Identification of novel therapeutics for complex diseases from genome-wide association data

The development of new therapeutics is essential to improve the human condition and lower the burden of disease. Due to our limited knowledge of the molecular basis of complex diseases, comparatively few gene targets for therapeutics have been identified to date. The standard approach to developing therapeutics involves testing many thousands of compounds against a known target in order to identify a lead compound. The lead compound can then be further refined in silico and in vitro before heading into the lengthy and costly clinical trials pipeline. This process, which consists of phases I, II, III and IV before final drug approval, involves 10-17 years of drug development, from target identification until FDA/EMEA approval, with only a 10% probability of success [1]. As a result, the pharmaceutical industry spends an average of about 1.2 billion US dollars to bring each new drug to market [2]. There is also a high risk associated with de novo drugs due to unforeseen adverse side effects, as seen in the case of Thalidomide, a drug used to treat morning sickness which resulted in devastating birth defects [3].

A novel approach to therapeutic development is to identify new applications for drugs that have already been approved, or have successfully completed phase I clinical trials which investigate toxicity [4, 5]. This process of "drug repositioning" aims not to develop drugs de novo, but associate existing therapeutics with new phenotypes. Here, we attempted to reposition existing drugs to treat common complex diseases using recently acquired Genome-Wide Association Study (GWAS) data.

Complex diseases are genetically intricate, polygenic and multifactorial [6]; and frequently arise as a consequence of interaction between genes and the environment. Recently, GWAS have begun to unravel the complicated genetic basis of complex diseases. Sheer statistical power has allowed GWAS to successfully identify some associations between Single Nucleotide Polymorphisms (SNPs) and complex diseases [7]. Despite high investment, far fewer genes have been identified than can account for the heritable component of complex diseases, and the clinical benefit remains limited to date [8]. A factor that contributes to the missing heritability is likely to be noisy genotype-phenotype association signals [9]. Also, analysis of GWAS data using highly stringent thresholds for statistical significance, by testing multiple isolated SNPs, has limited the scope of gene discovery based on existing data [10]. As shown in Manhattan plots, GWAS data obviously contain far more information than the most significant peaks, and more work needs to be done extracting data from slightly less significant peaks [9, 11].

Currently available gene discovery platforms can enhance candidate gene identification from GWAS data [9]. Candidate gene prediction tools are designed to find a needle in the genetic haystack. These tools are based on the assumption that genes with similar or related functions cause similar phenotypes [12]. Specific candidate gene prediction tools differ in the strategy adopted for calculating similarity, and the databases utilized for prediction [13, 14]. Gentrepid is one of the many bioinformatic tools developed to help geneticists predict and prioritize candidate genes [9, 15]. The Gentrepid tool and its knowledge base utilizes two independent methods: Common Pathway Scanning (CPS), a systems biology approach; and Common Module Profiling (CMP), a domain-based homology recognition approach, to prioritize candidate genes for human inherited disorders (see Methods for details). Compared to other prediction systems, Gentrepid is designed to make fewer, more conservative predictions which do not extensively extrapolate existing bioinformatic data i.e. it tends to be more specific than other systems [15].

We have previously developed protocols to analyze GWAS data using a multilocus approach which combines bioinformatic and genetic data [9, 16, 17]. To demonstrate the usefulness of these protocols, we reanalysed the well-studied Wellcome Trust Case-

Control Consortium (WTCCC) data for seven complex diseases [9]. Using a series of increasingly less conservative statistical thresholds, we attempted to discriminate the signal from the noise in the more statistically significant data (p ≤ 10-5, p ≤ 10-4, p ≤ 10-3). By incorporating bioinformatic data, we were able to predict 1,497 candidate genes for the seven complex diseases studied; namely, Type 2 Diabetes (T2D), Bipolar Disorder (BD), Crohn's Disease (CD), Hypertension (HT), Type 1 Diabetes (T1D), Coronary Artery Disease (CAD) and Rheumatoid Arthritis (RA) [9].

Here, we extend this pipeline to identify potential novel drug targets among the predicted candidate genes by associating drug information extracted from publicly available drug databases. The three databases sourced in this study were DrugBank [18], the Pharmacogenomics Knowledgebase (PharmGKB) [19] and the Therapeutic Target Database (TTD) [20]. The feasibility of this approach is again illustrated for the seven complex diseases investigated by the WTCCC [11]. This study shows that it is possible to identify therapeutics for treatment of specific complex diseases from genetic loci via the Gentrepid candidate gene prediction tool. Thus, in combination with drug target information, candidate gene prediction systems can be utilized as drug discovery tools to identify therapeutics which may be repositioned as novel treatments for complex diseases.

We implemented a computational workflow to enable repositioning of drugs by using Gentrepid as a bioinformatic candidate gene discovery platform, with drug data sourced from online databases (Figure 1). The two data sets integrated were:

1. A candidate gene data set obtained by integration of genotype-phenotype data from the WTCCC GWAS study on seven complex phenotypes [11], with bioinformatic data on structural domains and systems biology: identifying proteins that share common features, or participate in the same complex or pathway [21];

2. A drug-gene target association data set obtained from three drug databases namely TTD, DrugBank and PharmGKB [18–20].

In previous work, we predicted a total of 1,497 candidate genes for seven complex diseases by careful reanalysis of the WTCCC GWAS data [11] using the Gentrepid candidate gene prediction system [9].

In the original analysis, a highly stringent significance threshold (p ≤ 5x10^-7) was used in an attempt to correct for multiple testing [11]. This conservative statistical approach, combined with the selection of the nearest-neighboring gene to the significant SNP, resulted in identification of only a small number of loci associated with each phenotype, with modest cumulative heritability [9] (Additional file 1, Table S1).

We specifically addressed these two issues in our reanalysis of this noisy data by - (a) Considering a series of four thresholds of decreasing stringency, starting with the highly significant threshold used in the original study, and decreasing to weakly significant(WS - p ≤ 10^-3). This resulted in a series of four SNP sets containing up to 1064 SNPs being considered for each phenotype [9]. The number of loci and SNPs considered in the four data sets for each phenotype is shown in Table S1 (Additional file 1).

(b) Creating six different search spaces around each SNP-based locus, three of fixed-widths and three proximity-based, for analysis by our candidate gene prediction system [9].

Thus, for each of the seven phenotypes, twenty-four search spaces were constructed; using four SNP significance thresholds to obtain the loci, and six gene selection methods to construct the gene search spaces. In total, 168 search spaces ranging in size from 2 to 4,431 genes (up to 10% of the genome) were analyzed [9].

Gentrepid uses two modules: Common Pathway Scanning (CPS) and Common Module Profiling (CMP) to make candidate gene predictions.

The CPS module is based on the assumption that common phenotypes are associated with proteins that participate in the same protein complex or biochemical pathway [22]. Such systems biology methods are currently favored in candidate gene prediction because of the attractiveness of their basic thesis. Their weakness is the lack of coverage of the underlying systems biology knowledge bases [21]. Many tools attempt to ameliorate the deficits of the human systems biology knowledge base by extensive extrapolation of data from other species. Examples are GeneSeeker, ToppGene and Endeavour [13, 23–25]. Gentrepid CPS uses only human data to reduce the number of predicted false positives i.e. it makes fewer predictions which are more often correct compared to other prediction systems [15].

The other module, CMP, is a novel sequence analysis approach based on the principle that candidate genes have similar functions to disease genes already determined for the phenotype [26]. Gentrepid CMP differs from most candidate gene prediction systems which describe functional similarity via keywords, a procedure which also lacks good coverage of the human genome [21]. In CMP, sequences are parsed at the domain level, linking them directly to function [21]. Although CMP's performance was disappointing in our original benchmark using a set of nine oligogenic diseases with Mendelian inheritance [12], it produced a surprising number of statistically significant results when confronted with the GWAS data on seven complex diseases [9]. This result was robust when compared with simulations using random SNPs, and may arise from an important underlying role for homologous genes in complex diseases.

We compiled the drug-gene target data set from three publicly available drug databases: DrugBank [18], PharmGKB [19] and TTD [20]. Snapshots of these databases were taken in June 2012.

DrugBank is a freely available online database that combines detailed drug data and indication information with comprehensive drug-target associations [18]. From this database, we retrieved Drugbank IDs and drug names (generic and brand) to represent drugs, and the unique gene symbols to represent protein targets. We extracted 6,711 drug entries active against 3,410 unique drug targets from several species. We used the G-profiler conversion tool to separate human drug targets represented by official HUGO gene symbols [27], yielding 2,022 human drug targets associated with 3,910 drugs.

The Pharmacogenomics Knowledgebase (PharmGKB) is a drug knowledge base maintained by Stanford University, USA and funded by the US National Institute of Health (NIH). PharmGKB captures information about drugs, diseases/phenotypes and targeted genes [19]. From this database, we extracted the "drug-associated genes" field along with "description" which contains the disease information. This database contains around 3,097 drugs and 26,961 human genes, but not all these genes are associated with drugs. We retrieved 382 drugs for 566 human drug targets. For the PharmGKB database, the number of drug targets exceeds the number of drugs because some drugs target multiple genes.

The Therapeutic Target Database (TTD) is also a freely available online drug database which integrates drug data with therapeutic targets [20]. This database contains 17,816 drugs (approved, clinical and experimental) and 2,025 human and non-human (bacterial and fungal) drug targets. It describes synonyms of 3,167 drug names. We extracted "Drug names" along with "Disease" information, and "Uniprot accession numbers" for targets. UniProt accession numbers were replaced with official HUGO gene symbols using the G-profiler conversion tool [27]. Finally, we extracted 2,960 drugs for 544 unique human drug targets from TTD.

We identified potential therapeutic targets for the seven complex diseases from the Gentrepid predicted candidate genes generated by our reanalysis of the WTCCC data. In total, Gentrepid predicted 1,497 candidate genes for all seven diseases; comprising by phenotype: Type 2 Diabetes (291), Bipolar Disorder (212), Crohn's Disease (378), Hypertension (219), Type 1 Diabetes (358), Coronary Artery Disease (264) and Rheumatoid Arthritis (200) (Additional file 1 Table S1) [9]. We searched for these candidate genes in the drug-gene target files obtained from all three drug databases and found 452 potential therapeutic targets for the seven complex diseases (Table 1). This illustrates that almost 30% of the total number of predicted candidate genes by Gentrepid are potential targets for therapeutic treatments using currently available drugs (Figure 5). The disparity between the 8% of the human genome that is targettable (2,494 extracted targets - Figure 2) and the 30% of predicted candidate genes that are targettable (452 predicted targets - Figure 5) is interesting and should be investigated further. The enrichment of druggable targets in the candidate gene set might be a selection effect: either at the SNP level; or at the knowledgebase level: it might suggest that we already know more about disease genes than the genome in general. Alternatively, it has been previously suggested that the genome can be partitioned into "disease genes" and "non-disease genes". While such a Boolean distribution is likely to be overly simplistic, a spectrum of levels of disease association with specific gene subsets might explain this disparity.

PH	≠ TT	TI	RN	RTT	NTT	NV	RN
T2D	84	0.29	5th	7	77	0.92	5th
T1D	97	0.27	6th	2	95	0.98	2nd
RA	77	0.38	2nd	6	71	0.92	5th
HT	78	0.35	4th	5	73	0.94	4th
BD	59	0.27	6th	1	58	0.98	2nd
CD	135	0.36	3rd	0	135	1.00	1st
CAD	102	0.39	1st	4	98	0.96	3rd

Abbreviations - PH - Phenotypes; ≠TT - Number of Therapeutic Targets; TI - Targetability Index; NTT - Novel Therapeutic Targets; RTT - Replicated Therapeutic Targets; NV - Novelty; RN - Rank; T2D - Type 2 Diabetes; BD - Bipolar Disorder; CD - Crohn's Disease; HT - Hypertension; T1D - Type 1 Diabetes; CAD - Coronary Artery Disease; RA - Rheumatoid Arthritis.

Description of therapeutic targets and novel therapeutic targets. Total 452 unique therapeutic targets and total 428 unique novel therapeutic targets obtained for seven complex diseas

To drill a little further into the data, we assessed the therapeutic potential of each phenotype using currently available repositioned drugs. We calculated an empirical Targetability Index (TI), defined here as the ratio of the number of predicted targets to the number of predicted candidate genes for each phenotype (Table 1). The distribution was bimodal with four phenotypes (CAD > RA > CD > HT) being more targetable (TI = 0.35-0.39) than the other three (T2D > T1D ~ BD) (TI = 0.27-0.29). A factor which is likely to influence the targetability is our underlying knowledge of the phenotype. If the molecular pathways involved have been previously characterized, there is more likely to be drug-target information in the existing drug databases, even if the phenotype has not previously been associated with the molecular system. The low TIs for BD (0.28) and the diabetes phenotypes (0.27-0.29) likely arises from lack of knowledge of underlying pathways. More basic research in this area is required.

All three drug databases made significant contributions to target identification, with the highest contribution from DrugBank (400), followed by TTD (156) and PharmGKB (61). DrugBank is a chemical as well as a clinical drug database which contains broader coverage of drug targets and broader depth of information compared to the chemical drug database TTD and the clinical drug database PharmGKB. PharmGKB, being a clinical drug database, has a lower coverage of drug-target associations, but broader depth of information compared to TTD. To summarize, the total coverage of the predicted targets from all three databases was estimated to be 30% of the candidate genes predicted by Gentrepid, with the maximum contribution from DrugBank (Figure 5).

To identify novel drugs, we compared our phenotype of interest (from the pool of seven diseases considered in our study) with indications associated with the drug. In total, we retrieved 7,252 drugs associated with human drug targets from all three drug databases. We found 2,192 (30% of the extracted drugs) unique drugs that target the 452 potential therapeutic targets.

We retrieved the maximum number of drugs from DrugBank (1,618) and the remainder from TTD (735) and PharmGKB (91). In order to identify the novel drugs i.e. drugs not targeting our phenotype of interest, we filtered the above list of 2,192 drugs to retrieve 2,130 novel therapeutics. On a phenotype by phenotype basis, T1D and CAD had the maximum number of novel predicted therapeutics. Although CD had the highest number of novel targets, it had comparatively few novel therapeutics suggesting new drug development is needed for this phenotype. BD had the fewest therapeutics as expected based on the small number of predicted therapeutic targets. We found that the total percentage of drugs that may be repositioned towards identified novel targets was around 29% of the total number of extracted drugs.

Table 2 shows the 24 replicated targets with examples of replicated drugs found in our study. For example, the drug "Aleglitazar" is in phase III clinical trial for the T2D target PPARA, a predicted candidate gene for T2D. "Rosiglitazone" known to target PPARG as a therapeutic for diabetes mellitus, has a potential use in the related phenotype T1D.

FDA-approved and clinical targets

Identification of therapeutic targets targeted by approved and clinical trial drugs can help us to prioritize drugs for repositioning against phenotypes of interest. Both approved and clinical targets are potential drug targets, however, approved targets will undoubtedly be on the priority list for further experimental studies. We classified the predicted targets as FDA-approved and clinical targets for the seven complex diseases. An example depicted in Figure 7 shows comparison between T2D targets from the TTD database and targets predicted by Gentrepid for T2D. Of the 84 targets predicted for T2D by Gentrepid (Table 1), 28 are listed in TTD (Figure 7). Comparing these 28 targets with the 32 targets indicated for T2D in TTD, we found products of three genes (HSD11B1, PPARA, NR3C1) are targeted by drugs currently in clinical trials for T2D. In addition, PPARA is already targeted by FDA-approved drugs. Hence, we predicted 25 novel therapeutic targets from the TTD database for T2D. In total for the seven diseases, we found 291 approved therapeutic targets and 95% of these as novel approved targets. We also found 334 targets in clinical trials and 96% of these being novel (Table 4). To summarize, both approved and clinical novel targets are associated with therapeutics, which may be repositioned as novel treatments towards the cure of complex diseases.

PH	AT	NAT	CT	NCT
T2D	45	41	65	62
T1D	57	55	73	72
HT	71	68	43	40
RA	55	53	59	54
CD	93	93	135	135
CAD	63	61	80	76
BD	37	36	44	44
Unique sum	291	277	334	318

Abbreviations - PH - Phenotypes; AT - Approved Targets; NAT - Novel Approved Targets; CT - Clinical Targets; NCT - Novel Clinical Targets; T2D - Type 2 Diabetes; T1D - Type 1 Diabetes; HT - Hypertension; RA - Rheumatoid Arthritis; CD - Crohn's Disease; CAD - Coronary Artery Disease; BD - Bipolar Disorder.

Predicted therapeutic targets targeted by FDA-approved drugs and drugs in clinical trials.

The primary purpose of our work was to identify potential therapeutics and their targets by integrating publicly available genetic, bioinformatic and drug data using the Gentrepid candidate gene prediction platform. As the method involves repositioning of currently available drugs, it allows immediate translational opportunities for drug testing [8]. Other bioinformatic tools have been used to identify potential therapeutic targets for complex diseases and other conditions. For example, TARGET gene was used to identify and prioritize potential targets from hundreds of candidate genes for different types of cancer [34]. Another study identified potential drug targets for three neurological disorders - Alzheimer's disease, Parkinson's disease and Schizophrenia. This study involved the prediction of candidate genes using the ToppGene and ToppNet prediction systems [24, 35]. The repositioning tools could be used as an initial screening tool for potential drugs which can be used for further evaluation [34]. It is important to note that not all repositioning opportunities will be successful as there are always some limitations [36, 37].

There is a need to develop new approaches for the identification of therapeutic targets to accelerate the process of therapeutic drug discovery which has not kept pace with discoveries in genetics. In this study, we integrated detailed drug data with predicted candidate genes for seven complex diseases. We found 29% of the predicted candidate genes could serve as novel therapeutic targets and 29% of the extracted drugs are potential novel therapeutics for at least one of the seven complex diseases considered in our study. We have utilized both FDA-approved drugs and drugs in clinical trials. Further investigation is required to verify the action of these drugs. This study enables efficient identification of possible novel therapeutic targets and alternative indications for existing therapeutics. Hence, these drugs may be repositioned against seven phenotypes of interest, quickly taking advantage of already done work in pharmaceutics to translate ground-breaking results in genetics to clinical treatments. Gentrepid, thus can be utilized as a drug screening tool to save time and money spent on the initial stages of drug discovery.