Introduction

Target selection and prioritization are common goals for academic and commercial drug research organizations. While motivations differ, in all cases, the target selection task is fundamentally one of resource allocation in the face of incomplete information. Consequently, target selection strategies (and metric-based approaches to assess their success) remain complex1 and are hindered by multiple bottlenecks. Some bottlenecks pertain to the data themselves, such as disjointed, disparate data and metadata standards, data recording errors and accessibility issues; overcoming these issues will require human and computational efforts and coordination across multiple communities. Another set of bottlenecks pertains to the scientists involved. These include a tendency to focus on a small subset of well-known genes2 and the tendency to avoid riskier research paths, driven by poor research funding climates3.

For the purposes of this article, we define knowledge as the consensus of information aggregated from different sources and information as structured data, with a contextual layer that supports a broad range of data analytics. Data have quantity, quality and dimensionality (for example, genomic knowledge is defined in relation to associations with distinct entities such as molecular probes and disease concepts). Data, like facts, may also have an expiration date (Supplementary Box S1), and thus knowledge is subject to change. Yet, within a given time frame, knowledge provides context for interpretation and integration of emergent data, information and models.

Data-driven drug discovery strategies rely on the integration of proprietary and internal data with third-party resources — both public databases, such as PubMed, PubChem4, ChEMBL5 and The Cancer Genome Atlas (TCGA6), and commercial databases, such as Integrity. This integration requires fusion and reconciliation of heterogeneous and sometimes conflicting data sources and types. Although many of these resources are already partially interlinked, data heterogeneity, complexity and incompleteness, as well as contextual information and metadata capture, pose substantial barriers to reliable systematic analyses of all data required to address biomedical research questions, such as target prioritization in drug discovery1.

With the increasing scale and variety of data generation, collection and curation in the biomedical sciences, there is an unmet need for in-depth, accurate and truthful integration of multiple scientific domains across disciplines. Once successful, these data and knowledge integration efforts enable us to ask both global and fundamental questions about genes, proteins and the processes they are involved in. Integrated resources also allow us to address aspects of reproducibility7 via concordance of similar data types from unrelated sources and deficits in our knowledge of biological systems and their function. More generally, data integration facilitates our ability to quantify knowledge using an evidence-based approach.

Illuminating the Druggable Genome. “The reluctance to work on the unknown” (Ref. 2) is inherent to the scientific endeavour, partly due to our subconscious tendency to choose research subjects more likely to confirm what we already know or believe8. In a deliberate, strategic attempt to map the knowledge gaps around potential drug targets and to prompt exploration of currently understudied but potentially druggable proteins, the US National Institutes of Health (NIH) launched the Illuminating the Druggable Genome (IDG) initiative in 2014. As part of this broad, multimillion-dollar initiative, the IDG Knowledge Management Center (KMC) aims to systematize general and specific biomedical knowledge by processing a wide array of genomic, proteomic, chemical and disease-related resources (Box 1), with the explicit goal of supporting target hypothesis generation and subsequent knowledge creation, especially for genes and proteins that are not well studied.

In this article, we first define objective, evidence-based criteria for tracking target development levels (TDLs) for human proteins, using multiple sets of current knowledge. We discuss the data collected by the KMC on TDLs, which show the existence of a substantial knowledge deficit concerning a large portion of the human proteome (one out of three proteins). Reflecting the goal of illuminating the druggable genome, we then present spotlights on the TDL categories, as well as on key target classes, including G protein-coupled receptors (GPCRs), protein kinases and ion channels.

Knowledge-based protein classification

Target development levels. Most current protein classification schemes are based on structural and functional criteria. For any given protein, it is also possible to identify associated drugs and chemical or biologic modulators, and many types of experimental data can be associated with the protein, including publications, patents, gene expression data and experimental or modelled 3D structures.

For target prioritization and therapeutic development, it is useful to understand the quantity and diversity of data that are available for a given protein and to assign a qualitative knowledge metric that characterizes the degree to which a target is comparatively well studied or unstudied. To address this, we developed the TDL classification scheme, which categorizes proteins into four groups — Tclin, Tchem, Tbio and Tdark — with respect to the depth of investigation from a clinical, chemical and biological standpoint (Fig. 1; Table 1). Except for Tclin, TDL assignments were performed without human curation. Formal definitions for the TDL categories are as follows.

Figure 1: Target development level categories applied to the human proteome.
figure 1

a | Percentages of the whole proteome are shown in the inner ring. Percentages of each target development level (TDL) category for selected major protein families are shown in the outer ring, with the Tclin category expanded. Inner ring colours are as follows: Tdark, black; Tbio, red; Tchem, green; and Tclin, blue. b | TDL distribution across protein families, coloured by TDL category. Data show 3,644 proteins that have a confirmed disease association according to the Online Mendelian Inheritance in Man (OMIM) database. The enzyme category excludes kinases, which are considered separately. GPCR, G protein-coupled receptor.

PowerPoint slide

Full size image
Table 1 Current distribution of TDL categories by protein family for the druggable genome
Full size table

• Tclin (clinic) proteins are drug targets linked to at least one approved drug (that is, an active pharmaceutical ingredient) by mechanism of action (MoA) (this criterion supersedes any of the other parameters). Classification into this TDL category was achieved through exhaustive manual querying of primary literature and drug labels for MoA assignments with respect to molecular (protein) targets9; drug targets annotated as MoA-related proteins are categorized as Tclin (see further discussion below).

• Tchem (chemistry) proteins lack MoA-based links to approved drugs but are known to bind to small molecules with high potency. The interactions between proteins and small molecules (and sometimes approved drugs) are usually studied in the context of a disease and often arise from medicinal chemistry efforts. For inclusion in the Tchem category, we required the bioactivity of at least one small molecule to be above a specific cut-off chosen to include about 90% of the bioactivity values of drugs with a confirmed MoA for a target from that protein family (Supplementary Figure S2). Currently chosen thresholds are ≤30 nM for kinases, ≤100 nM for GPCRs and nuclear receptors, ≤10 μM for ion channels and ≤1 μM for other target families. Bioactivity values were extracted from ChEMBL5 and DrugCentral10.

• Tbio (biology) refers to those proteins that have a confirmed Mendelian disease phenotype in the Online Mendelian Inheritance in Man (OMIM) database11 (that is, at least two publications), have Gene Ontology (GO)12 leaf term annotations based on experimental evidence or meet two of the following three conditions: a fractional PubMed publications count13 above five; three or more National Center for Biotechnology Information (NCBI) Gene Reference Into Function (RIF) annotations; or 50 or more commercial antibodies, counted from data made available by the Antibodypedia database14. Tbio assignments imply that these proteins are not MoA-related drug targets (these are Tclin proteins). However, it does not follow that these proteins lack associations with bioactive molecules, including approved small-molecule drugs and biologics. It does, however, imply that given current levels of evidence, associated bioactivity values and clinical observations did not meet Tchem or Tclin criteria, respectively.

• Tdark (dark genome) refers to the remaining proteins that have been manually curated at the primary sequence level in UniProt15 yet do not meet any of the criteria for Tclin, Tchem or Tbio. Even for this category, evidence may be available concerning genome-wide association studies (GWAS), tissue location, dysregulation, inferred function via homology, etc. Many proteins in the Tdark category are not contextless sequences. However, these are proteins for which there is the least current knowledge and a low number of specific molecular probes available, and some represent unexplored opportunities within the druggable human genome. While evidence that approved drugs interact with some Tdark proteins may be available, the above criteria were observed for all Tdark assignments (Supplementary Table S3).

The knowledge deficit. Figure 2a summarizes the varying degree of available data (represented using a normalized count of occurrence) for seven different data types associated with individual targets and grouped by TDL. The first three groups illustrate category differences for three TDL defining criteria discussed above — namely, the fractional count of protein and/or gene mentions in PubMed abstracts, NCBI Gene RIF counts and antibody counts per protein. 'GO terms' examines the distribution of GO12 annotation counts per protein using data from UniProt15. 'R01 grants' examines the distribution of text-mined R01 grant counts detected for each protein using NIH RePORTER data (see below for further discussion). 'Patents' examines the distribution of text-mined granted patents for each protein using SureChEMBL16 data. Finally, the data availability score summarizes experimental information density per protein obtained from Harmonizome17 data — a resource developed independently for the KMC that provides an abstract representation of the many types of data associated with all human genes and proteins (Box 1).

Figure 2: Patterns of target development level distribution across different data: visualizing the knowledge deficit.
figure 2

a | The three criteria used in establishing the target development level are to the left, and their independent validation by four other data types are to the right. For PubMed abstracts, Gene Reference Into Function (RIF) annotations, antibodies, Gene Ontology, R01 grants and patents, the score for each target is the count of those entities associated with the target, normalized between 0 and 1. The values for the Harmonizome data availability score were computed differently, as described in the main text. See Fig. 1 for colour codes and Supplementary Table S4 for further details. b | Patterns of scientific curiosity: STRING database access counts by target development level (January–December 2016).

PowerPoint slide

Full size image

Whereas the first three data types were used to assign the TDL category for proteins in the Tbio and Tdark categories, the other four data types — derived from separate text corpora and repositories — provide independent validation of our criteria for categorization overall. Distribution trends within TDL categories are consistently reproduced across all data types in Fig. 2a and have statistically significant differences (Supplementary Table S4). Tdark proteins have the least amount of data associated with them regardless of source.

Increasing amounts of data are observed for proteins when progressing through the categories from Tdark to Tbio, Tchem and Tclin. For example, Tdark proteins tend not to be the object of study for many funded NIH R01 grants and are significantly less discussed in patents compared with proteins in other TDLs. Statistical significance breaks down when comparing Tclin and Tchem, but because successful clinical trials are required for the Tchem-to-Tclin progression, this evidence may not be well captured by the four data types highlighted in Supplementary Table S4. However, this is less surprising from a knowledge management perspective, since on average, the biochemistry and pharmacology of a protein are likely to be well studied upon reaching the Tchem development stage. It is important to note that the Tchem stage can be completely bypassed for targets of therapeutic antibodies and other biologics.

In summary, all the data, information and knowledge aggregated and processed within the IDG KMC archive (partially illustrated in Fig. 2a) confirm the existence of a knowledge deficit about many proteins, some of which could have therapeutic relevance. The bias towards well-described proteins2 is confirmed not only with respect to publications but also with respect to patents, NIH funding patterns, GWAS and mouse phenotype data (data not shown), availability of molecular probes such as antibodies and small molecules, and even queries in the STRING18 database (see below and Fig. 2b). Because of this bias, one out of three human proteins (Tdark) have been largely unstudied. Although the NIH acknowledged that illumination should directly target understudied proteins, scientists engaged in target selection are likely to remain risk-averse and perhaps systematically less inclined to study Tdark proteins.

Our classification provides overall insight into the current illumination levels and sizes the opportunity for drug targets from well-established and precedented druggable protein families. The natural progression is for proteins of potential therapeutic interest to migrate from Tdark to Tclin over time, and TDL monitors knowledge accumulation using multiple types of clinical, chemical and biological evidence, while providing an easily interpretable ranking scheme. We argue that proteins in Tdark and Tbio are understudied and more in need of illumination, and we discuss approaches for achieving this later in the article, after first overviewing knowledge on proteins in the Tclin and Tchem categories.

Spotlight on Tclin and Tchem

Evaluating protein target druggability — the ability of a protein to be therapeutically modulated by medicines — can involve complex assessments of a range of protein characteristics. Structural biology and computational and medicinal chemistry assessments of druggability largely focus on forecasting whether a target protein can bind to drug-like small molecules with high affinity and specificity19. However, druggability literature rarely mentions biologics, antibodies and other protein therapeutics, radiotherapy (Supplementary Box S5), gene therapy or stem cells. In this section, we discuss Tclin and Tchem proteins in the context of small-molecule drug discovery.

Tclin proteins. Ideally, unequivocal Tclin assignment (that is, identification of molecular drug targets) would require several layers of evidence: a full matrix of in vitro bioactivity for all prodrugs, drugs and active metabolites (active ingredients) assayed against all relevant human and non-human (for example, bacterial and viral) targets (such as the half-maximal inhibitory concentration (IC50), effector concentration for half-maximum response (EC50), inhibitory constant (Ki) and the dissociation constant (Kd)); on–off rate constants and other kinetic measurements performed at appropriately relevant concentrations in the tissue or tissues relevant for that particular disease context, preferably with matching in vivo data in humanized animal models (although human data are preferable); and phenotypic confirmation supported by pharmacodynamic data. In animal disease models lacking the gene or genes responsible for the MoA of the drug, the drug should lack therapeutic effect. Meeting these criteria would be needed in order to attribute the desired clinical outcome to a specific drug target interaction mechanism.

Because the above criteria are difficult to implement by automation, a previous analysis carefully curated MoA data from approved drug labels as well as primary literature, based on a rigorous definition of a drug target9. This ongoing process, performed in parallel by three teams, is anticipated to improve our ability to link drug responses to genetic variation and to help us understand the molecular basis of clinical efficacy, safety and adverse events. The interplay between target tissue expression under disease-specific conditions and the local concentration of the drug or its active metabolites at the relevant disease site is often difficult to ascertain, which is why we attributed a higher weight of evidence to data derived from multiple drugs belonging to the same therapeutic class. Indeed, we anticipate that efficacy target annotations will become more precise as our capability to colocalize target, disease and drug increases.

From this analysis, Tclin currently consists of ∼600 protein targets9, which is at the lower end of the original estimate of between 600 and 1,500 targets for the intersection between proteins in the druggable genome and disease-modifying genes by Hopkins and Groom20 (note, however, that Tclin includes targets of biologic drugs as well as small molecules, while the estimate was for small-molecule drug targets only20). So far, proteins in the Tclin category thus represent only a small fraction (3%) of the human proteome (Fig. 1a). From a commercial perspective, it is also noteworthy that most of the global revenues of the pharmaceutical industry are derived from drugs that target a relatively small number of the proteins in the Tclin category (Box 2; Table 2). The majority (259 or 79.7%) of these targets are single proteins, whereas 39 (12%) are complex multiprotein targets. Only 25 targets (8.3%) are comprised of multiple proteins for non-selective drugs; these include the muscarinic, α-adrenergic and oestrogen receptors, as well as cyclooxygenases and histone deacetylases.

Table 2 Financial (sales and NIH funding) activity for the top 20 mechanism-of-action targets
Full size table

Among the factors contributing to the small fraction of each major protein family in Tclin so far, one factor is that not all members of a protein family have drug-compatible or ligandable21 binding sites; for example, some nuclear receptors lack an (endogenous) ligand-binding domain or do not appear to be amenable to small-molecule perturbation. Another factor is that not all proteins can (or will) alter the course of disease via therapeutic intervention, perhaps in some cases owing to our lack of understanding of the underlying pathology.

Kubinyi pointed out that single proteins combine in vivo in ways that could lead to many more drug target combinations across multiple pathways — that is, a 'druggable proteome' (Ref. 22) — and there is now experimental evidence that alternative splicing, post-translational modification and heterogeneous oligomers produce functional isoforms with different interaction profiles, which may further result in increased diversity of the proteome23. It is also important to note that for many drugs, the precise MoA and contributing molecular targets remain cryptic, especially when polypharmacology (the simultaneous modulation of multiple targets by drugs) occurs. Shedding light on this would require data completeness24, namely, experiments across all proteins, in relevant physiological conditions, for all approved drugs. This remains a resource-intensive and costly task, which was partially accomplished25 by the NIH Molecular Libraries Initiative26.

Tchem proteins. Assignment to Tchem is based on compound activity thresholds originating from binding experiments for small molecules (Supplementary Figure S2). Selectivity, though important both in vivo and in vitro, could not be factored in for all Tchem targets (Supplementary Box S6 and Supplementary Table S7). Because, by definition, Tclin attribution requires supporting evidence for the MoA, many proteins known to interact with approved drugs, even with high affinity, remain in the Tchem category. Additional bioactivity data from, for example, patent literature and papers currently not indexed in ChEMBL may progress more targets to Tchem.

Many compounds that have reported activity against Tchem targets are also candidate drugs undergoing clinical trials. Based on an in-depth analysis of clinical trial data combined with data from ChEMBL, PubMed, the IUPHAR Guide to Pharmacology27 and ChemIDplus, we mapped 144 Tchem proteins to 356 clinical (phase I–III trial) candidates, for a total of 701 unique target–clinical candidate pairs. For 175 (25%) of these pairs, therapeutic indication data extracted from ChEMBL highlight the different distribution among protein families (Table 3). Most targeted proteins are kinases (93 unique enzymes), followed by GPCRs (31 unique receptors) and ion channels (13), with seven targets from other families, which is similar to the prior observation that most clinical candidates target the most druggable target families28. Analysis of the target–clinical candidate subset in which anticipated therapeutic indications are available shows that most of the kinase-targeting clinical candidates are aimed at oncology applications, whereas GPCRs and ion channel-targeting clinical candidates are aimed at central nervous system disorders.

Table 3 Summary of clinical candidates (phase I–III) with activity against Tchem proteins
Full size table

As noted above, target druggability is frequently estimated based on the ability to bind to small molecules20, and expectations of druggability typically diminish as the size of the binding pocket increases, with affinity and selectivity being major concerns. The challenge is even greater when the binding pocket is shallow and highly exposed to a solvent or when the therapeutic strategy involves disrupting the interaction between the targeted protein and other proteins. One approach to evaluating druggability is to focus on protein domains mined from the InterPro database29 and then to prioritize proteins that contain domains known to interact with approved drugs20 or bioactive small molecules30. Others have explored target druggability by evaluating side-effect similarity for known drugs31 or by performing combined chemical and target similarity queries32, followed by experimental confirmation of novel drug targets derived from clinical observation or computation. It is possible that induced binding sites in proteins in which a druggable pocket is not initially found may enable them to be targeted with drug-like small molecules, but identifying these binding sites with structural approaches is likely to be challenging, and phenotypic screens may be more useful.

An emerging approach harnessing so-called PROTACs (proteolysis-targeting chimaeras) may help substantially in addressing the issue of undruggability33, at least for proteins that are capable of selectively binding a small-molecule ligand, although not necessarily at a typical binding site34. Essentially, this strategy harnesses the endogenous ubiquitin–proteasome system to promote targeted degradation of desired proteins following binding of the PROTAC35; the mechanism for ternary ligase–PROTAC–target complex activation has been recently elucidated36. This technology may also be subcellular location-specific, which could be an additional advantage in some (but not other) cases. However, the oral bioavailability of PROTAC molecules may be constrained by their size.

Spotlight on Tbio and Tdark

A critical effort in addressing the knowledge deficit about Tbio and Tdark proteins is being undertaken by the Monarch Initiative37, which relies on informatics methods to identify phenotypically relevant disease models in research and diagnostic contexts based on integrated model organism and clinical research data. One of the main sources for the Monarch Initiative is phenotype data from the International Mouse Phenotype Consortium (IMPC), which was set up to generate and phenotypically characterize mouse knockout lines. Their recent analysis of 1,751 unique gene knockouts found that human disease genes are enriched for essential genes38.

The IDG KMC incorporates gene-centric mouse phenotype data and maps these data to the respective human orthologues. IDG coordinated with IMPC production centres to prioritize production of knockout mouse strains for druggable genes. As of November 2017, 568 new knockout strains had been produced: 166 GPCRs, 141 ion channels, 238 kinases and 23 nuclear receptors (see Supplementary Table S8). When ignoring olfactory GPCRs, these represent a little more than one-third of the druggable genes in these protein families. Phenotype data are available for 80% of these strains, with abnormalities detected in numerous biological systems, including those related to development, immune function, metabolism and behaviour. These IMPC strains provide evidence for biological systems that may be affected when a drug targets a gene with little-known function.

Of the 119 Tdark genes (51 GPCRs, 36 ion channels and 32 kinases) submitted by IDG to IMPC, 45 mouse lines were produced, with 41 phenotypes observed. For example, knockouts of the Tdark kinase gene Alpk3 have increased embryonic and perinatal lethality, with the surviving adults displaying severe heart defects (see Further information). Of 482 Tbio genes submitted by IDG (135 GPCRs, 133 ion channels, 200 kinases and 14 nuclear receptors), 184 mouse lines were produced, with 145 phenotypes observed. For example, knockouts of the Tbio GPCR gene Adgrd1 display reproductive defects, such as female infertility, and skeleton phenotype defects, such as decreased bone mineral density (see Further information).

Among 2,788 genes phenotyped in mice at the IMPC, 953 have at least one significant behavioural, neurological or other nervous system-related phenotype observation. Target Central Resource Database (TCRD) data from the GWAS Catalog39, OMIM11 and text-mined DISEASES13 databases confirmed human disease phenotypes for 191 (20%) of these 953 genes, ranging from neurological (for example, seizure disorders) to cognitive (for example, tauopathy) and psychotic affective disorders. Because only 9 of the 953 genes lack confirmed expression in any of the 34 neuro-related tissues tracked by IDG KMC (for example, GTEx40,41, HPA42 and HPM43), these data suggest that the remaining 80% of this set have the potential to be associated with human neurobehavioural phenotypes, paving the way for new research avenues in this direction (see Supplementary Table S8). Production of IMPC strains is set to continue for several more years, and so further knockout strains for druggable genes and their phenotype data are anticipated.

To further explore the characteristics of Tdark and Tbio proteins, we analysed their distribution in the L1000 gene set, as annotated in TCRD (Supplementary Table S3). L1000 is a set of 978 landmark genes that have been selected for their ability to predict a large portion of the total variability seen in large sets of microarray experiments. The proportion of Tdark proteins in the L1000 set (79 of 978; 8%) is substantially smaller than would be anticipated based simply on the a priori distribution of Tdark proteins (which make up 35% of the proteome), whereas Tbio targets (671 of 978; 69%) are more common than expected, as these make up 53% of the proteome. The proportions of Tclin proteins (41 of 978; 4%) and Tchem proteins in the L1000 set (187 of 978; 19%) are also higher than expected, as these make up 3% and 7% of the proteome, respectively. The L1000 TDL distribution data support the existence of a knowledge deficit.

To some extent, the data on Tbio and Tdark suggest a causality dilemma: are Tdark proteins underfunded because there is no scientific interest in this category, or is the lack of knowledge perpetuated by lack of funding? Although our data do not allow us to establish a causal relationship, we suggest that the absence of high-quality, well-characterized molecular probes is a root cause for this situation. Lack of tools leads to lack of interest, and lack of interest diminishes the probability of such tools being developed. A bibliometric evaluation by Edwards and colleagues2 examined how many newly sequenced proteins from several protein families were the subject of new studies 10 years after the completion of the human genome sequencing project. This analysis concluded that the process of druggable target selection is conservative and incremental and that limited progress has been observed with respect to understanding newly discovered proteins.

“If you don't know very much to begin with, don't expect to learn a lot quickly” (Ref. 44). Anecdotal evidence (summarized in Table 4) suggests that it is possible for proteins to migrate from Tdark to Tclin within 12–20 years. Data on the six protein targets highlighted indicate that proteins for which little information was available two decades ago (effectively Tdark) became attractive from a drug discovery perspective following key papers, namely, deorphanization and protein–disease association studies. Five of these six targets are modulated by at least one approved drug, which places them in the Tclin category.

Table 4 Examples of successful attempts of targeting the dark genome
Full size table

Successful 'promotions' across classes, such as those in Table 4, are currently rare. We expect the rate of knowledge accumulation for Tdark proteins to be low, at least initially. Well-studied proteins require multiple layers of management for diverse, rich sets of data, with information and knowledge stemming from corpora such as biomedical literature, patents and clinical trials. A paucity of data and lack of information for understudied proteins (Tbio and Tdark) affect both knowledge management and the decision-making process with respect to experimental planning, what research questions need to be asked (and in what order) and which methods may be better suited for each task. For example, we examined access counts for human proteins in STRING18 during 2016 (Fig. 2b). 'Counts by name' represents users that access the STRING website and type in a gene symbol. 'Counts by link' represents users accessing the network for a gene in STRING by linking to it from another resource (for example, GeneCards45 or UniProt15). Whereas 'Counts by link' shows a more comprehensive method to access the entire proteome, it also suggests that Tdark proteins have a lower probability of being recognized (input) by gene name. These data show a pattern similar to that observed in Fig. 2a: Tdark proteins are less likely to be the subject of scientific curiosity, which is a reflection of funding patterns and an overall lack of information and molecular probes. Indeed, the paucity of antibodies and small molecules (criteria that help define Tbio and Tchem, respectively) that could be used to interrogate Tdark proteins diminishes our ability to subject Tdark proteins to scientific inquiry.

Genomic and proteomic responses following radiation therapy are also understudied. One in vitro study46 suggested that as many as one-third of the 10,174 genes examined in immortalized B cells following ionizing radiation are radioresponsive (GSE26835 column46 in Supplementary Table S3). Of the 447 genes with significant fold changes in the GSE26835 set, only 26 are Tclin and 61 are Tchem, whereas the majority (268 Tbio, 92 Tdark) are understudied (see also Supplementary Box S5).

As many as 3,644 proteins have significant disease (confirmed OMIM11 phenotype) associations. Given their TDL assignments (335 Tclin, 543 Tchem and 2,766 Tbio), we examined the distribution of the TDL in relation to druggable protein family categories (Fig. 1b). It appears that Tbio–disease associations are quite rare for druggable families such as nuclear receptors, ion channels and GPCRs, as these families are more likely to be in the Tclin or Tchem category. Instead, Tbio assignments are quite frequent for transcription factors, epigenetic targets, transporters and unassigned protein families. The exception among druggable families are olfactory GPCRs, which appear to attract less interest from drug discovery programmes, despite some of these GPCRs being linked to metabolism and ageing47.

Concerted efforts focused on an entire target class have sometimes led to new drugs. For example, GlaxoSmithKline (GSK) had a comprehensive programme aimed at finding new ligands and characterizing the biology of nuclear receptors48. New insights into bile acid metabolism49 and xenobiotic transcription of cytochrome P450s50, mediated by nuclear receptors, were described. A bile acid receptor (FXR) agonist has reached the market since this programme started: obeticholic acid (Ocaliva), which was discovered by GSK in collaboration with the University of Perugia51 and subsequently developed by Intercept Pharmaceuticals. Currently, several FXR agonists are in clinical development52. Choosing the appropriate proteins as drug targets remains a complex process, where scientific factors need to be balanced against commercial factors (such as company investors and medical insurance companies) and societal factors (such as physicians and patients), as well as legal factors (such as the requirements of regulatory agencies)1.

Spotlight on G protein-coupled receptors

GPCRs are membrane-bound, cell-surface receptors that transduce signals via interactions with heterotrimeric G proteins, arrestins and other cellular transducers53,54. Alterations in GPCR signalling are implicated in the pathogenesis and treatment of neuropsychiatric55, immunological56, gastrointestinal57,58, cardiac59, renal, hormonal, infectious60 and many other disorders53,54. GPCRs represent the largest family of druggable targets in the human genome9, with between 20% and 30% of approved drugs acting on them53,54.

The number of publications per GPCR and the number of chemicals associated with that GPCR in ChEMBL were examined61 to determine which of the druggable, non-olfactory GPCRs are understudied: less than 100 citations and less than ten ChEMBL compounds define understudied, uninterrogated GPCRs. The number of publications, similar to the fractional PubMed publications count13, does not take into account large-scale (many proteins per paper) analyses. Counting ChEMBL compounds, a quantitative criterion similar to Tchem assignments, does not consider bioactivity values. However, this independent analysis validates the more general TDL criteria with respect to GPCR biological functions and corresponding chemical matter.

Tclin and Tchem G protein-coupled receptors. Currently 827 GPCRs — including 421 olfactory GPCRs — are tracked by IDG KMC; of these, 96 are Tclin and 113 are Tchem (none of which are olfactory). Slightly more than half of the non-olfactory GPCRs have annotated drugs and small molecules targeting them53,54,61; see also Table 1. A recent analysis indicates, however, that a handful of GPCRs — mainly biogenic amine, muscarinic and opioid receptors — represent the most abundantly targeted receptors for FDA-approved medications53,54. GPCRs also represent important off-targets for kinase inhibitors62,63, ion channel modulators64, anti-infectives65 and other classes of drug-like molecules53,54. As with other druggable target classes, off-target actions within the GPCR class can be associated with severe and life-threatening side effects. For example, valvular heart disease is associated with anorectic agents, such as fenfluramine, and antimigraine medications, such as ergotamine, via serotonin 5-HT2B receptor agonism66. Recent successes in structure-guided and cheminformatics-driven drug discovery show promise for creating safer and more effective medications targeting GPCRs.

Tbio and Tdark G protein-coupled receptors. Although 52 non-olfactory GPCRs are categorized as Tdark, the availability of new screening platforms to discover chemical matter for these GPCRs has begun the process of illumination64,67,68. Of 62 GPCRs for which significant phenotype calls have been reported by IMPC (Supplementary Table S8), 24 are Tbio and 7 are Tdark; of these, 15 are associated with neurological and behavioural phenotypes. Including olfactory GPCRs, 618 proteins are classified as Tdark or Tbio; of these, 126 non-olfactory GPCRs and 51 olfactory GPCRs have significant associations with human diseases via OMIM, GWAS and text mining. Whereas the majority of these associations (nearly 59%) stem from text mining69, 48 GPCRs have confirmed associations from at least two information channels. For example, Adgrb2-mutant mice (Tbio) showed significant antidepressant-like behaviour compared with wild-type mice70, whereas the association71 between schizophrenia and Fzd3 mutants (Tbio) remains controversial72,73.

Spotlight on protein kinases

Tclin and Tchem kinases. The ∼600-member human kinome (Supplementary Table S3) is made up of protein kinases, in addition to metabolic and lipid kinases, and is highly druggable using both competitive and allosteric small-molecule inhibitors. However, the functions of about one-third of the kinases in this family are poorly defined or unknown. The 634 human kinases were categorized as follows: Tclin, N = 50; Tchem, N = 390; Tbio, N = 163; and Tdark, N = 31 (Table 1). Tclin kinases are not exclusively protein kinases, and the number of FDA-approved small-molecule kinase inhibitors varies depending on inclusion criteria. Wu and colleagues74 found 38 small-molecule protein kinase inhibitors. Based on DrugCentral, we found 50 approved kinase inhibitors, including 40 small-molecule protein kinase inhibitors, of which 32 are FDA-approved, one FDA-approved protein kinase activator (ingenol mebutate) and the phosphoinositide 3-kinase subunit-δ (PIK3CD) small-molecule inhibitor idelalisib, which is also FDA-approved. An additional seven FDA-approved antibodies target the receptor tyrosine kinases human epidermal growth factor receptor 2 (HER2; also known as ERBB2), epidermal growth factor receptor (EGFR), vascular endothelial growth factor receptor 2 (VEGFR2) and platelet-derived growth factor receptor-α (PDGFRα), and there is also an FDA-approved HER2-targeting antibody–drug conjugate, trastuzumab emtansine (see Supplementary Table S9).

Tbio and Tdark kinases. A number of Tbio and Tdark kinases are known to interact with FDA-approved multikinase inhibitors. According to data in DrugCentral, sorafenib inhibits 114 kinases, of which only 9 are associated MoA-related targets, whereas sunitinib inhibits 263 kinases, of which 9 are MoA-related targets. Given the current state of kinase inhibitor chemistry, it is very likely that Tbio and Tdark kinases can be effectively therapeutically targeted with highly selective small-molecule inhibitors. Some of the characteristics shared by understudied Tbio and Tdark kinases include poorly defined integration of the kinase in signalling networks, poorly defined function and regulation, lack of activation-loop phospho-antibodies or immunohistochemistry-grade antibodies, and lack of selective chemical tools for functional characterization. Primary tools for knockout and/or altered expression are RNA interference (RNAi) and CRISPR–Cas9, and cDNAs for overexpression; kinase knockout or altered expression rarely provides readily assayable phenotypes (for example, growth, migration, apoptosis or in vivo function in mouse organ physiology). Currently, the IMPC has targeted 238 kinases with 114 knockouts having a significant phenotype; of the latter, 22 are of current interest for phase 2 of the IDG programme.

Many Tchem, Tbio or Tdark kinases are altered in expression or mutated in TCGA. Table 5 shows ten Tbio or Tdark kinases whose amplification is observed in the TCGA database, together with their RNA expression in triple-negative breast cancer (TNBC) cells. These understudied kinases are frequently altered in breast cancer. The potential increased expression of many kinases in primary human tumours suggests these understudied kinases have important functions for the tumour cell phenotype that have not been characterized to date. These represent unexplored kinases with possible therapeutic utility.

Table 5 Understudied kinases that are frequently altered in breast cancer
Full size table

The potential therapeutic importance of Tbio and Tdark kinases in the kinome is highlighted by a recent clinical study that assessed the response to trametinib, a MEK1 and MEK2 inhibitor, in TNBC patients75. Pretreatment needle biopsies and surgical tumour resections following 7-day trametinib treatment were used for RNA sequencing (RNA-seq) to analyse tumour transcriptomic changes in response to the drug. Pretreatment biopsies matched to post-treatment surgical specimens showed overall concordance of the transcriptional kinase response to trametinib, with FRK (Tchem) exhibiting the highest mean increase and cytoplasmic BMX (also Tchem) exhibiting the highest mean decrease among patients in response to a 7-day drug treatment. Among the kinases transcriptionally altered in the TNBC tumours were several understudied kinases, including MRCK-γ (also known as CDC42BPG), PRKACB, STK32B and leukocyte tyrosine kinase receptor (LTK). These findings demonstrate that in TNBC tumours in patients, members of the understudied Tbio and Tdark kinome are co-regulated transcriptionally with kinases from the Tclin and Tchem category, in a dynamic adaptive response to targeted inhibition.

Spotlight on ion channels

Ion channels mediate signalling within cells, between cells, and between cells and their environment. Defects in ion channels underlie many major disorders in humans, also known as channelopathies, including neuronal disorders76, diabetes77 and heart failure78. This makes ion channels an attractive target class for drug development. Ion channels are mostly heteromeric complexes that require optimal interactions with ligands at specific locations. Currently, 355 ion channel pore-forming and auxiliary subunits are tracked by IDG KMC (Table 1; Supplementary Table S3). About 100 ion channel modulators, including auxiliary subunits, are reported, but to our knowledge, a systematic list of cell type-specific auxiliary subunits for all ion channels is not available.

Tclin and Tchem ion channels. Many drugs are known to bind to ion channels. There are 217 drugs annotated in DrugCentral10 as acting through 125 (Tclin) ion channels for the MoA9. The number of drugs increases to 497 when querying how many drugs are known to interact with ion channels outside of the MoA-related constraint. Some of these interactions are likely to be responsible for side effects such as cardiac toxicity. An accurate understanding of MoA and side-effect assignment at the target (molecular) level is required if we are to improve upon available drugs. For example, the anaesthetic ketamine, which has been postulated to act as a noncompetitive N-methyl-d-aspartate (NMDA) antagonist79, has been used off-label as an antidepressant80. However, in-depth analysis of the antidepressant effects of ketamine found that its active metabolite (2R,6R)-hydroxynorketamine (HNK) does not block the NMDA receptor. Instead, HNK displays sustained activation of α-amino-3-hydroxy-5-methyl-4-isoxazole propionic acid (AMPA) receptors and lacks ketamine-related side effects81. This may pave the way for the development of novel, rapid-acting antidepressants. It is therefore conceivable that some ion channels currently categorized as Tclin or Tchem are in need of further illumination with respect to MoA and drug specificity. Indeed, the low bioactivity cut-off criterion for ion channels (≤10 μM) in Tchem (see also Supplementary Figure S2) may need revision, given that older drugs continue to reveal unexpected modes of action.

Tdark ion channels. A relatively small number of ion channels (31) are categorized as Tdark (Table 1). Part of the difficulty in illuminating dark ion channels is the replication of physiological context and expression of proteins in the appropriate heteromeric, pore-forming functional complexes. Currently, there are no scalable systems available to study the localization of functional complexes. Moreover, most ion channels have paralogues that function redundantly. Gene redundancy increases the difficulty of revealing phenotype and precise localization, both important elements for understanding physiological functions other than ion channel activity. This considerably delays our progress in understanding ion channel function in vivo and their role in human health. Unlike GPCRs or kinases, neither pore-forming subunits nor auxiliary subunits share characteristic motifs. Lack of specific protein sequence motifs makes it difficult to flag candidate genes for further study, even with computational assertions. There could be other ion channels, which perhaps should be categorized as Tdark.dark, to reflect our complete lack of knowledge, even by computational means, regarding these proteins.

The list of ion channel pore-forming subunits, as well as auxiliary subunits, continues to grow. For example, leucine-rich repeat-containing protein 8 (LRRC8) heteromers form82 volume-regulated anion channels (VRACs), and ORAI proteins assemble to form83 calcium-release-activated calcium channels (CRACs), whereas anoctamins are olfactory calcium-activated chloride channels (CACCs)84. Currently, LRRC8B, LRRC8C and LRRC8D subunits are classified as Tdark, with the exception being the subunit LRRC8A (Tbio); ORAI1 is annotated as Tchem, wheras the ORAI2 and ORAI3 proteins are annotated as Tbio. With the exception of anoctamin 1 (Tclin), all other anoctamins are labelled Tbio. These, and all other Tdark proteins that lack computational assertions, are in need of systematic genomic-scale studies.

Conclusions

Modern medicine often employs artificial distinctions in terms of what and how biological systems are studied: segregated by organ (for example, ophthalmology and cardiology) or by disease (for example, oncology and infectious diseases), medical specialty separations carry over into the research arena, both in academia and industry. This distinction breaks down in nature, as we are likely to observe the interplay between the same genes and pathways regardless of organ, albeit in a context-specific manner. These artificial divisions can prevent scientists from achieving a translational, integrative view of gene and protein function. We suspect this to be another reason why funding to study Tdark proteins is scarce: for functionally enigmatic proteins, or the 'ignorome' (Ref. 85), anticipating which organ, disease or phenotype is relevant may be far from trivial. To address this limitation, the NIH launched a series of high-risk programmes via the Common Fund resource, aimed to catalyse trans-disciplinary research. The IDG is one such Common Fund programme. The IDG programme's ostensible goal is to encourage and track the illumination of relatively understudied and unstudied parts of the genome. This implicitly requires the construction of a knowledge base objectively and in an unbiased manner, asserting what is currently believed to be true (a process that is explored metaphorically in a classic book by Italo Calvino, 'The Castle of Crossed Destinies'; see Supplementary Box S1). The IDG KMC enables us to quantitatively demonstrate the existence of a knowledge deficit with respect to dark and understudied proteins, which underscores the need for basic science and its major role in illuminating gene functions and roles in human disease. The TDL classification scheme provides a convenient way to partition human targets that highlights the focus (or lack thereof) of science and drug discovery efforts on different targets. Through the use of the TDL groupings, we can highlight knowledge accumulation, as well as deficits, for a variety of target families, with a common theme being that while much is known, there remains a large fraction of the proteome that is understudied. The IDG KMC, by collating and linking a plethora of disparate and diverse data sources and data types, aims to shed light on these dark regions with the hope that researchers will be empowered to use the data and knowledge presented by the KMC to jumpstart research programmes on these targets.

Confirmed associations with a specific disease, or receptor deorphanization (Table 4), remain major incentives to allocate resources and further study of Tdark proteins. As mentioned above, the only other deliberate targeted effort to study Tdark proteins in addition to IDG is the IMPC. As of March 2017, mouse lines corresponding to 4,165 human genes have been produced, with phenotypes available, 2,788 of which have resulted in statistically significant phenotype calls. Of these 2,788 proteins, 827 (436 Tdark) are not associated with any NIH-funded grants between 2000 and 2015 (Supplementary Table S8). By contrast, only 120 of the 1,961 proteins with significant IMPC phenotype calls and that are associated with NIH funding are Tdark. It was Edgar Allan Poe who once said, “the enormous multiplication of books in every branch of knowledge is one of the greatest evils of this age, since it presents one of the most serious obstacles to the acquisition of correct information”. Poe's 19th century line of thought is remarkably apt in the context of current KMC activities, since the “acquisition of correct information” remains the largest challenge.

Another challenge relates to an area of knowledge largely neglected in the scientific literature: the large-scale capture of negative results. Due to confirmation bias8, scientists have a tendency to primarily publish successful accounts of research. Although there are attempts to overcome this problem86, we are not aware of the existence of an unbiased, easy mechanism to capture negative results. The aphorism “absence of evidence is not evidence of absence” illustrates practical limitations of knowledge management systems: does lack of evidence imply that the study was conducted, but nothing was found, or does it imply (more often) that the measurement was not carried out? Proper archiving of negative results (for example, “protein P is not expressed in cell type CT” or “gene mutation Gm does not play a role in disease D”) would benefit the community at large and would improve our specific knowledge about proteins. However, such non-positive facts fit poorly to current publishing and citation paradigms. One possibility for archiving such statements could be nanopublications87, which would be amenable to large-scale integration into systems such as TCRD–Pharos.

Finally, a key challenge faced by IDG KMC, and perhaps by other data analysts working in drug discovery, is that of reliable predictions: when examining Tdark proteins in particular, experimentalists would like to know what experiment to do next, what phenotypic changes should be examined first and which pathway is relevant in a specific disease. These, and other similar questions, have yet to find a computer-driven, reliable answer. To paraphrase William Gibson, “The truth is already here — it's just not very evenly distributed”.

Publisher's note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.