Use of big data in drug development for precision medicine

Summary

Drug development has been a costly and lengthy process with an extremely low success rate and lack of consideration of individual diversity in drug response and toxicity. Over the past decade, an alternative “big data” approach has been expanding at an unprecedented pace based on the development of electronic databases of chemical substances, disease gene/protein targets, functional readouts, and clinical information covering inter-individual genetic variations and toxicities. This paradigm shift has enabled systematic, high-throughput, and accelerated identification of novel drugs or repurposed indications of existing drugs for pathogenic molecular aberrations specifically present in each individual patient. The exploding interest from the information technology and direct-to-consumer genetic testing industries has been further facilitating the use of big data to achieve personalized Precision Medicine. Here we overview currently available resources and discuss future prospects.

1. Introduction

Drug development has been a protracted and costly multi-step process. Traditionally, candidate drug targets are defined and assessed based on a reductionist approach or a close-world assumption (CWA), relying on a highly simplified view of biology and restricted to manipulating one molecule or molecular pathway. Due to our limited understanding of biology at the systems level, each step in drug discovery and development has been suffering from a high level of uncertainty, resulting in an extremely low success rate. It takes billions of investment dollars and an average of about 9–12 years to bring a new drug to the market [1, 2]. Despite this, failures throughout the drug development pipeline make improving drug research and development (R&D) productivity an overarching priority. In fact, according to one study, the clinical approval success rate of drugs originating from pharmaceutical companies in the United States was 16% [3, 4].

Current challenges in drug development include (i) a limited capability to comprehensively characterize and/or monitor biological contexts of interest, (ii) a lack of adequate experimental models to test candidate drugs/perturbations and their clinical relevance, (iii) a range of accompanying multiple off-target effects of the drugs that are generally unknown (can be referred to as “polypharmacology”), and (iv) the genetic and environmental diversity across individuals in drug metabolizing enzymes, on/off-target molecules, and other drug response-related factors. The increasing rate of attrition at the late stages of drug development despite mounting R&D spending underscores the need for innovative and network-based approaches to drug discovery.

The recent emergence of “big data” in biology has fundamentally revolutionized how we study molecular biology and drug development. We are now at the tipping point to resolve some of the aforementioned drug discovery obstacles and facilitate the delivery of new drugs into clinical practice as evidenced by the initiation of the Precision Medicine Initiative [5]. In this review, we overview recent attempts of utilizing big data specifically in the field of drug development, and discuss exciting future prospects.

2. Big data

The word “big data”, generally refers to an approach where comprehensive data are first collected in an unbiased manner, without a priori hypotheses, and subsequently analyzed by data mining algorithms to formulate novel ideas [6]. In molecular biology, this approach has been led by the analysis of nucleic acid and protein sequences accumulated in public databases as well as DNA microarray-based genome-wide transcriptome and DNA structural variation data [7]. With unprecedented levels of development of genomic assay technologies and the massive growth of global information storage and sharing technologies, the big data approach has been further expanded and integrated with other data types such as epigenomic features, bio-ontologies, chemical structures, protein structures, electronic medical records, clinical trial registries, and clinical toxicity databases, at human population level. In parallel, data analysis algorithms and infrastructure tailored to analyze this type of data have been actively developed [8–10].

A major challenge in the analysis of big data has been to adapt statistical techniques to the added scale and complexity of datasets. For example, correction of p-values for multiple hypothesis testing (MHT) is a key issue in order to control false discovery, while minimizing false negatives [11]. In addition, high-dimensional data often requires, dimensionality reduction using data projection techniques such as extraction of principal components, non-negative matrix factorization, and molecular pathway-level abstraction of the data [12, 13]. A variety of unsupervised and supervised machine learning algorithms has been adapted to the analysis of biomedical big data to identify unknown subpopulations of diseases, elucidate novel disease targets, and predict clinical outcome based on genomic and clinical data [14].

3. Big data analysis for drug development

3.1 Therapeutic target discovery

Our rapidly expanding capability to comprehensively characterize biomolecules at the whole genome level has led to a paradigm shift in the field of therapeutic target discovery. Over the past decade, the analysis of genomic DNA structural alterations within or near protein-coding regions as well as transcriptome profiles have led to the development of novel methodologies to enable big data-based, unbiased target exploration. Now we can also measure genome-wide DNA methylation, histone protein modifications, splicing variants, transcription factor binding sites, protein abundance and modifications such as phosphorylation, non-coding RNA expression, and structural alterations in non-coding regions of genomic DNA. These technological developments have resulted in ever-growing big data repositories for biomedical research and drug development that allow hypothesis-free, unbiased target discovery (Table 1).

Table 1.

Resources for big data-driven therapeutic target identification

Type	Focus	Resource	Assay/datatype	Species	URL
Projects/initiatives	Genome, phenome, microbiome	Personal Genome Project (PGP)	DNA-seq, clinical phenotypes	Human	www.personalgenomes.org/Harvard
	Genomic DNA variations across populations	1000 Genomes	DNA-seq	Human	www.1000genomes.org/data
	Genomic DNA variations across populations	HapMap	SNP array	Human	hapmap.ncbi.nlm.nih.gov/downloads/index.html.en
	Regulatory elements in genome	The Encyclopedia of DNA Elements (ENCODE)	ChlP/RNA/DNA-seq, DNA microarray	Various	www.encodeproiect.org
	Multi-omic data in cancer	The Cancer Genome Atlas (TCGA)	DNA-seq, RNA-seq, methylation array, clinical phenotypes	Human	tcga-data.nci.nih.gov/tcga
	Multi-omic data in cancer	The International Cancer Genome Consortium (ICGC)	DNA-seq, RNA-seq, methylation array, clinical phenotypes	Human	dcc.icgc.org
	Molecular signatures of gain/loss of genes	Library of Integrated Network-Based Cellular Signatures (LINCS)	Bead-array with informatic inference	Human	www.lincsproiect.org/data/tools-and-databases
	Expression/DNA variants across organs	Genotype-Tissue Expression (GTEx)	RNA-seq, SNP array	Human	www.gtexportal.org/home
	Expression/DNA variants in cancer cell lines	Cancer Cell Line Encyclopedia (CCLE)	Expression array, targeted DNA-seq	Human	www.broadinstitute.org/ccle
	Presence/variations of microorganisms in human	Human Microbiome Pro) ect (HMP)	DNA-seq, reference genome, clinical meta-data	Microorganisms	hmpdacc.org/resources/tools_protocols.php
Data repository	Somatic DNA mutations in cancer	Catalogue of Somatic Mutations in Cancer (COSMIC)	Various omic types from literature/databases	Human	cancer.sanger.ac.uk/cosmic
	Various	Gene Expression Omnibus (GEO)	Various omic types with or without clinical phenotypes	Various	www.ncbi.nlm.nih.gov/geo
	Various	ArrayExpress	Various omic types with or without clinical phenotypes	Various	www.ebi.ac.uk/arrayexpress
	Expression/DNA variations, phenotypes	European Genome-phenome Archive	Various omic types, clinical phenotypes	Human	www.ebi.ac.uk/ega/home
	Annotated gene sets	Molecular Signature Database (MSigDB)	Gene expression, knowledgebase, genomic/genetic structural	Various	www.broadinstitute.org/msigdb
	Expression/DNA variations, phenotypes in cancer	Oncomine	Various omic types with or without clinical phenotypes	Human	www.oncomine.org/resource/login.html
	Multi-omic data in mouse models	Mouse Genome Informatics (MGI)	Various omic types from literature/databases	Mouse	www.informatics.jax.org
	Clinically relevant DNA variations	Clinical Genome Resource (ClinGen)	Various DNA variant assays, clinical phenotypes	Human	www.clinicalgenome.org

Multiple cancer genome sequencing projects have yielded a catalog of somatic DNA structural alterations such as single nucleotide variations, small insertions and deletions, copy number alterations, and genomic translocations for each cancer type as potential cancer drivers and new therapeutic targets [15, 16] (Table 1). Detailed catalogs of transcriptional regulation machinery such as genome-wide transcription factor binding sites for major transcription factors as well as epigenomic marks such as histone protein acetylation have been also developed to help identify biologically functional drug targets [17–21]. Transcriptome databases have been extensively utilized. Together with the rapid expansion of highly selective chemical libraries, these databases have substantially shortened the potential time period required from target discovery to clinical deployment (traditionally a few decades).

An example of such big data-driven translational success story pertains to the treatment of non-small cell lung cancer (NSCLC) patients. Approximately 5% of NSCLC cases were identified to harbor somatic rearrangements in the anaplastic lymphoma kinase (ALK) gene [22], and a selective ALK inhibitor, crizotinib, showed improved progression-free survival in the ALK gene rearrangement-positive NSCLC patients [23, 24]. Based on these results, crizotinib as well as companion biomarker assays detecting the gene fusion were FDA-approved and incorporated into clinical practice guideline, replacing the traditional combination chemotherapy only a few years after the initial discovery of the ALK rearrangements [25]. Similarly, pharmacologically targetable epidermal growth factor receptor (EGFR) gene mutations have been successfully utilized in the clinical deployment of selective EGFR inhibitors [26].

Despite these promising achievements, the relatively low prevalence of such somatic gene mutations in cancer often obscures the therapeutic benefit observed when all, unselected, patients enrolled in clinical trial are analyzed. Also, prediction of functional consequences of such genetic alterations remains challenging as recently highlighted in a follow-up case study of a patient with ALK-rearranged lung cancer, who developed a crizotinib resistance-conferring mutation, requiring the use of a third generation ALK inhibitor, but subsequently resensitizing to crizotinib when a further mutation developed negating the initial resistance [27]. This example suggests that the therapeutic consequence of molecular targeted drug treatment depends on each specific mutation even if the same target gene is mutated. In other words, the true therapeutic effect of a molecular-targeted drug on a specific molecular aberration in the target will never be known unless it is tested in patients characterized for the molecular target. However, it is practically infeasible to test the astronomical number of potential combinations of a drug-target mutation in the traditional format of a clinical trial, testing each single therapy at a time. Several attempts have been proposed and/or tested to overcome the challenges with new clinical trial designs. For example, integrating information regarding predictive biomarker-based enrichment of potential drug responders to clinical trials is expected to improve statistical power, while reducing sample size [28]. This approach has been further expanded to a variety of specific scenarios such as two-stage patient enrichment to cope with biomarker misclassification [29] and enrichment for time to event endpoint [30]. Nevertheless, running a clinical trial for each one of drug-target pairs is unrealistic given the vast number of such pairs. In addition, evaluation of multiple drugs in parallel in a trial has been practically infeasible due to many reasons, including conflict in intellectual property. To address this long-standing challenge, the umbrella trial design, first screens multiple biomarkers (most of them are expected to be less prevalent markers) at a central platform/infrastructure, then distributes each patient to a matched sub-study, assessing an agent expected to target the positive biomarker, focusing on a single tumor type or histology. The basket trial design expands the strategy to multiple tumor types sharing the same molecular aberrations to increase the chance of finding therapies for rare genetic variants/diseases [31]. The National Cancer Institute-Molecular Analysis for Therapy Choice (NCI-MATCH) is a basket trial, in which up to 3,000 cancer patients will be screened for “actionable” DNA mutations and assigned to either of 20 treatments [32]. Exploration of “exceptional responders” followed by mechanistic interpretation is another alternative approach [33]. Co-clinical trial, simultaneously assessing therapeutic effect in human and mechanisms in animal model, was also proposed [34]. Yet another approach to account for intersubject variability in therapeutic response is the n-of-1 trial, where the study focuses on response of an intervention in one subject [35, 36]. Detailed data is collected for one participant during an intervention with the possibility of interrogating mechanisms of extreme drug response, testing another intervention after a suitable “wash-out” period, etc. This approach allows for example to explore molecular effects of a new compound at early stages of drug development, but also to draw inferences about population-wide effectiveness when combining multiple n-of-1 trials. In addition, this also allows taking into account individual factors linked to toxicity and may help to reduce attrition rate due to drug safety problems in subsequent traditional larger clinical trials. These new strategies have been rapidly evolving in the field of cancer therapeutics, which could be applicable to similarly heterogeneous polygenic diseases.

Further, the genome sequencing initiatives have now shifted the focus to population-based studies that provide deeper insight into rare pathogenic variants [37] (Table 1). The rapid dissemination of sequencing as part of routine clinical practice is further expanding the resource of population-level genomic variation data [38]. Lastly, there is a progressive transition to the use of electronic medical record (EMR) for unbiased mining of clinical phenotypic information, which has the potential to identify novel disease subtypes and potential therapeutic targets [39], although the establishment of a unified and interchangeable infrastructure is still challenging [40].

One of the key challenges facing researchers will be to properly integrate the multiple layers of data into a manageable and systematic unit for better informing drug development and patient care. Integration of different “omics” data (for example, genomic, epigenomic, transcriptomic and proteomic) with clinical phenotype information recorded in EMR is particularly important to delineate clinically-relevant pathogenic molecular alterations as drug targets. For instance, topology-based patient-patient networks based on integrative clinical and genomic data from 11,000 individuals could identify 3 novel subgroups of type 2 diabetes [39]. Properly leveraged and harnessed, the integration of dynamic genomic and phenomic information, medical data and scans, as well as socioeconomic and environmental variables will revolutionize clinical medicine in the near future [41].

3.2 Drug discovery

Big data resources related to chemical substances have been utilized for computational virtual screening of compounds. Chemical structure-based exploration of small molecule compounds is a traditional way to identify refined compounds with similar biological activity. This approach is based on an assumption that compounds with similar structure share similar biological properties, although it does not hold in many cases (Table 2). Nevertheless, structure family-based prioritization and/or pooling of compounds to be screened is a commonly used strategy to reduce complexity and cost of drug screening. When structure of target protein is available together with biochemical properties of each amino acid residue, informatic estimation of physical interaction with small molecules has been widely performed as another popular strategy of drug discovery. Previously determined protein structures from individual studies and large-scale projects/initiatives such as the Protein 3000 Project [42], have been assembled and made avialable in public databases (Table 2).

Table 2.

Resources for big data-driven drug identification.

Type	Focus	Resource	Assay/data type	Species	URL
Data repository	Chemical property/structure/ biological function	ChemBank	Chemical properties, phenotypic readouts	Various	chembank.broadinstitute.org
		PubChem	Chemical properties, phenotypic readouts	Various	pubchem.ncbi.nlm.nih.gov
		ChEMBL	Chemical properties, phenotypic readouts	Various	www.ebi.ac.uk/chembl
		Phamacogenomics Knowledgebase (PharmGKB)	Drug annotations, drug-gene association	Human	www.pharmgkb.org
		KEGG DRUG	Drug annotations, drug-gene association		www.genome.jp/kegg/drug
		DrugBank	Drug and drug target information	Various	www.drugbank.ca
		FDA Orange Book	List of FDA-approved drugs	Human	www.accessdata.fda.gov/scripts/cder/ob
	Protein properties	UniProt/Swiss-Prot	Protein sequence, structure, function, ontology	Various	www.uniprot.org
		Protein Data Bank (PDB)	Protein sequence, structure, function, ontology	Various	www.wwpdb.org
		Max-Planck Unifed Proteome Database (MAPU)	Protein sequence, structure, function, ontology	Human, mouse	mapuproteome.com
		ExPASy	Protein sequence, structure, function, ontology	Various	www.expasy.org
		Protein Information Resource (PIR)	Protein sequence, structure, function, ontology	Various	pir.georgetown.edu
		CATH	Classification of protein structures	Various	www.cathdb.info
		Plasma Proteome Database (PPD)	List of plasma/serum proteins	Human	www.plasmaproteomedatabase.org
	Protein-protein/chemical/genetic interaction	Database of Interacting proteins (DIP)	Experimentally determined protein-protein ineraction	Various	dip.mbi.ucla.edu/dip
		STITCH	Chemical-protein interaction	Various	stitch.embl.de
		BioGRID	Protein, genetic, chemical interactions/associations	Various	thebiogrid.org
		Therapeutic Target Database (TTD)	Protein, genetic, chemical interactions/associations	Human	database.idrb.cqu.edu.cn/TTD
		Potential Drug Target Database (PDTD)	Protein, genetic, chemical interactions/associations	Human	www.dddc.ac.cn/pdtd
Tools		NIH Small Molecule Repository	Compound libraries for screening	-	nihsmr.evotec.com/evotec
		ViCi	in silico ligand-based drug design	-	www.embl-hamburg.de/vici

Similarly, inhibition of protein-protein interaction can also be comptationally explored [43]. Although X-ray crystallography was historically applicable to a limited classes of proteins, recent technical developments have been expanding our capability to crystalize membrane and soluble proteins and complexes proteins as well as small moleculaes [44]. Computational protein structure prediction from amino acid sequence has been an outstanding problem in biomedical research that has been tackled by long-standing community efforts such as The Critical Assessment of protein Structure Prediction (CASP) (www.predictioncenter.org). Imaging reconstruction-based methods such as cryo-electron microscopy (cryo-EM) are potential alternative approaches addressing the problem from a completely different angle [45]. Large-scale machine learning predictions of on/off-target effects based on structure of receptor ligand (instead of target receptor protein itself) was proposed to capture ligand-based similarities of otherwise unrelated proteins [46, 47].

Transcriptome is increasingly recognized as a common space, where the functional consequence of various types of molecular dysregulations and aberrations as well as drug treatment can be directly compared and linked based on the similarity of transcriptional changes measured as gene signatures [48]. This approach, systematically merging disease-associated gene signature(s) and drug signatures, has enabled high-throughput search of compounds potentially reversing the disease gene signature and therefore treating the disease, as an alternative to the traditional target-based drug discovery [49]. The simple computational unbiased screening of gene signature similarity often identifies unexpected compounds outside known mechanisms of drug action and biology [50], and has been successfully utilized for a variety of diseases such as cancer and inflammatory diseases [51–53]. Application of this method for clinically approved compounds or compounds with halted development for certain indications at early stages could identify other indications of drugs already used in clinical practice. This “drug repurposing” approach has increasingly gained popularity [43, 54] and is getting more attention from pharmaceutical industry and regulatory authorities because it could substantially reduce the cost and time for drug development [55]. A database of transcriptome signatures of >1,000 Food and Drug Administration (FDA)-approved compounds (connectivity map) has enabled convenient compound search similar to the internet search engines [56]. The database is being expanded to bioactives and tool compounds beyond the FDA compounds as well as genetic manipulation (open reading frame [ORF] expression and shRNA-based knockdown) in multiple cellular contexts under the National Institute of Health (NIH) Library of Integrated Network-based Cellular Signatures (LINCS) program to further utilize the approach to probe the highly inter-connected wiring of cellular signaling. Addition of transcriptome profiles of other class of compounds such as diversity-oriented synthesis (DOS) library and collections of natural products could further increase the chance of novel drug discovery and facilitate drug development [57]. Another unique approach is to use similarity in side-effect profiles [58]. The assembly of large compound libraries has enabled high-throughput compound library screening, and yielded big data of phenotypic readouts of pharmacological interventions in addition to transcriptome such as cell survival, abundance of target molecules, enzymatic activity, and morphological changes, many of which are deposited in the public domain [59–61] (Table 2). Similar resources are expected to evolve into more integrated resources for systematic therapeutic discovery [62].

3.3 Drug toxicity

From a pharmaco-therapeutics perspective, one of the major goals of precision medicine is to improve the therapeutic index of drug products. However, a recent report has underlined that only approximately one in ten drugs entering phase I clinical trials will eventually be approved by the FDA, an approval rate that was even lower for new molecular entities reaching only 7.5% rate of FDA approval [4]. Even when examining the filings for new drug applications and biologic license application, a relatively advanced stage in drug development, safety concerns were the root cause for drug development suspension in 31% of programs, the second cause behind a lack of efficacy [4]. Therefore, there is an increasing interest to develop novel early detection systems for detecting drug toxicity and in silico approaches to model toxicity provide a high-throughput and low-cost alternative to currently accepted drug development processes. These methods aim to complement in vitro and in vivo toxicity tests to reduce the cost and time of toxicity testing, minimize the requirement to animal testing and improve overall safety assessment.

An ever-expanding array of methodologies is used for in silico toxicity modeling. Although it is beyond the scope of this text to give a comprehensive overview of the field, we underline major themes and methodologies currently being used in this dynamic field. Quantitative structure-activity relationship (QSAR) denotes a family of models seeking a relationship between molecular characteristics of a given molecule and its toxicity, under the assumption that chemicals that fit the same QSAR model may work through the same mechanism [63, 64]. Advantages of QSAR include ease of interpretation and modeling of categorical and continuous toxicity endpoints but it requires large datasets and cannot be extrapolated between species. QSAR models are implemented in a number of tools including OECD QSAR toolbox and TopKat (Table 3). Alternatively, “expert systems” are software programs codifying a set of rules such as Oncologic Cancer Expert System, Toxtree, and DEREK Nexus [65–67]. An example of expert systems includes structural alerts, a list of chemical structures that indicate or are associated to toxicity. Other models include read-across prediction, using compounds with known toxicity and similar chemical structures, to predict toxicity of investigational compounds.

Table 3.

Resources for computational prediction of drug toxicity

Focus	Resource	Assay/Data type	URL
Computational prediction of drug toxicity	DEREK Nexus	Toxicological profile based on structure	www.lhasalimited.org/products/derek-nexus.htm
	ToxTree	Toxicological profile based on structure	toxtree.sourceforge.net
	HazardExpert	Toxicity profile based on toxic fragments	www.compudrug.com/hazardexpertpro
	TOPKAT	Toxicological profile based on structure. Predicted ADMET properties.	accelrys.com/products/collaborative-science/biovia-discovery-studio/qsar-admet-and-predictive-toxicology.html
	CASE Ultra	Toxicological profile based on structure	www.multicase.com/case-ultra
	OECD QSAR	Toxicological profile based on structure	www.oecd.org/chemicalsafety/risk-assessment/theoecdqsartoolbox.htm
Database of bioactive molecules and targe	ChEMBL	Chemical properties, phenotypic readouts	www.ebi.ac.uk/chembl
	DrugBank	Drug and drug target information	www.drugbank.ca
	Drugs@FDA Database	FDA-approved drugs	www.fda.gov/Drugs/InformationOnDrugs/ucm135821.htm
	PubChem	Chemical properties, phenotypic readouts	pubchem.ncbi.nlm.nih.gov
	SWEETLEAD	Chemical structure database	simtk.org/home/sweetlead
	The NCGC Pharmaceutical Collection (NPC)	Chemical structure database	tripod.nih.gov/npc
	Chemical Entities of Biological Interest (ChEBI)	Small compound database	www.ebi.ac.uk/chebi
Toxicology databases	SIDER	Adverse drug reaction database	sideeffects.embl.de
	Comparative Toxicogenomics Database	Chemical-protein and chemical-phenotype interactions	ctdbase.org
	Environmental Protection Agency Aggregated Computational Toxicology Resource (ACToR)	Chemical toxicity database	actor.epa.gov
	FDA Adverse Event	Adverse drug reaction database	www.fda.gov/Drugs/GuidanceComplianceRegulatoryInformation/Surveillance/AdverseDrugEffects
	OpenTox	Multiple toxicological resources (data, computer models, validation and reporting)	www.opentox.org
	Phamacogenomics Knowledgebase (PharmGKB)	Drug annotations, drug-gene association	www.pharmgkb.org
	T3DB	Toxin-target database	www.t3db.ca
	TOXNET	Toxicology resources	toxnet.nlm.nih.gov
Chemical property/structure	Chembank	Chemical properties, phenotypic readouts.	chembank.broadinstitute.org

Adsorption, distribution, metabolism and excretion (ADME) properties are key components to the drug for understanding therapeutic effects, but also for better understanding toxicity related, for example, to drug accumulation and metabolism. Among the components of ADME, metabolism has been among the most studied, due to its crucial role in drug toxicity [68]. Computational approaches to assess drug metabolism is a complex process, but can be generally classified as specific or comprehensive tools [69]. Specific approaches focus on specific molecules of pathways, while comprehensive tools take a more general approach and apply to a wider variety of biological systems. The methodologies used in modelling metabolic systems include expert systems, QSAR, data mining and others, but it has become clear that the key is the integration of a number of different methods and resources. Resources useful for predicting toxicological effects of metabolites include DEREK Nexus, ToxTree and HazardExpert among others [66, 67, 70]. Increasing efforts attempt to integrate ADME predictions into pharmacokinetic and pharmacodynamics models. These models attempt to calculate concentrations at a given time (pharmacokinetics) or the effect of a drug at a given concentration (pharmacodynamics) and therefore quantify ADME processes. Toxicokinetic studies attempt to correlate drug concentrations in various tissue compartments to the timing of toxic responses [71]. Although we do not directly consider these here, it is notable that modelling of off-target side-effects has also been attempted [72]. One of the recurring problems in in silico modelling of toxicity is the lack of data available to enter in these data-hungry models, in particular in the case of heterogeneous compound modelling. Further addition of experimental data and additional integration of these different approaches will ensure the growth of the field of drug toxicity modelling [73].

An additional toxicological concern that has raised attention in the past decades is the presence of pharmaceuticals in the environment [74]. For example, environmental toxicity of ethinylestradiol has been linked to feminization of male fish in effluent-dominated rivers [75] and altered behavior and feeding rate of the wild fish Perca fluviatilis was shown to be induced by oxazepam at environmentally relevant concentrations [76]. QSAR-based modelling and quantitative structure-toxicity relationships (QSTR) has been used to model ecotoxicity of organic chemicals and fungicides [77, 78]. In addition, the US Environmental Protection Agency (EPA) has developed a high throughput screening program, ToxCast, to predict potential for toxicity and environmental hazard of various chemicals [79]. This approach has been used to develop a computational network model that integrates in vitro screening assays measuring estrogen receptor activation and applied to a library of 1,812 commercial and environmental chemicals identified 52 environmental chemicals predicted to be strongly activate or inhibit the estrogen receptor and therefore prioritize environmental chemicals for additional in vivo endocrine testing [80].

3.4 Combination therapy

Combination drug therapy was developed in the mid-20^th century, in the context of cancer chemotherapy, to increase effectiveness and mitigate side effects of anti-cancer therapy [81] [82]. However, in recent decades, with an increased focus to treating disease of multifactorial etiology, such as diabetes, obesity, and cardiovascular disease it is becoming more apparent that the ‘one drug-one target’ approach may be over simplistic, and that combination drug therapy may be reexamined, in a broader perspective than only cancer therapy [83]. In addition, linked to the emergence of modern genomics and a systems-medicine approach, the classic “mechanism of action” has transitioned to a broader “signature”-based prediction achieving powerful insight to examine drug mechanism [84].

Potential mechanisms of action of combination drug therapy include complementary actions where 2 or more drugs target multiple targets along the same protein or pathway, anticounteractive actions where a drug targets the biologic response to a first drug and facilitating actions where the second drug promotes the activity of the first drug. Recent successful examples of combinatorial drug design include targeting multiple immune checkpoints in melanoma immunotherapy [85], the combination of two human epidermal growth factor receptor 2 targetting drugs, pertuzumab and trastuzumab with docetaxel in breast cancer [86] and the combination of a BRAF inhibitor (dabrafenib) and a MAPK/ERK kinase (MEK) inhibitor (trametinib) for melanoma [87].

Strategies for computational screening for combinatorial drug treatment are still being developed but include methods based on empirical models, for example network-based methods centered around the concept of gene neighborhood or methodologies based on physiological models in cases where the pathophysiology of a disease process is well understood [83]. Nevertheless, despite a better understanding of these approaches, multiple challenges remain. In particular our limited biological understanding of the complex biological systems involved and the increased potential of toxicity by using combinatorial drug therapy has in part limited this approach.

4. Expert commentary and five-year view

Big data-driven drug development has been gaining popularity and rapidly evolving over the past decades. The field started from proof-of-concept studies, followed by the development of new methodologies spanning multiple disciplines, accumulated experimental data for public use and testing, and is now further growing to cover more specific biological contexts and clinical scenarios for better applicability in real-world drug development challenges. Big data-driven drug development will facilitate identification of therapeutic strategies for rare subtypes of common diseases, rare orphan diseases, and disease conditions specific to genetically minor populations. The direct-to-consumer genetic testing industry has generated multiple large-scale population-level genetic and clinical databases outside traditional medical institutions and clinical practice [88]. The ever-growing interest and involvement of the information technology (IT) industry is transforming the field by providing infrastructure for data storage, sharing, and analysis, even without the involvement of traditional medical institutions, via wearable devices directly connected to cloud-based commercial storage/analysis centers [89, 90]. For instance, as part of the Precision Medicine Initiative, the NIH is seeking to create a longitudinal cohort of 1 million or more Americans, with extensive characterization of biologic specimens and sophisticated exposure assessment using mobile devices or wearable sensors [91]. The integration of this data with the multitude of other data points generated for each individual may allow to achieve the goal of precision medicine, “to work together toward development of individualized care [92].”

Key issues.

• Accumulated drug knowledge bases, experimental data, clinical data, and exponentially expanding IT infrastructure now enable big data analysis to systematically explore new therapeutics.

• Omics data accumulation covering a variety of biological systems/contexts and variations across human populations and individuals, is becoming a resource to demonstrate the proof-of-concept of Precision Medicine.

• Experimentally or computationally determined chemical, genetic, and proteomic interactions connect cross-domain knowledge, and facilitate drug discovery.

• Not only chemical/protein structures and interactions, but also transcriptome, enzymatic activity, ligands, other functional readouts, and even clinical toxicities are now available as tools for high-throughput identification of therapeutic targets and strategies.

• Drug toxicity prediction may bypass less-reliable rodent models, through the leveraging of big data-format clinical phenotypic information.

• Rapidly increasing involvement of direct-to-consumer genetic testing as well as information technology industries has been drastically accelerating the big data-based approach, further promoting individualized precision medicine and improved health.