Recommended Guidelines for Developing,Qualifying,and Implementing Complex In Vitro Models (CIVMs) for Drug Discovery

Abstract

The pharmaceutical industry is continuing to face high research and development (R&D) costs and low overall success rates of clinical compounds during drug development. There is an increasing demand for development and validation of healthy or disease-relevant and physiological human cellular models that can be implemented in early-stage discovery, thereby shifting attrition of future therapeutics to a point in discovery at which the costs are significantly lower. There needs to be a paradigm shift in the early drug discovery phase (which is lengthy and costly), away from simplistic cellular models that show an inability to effectively and efficiently reproduce healthy or human disease-relevant states to steer target and compound selection for safety, pharmacology, and efficacy questions. This perspective article covers the various stages of early drug discovery from target identification (ID) and validation to the hit/lead discovery phase, lead optimization, and preclinical safety. We outline key aspects that should be considered when developing, qualifying, and implementing complex in vitro models (CIVMs) during these phases, because criteria such as cell types (e.g., cell lines, primary cells, stem cells, and tissue), platform (e.g., spheroids, scaffolds or hydrogels, organoids, microphysiological systems, and bioprinting), throughput, automation, and single and multiplexing endpoints will vary. The article emphasizes the need to adequately qualify these CIVMs such that they are suitable for various applications (e.g., context of use) of drug discovery and translational research. The article ends looking to the future, in which there is an increase in combining computational modeling, artificial intelligence and machine learning (AI/ML), and CIVMs.

Keywords

complex in vitro model 3D cell culture organoids spheroid microphysiological systems 3D bioprinting efficacy functional genomics screening safety

Introduction

The pharmaceutical industry is looking for opportunities to significantly accelerate drug discovery, because the financial impact of late-stage failures of drug candidates has driven the demand for faster and more predictive in vitro and in silico models. The industry demand is focused on development and validation of disease-relevant and physiological human cellular models that can be implemented in early-stage discovery, thereby identifying risks early to enable us to focus on molecules that are less likely to fail through efficacy or safety concerns. The current drug discovery process involves lengthy and costly lead discovery and optimization campaigns, often on poorly validated targets in cellular models with weak translational relevance to human disease. Between 2005 and 2014, the primary reason for termination was lack of human efficacy at phase II/III (35%), and in the preclinical/discovery context this was attributed to animal and safety tolerability (47%).¹ This exemplifies an inability to effectively and efficiently reproduce human disease-relevant states to steer target and compound selection. Therefore, a fundamental question is: How do we capture the human biological complexity of disease states in robust translational in vitro assays to mitigate late-stage termination of drug discovery programs?

Complex in vitro models (CIVMs) can be used to better tackle the challenges of understanding the right target, right tissue, right safety, and right patient.² For this perspective article, we define CIVMs as either spheroids, organoids, three-dimensional (3D) bioprinted tissue, organs-on-a-chip or a microphysiological system (MPS), or a multi-organ system or human body on a chip (Fig. 1 and Table 1). Several companies including GlaxoSmithKline (GSK) and AstraZeneca have reported that choosing genetically validated drug targets can improve the success rates in the clinic up to twofold.³ For that to occur, well-qualified cellular disease models need to be created to link the right target with the right tissue and disease setting. Moreover, having the right safety profile will also depend on having CIVMs that are complementary to in vivo models that show clear safety margins in healthy and possibly diseased tissue and the ability to understand any target liability and secondary pharmacology.

Figure 1.

Examples of three-dimensional cell culture tools and technologies. Going clockwise, there is increasing complexity and cost.^78–85

Table 1.

Various Complex In Vitro Models That Could Be Applied to Drug Discovery.

	Organ Structure and Functionality (Representative of Native Human Tissue)	Multicell	Throughput	Amenable to High-Content Screening	Drug Discovery Phase(s)
Spheroids	•	•••	•••	•••	All
Organoids	••	•••	••	••	All
Microphysiological systems	•••	•••	•	•	Target identification and validation to candidate selection [pharmacokinetic/pharmacodynamic (PK/PD)]
Multi-organ or human body on a chip	•••	•••	•	•	Candidate selection (PK/PD)
Tissue slices	•••	•••	•	•	Target identification and validation, and lead optimization to candidate selection

•

: Low; ••: medium; •••: high.

The MPS environment is challenging due to the plethora of innovative in vitro technologies (3D cell cultures, spheroids, organoids, matrices, microfabrication devices, bioprinting, and artificial scaffolds) that are being explored and developed by an increasing number of academic institutions, start-ups, and biotech and pharmaceutical companies. There is, however, a gap between development and qualification of such in vitro models; without qualification of the model, end-users cannot be confident in the physiological relevance to the in vivo setting and, subsequently, any data generated from the model. Because multiple companies are now starting to incorporate CIVMs (e.g., tissue-on-a-chip or organoid models) into every aspect of the drug discovery process,^4,5 it is crucial to ensure architectural, physiological, and functional features relevant to the tissue in vivo are recapitulated for both healthy and disease states.

We will outline an approach that considers the patient at every stage of the drug discovery process, with the end goal of reducing clinical attrition. This includes using better translational models, combined with evolving technologies such as artificial intelligence (AI), machine learning (ML), gene modification using clustered regularly interspaced short palindromic repeats (CRISPR)/Cas9, and quantitative systems pharmacology (QSP) modeling.

Paradigm Shift in Using CIVMs throughout the Drug Discovery Process

In the current drug discovery paradigm, simplistic or reductionist approaches to modeling disease have been the convention for target identification and validation. Efforts to run screens with 10,000–1,000,000s of compounds has limited the ability to develop clinically relevant primary cells and model platforms. Often, it is not until the lead optimization (LO) and candidate selection phases of the drug discovery pipeline when more complex cellular disease models are developed and used; it is also rare that these are coupled with investigating safety liabilities in parallel. Animal models are used with increasing numbers in the late LO phase and particularly in the candidate selection phase for pharmacokinetic/pharmacodynamic (PK/PD) and efficacy, and then also used in regulatory required safety studies. Therefore, there are opportunities throughout the drug discovery process to incorporate more translationally predictive cellular models, or CIVMs, to both reduce animal use aligned to our 3Rs commitment (replacement, reduction, and refinement) and provide data that better translate to the clinic, which ultimately results in better medicines for patients ( Fig. 2 ). To achieve this stepwise change, it is time to stop using (1) poorly defined cellular models for target identification, (2) nonrelevant cell-based screening or LO assays with immortalized cell lines, and (3) in vivo models that have shown low predictive value or translational relevance ( Fig. 2 ).

Figure 2.

Opportunities for more predictive complex in vitro models (CIVMs) earlier in research and development. Modified from Marx et al.⁸⁶

Identification and validation of novel targets, with increased confidence that the target will be efficacious and safe in the clinic, require cellular models to be fully qualified. Having qualified models⁶ at the target validation phase, which would include CIVMs of high complexity but low throughput, would enable us to confidently progress in the selection of translational targets. The translatability of such qualified models should be benchmarked against transcriptomic, proteomic, metabolomic, and pathological studies with reference to human healthy and diseased states, in addition to functional readouts. Of note, however, is that no single model is likely to recapitulate all aspects of complex diseases. By qualifying a model, it will be clear which aspects of biology the model can accurately describe. To be transformational, a drug discovery program should invest in developing multiple models to recapitulate all mechanisms of interest. For example, chronic obstructive pulmonary disease (COPD) manifests itself in the upper and small airways and the alveoli. The development of airway models (upper and lower) could serve as the foundation to adapt to other epithelial barrier models, such as gut epithelial models. Throughout the drug discovery process, these models could be modified for genetic screening, efficacy, and/or toxicity screening (lower complexity but high throughput) ( Fig. 3 ).

Figure 3.

Incorporation of complex in vitro models (CIVMs) into the drug discovery process.^{75,84,85,87,88}

In LO and candidate selection phases, human-relevant in vitro data could support computational and modeling functions to build in silico models to support physiologically based pharmacokinetic (PBPK)/quantitative systems toxicology (QST) and QSP. Such models can be used to calculate the therapeutic window and to predict the minimum anticipated biological effect level (MABEL). Recent publications have shown that integration of advanced, human-relevant complex models may provide improved translation of preclinical findings to clinical outcomes.⁷ In comparison to traditional ADME/PK models, it provides more mechanistic knowledge of biological and pharmacological processes ( Fig. 3 ).

Selecting the right patient population for a drug (precision medicine) can be achieved through integrating patient tissue samples, primary cells, patient-derived induced pluripotent stem cells (iPSCs), or adult organoids with the most translationally relevant CIVM platform. This approach would enable the individual genetic makeup, combined with an understanding of mechanisms linked to a particular disease endotype, to be factored into validating new targets and drugs, allowing stratification of patient populations, or even individuals, who are most likely to benefit or least likely to suffer from adverse reactions to a drug (Figs. 2 and 3).

A SWOT (strengths, weaknesses, opportunities, and threats) breakdown of the high-content analysis of 3D cellular models has recently been published that identifies many relevant points applicable in the safety space.⁸ Building on this, we have conducted a SWOT analysis from a GSK perspective, on the role of CIVMs in pre-drug discovery. The following sections will detail opportunities, challenges, and limitations as supported by specific examples throughout phases of the drug development process ( Table 2 ).

Table 2.

SWOT of Complex In Vitro Models in Pharmaceutical Development.

Strengths
Human biological relevance improved with translational readout:
• Ability to generate human disease-relevant complex models and disease-specific modeling; biggest impact in areas such as neuroscience, in which previously rodent models were the only option available
• Enables models supporting the 5 Rs (principles of successful drug discovery): the right target, patient, safety, tissue, and commercial potential
• Allows for modalities otherwise impossible to be assessed, such as study of paracrine interactions, impact of secretory events, or tumor invasion
• Provides higher-throughput systems to identify human-specific toxicological events earlier in drug discovery
• Enables multiplexed endpoints to link together phenotypes to cellular-signaling events; provides ability to explore multiple pathway interactions simultaneously
• Cell-specific and patient-derived analysis
• Enables comparison between species
• Enables reduction of use of animal models and in vivo testing• Ease of miniaturization and automation of system

Weaknesses
• Gap between development and qualification of models
• Still unable to recapitulate full extent of pathophysiology
• Access to cell types
• Limitations of biological understanding; there is a risk of a lack of translation to the human clinical response.
• Long generation time (particularly with higher-complexity models, such as spheroids and organoids)
• Potential off-target effects
• Sterility risk
• Typically endpoint analysis
• Limited lifespan of system
• Difficulty in balancing complexity with meaningful data
• Special and complex machinery often required

Opportunities
• Establish consortiums and cross-industry partnerships that develop and establish domain of validity for CIVMs.
• Link CIVMs with computational modeling and use in PK/PD.
• Development in QST mathematical models to increase predictivity
• Reduced reliance on animal models
• Combining immune system, stem cells or adult organoids, bioprinting, and MPS technologies
• Multiple organ systems, species, and stages
• These models are starting to demonstrate therapeutic potential—more than just a discovery tool.

Threats
• Overhyped CIVMs that lack adequate validation and determination of domain of validity for safety or efficacy
• FDA and other regulation requirements
• Not amenable to scale, thus limiting applications to low throughput (early target validation or late lead optimization)
• Pace of field accelerating and other disruptive technologies (e.g., in silico modeling): Are we able to maintain competitiveness and the challenge to keep pace with the latest and greatest models?
• High cost of CIVMs
• Requirement of validated models at speed
• Throughput needs to be pitched against throughput of more traditional approaches (e.g., animal models).

5 Rs: The right target, right tissue, right safety, right patient, and right commercial potential; CIVM, complex in vitro model; FDA: US Food and Drug Administration; MPS, microphysiological system; PK/PD: pharmacokinetic/pharmacodynamic; QST, quantitative system toxicology; SWOT: strengths, weaknesses, opportunities, and threats.

Alignment of CIVM Development for Target Identification and Validation and Lead Optimization Phases

Drug development is a stepwise process in which each step depends on the quality of the previous ones; therefore, it is critical to have the highest confidence in the choice of the targets. It is easy to imagine the consequences of starting a drug development program with the wrong target: It will not lead to the development of a successful drug, but it may take a lot of time and resources before recognizing that the modulation of the target doesn’t cure the disease or is toxic. To obtain the maximum confidence in targets, qualified human-relevant models providing as much information as possible about the physiology of the organ or disease are required. Often, more than one CIVM is used to validate targets because currently most models reflect only a part of the physiology of the organ or disease. The choice of human-relevant models in the target validation step is mainly driven by the quality and quantity of information provided by models, and the throughput is secondary. In contrast, during the screening phase, high-throughput capacity of models is one of the main drivers in the choice of models. Because each step has specific needs, different models are used during drug development, but all the phases require human-relevant models. The use of inadequate human-relevant models at a single step strongly derails the chance of developing a successful drug. Before using a model, it should be qualified by determining its domain of validity as defined by Scannell and Bosley⁹ based on rigorous scientific data, and not on assumptions such as “Primary cells are better than immortalized cell lines” or “The more complex a model is, the more human relevant it is.” Identification of the right model requires it to be fully qualified with regard to its attributes in comparison to the in vivo situation. Until recently, model characterization has relied on techniques looking at a limited number of physical and biological attributes, including histology assessment and measurement of gene and protein expression at the subpopulation level, using reverse transcriptase–quantitative polymerase chain reaction (RT-qPCR)-based and enzyme-linked immunosorbent assay (ELISA)-based technologies. The comparison with fresh tissue is key. Advances in “omics” technologies, including metabolomics, transcriptomics, and proteomics, at the single-cell level are enabling a more in-depth characterization of models and an ability to score a model and assess whether it is fit for purpose. One example by Huh et al.¹⁰ demonstrated a well-defined domain of validity for an in vitro pulmonary edema–on-a-chip model. The lung-on-a-chip paper used a polydimethylsiloxane (PDMS) device that generated mechanical cyclic stretch in combination with interleukin-2 (IL2) and was able to recapitulate human pulmonary edema with hallmark human pathophysiology that included fibrin clots, vascular leakage, and accumulation of fluid in the alveolus space. Pharmacological agents were used in the study that demonstrated clinical translatability by showing suppressed pulmonary vascular leakage. Similar findings in barrier permeability changes in the in vitro pulmonary edema model were demonstrated in an IL2-administered mouse with mechanical ventilation. A gap in this paper is that no transcriptomic profiling comparing in vitro pulmonary edema, mouse-induced IL2 edema, and human in vivo edema was performed. A recent paper by Kasendra et al.¹¹ showed the power of transcriptomic profiling, because the data showed there was increased similarity in gene expression profiles when comparing a microfluidic intestinal organoid to adult duodenal tissue than to static duodenal organoids. The transcriptomic profiling increases the confidence in which the model better replicates the in vivo tissue. The paper demonstrated a domain of validity for the model for possible drug transport metabolism and drug–drug interaction studies, because it showed evidence comparing the duodenal intestine–on-a-chip to duodenum tissue. The authors showed that the intestine-on-a-chip had a polarized cell architecture, intestinal barrier function, the existence of specific cell populations, and in vivo–like expression, localization, and function of key intestinal drug transporters that are key requirements for intestinal drug metabolism cellular models.

Cell Lines in CIVMs

Of highest priority is the qualification of the human-relevant cellular model and the tools available that will determine the selection of the model. Availability of immortalized cell lines, coupled with advances in molecular biology techniques in the 1980s and 1990s, led to the widespread use of engineered models in which targets of interest could be readily overexpressed. It is clear now that studying the function and activity of a target in an artificial context does not necessarily translate to the physiological situation.¹² Cell lines can be incorporated into a CIVM and are a potentially useful model for target validation and drug discovery work, but as discussed previously, these need to be qualified for what they are, and are not, to establish if they are fit for purpose.¹³ With the aim of determining the relevance of the cells for use in research, an Open Targets project in collaboration with scientists at the Wellcome Trust Sanger Institute (WTSI) has taken a number of commonly used cell lines in research and performed a detailed analysis to investigate gene expression with reference to primary cell data.¹⁴ Cell lines have the advantage in that they can be obtained in large numbers, which makes them attractive for use in CIVMs for target validation work. The results from the Open Targets study will provide a useful resource for assessment of cell lines and physiological relevance with respect to different tissue types for use in CIVMs.

Human Primary Cells in CIVMs

Human primary cells are assumed to represent a better alternative for target validation compared to cell lines. Depending on the disease of interest, accessing primary cells and tissues can be challenging for differing reasons. For example, a range of immune cells can be obtained from peripheral blood and are routinely used for in vitro studies throughout drug discovery for inflammatory diseases, as evident in widespread literature. A major limitation is, however, a lack of control of the activation status—for example, monocytes derived from donors—that can easily skew assay readouts. Other primary cells are not so easily accessible (e.g., neuronal cells for validating targets for diseases, including Alzheimer’s and Parkinson’s disease). Challenges of working with primary cells include quality of the primary cells or biological samples, reliable supply, availability of methods to isolate pure subpopulations from tissue samples, maintaining viability and the in vivo phenotype when in culture, limitations on cell numbers, and donor intra- and intervariability. The minimum number of donors that should be evaluated during assay development or cross-comparison between donors using primary cells is three, but this could vary depending on the robustness of the assay, the variability of the primary cells,¹⁵ and the heterogeneity of the disease. As for cell lines, it is critical to qualify primary cells to ensure that the phenotype and functionality are maintained in cultures and that the cells are fit for their intended purpose. An important gap relating to developing relevant complex models very close to in vivo tissue is the lack of an appropriate immune component and microbiome. Because these two aspects are crucial for health and implicated in certain disease states, careful consideration should be given to the inclusion of these when applicable.

Stem Cells in CIVMs

The use of stem cell technologies in CIVMs has the potential to transform genetic target validation and our understanding of gene variant to function. iPSCs or adult stem cells bearing a genetic variant can be differentiated into a variety of mature cell types, including neuronal cells, cardiomyocytes, and epithelial cells, enabling functional studies to be performed.¹⁶ Stem cells offer a number of unique advantages to drug discovery in that they provide a source of cells that retain the genetic information of the donor, but are scalable and amenable to gene editing (unlike a number of primary immune cell types) and represent an infinite source of cells for target validation in CIVMs. For example, we have generated a robust and scalable iPSC–macrophage platform that is amenable to bespoke editing using CRISPR/Cas9. The majority of diseases include an immune cell compartment, and stem cell technology could facilitate the development of CIVMs in which the function of the genetic variant can be assessed in the context of cell–cell interactions that occur in vivo (e.g., inflammation in lung tissue). In allergic asthma, there is evidence of immune T cell–epithelial cell crosstalk in the lung.¹⁷ Stem cell technology would facilitate development of syngeneic cell populations for experiments, avoiding artifacts that could occur through allogeneic T cell activation in CIVMs, where cells are derived from two different genetic backgrounds. A secondary approach for the use of both iPSCs and adult stem cells includes the use of such cells for originator samples to generate organoid cultures; an example includes the use of gut-derived stem cells obtained from intestinal crypts, which have been gene-edited and then differentiated and cultured as organoids, enabling the assessment of the gene-edited target on epithelial barrier function in a model that includes multiple epithelial cell subtypes.¹⁸

CIVMs Generated from Organoids

Organoids are currently considered to be the highest order of complexity for an in vitro system that mimics more closely the development and niche observed in vivo, largely due to their self-assembly. Organoids develop from either tissue-restricted adult stem cells or pluripotent stem cells, including induced pluripotent and embryonic stem cells. One downside with these organoid models is that they are inward facing: As an example, the epithelial layer in gut organoid models typically faces the interior of the 3D structure and is embedded in an expensive matrix, reducing the ability to perform experimental manipulation. Stem cells have the capacity to differentiate into different cell types and organize themselves in a manner consistent with the organ they are derived from; for example, gut organoids representing a model of the small intestine develop an internal lumen and structures resembling the crypts and villi that are seen in vivo.¹⁹ Researchers have generated organoids from a large variety of both adult and iPSC-derived material, namely (but not exhaustively), brain, intestine, liver, kidney, and lung.^20–23 The first reported use of CRISPR in human intestinal organoids showed feasibility for CRISPR-mediated homologous deficient recombination (HDR) repair of the monogenic, disease-causative cystic fibrosis transmembrane conductor receptor (CFTR) gene.²⁴ Repair of the CFTR mutation reversed disease-associated phenotypes, including forskolin-mediated swelling, to organoids derived from seemingly healthy individuals.^24,25 These seminal publications have driven the use of gene editing in complex systems for disease modeling in target identification, validation, and to a lesser extent screening. Subsequent CRIPSR/Cas9 studies have opened avenues for cancer research and functional genomic screening, demonstrating the utility of generating novel in vitro cancer models. For example, effective CRISPR-mediated targeting in colon cancer organoids was used to generate commonly known mutations associated with the disease (including disruption of tumor suppressor APC, TP53, and SMAD4). These organoids displayed altered biological changes aligned to those seen in patients, such as chromosomal instability.^26,27 Moreover, Drost et al. recently demonstrated the use of CRISPR/Cas9 to derive mutational signatures within colon organoids seen in patient cohorts, to enable a scalable culture model with clear clinical translation.²⁸

CIVMs Generated from Microphysiological Systems

One of the biggest challenges in the development and implementation of CIVMs is the limitations imposed by static cellular models. In static culture, supply of nutrients and buildup of waste products can become limiting to cell growth and function. Incorporating fluidics into models to mimic blood supply and interstitial flow is likely to increase the physiological relevance of the model.²⁹ To address this, a number of organ-on-a-chip models have been developed that incorporate microfluidics.³⁰ An organ-on-a-chip, or MPS, is defined as a model in which the biology and physiology of an organ are replicated³¹ ( Fig. 2 ). GSK has been working with external collaborators on a number of organ-on-a-chip models, including a lung-on-a-chip in collaboration with the University of Southampton and SAL Scientific. This model incorporates impedance measurements in real time to assess barrier integrity, avoiding manipulations that disrupt the culture and enabling target function to be assessed in longitudinal studies. The organ-on-a-chip platforms are readily amenable to interrogating target function via small molecules or biological reagents, introduced via the fluidics.

CIVMs Generated from 3D Bioprinting

Another challenge is to try to replicate in vivo architecture of tissues. Technological advances such as bioprinting have reproducibly demonstrated the ability to spatially control deposition of multiple cell types and gels, resulting in the construction of tissues with architectures closer to those of organs than two-dimensional (2D) culture systems.³² The reduced throughput and the ability to miniaturize bioprinted tissue can be challenges, however, potentially limiting bioprinting’s application to target validation or LO. The physiology of an organ and its perturbation in a disease state are the result of individual cell activities and cell–cell interactions, including immune cells. Nonalcoholic steatohepatitis (NASH) is a great example of a multicellular disease in which hepatocytes accumulate excessive amounts of fat; stellate cells become activated, losing their retinol content and secreting large amounts of collagen; and Kupffer cells often play a central role in inflammation. To be physiologically relevant, in vitro models of NASH require the presence of at least three liver cell types—hepatocytes, stellate cells, and Kupffer cells—preferably in a 3D organization. Then, models need to be validated using human data from healthy and disease liver tissues as references, before being used for target identification and validation and LO. A recent review by Retting et al. described how liver fibrogenesis could be observed in 3D bioprinted liver tissue that contains human hepatocytes, hemopoietic stem cells (HSCs), endothelial cells, and Kupffer cells after exposure to known fibrogenic agents such as low doses of methotrexate, thioacetamide, or transforming growth factor-β (TGFβ).³³ In addition, a NASH-like phenotype—including steatosis, ballooning hepatocytes, excessive collagen production, and inflammation—could be observed after exposing the 3D printed tissues to high levels of sugar and fatty acids, mimicking a diet regimen typically associated with obesity.^33,34 Comparisons of gene expression profiles between healthy and diseased human liver tissues and vehicle versus TGFβ-treated 3D bioprinted liver tissues showed a 5% overlap between human organs and 3D printed tissues.³³ Subsequently, understanding the domain of validity for any cell model is critical to deciding how to apply the model for target validation and development of drugs to treat liver fibrosis.

CIVMs Generated from Ex Vivo Tissues

Tissue slices have the advantage that multiple cell types are represented, and they have been used for target profiling and LO in GSK studies. The current limitation for using tissue slices for longitudinal target validation studies is their limited viability in culture. Another caveat of tissue slices is that although they are amenable to target validation and LO using chemical perturbation (small molecules, peptides, and antibodies), they are not amenable to gene-editing studies.

Gene Editing of CIVMs

A fundamental approach in our attempts to help to delineate complex cellular function is using chemical and genetic modifiers to perturb the target function in healthy cells and looking to see whether this induces a disease state, or vice versa when working with a disease model and seeing whether modulating the target restores a healthy state. Recent significant developments in the field of genome editing are powering a revolution in the understanding of biology and disease via modeling the effects of genetic variants, exploring target pathways and mechanisms, and understanding gene structure–function effects. Tools to modulate the function of a target include the use of modulators of pathways such as small molecules, RNA interference (RNAi), and technologies such as precise gene-editing technologies. The challenge we face is linking these genetic traits to the protein malfunction in disease and linking genotype to phenotype in the most complex cellular system that mimics the in vivo disease state. The generation of isogenic knockout (KO) human cell lines for comparative genomics is commonplace, and gene KO by CRISPR/Cas9 has proved effective in almost all cell types, including iPSCs, tumor spheroids, organoids, and a range of primary cells, in which editing is often more challenging. Furthermore, “knock-in” of mutant alleles by HDR permits assessment of disease-associated mutations in an isogenic background.²⁴ Recent advances in the field have included novel CRISPR systems that are able to activate or repress gene expression, and modulate the epigenome and RNA abundance.³⁵ As with all technologies, there are limitations to each of these systems, but this will not be discussed in the context of this review.

A transformational opportunity exists to industrialize genome-editing capabilities for gene knockout, activation, repression, or perturbation and to integrate with robust qualified iPSCs and cell-based phenotypic and mechanistic models to accelerate target identification and validation. To perform genome-editing studies, capabilities in robust cellular models require extensive analysis of gene expression in human tissue [laser capture microdissection (LCM), single-cell technologies, spatial transcriptomics, and proteomics] to enable (1) data-driven decisions on cell model relevance, (2) expansive multiparameter assay endpoints, (3) delivery of high-quality targets with strong biological evidence, and (4) the application of artificial intelligence and machine learning (AI/ML) techniques for target discovery and validation. An area of key focus for us, and others, is immuno-oncology (IOC). This use of the immune system to fight cancer is with a view to learning how we can “train” myeloid and lymphoid cells to identify and destroy tumor cells. There are several ways in which gene editing can be applied to solid tumor CIVMs: directly editing the tumor spheroids or organoids itself; specifically targeting the immune cell component, the primary human T cell; or dual CRISPR targeting in two or more cell types for use in co-culture models and downstream assays. Despite anticipated challenges faced with introducing modifications to immune cells, two recent publications have successfully demonstrated the utility of CRISPR/Cas9 in T cells.^36,37

CIVMs in Lead Optimization

Lead optimization and candidate selection are often when animal models for PKPD and efficacy may be used. While improvement in PBPK modeling simulations is reducing reliance on animals, the efficacy space remains challenging. Efficacy models in animals are often costly and time-consuming, and, depending on the mechanism and/or disease area of interest, may have limited translational relevance to the human patient; or the depth of qualification for a model for a specific pathway may not have been conducted and remain a significant unknown. Real opportunities exist to understand the weaknesses in available animal models to strategically drive investments in CIVMs.³⁸

Development of in vitro disease models for target validation and LO is challenging. Most diseases with unmet need are multifactorial. CIVMs mimicking the pathophysiology of the human disease used in target validation can then be used or modified for later drug development for PD studies, to determine efficacy (a drug’s action on a target in a tissue or organ), toxicity (an unwanted side effect of a drug), and drug–drug interactions (effects of drug combinations) ( Fig. 3 ). Incorporating these more physiologically relevant models to generate evidence for a lead molecule engagement with a molecular target and the consequential effect on the biology of the target cell should aid in reducing project attrition rates. CIVMs that incorporate flow and are vascularized should improve the determination of parameters for dose–effect (D-E) relationships by mimicking the drug delivery via blood to an organ in vivo. Currently, D-E relationships are usually generated using in vivo or in vitro drug titration assays; these methods are unsatisfactory in predicting drug efficacy and selectivity in vivo, because the drug concentration at the target site typically changes throughout time due to blood flow and the binding process is maintained at a steady state. In a number of examples, CIVMs with fluid flow have been used for modeling PD studies that include pulmonary edema in a lung-on-a-chip and protection by transient receptor potential cation channel subfamily V member-4 (TRPV4) inhibitors,¹⁰ small airway–on-a-chip for asthma,³⁹ vascularized tumor–on-a-chip,^40–43 hepatitis B virus (HBV) infection in a liver-on-a-chip,⁴⁴ thrombosis,⁴⁵ inflammatory bowel disease⁴⁶ and contribution of microbiome,⁴⁷ and Parkinson’s disease–specific dopaminergic neurons.⁴⁸

A current limitation in highly complex in vitro models is the opportunity to industrialize to the scale that might enable medium- to high-throughput chemical or genetic screening. A recent publication, however, demonstrated the use of a spheroid-based (3D) human pancreatic ductal adenocarcinoma cell line in 384-well format, to screen selective inhibitors of mutated KRAS. The authors identified selective inhibitors that were differentially active between 2D and 3D cultures, signifying the importance of screening in 3D culture systems to better replicate human translation.⁴⁹ There are opportunities to develop CIVMs that have less complexity but higher throughput, which will be discussed in the next section.

CIVMs in Hit/Lead Discovery for Efficacy and Toxicity Testing

In recent years, automation and miniaturization of cell-based assays have become an important focus in high-throughput screening (HTS) techniques. For the most part, HTS assays using monolayer or suspension cultures still show very limited predictive value for benefits of clinical efficacy of small molecules. Advances in CIVMs, for example the introduction of 3D cultures⁵⁰ or pluripotent-derived stem cell–derived systems,⁵¹ have allowed the improvement of capturing more biologically relevant information relative to conventional HTS cell-based assays.

Due to the requirement to screen hundreds of thousands of compounds, a number of limiting factors need to be considered for using a CIVM. These include the higher costs of plates and biological reagents, the larger variability and heterogeneity within assays, and the additional data-handling requirements, which can be extensive for more complex models. For hit discovery, the development of CIVMs that mimic disease-relevant biology increases all these challenges and currently limits their use in HTS. When evaluating the cost and length of an HTS, these factors drive many pharmaceutical companies down the route of using 2D recombinant cellular systems. The difficult decision is to prosecute 10 screens using recombinant models, compared to one screen using a CIVM, and it is often hard to justify the costs despite the fact that in later stages of the drug discovery process, molecules identified using recombinant systems are more likely to fail through lack of efficacy in the clinic. Through use of automated systems, these “per-well” costs can be offset against the richness of the resulting data, which have the potential to enable earlier, more robust pipeline progression decisions. With cost and availability of tissues a major consideration, it is necessary to reduce the number of experimental tests to a minimum. Design of experiments has been successfully adopted at GSK, reducing reagent requirements and cycle times for assay development. Latest-generation liquid handlers are the key factor in our transition to this approach because they can address each well independently with minimal dead volume. For example, the Formulatrix Mantis and the higher-throughput Tempest can dispense a wide dynamic range, from as low as 100 nL volume. This, coupled with easy-to-set-up liquid handlers such as Andrew Alliance and Cybi Felix for bulk reagent addition, allows us to move away from larger, less flexible systems for this more bespoke type of experiment. The low volumes and precision of the HP D300 have been instrumental in allowing us to quickly and accurately define kinetic parameters, by dispensing with far greater accuracy than is achievable manually.

Media exchange, often coupled with test compound dosing, is a necessity for maintaining viability over longer-term studies. It is a repetitive, low-value-added activity, which has in the past made large-scale CIVM experiments logistically difficult and at times prohibitive. At GSK, we have found the Agilent Bravo to be a useful tool, enabling careful liquid extractions and additions with minimal disturbance to delicate samples. Similar technologies on the horizon include the Yamaha Cell Handler System and the TTP Dragonfly.

In terms of data handling, rapid advances in ML are enabling faster processing of the significant volume of data generated by more complex cell models. For example, neural networks automating analysis of experimental data are being applied to phenotypic screening using imaging systems with multiparametric readouts. CIVMs such as spheroids and organoid cultures have been miniaturized from 96-well formats^52,53 to microHTS 1536- and 386-well formats,^54,55 while advances in microfluidic systems allow us to work routinely from 24–96-well formats. Microarrays and microwell platforms that could include hydrogel, micropad, or inverted droplet arrays⁵⁶ could be cost-effective methods to rapidly generate heterogeneous 3D cell microenvironments through 3D bioprinters or robotic liquid handlers. Developing methods for higher-throughput compound screening with CIVMs on these platforms requires time.

While spheroids have been used in high-throughput microtiter plates for screening,⁵⁴ there are major challenges in using 3D organoid models for HTS. As well as cost, the complexity of screening in complex systems often poses a challenge for miniaturization and automation. One of the biggest challenges is to dispense organoids in HTS formats. Two methods are commonly used: One is the “sandwich method,” in which the organoid cell suspension is added on top of a polymerized matrix layer,⁵⁵ and the other method is to mix the organoid cell suspension with Matrigel matrix and dispense while the plates are kept on a cool rack.⁵⁷ These methods can be non-homogeneous and expensive for HTS purposes, and for this reason, GSK is developing a new methodology to deliver small numbers of organoids suspended in Matrigel into 384- and 1536-microtiter plates using sub-microliter volumes per well, reducing the amount of Matrigel and organoids by a factor of 100.

In oncology programs, many 3D HTSs use cell viability readouts that measure metabolic activity or cellular adenosine triphosphate (ATP) as surrogate markers. Such assays are limited in mechanistic information and may not translate to predictive drug effects. Some studies have shown that 3D organoids are significantly less sensitive than monolayer culture systems,^58,59 and drug responses can be more diverse in 3D tumor organoids than in 2D cultured cells.⁵² GSK and others are working to generate organoids that are HTS compatible for studying differentiation and phenotyping using high-content imaging as readouts to deconvolute drug effects. A screening 3D tumor organoid model is in development using high-content imaging to characterize the effects of drugs measuring a number of phenotypic readouts, including apoptosis, cell death, and cell health.

Another area in the hit/lead space in which CIVMs could be used is in the early identification of safety liabilities. The second highest cause for failure of candidate molecules in the clinic is due to safety liabilities, and the underlying factors are the lack of cellular systems that translate to humans. Hepatotoxicity is the most frequently cited toxicity responsible for drug withdrawal from the market.⁶⁰ Early deselection of drug candidates likely to cause hepatotoxicity could improve patient safety, decrease the rate of attrition, and cut the cost of drug development. To address this, GSK has invested in the development and industrialization of an in vitro primary human hepatocyte (PHH) culture model that could improve early hepatotoxicity prediction.

GSK recognized a gap in the predictivity of conventional in vitro cellular models for hepatotoxicity, because they often lack biological relevance and are limited to short-term studies. It became clear that moving cultured cells from 2D monolayers into 3D systems improves their biological relevance and response.⁶¹ The basis of the 3D PHH model was conceived and characterized in a precompetitive, Innovative Medicines Initiative (IMI)-sponsored public/private consortium, Mechanism-Based Integrated Systems for the Prediction of Drug-Induced Liver Injury (MIP DILI), a collaboration with industry and academic leaders in the field. Throughout several years of research, the data generated demonstrated that this model had improved phenotype and translational relevance more than other cell types and models tested.¹⁵ Within GSK, we further characterized and industrialized the model into a screening platform, enabling the first 3D complex cell model in GSK to enter production as part of the early panel of safety assays. This 3D spheroid model consists of PHHs that have a compound repeatedly dosed throughout a period of 14 days. A direct comparison of the predictivity of this model in comparison to an immortalized HepG2 monolayer assay was completed and demonstrated an improvement in predictivity of known hepatotoxins ( Fig. 4 ). When running the 3D model at an early time point (3 days), no improvement was seen over the 2D HepG2 assay. This suggested the differences seen in this model are not actually due to cellular background but instead are down to the fact that the 3D models enable us to assess the effects of chronic compound exposure. Unlike the 2D assay format, this model is also metabolically active, and this means, for the first time, we will identify any safety liabilities of metabolites as well as the parent compound itself. With further development, the model offers the potential for richer mechanistic readouts through use of technologies such as high-content imaging and mass spectrometry that, in combination, will enable us to evaluate compound metabolism and develop insight into which pathways or cellular functions are being modulated by compound exposure.

Figure 4.

Three-dimensional (3D) primary human hepatocyte spheroids show improvement over two-dimensional (2D) HepG2 to identify known hepatotoxins. 173 compounds tested in 3D hepatocyte spheroid and 2D HepG2 demonstrated a 28% improvement in the ability of the model to identify known hepatotoxins. Predictivity is the percentage of most-concern compounds identified, and sensitivity is the percentage of successful classification of no-concern compounds. For the 2D HepG2 assay, it was performed as previously described.⁸⁹ For the 3D hepatocyte assay, it was performed as previously described,¹⁵ except it was performed at the 384-well scale. When generating the data for the 173 compounds, compound-only controls were also conducted to control for compound interference. IC₅₀ values, corresponding to the concentration of the compound causing a 50% reduction in viability, were computed for each compound. The human biological samples were sourced ethically, and their research use was in accord with the terms of the informed consents under an Institutional Review Board/Ethics Committee (IRB/EC)-approved protocol.

We currently stand at an inflection point where the pharmaceutical industry, academia, and microfluidic vendors must work together to develop industrialized systems that are capable of practically useful throughput of human-relevant system studies. GSK is actively working to link our newer safety liability cellular models with our preclinical safety and toxicity models.

CIVMs in Preclinical Safety

Attrition specifically due to safety issues represents a huge liability for the drug development process, and implementing complex cellular systems to investigate or screen for safety assessment has the potential to transform the in vitro support of the safety process of drug development. Being able to model human organs, cellular and fluid interactions, and pathological mechanisms can generate more valuable information, encourage greater refinement of in vivo studies (supporting the 3 Rs), improve toxicity and dose predictions, support in silico QST modeling, and reduce attrition, which will overall reduce costs and time to market.

During drug development, there are many ways that CIVMs can support toxicological assessments and investigations, performing a vital role in drug progression. Many toxic responses seen in vivo are the result of complex interactions between multiple cell types. CIVMs could serve to recapitulate adverse events seen in in vivo preclinical safety studies and be used to either rank compounds for progression or investigate the pathological mechanisms and evaluate their relevance in translatability from the in vivo toxicology species to clinical studies. Alternatively, the model may serve to develop monitorable biomarkers to take into the clinic that can either (1) predict early events that occur before damage manifests clinically or (2) assess the progression and/or severity of an adverse event. There is an underlying assumption that higher-complexity systems coupled with an in vivo–like scaffold or matrix will lead to better predictivity than more simplified 2D systems. At GSK, it is imperative that we not only develop the next generation of 3D models but also qualify the usefulness of the data they provide. There is an urgent need to increase our understanding of preclinical toxicities translating into human adverse events, and CIVMs are a key part of this endeavor.

To continue to advance the safety predictive capability of preclinical models and their acceptance in the field, the US Food and Drug Administration (FDA) has recently used the agency’s Toxicology Working Group to develop an FDA predictive toxicology roadmap. The FDA clearly identified “Microphysical systems like tissues or organs on a chip” as well as “in vitro alternatives” on their list of “Promising New Technologies in Predictive Toxicology,” highlighting the value in supporting and developing these systems. The alternative models or assays focus on pathways and risk assessments at multiple levels, including genes, proteins, biochemical pathways, and cell and organ function.⁶² They define the importance of qualifying the model or assay against a specific biological question and setting boundaries such as relevance, reliability, reproducibility, and sensitivity.⁶² Each model or assay test should have a specific interpretation and application in the drug development pipeline and in regulatory decision making. Further advice from the FDA on acceptance of new toxicology methods includes generating sufficient convincing data as well as maintaining continuous dialogue and feedback among all stakeholders throughout development, qualification, implementation, and acceptance by the regulatory authorities.

The Innovation and Quality Microphysiological Systems (IQ MPS) affiliate⁶³ brings together more than 20 pharmaceutical and biotech companies with a goal to expedite the potential of MPS for drug discovery and development. This has been through interactions with the FDA, National Institutes of Health (NIH), and Defense Advanced Research Projects Agency (DARPA) and the National Center for Advancing Translational Sciences (NCATS) Tissue Chips for Drug Screening programs. These wide-reaching partnerships are of great value to GSK and the other pharmaceutical and biotechnology companies, because they facilitate access to collaborations; provide a resource and expert guidance on reference sets of tested compounds, assays, and biomarkers; increase the confidence in marketed models; and could be a path for collaborations for building confidence in safety, ADME, and pharmacology models.

A key point in the FDA toxicology roadmap is the importance of collaboration and transparency about new technologies so we can further advance safety models. Safety models can be more difficult to design than efficacy models. In efficacy models, generally there is a hypothesis, target, and marker or assay that can measure an output. In safety models, there is often a need to look at every possible mechanism of toxicity because it is unknown why the drug causes adverse events, or there may be a requirement to determine at what levels the drug will cause adverse events. This complexity is why it is critical to cultivate collaborations and sharing of data and ideas.

One external example of a successful safety model has been demonstrated by the collaboration between Genentech and AstraZeneca.⁴ In this case, a 3D liver spheroid model comparing multiple donors was determined to be applicable because it showed enhanced liver phenotype, metabolic activity, and stability compared to the typical 2D models. Hepatotoxicity risk is difficult to predict because of the number of different factors that can cause drug-induced liver injury (DILI) in humans. There are many pathways affected in DILI and many assays available to measure the effects of perturbing a single pathway. The 3D model was successful because it contained a small number of human liver cells and has the potential to be high throughput, as well as being more physiologically relevant, containing relevant transporters and metabolic functionality. The 3D model was overall more reproducible, stable, and sensitive, and it allowed for the investigation of predictive biomarkers for liver injury.⁴

GSK has recently begun the qualification of the EpiIntestinal models within the preclinical safety space. Investigation of these prospective models includes compiling a checklist of requirements ( Table 3 ), which reflects the FDA initiatives described earlier, including investigating specific biological questions and having key endpoint assays defined. GSK has generated several of these qualification and characterization checklists per organ or tissue type to help scientists qualify their models, ensuring a consistent approach to allow comparison of models throughout the company. It is not expected that all the criteria outlined in the checklists are reproduced in any one model, or that all of the criteria should be selected to qualify a single model, because it is anticipated there will be limitations depending on model complexity. If, however, there are specific cell types that are expected to exist in the model and are thought to be essential for its future purpose, then it would be fundamental to evaluate their presence, morphological architecture, functionality, and longevity in the system. If the model has also been selected to investigate modulation of a specific biochemical pathway or for target validation, then it would be necessary to confirm levels of expression of proteins and genes in that pathway, and the biological target of interest, and confirming these exist in the appropriate cell types. Scientists should therefore understand the domain of validity associated with the practical use of the system and consider whether the platform is robust enough for the question it is being used to answer.

Table 3.

Markers for Qualification and Characterization of Small-Intestine CIVMs.

Qualification Type	Marker	Measurement
Cells	Cell types
	Enterocytes	IHC/EM/H&E/imaging
	Paneth cells	IHC/EM/H&E/imaging
	M cells	IHC/EM/H&E/imaging
	Tuft cells	IHC/EM/H&E/imaging
	Goblet cells	IHC/EM/H&E/imaging
	Stem cells (Lgr5+)	IHC/EM/H&E/imaging
	Cytokeratin 18/19 or pan-cytokeratin cocktail to identify epithelial cells	IHC/IF/imaging
	Presence of intestinal crypts and villi	IHC/EM/H&E/imaging
	Presence of brush border, and brush border proteins	IHC/EM/H&E
	NaK-ATPase: marker for polarization status of the intestinal model (should be mainly expressed on basolateral side of enterocytes)	IHC/IF
	Other cell types (e.g., fibroblasts and endothelial cells)	IHC/EM/H&E/imaging
Functional	Tight junctions
	TEER	Physical measurement using electrodes
	Tight junction/desmosomes	EM morphology
	Claudins/occludin/ZO-1	IHC/IF
	Passive intestinal barrier permeability (paracellular transport)	Assess leakage of fluorescence particles (apical and basal) (e.g., Lucifer yellow)
	Ca2+ transport (paracellular route)	Arsenazo III method in combination with Lucifer yellow
	Mucins
	MUC2	IHC/IF/imaging
	Components of mucin: CLCA1, FCGBP, AGR2, ZG16, and TFF3	IHC/IF
	Gut metabolizing enzymes
	Phase I: CYPs 3A4 (82%), 2C9 (14%), 2C19 (2%), 2J2 (1.4%), and 2D6 (0.7%) initially	Depending on type of system:1. Gene expression2. Use of substrates for CYP 3A4 (Midazolam, Nifedipine, Atorvastatin) and CYP 2C9 (Diclofenac), and assess metabolites using LCMS
	Phase II: UGT1A1 substrate estradiol	Depending on type of system: 1. Gene expression2. Use of substrate estradiol and assess using LCMS
	Drug transporters (clinically relevant)
	MDR1/P-gp, MRP2, MRP3, OATP2B1, PEPT1, PEPT2, BCRP, and OCT initially	Depending on type of system:1. Gene expression2. LCMS with the use of relevant substrates and inhibitors to assess efflux ratioPgP: Digoxin (substrate)BCRP: Rosuvastatin (substrate)PepT1: Valcyclovir (substrate)GF120918 as an inhibitor for PgP and BCRP and relevant controls (propranolol, Atenolol, and luciferin)
Cell health	Viability throughout time in culture
	ATP	ATP assay
	TEER	Physical measurement using electrodes
	Protein content	BioRad assay
	LDH leakage	Biochemical assay
Inflammation/irritancy	Cytokine release
	IL6, IL8, IL7, stem cell factor, IL10, IL15, IL18, Gro-α, GM-CSF, CXCL5. This list is not exhaustive because there are several other chemokines that could also be assessed.Exposure to inflammatory cytokines IL1β and TNFα, and any GSK molecules that cause an inflammatory response in the small intestine, to measure increased levels of cytokine production.	ELISAs/MSD/gene expression
Intestinal virus infection and innate immune response	Can the model support infection of intestinal pathogens such as rotavirus or norovirus?	IF of viral antigens
	Can the model mount an innate immune response to virus mimics such as Poly (I:C)?	RT-qPCR of interferon-stimulated genes

ATP: Adenosine triphosphate; CIVM: complex in vitro model; CYP: cytochrome P450 enzyme; ELISA: enzyme-linked immunosorbent assay; EM: electron microscopy; GSK: GlaxoSmithKline; H&E: hematoxylin & eosin staining; IF: immunofluorescence; IHC: immunohistochemistry; IL: interleukin; LCMS: liquid chromatography–mass spectrometry; LDH: lactate dehydrogenase; MSD: meso scale discovery; RT-qPCR: reverse transcriptase–quantitative polymerase chain reaction; TEER: transepithelial electrical resistance; TNF: tumor necrosis factor.

Gastrointestinal (GI) models are being developed for two potential projects. One project assesses the safety liability of disrupting stem cell activity in the gut due to targeting a molecule that is both present in the target organ and also associated with gut stem cell signaling. We are evaluating a commercially available intestinal transwell model containing a layer of fibroblasts that has been shown to model some aspects of wound repair and has the potential to support testing of compounds that may have the on-target but off-organ or -tissue effect of disrupting stem cell function and tissue renewal.

A second application for the GI model involves support of a cell and gene therapy anticancer project in which there is limited availability of relevant preclinical species. Using human CIVMs is valuable for preclinical safety assessment, particularly monitoring whether the product would interact with normal nontarget human tissue. For an intestinal model to be of value in both projects, there should be the presence of crypts, villi, and tight junctions to mimic the potential exposure of a particular cell surface, intracellular, or tight junction protein to the cell and/or gene therapy. A question that needs to be addressed in an intestinal transwell model is whether realistic compound or cell movement among the cell layers in the CIVM is accurately reproduced.

To enable robust safety assessment using CIVMs, the qualification process is vital to ensuring selection of an appropriate model. Currently, there is generally a compromise between cellular heterogeneity versus throughput. Toxicity measurements can easily be applied to CIVMs, such as cell death and barrier integrity, and can be scaled up for higher-throughput systems. The challenge is to maintain the required complexity of multiple cell types and structures. To help address this, GSK has partnered with academic innovators and groups such as the open innovation platform CRACKIT NC3Rs initiative, an IQ-MPS affiliate, to support the development and advance the technology for CIVMs for safety assessments.

Future Perspective and Conclusion

During the past two decades, advances in complex cell-based models with microfluidics and virtual computer-based models of human tissues and organs have enabled new tools and methods for analyzing drug responses with human cell and tissue sources incorporating multiple cell types and complex architectures under dynamic conditions.⁶⁴ The rise of big data and the development of large-scale computing capabilities for analysis using artificial intelligence (AI) such as machine-learning (ML) and deep-learning (DL) methods are anticipated to revolutionize the field of drug discovery.

Leading pharmaceutical companies including Pfizer, Sanofi, Genentech, and GSK are proactively implementing and adopting systems that use AI, ML, and DL to power the search for new drug candidates.⁶⁵

In the literature, experimental and data scientists in drug discovery have shown diverse applications of ML and DL.^66–73 For both AI–ML–DL and systems biology–inspired modeling space, the prediction of DILI or hepatotoxicity is one of the most notable areas in which these modeling efforts are being applied. DILI is a major clinical and pharmaceutical concern leading to the termination of potential drug candidates. A ML-based platform called CANscript, developed by Mitra Biotech, is a good example of how in vitro or ex vivo models enable accurate recapitulation of the 3D architecture of the tumor and the tumor microenvironment, including the immune component. The data produced by the experimental platform can be used to train a ML algorithm to give a predictive translatability score (“M-score”) or clinical efficacy for the drug response for each drug tested.⁷⁴ This shows the potential for building more human-relevant complex in vitro models and the use of ML for generating more clinically relevant data.

Another notable effort in computational modeling and application for drug discovery is multiscale systems biology modeling–integration of QSP or QST. Edington et al. demonstrated how a QSP approach can be applied to plan and interpret pharmacological experiments in multi-MPS via proof-of-principle experiments comparing experimental results with the theoretical distribution values calculated from the PBPK models looking at endogenous albumin distribution in a four-way MPS platform and exogenous oral drug distribution in a seven-way MPS platform.⁷⁵ The author noted that QSP approaches using PBPK models of multi-MPS platforms are critical for experimental design and operational strategies when defining dose selection, sampling time points, and setting other MPS parameters.^76,77

There is no doubt that the integration of computational models, continuously optimized and retrained by more human-relevant datasets, will synergistically accelerate research in drug discovery. Data-driven models (AI-based) and systems biology–based models (QST and QSP) feeding information to each other and cross-validating their individual predictions have the scope to enable multiparametric optimization and multiscale predictions of human drug responses, which will support the design, optimization, and validation of new future emerging complex in vitro models, especially for qualifying and validating the clinical translatability of the models. The aspiration is that integrating in vitro (including CIVMs), in vivo, and in silico models, and implementing them in every phase of drug discovery with confidence in translatability, will accelerate our current sequential, inefficient, and costly drug discovery process with a higher success rate of development of our next generation of efficacious and safe drugs for our patients.

Footnotes

Declaration of Conflicting Interests

All authors were employed by GlaxoSmithKline, and their research and authorship of this article was completed within the scope of their employment with GlaxoSmithKline.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Jason E. Ekert

Wendy Rowan

References

Hay

Thomas

D. W.

Craighead

J. L.

; et al. Clinical Development Success Rates for Investigational Drugs. Nat. Biotechnol. 2014, 32, 40–51.

Morgan

Brown

D. G.

Lennard

; et al. Impact of a Five-Dimensional Framework on R&D Productivity at AstraZeneca. Nat. Rev. Drug Discov. 2018, 17, 167–181.

Nelson

M. R.

Tipney

Painter

J. L.

; et al. The Support of Human Genetic Evidence for Approved Drug Indications. Nat. Genet. 2015, 47, 856–860.

Proctor

W. R.

Foster

A. J.

Vogt

; et al. Utility of Spherical Human Liver Microtissues for Prediction of Clinical Drug-Induced Liver Injury. Arch. Toxicol. 2017, 1–15.

Ewart

Fabre

Chakilam

; et al. Navigating Tissue Chips from Development to Dissemination: A Pharmaceutical Industry Perspective. Exp. Biol. Med. (Maywood) 2017, 1535370217715441.

Jeong

C. G.

Dal Negro

Getsios

; et al. Application of Complex In Vitro Models (CIVMs) in Drug Discovery for Safety Testing and Disease Modeling. In Microfluidic Cell Culture Systems; 2nd ed.; Borenstein

J. T.;

Tandon

V.;

Tao

S. L.

; et al., Eds.; Elsevier: Amsterdam, 2019; Chapter 5, pp. 121–158.

Maass

Stokes

C. L.

Griffith

L. G.

; et al. Multi-Functional Scaling Methodology for Translational Pharmacokinetic and Pharmacodynamic Applications Using Integrated Microphysiological Systems (MPS). Integr. Biol. (Camb.) 2017, 9, 290–302.

Carragher

Piccinini

Tesei

; et al. Concerns, Challenges and Promises of High-Content Analysis of 3D Cellular Models. Nat. Rev. Drug Discov. 2018, 17, 606.

Scannell

J. W.

Bosley

When Quality Beats Quantity: Decision Theory, Drug Discovery, and the Reproducibility Crisis. PLoS One 2016, 11, e0147215.

10.

Huh

Leslie

D. C.

Matthews

B. D.

; et al. A Human Disease Model of Drug Toxicity-Induced Pulmonary Edema in a Lung-on-a-Chip Microdevice. Sci. Transl. Med. 2012, 4, 159ra147.

11.

Kasendra

Luc

Yin

; et al. Duodenum Intestine-Chip for Preclinical Drug Assessment in a Human Relevant Model. Elife 2020, 9.

12.

Moriya

Quantitative Nature of Overexpression Experiments. Molec. Biol. Cell 2017, 26, 3893–4181.

13.

Geraghty

R. J.

Capes-Davis

Davis

J. M.

; et al. Guidelines for the Use of Cell Lines in Biomedical Research. Br. J. Cancer 2014, 111, 1021–1046.

14.

Najgebauer

Yang

Francies

; et al. CELLector: Genomics Guided Selection of Cancer In Vitro Models. BioRxiv 2018.

15.

Bell

C. C.

Dankers

A. C. A.

Lauschke

V. M.

; et al. Comparison of Hepatic 2D Sandwich Cultures and 3D Spheroids for Long-Term Toxicity Applications: A Multicenter Study. Toxicol. Sci. 2018, 162, 655–666.

16.

Rowe

R. G.

Daley

G. Q.

Induced Pluripotent Stem Cells in Disease Modelling and Drug Discovery. Nat. Rev. Genet. 2019.

17.

Gordon

E. D.

Locksley

R. M.

Fahy

J. V.

Cross-Talk between Epithelial Cells and Type 2 Immune Signaling: The Role of IL-25. Am. J. Respir. Crit. Care Med. 2016, 193, 935–936.

18.

Fujii

Clevers

Sato

Modeling Human Digestive Diseases with CRISPR-Cas9-Modified Organoids. Gastroenterology 2019, 156, 562–576.

19.

Gehart

Clevers

Tales from the Crypt: New Insights into Intestinal Stem Cells. Nat. Rev. Gastroenterol. Hepatol. 2019, 16, 19–34.

20.

Lancaster

M. A.

Renner

Martin

C. A.

; et al. Cerebral Organoids Model Human Brain Development and Microcephaly. Nature 2013, 501, 373–379.

21.

Sato

Vries

R. G.

Snippert

H. J.

; et al. Single Lgr5 Stem Cells Build Crypt-Villus Structures In Vitro without a Mesenchymal Niche. Nature 2009, 459, 262–265.

22.

Takasato

P. X.

Becroft

; et al. Directing Human Embryonic Stem Cell Differentiation towards a Renal Lineage Generates a Self-Organizing Kidney. Nat. Cell Biol. 2014, 16, 118–126.

23.

Dye

B. R.

Hill

D. R.

Ferguson

M. A.

; et al. In Vitro Generation of Human Pluripotent Stem Cell Derived Lung Organoids. Elife 2015, 4.

24.

Schwank

Koo

B. K.

Sasselli

; et al. Functional Repair of CFTR by CRISPR/Cas9 in Intestinal Stem Cell Organoids of Cystic Fibrosis Patients. Cell Stem Cell 2013, 13, 653–658.

25.

Dekkers

J. F.

Wiegerinck

C. L.

de Jonge

H. R.

; et al. A Functional CFTR Assay Using Primary Cystic Fibrosis Intestinal Organoids. Nat. Med. 2013, 19, 939–945.

26.

Drost

van Jaarsveld

R. H.

Ponsioen

; et al. Sequential Cancer Mutations in Cultured Human Intestinal Stem Cells. Nature 2015, 521, 43–47.

27.

Matano

Date

Shimokawa

; et al. Modeling Colorectal Cancer Using CRISPR-Cas9-Mediated Engineering of Human Intestinal Organoids. Nat. Med. 2015, 21, 256–262.

28.

Drost

van Boxtel

Blokzijl

; et al. Use of CRISPR-Modified Human Stem Cell Organoids to Study the Origin of Mutational Signatures in Cancer. Science 2017, 358, 234–238.

29.

Blume

Swindle

E. J.

Dennison

; et al. Barrier Responses of Human Bronchial Epithelial Cells to Grass Pollen Exposure. Eur. Respir. J. 2013, 42, 87–97.

30.

Bhatia

S. N.

Ingber

D. E.

Microfluidic Organs-on-Chips. Nat. Biotechnol. 2014, 32, 760–772.

31.

Cirit

Stokes

C. L.

Maximizing the Impact of Microphysiological Systems with In Vitro-In Vivo Translation. Lab Chip 2018, 18, 1831–1837.

32.

Vijayavenkataraman

Yan

W. C.

W. F.

; et al. 3D Bioprinting of Tissues and Organs for Regenerative Medicine. Adv. Drug Deliv. Rev. 2018, 132, 296–332.

33.

Retting

Carter

Crogan-Grundy

; et al. Modeling Liver Biology and the Tissue Response to Injury in Bioprinted Human Liver Tissues. Appl. In Vitro Toxicol. 2018, 4, 288–303.

34.

Nguyen

D. G.

Funk

Robbins

J. B.

; et al. Bioprinted 3D Primary Liver Tissues Allow Assessment of Organ-Level Response to Clinical Drug Induced Toxicity In Vitro. PLoS One 2016, 11, e0158674.

35.

Gilbert

L. A.

Horlbeck

M. A.

Adamson

; et al. Genome-Scale CRISPR-Mediated Control of Gene Repression and Activation. Cell 2014, 159, 647–661.

36.

Roth

T. L.

Puig-Saus

; et al. Reprogramming Human T Cell Function and Specificity with Non-Viral Genome Targeting. Nature 2018, 559, 405–409.

37.

Seki

Rutz

Optimized RNP Transfection for Highly Efficient CRISPR/Cas9-Mediated Gene Knockout in Primary T Cells. J. Exper. Med. 2018, 215, 985–997.

38.

Wendler

Wehling

Translatability Score Revisited: Differentiation for Distinct Disease Areas. J. Transl. Med. 2017, 15, 226.

39.

Benam

K. H.

Villenave

Lucchesi

; et al. Small Airway-on-a-Chip Enables Analysis of Human Lung Inflammation and Drug Responses In Vitro. Nat. Methods 2016, 13, 151–157.

40.

Phan

D. T. T.

Wang

Craver

B. M.

; et al. A Vascularized and Perfused Organ-on-a-Chip Platform for Large-Scale Drug Screening Applications. Lab Chip 2017, 17, 511–520.

41.

Hassell

B. A.

Goyal

Lee

; et al. Human Organ Chip Models Recapitulate Orthotopic Lung Cancer Growth, Therapeutic Responses, and Tumor Dormancy In Vitro. Cell Rep. 2018, 23, 3698.

42.

Jeon

J. S.

Bersini

Gilardi

; et al. Human 3D Vascularized Organotypic Microfluidic Assays to Study Breast Cancer Cell Extravasation. Proc. Natl. Acad. Sci. USA. 2015, 112, 214–219.

43.

Boussommier-Calleja

Chen

M. B.

; et al. Microfluidics: A New Tool for Modeling Cancer-Immune Interactions. Trends Cancer 2016, 2, 6–19.

44.

Ortega-Prieto

A. M.

Skelton

J. K.

Wai

S. N.

; et al. 3D Microfluidic Liver Cultures as a Physiological Preclinical Tool for Hepatitis B Virus Infection. Nat. Commun. 2018, 9, 682.

45.

Barrile

van der Meer

A. D.

Park

; et al. Organ-on-Chip Recapitulates Thrombosis Induced by an Anti-CD154 Monoclonal Antibody: Translational Potential of Advanced Microengineered Systems. Clin. Pharmacol. Ther. 2018, 104, 1240–1248.

46.

Shin

Kim

H. J.

Pathomimetic Modeling of Human Intestinal Diseases and Underlying Host-Gut Microbiome Interactions in a Gut-on-a-Chip. Methods Cell Biol. 2018, 146, 135–148.

47.

Kim

H. J.

Collins

J. J.

; et al. Contributions of Microbiome and Mechanical Deformation to Intestinal Bacterial Overgrowth and Inflammation in a Human Gut-on-a-Chip. Proc. Natl. Acad. Sci. USA. 2016, 113, E7–E15.

48.

Bolognin

Fossepre

Qing

; et al. 3D Cultures of Parkinson’s Disease-Specific Dopaminergic Neurons for High Content Phenotyping and Drug Testing. Adv. Sci. (Weinh.) 2019, 6, 1800927.

49.

Kota

Hou

Guerrant

; et al. A Novel Three-Dimensional High-Throughput Screening Approach Identifies Inducers of a Mutant KRAS Selective Lethal Phenotype. Oncogene 2018, 37, 4372–4384.

50.

Kunz-Schughart

L. A.

Freyer

J. P.

Hofstaedter

; et al. The Use of 3-D Cultures for High-Throughput Screening: The Multicellular Spheroid Model. J. Biomol. Screen. 2004, 9, 273–285.

51.

Rowntree

R. K.

McNeish

J. D.

Induced Pluripotent Stem Cells: Opportunities as Research and Development Tools in 21st Century Drug Discovery. Regen. Med. 2010, 5, 557–568.

52.

Jabs

Zickgraf

F. M.

Park

; et al. Screening Drug Effects in Patient-Derived Cancer Cells Links Organoid Responses to Genome Alterations. Mol. Syst. Biol. 2017, 13, 955.

53.

Ekert

J. E.

Johnson

Strake

; et al. Three-Dimensional Lung Tumor Microenvironment Modulates Therapeutic Compound Responsiveness In Vitro: Implication for Drug Development. PLoS One 2014, 9, e92248.

54.

Madoux

Tanner

Vessels

; et al. A 1536-Well 3D Viability Assay to Assess the Cytotoxic Effect of Drugs on Spheroids. SLAS Discov. 2017, 22, 516–524.

55.

Boehnke

Iversen

P. W.

Schumacher

; et al. Assay Establishment and Validation of a High-Throughput Screening Platform for Three-Dimensional Patient-Derived Colon Cancer Organoid Cultures. J. Biomol. Screen. 2016, 21, 931–941.

56.

Feng

Ueda

Levkin

P. A.

Droplet Microarrays: From Surface Patterning to High-Throughput Applications. Adv. Mater. 2018, 30, e1706111.

57.

van de Wetering

Francies

H. E.

Francis

J. M.

; et al. Prospective Derivation of a Living Organoid Biobank of Colorectal Cancer Patients. Cell 2015, 161, 933–945.

58.

Loessner

Stok

K. S.

Lutolf

M. P.

; et al. Bioengineered 3D Platform to Explore Cell-ECM Interactions and Drug Resistance of Epithelial Ovarian Cancer Cells. Biomaterials 2010, 31, 8494–8506.

59.

Lee

J. M.

Mhawech-Fauceglia

Lee

; et al. A Three-Dimensional Microenvironment Alters Protein Expression and Chemosensitivity of Epithelial Ovarian Cancer Cells In Vitro. Lab. Invest. 2013, 93, 528–542.

60.

Onakpoya

I. J.

Heneghan

C. J.

Aronson

J. K.

Post-Marketing Withdrawal of Anti-Obesity Medicinal Products Because of Adverse Drug Reactions: A Systematic Review. BMC Med. 2016, 14, 191.

61.

Bell

C. C.

Hendriks

D. F.

Moro

S. M.

; et al. Characterization of Primary Human Hepatocyte Spheroids as a Model System for Drug-Induced Liver Injury, Liver Function and Disease. Sci. Rep. 2016, 6, 25187.

62.

FDA. FDA’s Predictive Toxicology Roadmap. https://www.fda.gov/media/109634/download.

63.

IQ MPS. IQ Microphysiological Systems Affiliate. https://www.iqmps.org.

64.

Knudsen

T. B.

Keller

D. A.

Sander

; et al. FutureTox II: In Vitro Data and In Silico Models for Predictive Toxicology. Toxicol. Sci. 2015, 143, 256–267.

65.

Fleming

How Artificial Intelligence Is Changing Drug Discovery. Nature 2018, 557, S55–S57.

66.

Lavecchia

Di Giovanni

Virtual Screening Strategies in Drug Discovery: A Critical Review. Curr. Med. Chem. 2013, 20, 2839–2860.

67.

Alqahtani

In Silico ADME-Tox Modeling: Progress and Prospects. Expert Opin. Drug Metab. Toxicol. 2017, 13, 1147–1158.

68.

Korotcov

Tkachenko

Russo

D. P.

; et al. Comparison of Deep Learning with Multiple Machine Learning Methods and Metrics Using Diverse Drug Discovery Data Sets. Mol. Pharm. 2017, 14, 4462–4475.

69.

Norris

J. L.

Farrow

M. A.

Gutierrez

D. B.

; et al. Integrated, High-Throughput, Multiomics Platform Enables Data-Driven Construction of Cellular Responses and Reveals Global Drug Mechanisms of Action. J. Proteome Res. 2017, 16, 1364–1375.

70.

Christiansen

E. M.

Yang

S. J.

Ando

D. M.

; et al. In Silico Labeling: Predicting Fluorescent Labels in Unlabeled Images. Cell 2018, 173, 792–803 e19.

71.

Chen

C. L.

Mahjoubfar

Tai

L. C.

; et al. Deep Learning in Label-Free Cell Classification. Sci. Rep. 2016, 6, 21471.

72.

Gopakumar

Hari Babu

Mishra

; et al. Cytopathological Image Analysis Using Deep-Learning Networks in Microfluidic Microscopy. J. Opt. Soc. Am. A Opt. Image Sci. Vis. 2017, 34, 111–121.

73.

Kourou

Exarchos

T. P.

Exarchos

K. P.

; et al. Machine Learning Applications in Cancer Prognosis and Prediction. Comput. Struct. Biotechnol. J. 2015, 13, 8–17.

74.

Majumder

Baraneedharan

Thiyagarajan

; et al. Predicting Clinical Response to Anticancer Drugs Using An Ex Vivo Platform That Captures Tumour Heterogeneity. Nat. Commun. 2015, 6, 6169.

75.

Edington

C. D.

Chen

W. L. K.

Geishecker

; et al. Interconnected Microphysiological Systems for Quantitative Biology and Pharmacology Studies. Sci. Rep. 2018, 8, 4530.

76.

Cilfone

N. A.

Large

E. M.

; et al. Quantitative Systems Pharmacology Approaches Applied to Microphysiological Systems (MPS): Data Interpretation and Multi-MPS Integration. CPT Pharmacometrics Syst. Pharmacol. 2015, 4, 585–594.

77.

Tsamandouras

Chen

W. L. K.

Edington

C. D.

; et al. Integrated Gut and Liver Microphysiological Systems for Quantitative In Vitro Pharmacokinetic Studies. AAPS J. 2017, 19, 1499–1512.

78.

Bauer

Wennberg Huldt

Kanebratt

K. P.

; et al. Functional Coupling of Human Pancreatic Islets and Liver Spheroids on-a-Chip: Towards a Novel Human Ex Vivo Type 2 Diabetes Model. Sci. Rep. 2017, 7, 14620.

79.

Blume

Reale

Held

; et al. Cellular Crosstalk between Airway Epithelial and Endothelial Cells Regulates Barrier Functions during Exposure to Double-Stranded RNA. Immun. Inflamm. Dis. 2017, 5, 45–56.

80.

Alepee

Bahinski

Daneshian

; et al. State-of-the-Art of 3D Cultures (Organs-on-a-Chip) in Safety Testing and Pathophysiology. ALTEX 2014, 31, 441–477.

81.

Zhang

Gao

; et al. 3D Bioprinting: A Novel Avenue for Manufacturing Tissues and Organs. Engineering 2019, 5, 777–794.

82.

Knight

Przyborski

Advances in 3D Cell Culture Technologies Enabling Tissue-Like Structures to Be Created In Vitro. J. Anat. 2015, 227, 746–756.

83.

van Duinen

Trietsch

S. J.

Joore

; et al. Microfluidic 3D Cell Culture: From Tools to Tissue Models. Curr. Opin. Biotechnol. 2015, 35, 118–126.

84.

Howes

A. L.

Richardson

R. D.

Finlay

; et al. 3-Dimensional Culture Systems for Anti-Cancer Compound Profiling and High-Throughput Screening Reveal Increases in EGFR Inhibitor-Mediated Cytotoxicity Compared to Monolayer Culture Systems. PLoS One 2014, 9, e108283.

85.

Pastula

Middelhoff

Brandtner

; et al. Three-Dimensional Gastrointestinal Organoid Culture in Combination with Nerves or Fibroblasts: A Method to Characterize the Gastrointestinal Stem Cell Niche. Stem Cells Int. 2016, 2016, 3710836.

86.

Marx

Andersson

T. B.

Bahinski

; et al. Biology-Inspired Microphysiological System Approaches to Solve the Prediction Dilemma of Substance Testing. ALTEX 2016, 33, 272–321.

87.

Hunziker

Choudhury

Engineering Microfluidic Organoid-on-a-Chip Platforms. Micromachines (Basel) 2019, 10.

88.

Rezaei Kolahchi

Khadem Mohtaram

Pezeshgi Modarres

; et al. Microfluidic-Based Multi-Organ Platforms for Drug Discovery. Micromachines (Basel) 2016, 7.

89.

O’Brien

P. J.

Irwin

Diaz

; et al. High Concordance of Drug-Induced Human Hepatotoxicity with In Vitro Cytotoxicity Measured in a Novel Cell-Based Model Using High Content Screening. Arch. Toxicol. 2006, 80, 580–604.