Assessing drug target suitability using TargetMine

Implementation

TargetMine⁸ is based on the InterMine framework, an open-source data warehouse system designed for biological data integration⁹. In this update, we added a few customized data sources by defining new data models and implementing new data parsers. Details of how we designed the data models are described in the following sub-sections.

GWAS catalog

The GWAS catalog, founded by NHGRI, is a curated archive of the published genome wide association studies¹⁰. We had tried to associate genes to related diseases using the GWAS catalog in the former release of TargetMine¹¹. To annotate disease terms to a trait or study, we first chose the disease ontology (DO)^12,13 and then manually assigned the terms with the assistance of some text matching approaches. However, this process required some knowledge and involved a lot of manual examinations. Thus, it became difficult to keep updating regularly. Fortunately, the curation team started to use experiment factor ontology (EFO)¹⁴ to describe the curated GWAS traits in the recent implementation¹⁵. EFO covers several domain-specific ontologies that facilitate easier data integration. In our new implemented model, we replace DO terms with EFO terms and also incorporate some more information from each study (Figure 2). SNP annotations and details of EFO terms are retrieved from the dbSNP database and EFO, respectively.

Figure 2. The new implemented data model.

The colored lines indicate how the genes and diseases/phenotypes are associated in the post processing step.

ClinVar

ClinVar is a public archive of the relation between human variations and phenotypes^16,17. As defined by ClinVar, a “Variation” could be a single variant, a compound heterozygote, or a complex haplotype. If a haplotype consists of multiple alleles, each allele is assigned with an independent identifier. On the other hand, the same allele could be the member of a different haplotype, thus the relation between the “Variation” and “Allele” is a many-to-many association. An “Allele” is supposed to describe a specific change of a variation, e.g. G>A. However, the SNP entries in dbSNP sometimes merge different combinations of variations (alleles) together if the variations occur at the same genomic position. Thus, an “SNP” entity may contain multiple “Allele” entries in the data model (Figure 2). Here, we only retrieve the SNP identifier, and the rest of the annotations are integrated from the dbSNP database. The structural variations which reference the dbVar records are not included in the current version. In addition, those alleles which were not assigned with any dbSNP or dbVar identifiers were treated as SNP entities and were stored in TargetMine⁸ using the information provided by ClinVar. Most of the data were processed from tab delimited files, while some information that were not available in the tab delimited files were processed from XML files. MedGen terms, which are used to integrate the human medical genetic information at NCBI (https://www.ncbi.nlm.nih.gov/medgen/), were adopted to describe diseases and phenotypes.

dbSNP

dbSNP is a database which archives short human genetic variations. We first performed a whole data dump to a relational database, and then made queries to extract the necessary information into a flat table. These data include genomic position (based on genome assembly GRCh38), reference mRNA, nucleotide variation, reference protein, and amino acid variation, if available. SNP to gene is a many-to-many relationship, thus we introduce an intermediate class named “VariationAnnotation” to associate them together (Figure 2). Although the InterMine framework is capable of incorporating whole SNP entries in dbSNP, the integration takes a few days to finish. Considering the frequency that we update TargetMine⁸ (once a month), it is not very practical to spend a few days doing the integration. As a tradeoff, we decided to store only a subset of SNPs. Only those SNPs which are related with GWAS associations or clinical assertions, or those where there is an associated publication, are selected for storage in TargetMine⁸.

Frequency data

Population specific genetic variation frequency is important for evaluating drug efficacy. We preprocessed the frequency data from several data sources, including the Human Genetic Variation Database (HGVD)¹⁸, the integrative Japanese Genome Variation Database (1KJPN)¹⁹ (download from the archive in National Bioscience Database Center), the Exome Variant Server (EVS)²⁰, and the 1000 Genomes Project (1KGP)^21,22. At the moment, we only incorporate the population specific frequency for those SNPs stored in TargetMine⁸.

Post-processing the integrated data

Our implementation allows us to associate the genetic phenotype (disease) and the gene via the GWAS or ClinVar dataset, or moreover the relation that is implied from the disease related MeSH (Medical Subject Headings, https://www.ncbi.nlm.nih.gov/mesh) terms assigned to the correlated publications of the SNPs. In order to make a shortcut and to summarize the available information, we perform post-processing and store the results using a new class named “GeneDiseasePair”. At the moment, there are three types of shortcuts. Gene to SNP to GWAS to EFO terms for GWAS catalog data (the red lines in Figure 2). Gene to SNP to clinical assertions to disease (MedGen) terms (the green lines in Figure 2). And Gene to SNP to publication to MeSH terms (the blue lines in Figure 2). The “GeneDiseasePair” class also includes correlated information including ontology terms, studies, SNPs and publications. These improvements in the data model facilitate quick access from a gene to the associated diseases, annotated by different data sources.

Operation

TargetMine⁸ is a Java-based web application that runs on Apache Tomcat. The user interface communicates with the integrated data stored in PostgreSQL, a relational database.

Querying linkage to disease with TargetMine

To demonstrate the effectiveness of the new version of TargetMine⁸ in evaluating linkage to disease, we conducted a feasibility study, taking human PCSK9, proprotein convertase subtilisin/kexin type 9, as a typical case. The PCSK9 gene encodes a protein that promotes degradation of low-density lipoprotein (LDL) receptors in hepatocytes, thereby elevating or maintaining LDL cholesterol levels in the blood. Mutations in this gene are shown to be associated with familial hypercholesterolemia²³, and monoclonal antibodies to PCSK9 have been launched on the market as drugs for hypercholesterolemia with and without genetic predispositions^24,25.

Figure 3A demonstrates a schematic representation of the searching protocol for genetic disease associations with TargetMine⁸. We first went to a gene report page by searching for the PCSK9 gene from the top page of TargetMine⁸ (not shown). From the gene report page, we got information of genetic disease associations (Figure 3B) as well as many other basic or advanced characteristics such as orthologous genes and upstream transcription factors. The results table of genetic disease associations for PCSK9 enabled us to confirm that a number of SNPs relevant to this gene have been reported to be associated with plasma LDL cholesterol levels, hypercholesterolemia, or coronary artery disease. By clicking the record of association between “low density lipoprotein cholesterol measurement” and PCSK9 in the GWAS catalog section (Figure 3B), we moved to a “gene disease pair” page and checked the details of the GWAS record, including the information on samples, statistical significance and publications (Figure 3C). Clicking on the SNP identifier (e.g., rs2479409) redirected us to an SNP report page containing the individual SNP basic information (allele, function, literature) and allele frequencies of different human populations (from 1000 Genome Project²⁶ and others, not shown in the figure). Similarly, we examined the associations between “Hypercholesterolemia, autosomal dominant, 3” and PCSK9 from the ClinVar section in the table (Figure 3B) and got the details of the ClinVar record such as clinical assertions and publications (Figure 3D). The publications here reported mutations in PCSK9 as a cause of autosomal dominant hypercholesterolemia²³ (not shown), as mentioned above.

Figure 3. Searching information about linkage to disease with TargetMine.

(A) Outline of the procedure for searching. (B) A screenshot of the summary of Genetic disease associations of PCSK9. (C) GWAS records of a pair of PCSK9 and low density lipoprotein cholesterol measurement. (D) ClinVar records of a pair of PCSK9 and hypercholesterolemia, autosomal dominant, 3.

Querying target tractability for small molecule drugs with TargetMine

We performed another feasibility study to examine whether TargetMine⁸ provides informative evidence to assess target tractability for small molecules. In this case we also used PCSK9 as an example because no potent small molecule inhibitors for this protein have been reported so far in spite of the intensive research activities of many laboratories²⁷, indicating that PCSK9 is not a highly tractable target.

Figure 4A shows a schematic diagram of the procedure of querying tractability with TargetMine⁸. We first went to the protein report page of PCSK9 and found the bioactive compounds targeting this protein. As we expected, it was revealed that no potent compounds could be found in the ChEMBL database, and the lowest IC50 value was 440 nM (CHEMBL3923422) (Figure 4B). On the PCSK9 protein report page, we also checked the experimentally determined 3D structures, referred to as “protein structure regions” in TargetMine⁸, and identified several Protein Data Bank (PDB) entries for this protein (Figure 4C). Then, we moved to the “Protein Structure” page of a specified PDB ID (2p4e in this case) and found that in the “DrugEBIlity” table (from the DrugEBIlity database), some domains of the PCSK9 protein had positive Ensembl scores (Figure 4D), which are not ligand-based, but structure-based tractability scores. This result indicates that PCSK9 protein may contain some sites/pockets that can bind small molecules, although Ensemble scores of DrugEBIlity may need to be further validated.

Figure 4. Searching information about target tractability for small molecule drug with TargetMine.

(A) Outline of the procedure for searching. (B) Protein structure regions and their Ensembl scores calculated by DrugEBIlity. (C) Compounds with bioactivity for PCSK9.

Collectively, we were able to confirm that the new version of TargetMine⁸ can quickly provide lines of evidences to assess linkage to disease and target tractability of PCSK9, and that the gathered data correctly reflected the real world situation; namely, it has been a challenge to obtain potent small molecule inhibitors for PCSK9, whereas antibody drugs for this protein have been successfully developed and marketed recently.

Gathering and prioritizing candidate drug target genes

To assess the utility of the new update of TargetMine⁸ for prioritizing candidate targets, we conducted a case study where we employed a list of genes associated with hypercholesterolemia in literature. We tentatively defined three key properties of a novel drug target suitable for small molecules as follows: (1) being associated with hypercholesterolemia via SNPs (GWAS catalog, ClinVar, or dbSNP-Pubmed; see Materials and Methods), (2) having greater than or equal to 50% of protein 3D structures with positive Ensemble scores (DrugEBIlity), and (3) having no reported (ChEMBL) potent small molecule inhibitors (IC50 or EC50 ≤ 100 nM).

We first searched PubMed using the term “hypercholesterolemia” (from 2017/1/1 to 2018/9/10) and curated the resultant literature with the “Pubtator” text-mining tool²⁸, resulting in 510 human genes (Figure 5A). We then selected the genes meeting the requirements defined above using the “Query Builder” in TargetMine⁸. Figure 5B shows an overview of the actual query, which aimed to extract the genes with gene evidences obtained from the GWAS catalog, where “Mapped Trait” contained “LDL cholesterol”, “total cholesterol”, or “low density lipoprotein cholesterol”, from ClinVar where “Reported phenotype Info” contains the term “Hypercholesterolemia”, and from dbSNP where “Mesh Terms Name” of related articles contains “Hypercholesterolemia”. Thus, the new implementation enabled us to filter objects on complex conditions with a user-friendly, intuitive graphical interface.

Figure 5. Gathering and prioritizing candidate drug target genes for hypercholesterolemia.

(A) Gathering hypercholesterolemia-related genes from article information in PubMed. (B) Prioritizing hypercholesterolemia-related genes with TargetMine to identify novel targets for small molecule drugs. Top prioritized genes were defined as those that met all of the following three requirements: 1) more than or equal to 50% of protein 3D structures (PDB IDs) having positive Ensembl scores, 2) no potent bioactive compounds (EC50 or IC50 ≤ 100 nM in ChEMBL) and 3) having genetic associations with hypercholesterolemia (for more details, see the Use Cases section).

Genes that satisfied all three requisites above are presented in Figure 5C (CYP7A1, FABP2, LDLR, MYLIP, PCSK9, SREBF2 and STAP1). Among the seven genes we found MYLIP and STAP1. MYLIP is an E3-ubiquitin ligase that degrades LDL receptors in the liver, which are therefore considered to be a potential therapeutic target for dyslipidemia²⁹. Similarly, the STAP1 gene has been recently annotated as a fourth locus associated with autosomal-dominant hypercholesterolemia, and might be a novel target for therapeutic development of hypercholesterolemia³⁰. This result suggests that the new version of TargetMine⁸ allows us to effectively prioritize target candidate genes in terms of linkage to disease, tractability and competitors. On the other hand, the list includes intractable targets such as PCSK9 and LDLR, indicating the need for improvement of the data and/or the thresholds with which tractable proteins are selected (in this study, ≥50% of protein 3D structures have positive Ensemble scores in DrugEBIlity database).