Introduction

Target identification and target validation are involved with the identification and validation of a target that can be modulated for a therapeutic purpose. Whilst there is much variation on what a target is and how to define one, we can consider a therapeutic target to be one that is involved with the mechanisms of a particular disease and modulation of it will affect the disease in a beneficial way. It is important to also note that when a drug fails, it tends to be due to safety concerns or lack of efficacy.¹

There are several computational methods that are applied within the target identification stages of drug discovery, including cheminformatic methods, such as similarity searching and network-based methods.² However, much information is also contained within databases and scientific publications.

Whilst here we primarily focus on small molecules as the main drug modality, it must be noted, that there are wide variety of modalities beyond small molecules and antibodies, and therefore, it remains to be seen how much of the genome can be drugged effectively. Efforts to expand further into and utilize more of the druggable genome have been undertaken such as the initiative called Illuminating the Druggable Genome of which started in 2013.³ However, it is important to note the vast array of modalities that are now available. The druggable genome is the genes that encode proteins that can potentially be modulated. It is thought that of the approximately 20,000 genes that make up the human genome, approximately 3000 of these are considered part of the druggable genome despite only a few hundred targets being currently exploited.³ The program aims to create a resource for the study of protein families.³ The program along with Nature Reviews Drug Discovery publish content on understudied protein targets frequently and have included examples like K_na1.1 channels as potential therapeutic targets for early-onset epilepsy and PIP5K1A as a potential target for cancers that have KRAS or TP53 mutations.⁴ Again, it is important to note that developments in technology have allowed for new modalities to be used as therapeutic agents.⁵ Such an example is the approvals of ASOs (two) by the FDA in 2016.⁶

Animal models are another approach to understand human biology. A series of challenges named systems biology verification for Industrial Methodology for PROcess Verification in Research (sbv IMPROVER) was designed to address translatability between rodents and humans.⁷ This initiative saw scientists apply computational methodologies on datasets containing phosphoproteomics, transcriptomics, and cytokine data from both human and rat tissues that were exposed to various stimuli.⁷ A key finding was that even similar computational method had diverse performance and there was no clear “best practise”.⁷ Despite this, the computational methods were able to translate some of these stimuli and biological processes between the human and rat.⁷

Despite such successes in translatability, animal studies have been considered poor predictors of human responses for a very long time. Bracken explored this idea of why animal studies are often poor predictors of human reactions to exposure.⁸ Bracken summarized the reasoning as to why there is often a lack of translatability between animal studies and human trials is that many animal experiments do not appear to be designed and performed well.⁸ There may be other issues such as limitations in methods and biases.⁸ Despite this, Denayer and Van Roy explain that animal models allow for a greater understanding of biology, disease, and medicine.⁹

Targeting the wrong target can lead to failure of the development of the drug and therefore result in cost, not just financially, but in time as well.¹ Failli and co-authors discuss that to improve this process, many developed platforms rely on estimates of efficacy that have been derived from computational methods that create scores of target-disease associations.¹ Failli and co-authors also present novel methods that consider both efficacy and safety by using gene expression data, transcriptome based, as well as tissue disease-specific networks.¹

Defining what a drug target is crucial but is not well defined.¹⁰ Santos and co-authors¹⁰ highlighted the need to keep accurate and current record of approved drugs and the target in which the drugs modulation results in therapeutic effects.¹⁰ The authors note several resources including databases and papers that contain this information but note that it is still difficult to find a complete picture.¹⁰ Owing to this, the authors attempt to annotate the current list of FDA-approved drugs (up to when published).

The understanding of what makes a good drug target is important in target identification and the drug discovery process as a whole. Gashaw and co-authors¹¹ have previously discussed what a good drug target is and also detail what the target evaluation at Bayer HealthCare involved at time of writing. The authors highlight that first a disease is identified that does not currently have treatment and then the identification of a drug target follows. Target assessment follows comprising of experimental target assessment methods, theoretical drugability assessment, and thoughts on target-related/stratification biomarkers. Alongside this adverse event evaluation and intellectual property is carefully considered.¹¹ These steps are important and investigate key areas of a target such as whether the target actually has a role within the disease of interest as well as whether the target can be modulated without adverse events.

Predicting drug target interactions is an area of great interest in the drug discovery process. Predicting such interactions can help to save time and money. The term drug prediction can be an ambiguous term. It can refer to the identification of a target of which will modulate the disease of interest (related to target focused approaches), but also can refer to the identification of a target that a drug is binding to (more related to phenotypic approaches). Unlike in targeted drug discovery approaches, phenotypic approaches do not have prior knowledge of a specific target or what role that target places within a disease.

The large number of resources available for aiding in the discovery of a drug target have been summarized by an online article by BiopharmaTrend of which listed 36 different web resources for identifying targets.¹² This list includes databases such as WITHDRAWN,¹³ which is a resource for withdrawn and discontinued drugs, DisGeNET,¹⁴ which is focused on target-disease relationships as is Open Targets.^{15, 16}

Open Targets uses human genetic and genomic data to aid in target identification and prioritization. The website describes that the Open Targets Platform has public data to aid in this task and the Open Targets Genetics Portal uses GWAS (Genome Wide Association Studies of which aims to associate particular variations with diseases) and functional genomics to identify targets. The data integration is complemented with large-scale systematic experimental approaches.¹⁵

Drug-target databases mentioned by Santos and co-authors included the Therapeutic Targets Database,¹⁷ of which focuses on known and explored therapeutic protein and nucleic acid targets, as well as the diseases, pathway information and any corresponding drugs. DrugBank,¹⁸ contains bioinformatic and cheminformatic resources for combining drug data and drug target information. Next, the authors mentioned SuperTarget¹⁹ which includes various drug information for target proteins. Finally, the authors mention the International Union of Basic and Clinical Pharmacology and British Pharmacology Society Guide to Pharmacology Database.²⁰ This resource aggregates target information and separates targets into eight different target classes. Namely, GPCRs, Ion channels, Nuclear hormone receptors, Kinases, Catalytic receptors, Transporters, Enzymes, and other protein targets. It has been argued that focusing on a class of targets that are known to have key roles can provide beneficial outcomes in early drug discovery phases.²¹ The thought is that it will reduce changes for technical failure due to the diversification of the target space, and thus available binding pockets.²¹ This in turn increases potential clinical applications.²¹

These resources are important for understand the mechanisms of disease. Understanding the mechanism of disease is an important step in identifying the targets involved. However, a difficulty with understanding disease is that regardless of the technology used, capturing the complete biological picture, with all of its complexities, is not possible in most human diseases.²² The use of integrating multiple technologies such as multiomics is used to capture a more complete picture. For example, genetic data and transcriptomic data have been combined for the purpose of using gene expression profiles to identify disease-related genes for prioritization and drug repositioning.^{23, 24}

Target identification predictions

As previously mentioned, target identification can relate to two distinct questions. First, we have a compound, what is the protein it is associated with where we have a linkage between a compound and a protein. Second, we have a disease what is the protein involved where there is a linkage between protein and disease.

Gene-disease association data have been used to predict therapeutic targets using a variety of classifiers and validating the results by mining scientific literature for the proposed therapeutic targets.²⁵ Protein-protein interactions have also been used to predict genes involved in disease.²⁶ Predicting mutations that are associated with disease is also of interest such as is demonstrated with the tool BorodaTM (Boosted RegressiOn trees for Disease-Associated mutations in TransMembrane) proteins.²⁷

Several tools have been developed to aid in drug-target prediction with many different strategies in place. For example, Bayesian approaches, such as the tool BANDIT for predicting drug binding targets²⁸ as well as deep learning approaches, such as the tool deepDTnet that is network based.²⁹ Many methods have been summarized by Mathai and co-authors in their paper discussing validation strategies for target prediction methods.³⁰ Here we have picked out a handful of key methods to discuss.

The SwissTargetPrediction tool uses multiple logistic regression to combine the different similarity values from shape and fingerprint compound similarities.³¹ HitPick is another online service³² for target prediction where a user uploads their compounds in SMILES format and uses 2D molecular similarity-based methods such as nearest neighbor similarity searching and machine learning methods that are based on naive Bayesian models (Laplacian modified).³² The chemical similarity is measured using Tanimoto coefficients. The tool SEA³³ stands for the similarity ensemble approach and performs the predictions based on the chemical similarity of the compounds. The compounds have been annotated into sets (based on the receptor that the ligand modulates) and then the Tanimoto coefficient similarity score between these sets was calculated. Another method named SPiDER³⁴ use self-organizing maps to identify targets. SLAP is a tool that uses a network approach and is able to make links between related data and these data can be of differing types.^{35, 36} Finally, PIDGIN,^{37, 38} which stands for Prediction Including Inactivity uses nearly 200 million bioactivity data points containing both active and inactive compounds from ChEMBL and PubChem for target prediction.

The validation strategies that are used in target prediction have been previously discussed and note that most strategies that are used are not accounting for biases in the data. The authors³⁰ suggest external validation of the parameterized models and compared with the results from the internal validation.³⁰

Mathai and co-authors³⁰ summarized such validation strategies for target prediction methods. The authors note that it is important to understand how the data were validated, based on the data partitioning schemes (how the data is split into train and test sets) used, as well as how the performance was measured, based on the metrics used). The authors note that performance assessment for target prediction methods do not often take into account biases in the data and that new metrics are needed. However, they do suggest several steps that can be taken to represent a more realistic view on the prediction performance. Examples of their suggestions include using a combination of metrics and methods should be used, the minima, maxima, and distributions should be reported. The use of stratified sampling and external data should be used.³⁰ The authors also discuss the validation strategies used by various tools including most of the examples shown above in Table 1.

When validating targets, methods are required to identify the most complexing targets that have been identified. MANTIS_ML⁴¹ is a tool that uses genetic screens from HTS that uses machine learning to rank disease-associated genes where the authors showed good predictive performance when applied to chronic kidney disease and amyotrophic lateral sclerosis. The tool works by compiling features from multiple sources, followed by an automated machine learning preprocessing step with exploratory data analysis and finally stochastic semisupervised learning.⁴¹ These features are derived from sources such as disease-specific resources such as CKDdb,⁴² a database containing multiomic studies of chronic kidney disease, data filtered by tissue and disease such as from the GWAS catalog⁴³ and resources that are more generic like the Genome Aggregation Database (GnomAD)⁴⁴ and the Molecular Signatures database (MSigDB),⁴⁵ of which has the purpose of combining exome and genome sequencing data from large-scale sequencing projects.⁴¹

In some cases, a compound will have been discovered to modulate a protein for a particular disease, but the actual target is not known. For example, the observed change in protein expression that is the result of a target upstream interacting with a compound. The structure of the target can also be predicted, and deep learning has been applied to such effect. The distance between pairs of residues can be predicted via neural networks even with fewer homologous sequences.⁴⁶

Gene prioritization methods

After identifying a list of genes, it is often useful to prioritize this list to focus the search, ideally starting with the most likely target gene. Methods for gene prioritization vary in their approaches from network-based methods to matrixes. Moreau and Tranchevent⁴⁷ previously reviewed computational tools for prioritizing candidate genes and noted that the overall aim is to ingrate data for the purpose of identifying the most promising genes for validation. It often utilizes prior knowledge such as in the form of keywords that describe the phenotype or known gene associations to a particular process or phenotype. Furthermore, with a number of gene prioritization tools available, methods to benchmark such gene prioritization tools have been undertaken with the authors⁴⁸ (Guala and Sonnhammer) recommending the use of Gene Ontology⁴⁹ data alongside the tool FunCoup.^{50, 51} Because of the cost of identifying and validating new genes that are involved in diseases, many of the algorithms have had the performance measured by using known disease genes.⁵²

Many of these tools rely on different types of data and algorithms to perform gene prioritization. For example, the use of genomic data fusion, such as is the case with the gene prioritization tool ENDEAVOUR.^{53, 54} Other approaches include Beegle,⁵⁵ of which mines the literature to find genes that have a link to the search query. It then uses ENDEAVOUR to perform the gene prioritization step. Cross-species approaches have been proposed such as in the case of GPSy.⁵⁶ The use of Bayesian matrix factorization using both genomic and phenotypic information has also been implemented.⁵⁷ Network based approaches are also used in gene prioritization tasks.

Such network-based tools included Rosalind which uses knowledge graphs to predict linkages between disease and genes that uses data integration (in graph format) and relational inference.⁵⁸ MaxLink, which is a guilt-by-association network which aims to identify and rank genes in relation to a given query list.⁵⁹ NetRank is based on the algorithm used in the ToppGene Suite^{60, 61} and uses the Page Rank with priors algorithm where a random surfer is more likely to end up in nodes that are initially relevant.^{60, 62} Other algorithms include, NetWalk, which is based on the random walk with restarts algorithm which operates by walking from a node to another randomly but at any time the walk can be restarted (based on probability).⁵²Using not only the shortest path interaction but also taking into the global interactome structure has been shown to have performance advantages.⁵² NetProp is based on the network propagation algorithm of which was designed to address the issue of many approaches using only local network information.

GUILD⁶³ (Genes Underlying Inheritance Linked Disorders), uses four network-based disease-gene prioritization algorithms, namely, NetShort (considers node importance for a given phenotype will have a shorter distance to other seed nodes within the network), NetZcore (normalizes the scores of nodes in the network in relation to a set of random networks with similar topology and therefore considers the relevance of the node for a given phenotype), NetScore (considers multiple shortest paths from source to target) and NetCombo (averaged normalization score for each of the other three prioritization algorithms). GUILD also includes NetRank, NetWalk, NetProp, and fFlow of which is based on the FunctionalFlow algorithm.^{60, 64} This algorithm generalizes the guilt-by-association method by considering groups of proteins that may or may not physically interact with each other.⁶⁴ In fFlow, the scores annotated from nodes with higher scores “flow” toward nodes with lower scores depending on the edge capacity.⁶⁰ The authors who describe the FunctionalFlow algorithm compare it with three others, Majority^{64, 65} (consider all neighboring proteins and calculate how many times each annotate occurs for each protein), Neighborhood^{64, 66} (considers all proteins within a radius and then for each function consider if it is over-represented), and GenMultiCut. The authors note there are different considerations of GenMultiCut with two instances agreeing that functional annotations on interaction networks should be made with the purpose of reducing the occurrences of different annotations that are associated with neighboring proteins.⁶⁴

Machine learning is also used in gene prioritization tasks. In the case of Amyotrophic Lateral Sclerosis, Bean and co-authors published a knowledge-based machine learning approach that used a variety of available data to predict new genes involved in Amyotrophic Lateral Sclerosis.⁶⁷ Isakov and co-authors used machine learning gene prioritization tasks to identify risk genes for inflammatory bowel disease and the authors performed a literature search to identify whether particular genes had been published alongside inflammatory bowel disease.⁶⁸ Prioritizing loci from genome wide association studies (GWAS) studies is another important prioritization task that machine learning has aided in.⁶⁹ Nicholls and co-authors reviewed the machine learning approaches used in loci prioritization and noted that GWAS results require aid in sorting through noise as well and understanding which are the most likely causal genes and variants.