bleomycin-induced mouse model and it has been proposed for the treatment of IPF (Knoerzer et al., 2013). Results of in vitro studies and animal models show that receptor tyrosine kinases, such as PDGFR, VEGFR and FGFR, and non-receptor tyrosine kinases, such as the Src family, play crucial roles in the pathogenesis of IPF (Grimminger et al., 2010; Richeldi et al., 2011).

Dabrafenib is approved by FDA for the treatment of patients with unresectable or metastatic melanoma. Recent clinical studies demonstrate that the extracellular signal regulated kinase (ERK) and mitogen-activated protein kinase (MAPK) are up-regulated in lung tissues of patients with IPF (Madala et al., 2012; Yoshida et al., 2002). In particular, results of studies on MAPK signaling pathways show that the level of serine/threonine-protein kinase B-Raf (BRAF) is increased in patients samples compared to the normal ones, suggesting the potential therapeutic effects of MEK/ERK inhibitors for pulmonary fibrosis (Madala et al., 2012; Olsen et al., 2014). This supports the idea that the BRAF inhibitor Dabrafenib may have atherapeutic effect on IPF.

Nilotinib is another FDA-approved drug we propose to be repurposed for the treatment of IPF. Nilotinib is a transduction inhibitor targeting BCR-ABL, c-kit and PDGF, that is approved by FDA for treatment of patients who are newly diagnosed with Philadelphia chromosome positive chronic myeloid leukemia (Ph + CML). It is also used for treatment of patients with Ph + CML in chronic phase and accelerated phase if they were resistant (or intolerant) to previous treatments. The potential roles of PDGFs in IPF have been shown by many studies (Antoniades et al., 1990; Allen and Spiteri, 2001; Cao et al., 2000; Homma et al., 1995; Wollin et al., 2015). The advantage of PDGF inhibition in IPF is well studied and supported by several studies (Abdollahi et al., 2005; Chaudhary et al., 2007; Wollin et al., 2015). Authors of (Grimminger et al., 2015; Rhee et al., 2011) confirmed the potential effect of Nilotinib in decreasing the extent of pulmonary fibrosis in a mouse model.

The phosphatidylinositol 3 kinase (PI3K) inhibitors Buparlisib and GDC-0941 are undergoing clinical trials for a number of diseases. Buparlisib is in Phase III of clinical trials for treatment of breast cancer and in and Phase II for several other solid tumors. GDC-0941(Pictilisib) has been used in clinical trials for the treatment of several cancers, including breast cancer. Preclinical studies proved that PI3K inhibitors have potential roles in treatment of IPF by interfering with the fibrogenic effects of $TGF-\beta 1$ signaling (Beyer and Distler, 2013; Conte et al., 2013; Hsu et al., 2017; Mercer et al., 2016). Based on this evidence, Buparlisib and GDC-0941 may have potential therapeutic effects on IPF.

The tyrosine kinase inhibitors Saracatinib and Linifanib are also highly ranked by our approach for treatment of IPF. Saracatinib (AZD0530) is an oral, tyrosine kinase inhibitor selective for Src. It underwent clinical tests at AstraZeneca for the treatment of cancer (Gucalp et al., 2011; Lara et al., 2009; Messersmith et al., 2010; Poole et al., 2010). However, it failed to show a sufficient efficacy in these studies. Subsequently, it was proposed for other usages such as Alzheimer’s disease (in Phase II; Nygaard et al., 2015). Linifanib (ABT-869) is also a multi-targeted receptor tyrosine kinase inhibitor that is intended to suppress tumor growth. It is investigated for treatment of leukemia (myeloid), myelodysplastic syndrome and solid tumors (Chen et al., 2016; Chiu et al., 2013; Wang et al., 2012). The efficiency and tolerability of Linifanib versus Sorafenib has been assessed in patients with advanced hepatocellular carcinoma (Cainap et al., 2013). The tyrosine kinase inhibitors are proven to be effective in treatment of IPF (Adamali and Maher, 2012; Beyer and Distler, 2013; Grimminger et al., 2015; Richeldi et al., 2011; Wollin et al., 2014). In particular, the Src kinase inhibitor Saracatinib is reported to be useful in treatment of IPF through targeting the $TGF-\beta$ signaling pathway (Hu et al., 2014). These represent additional and independent evidence supporting out findings that Linifanib and Saracatinib are expected to be useful in the treatment of IPF.

Alvocidib is a cyclin-dependent kinase (CDK) inhibitor that is undergoing clinical trials for a number of cancers: esophageal cancer, leukemia, lung cancer, liver cancer and lymphoma. Studies of murine models show that the CDK inhibitors block the epithelial apoptosis and decrease the tissue fibrosis in pulmonary fibrosis (Inoshima et al., 2004; Leitch et al., 2009). As a result, CDK inhibitors have been suggested as a novel therapeutic strategy against IPF (Zhou et al., 2014).

4 Conclusion

In this paper, we presented a systems biology approach to discover new uses of existing FDA-approved drugs. We take advantage of known knowledge of disease-related genes, drug targets information and signaling pathways to discover drugs with the potential desired effects on the given disease. We estimate a network of genes potentially perturbed by drugs and integrate this network with drug and disease gene expression signatures to conduct a more powerful analysis at system level. To evaluate the proposed approach for drug repurposing, we analyzed four different diseases (IPF, NSCLC, prostate cancer and breast cancer) using three approaches: proposed, drug-disease and anti-correlation. For each disease, there is at least one FDA-approved drug that is used to treat that disease in our input drug data. The already FDA-approved drugs for a given disease are considered as the gold standard because such drugs successfully passed all the pre-clinical and clinical trials for that disease and were demonstrated to be efficacious in each disease. The approach is validated by its ability to identify drugs that are already approved by FDA for these conditions. The proposed approach was able to find such drugs based on the molecular profile alone, while existing repurposing approaches failed to do so. Specific drugs have been identified as repurposing candidates for the four diseases studied here. Although the proposed approach is studied in the context of drug repurposing, it also can be used to identify novel targets for FDA-approved drugs and understanding their mechanism of action.