Changing demands in global healthcare over the past 15 years have led to greater complexity and spiralling costs in drug development. The average price tag of taking a new drug from discovery to completion of Phase III clinical trials is now $2.87 billion (1), which means informed decisions need to be made early on about which compounds to pursue.
AstraZeneca’s new R&D framework, termed the ‘5Rs’, has introduced an increased scientific rigour and emphasis on quality, driving an almost five-fold increase in R&D productivity. AstraZeneca has surpassed the industry norm in recent years, moving from a 4% success rate in molecules progressing from candidate nomination to completion of Phase III trials in 2010, to more than 19% in 2017 (2).
In our pursuit for continual improvement, the question now becomes how far can we push an even greater improvement in drug discovery productivity? One approach is by adopting new and evolving preclinical technologies to further improve clinical translation and reduce clinical attrition.
R&D productivity across the industry has declined from 9% in the 1990s to 4% in 2009 (3). AstraZeneca recognised the need to make bold changes; in 2010, we conducted an in-depth review of our R&D to identify critical ‘success factors’. As a result, we launched the ‘5R framework’, which champions quality over quantity and has transformed the culture of our organisation.
This is a new model of working, based on ensuring each project team focuses on improving their understanding around a key set of criteria which we believe increase the probability of success: right target, right tissue, right safety, right patient, and right commercial (see Table 1).
Between 2005 and 2010, AstraZeneca’s preclinical pipeline contained around 200 projects at any time. After the implementation of the 5R framework, the number of projects halved, with the remaining projects having stronger validation as deemed by the 5R criteria; while the number of projects decreased, the probability of success increased. This quality-over-quantity approach led to decreased clinical attrition, a fuller clinical pipeline and an increase in overall R&D productivity.
Today, our projects rarely fail for safety reasons, or missing a proof of mechanism. Instead, the major cause of failure is a scientific hypothesis that turns out to be incorrect. If we can improve our understanding of disease biology and how best to modulate validated biological pathways, we believe this will further improve our R&D productivity. In recent years, we have adopted methodologies based on advanced ‘omics’ and humanised models, allowing us to develop our ability to rigorously test hypotheses and translate science into medicine.
Here we review four technologies that we are investing in to improve predictability of preclinical data: CRISPR-Cas9 system for efficient target validation; patient-derived xenograft (PDX) models to model human tumours; mass spectrometry imaging (MSI) to gather vast datasets on multiple predictive criteria; and humanised models that realistically simulate the tissue environment in vitro.
While in vivo animal systems provide a lifelike system in terms of architecture and complex physiological interactions, in many cases the effects of drugs on animal tissue are not translatable to humans. This lack of translatability is one reason for drug trial failures. Humanised models bridge this gap, providing an environment in which human cells behave more closely to how they would in the body. These technologies have the added advantage over 2D and 3D cell culture in that they mimic aspects of the tissue environment to provide human-relevant data on multiple parameters including toxicity, efficacy and PK/PD determination prior to compounds ever being tested in patients.
Patient-derived xenograft (PDX) models
Clinical cancers often have complex, hard-to-predict phenotypes that are not adequately reflected in available cell lines. Tumour complexity and heterogeneity can, however, be recreated using PDX models – in vivo disease models made by grafting tumour tissue directly into immunodeficient mice. This provides a translatable system for testing drugs in the ‘right tissue’, giving a stronger prediction of how drugs might behave in a real human tumour than classical human in vitro or animal in vivo models.
PDX models are of particular value in studying late-stage cancers and modelling resistance mechanisms. Samples taken from late-stage tumours and patients relapsing on available therapies provide new models to explore next-generation approaches that could treat or even prevent resistance emergence. As part of an academic collaboration, we recently published a study using a series of PDX models of colorectal cancer. We showed that a novel inhibitor of DNA damage repair protein, ATM, could resensitise chemotherapy-resistant tumours to respond to therapy in preclinical models (4). This is an encouraging lead, and supports further research into the inhibitor as part of a combination therapy.
As part of a wider collaboration with several partners, we are testing drugs in a biobank of breast cancer PDX models (5). These heterogeneous models can also be used to generate short-term organ cultures to test drug response at higher throughput. These systems enable us to measure drug effects across a diverse cancer model population, simulating a clinical trial population.
To further these innovative approaches to candidate screening, we are currently researching ways to incorporate a human immune system into mouse models. Immune interactions are fundamental in shaping cancer progression and response to therapy, so this would provide a translatable preclinical model that will support the development of next-generation immuno-oncology drugs.
PDX models provide a unique means of simulating multiple aspects of real-world heterogeneity and tumour behaviour outside of the patient. The predictability of these models when considering the behaviour of therapeutic candidates is already proving to be superior when compared to homogenous cell lines, or cell line-derived xenografts. As improvements are made to model the tumour environment they will provide even stronger validation packages for innovative treatments and combinations, especially in the arena of drug resistance.
Organs-on-Chips, also known as microphysiological systems, are in vitro systems that recreate the in vivo microenvironment (6).
The chips contain tiny channels lined with extracellular matrix, which can be populated with almost any cell type. Cell media is flowed through the channels, creating the mechanical shear stress present in all tissues. These complex systems behave like real tissue in many aspects, and can even represent elements of organ functionality. The Lung-Chip, for example, can simulate breathing processes, with airflow across a surface of cells and blood flow beneath, allowing for gaseous exchange. This provides a realistic system for modelling effects of interest, such as properties of inhaled compounds. Organs-on-Chips can also contain immune cells and blood components, enabling interactions that influence disease and compound behaviour to be replicated.
Establishing compound safety in relevant tissue is a major application of our Organs-on-Chips. If a toxic response is seen in an animal model, we may not know whether it is generally toxic, or a species specific effect. This limit in translatability between different animal species and humans creates a critical and difficult decision-making stage in drug discovery. In order to bridge the gap between animal in vivo and human in vitro systems, we have created rat Liver-Chips and dog Liver-Chips, and are currently working to confirm that they give an equivalent toxicity response to their in vivo counterparts. Once we establish this, we will be able to compare data across human, dog, and rat Liver- Chips to determine whether toxic responses in animals are species specific.
Advancing this technology even further, we are starting to investigate multi-organ systems. These can be used, for example, to model complex disease processes and allow us to rapidly assess the functional outcome of targeting in a human-relevant system. We are currently developing a model of type 2 diabetes using a Liver-Chip coupled to a Pancreas-Chip (7). Preliminary results show that when the Liver-Chip is insulin-resistant, it can feed back to the pancreas affecting insulin release, as is the case in type 2 diabetes. Once the system is fully tested, we will be able to use this as a diabetes model for target validation.
AstraZeneca is committed to the adoption and application of Organs-on-Chips technology, and recently announced a partnership with Emulate Bio, a leading producer and founded by pioneers of the technology, to embed Organ-Chips within our drug safety laboratories and eventually across our entire pipeline.
Humanised miniature organs
Cardiovascular disease is a priority therapy area for AstraZeneca, so translatable heart models are important for our early science prediction. Mechanical and electrical forces in a beating heart are difficult to realistically simulate in vitro. We recently published a proof-of-principle paper in which we developed a humanised heart model that mimics the complexity of heart function (8).
By ‘decellularising’ a rat heart to leave a scaffold, we were able to recreate a realistic model of a functional heart by repopulating it with human iPSC-derived heart cells. The model had features such as vasculature, valve function and the ability to beat. Accurately predicting cardiac drug effects and disease biology requires a model that mimics all the complexity of the heart. The model is still in development, but there is hope that it will be useable as an ex-vivo translational model for drug discovery and understanding cardiac disease.
3D bioprinting involves printing-like techniques, where ‘ink’, often silicon, printed on to a gelatin/fibrogen matrix is used to build scaffolds that are populated with cells during or after the printing process to create biological structures. These can include organ components in which the heterogeneity and architecture of their biological counterparts are preserved, providing a more translatable system than 2D models.
This has been of use in the kidney field, where previously there has been a lack of translatable models. We have been collaborating with a group at Harvard University who have published a bioprinting method for creating 3D human renal proximal tubules (9). This provides a realistic model for studying cellular crosstalk, drug uptake, delivery and toxicity, and to support biomarker discovery, target identification and validation. The aim for the future will be to build on the complexity of these models to mimic the in vivo environment.
Other structures, such as vasculature, could be printed and linked to simulate cellular crosstalk, and eventually components of a whole kidney could be created. An immediate next step is to use patient-derived IPS cells in place of cell lines to mimic real patients’ disease. As with other human tissue-based models, 3D-bioprinted systems can be used to understand complex, system-wide interactions and simulate the disease mechanisms and compound behaviour early in the drug discovery process.
Mass-spectrometry imaging (MSI)
MSI combines mass spectrometry data with spatial information to map the molecular composition of a sample surface in exquisite detail. MSI can show distribution and behaviour of almost all molecule types in a sample. This includes drug candidates and their metabolites, as well as proteins and lipids. This depth of information provides unparalleled insights into how drugs and targets interact in a tissue. It can be used to give predictive information on compound safety and efficacy, as well as guiding drug delivery methods and formulations.
The predominant use of MSI at AstraZeneca is in early discovery projects to inform preclinical decision making. A key area for us where this has had significant impact is in understanding bloodbrain barrier (BBB) penetration. Due to the complexity of systems, this is a challenge to model and a ‘blind spot’ in terms of predicting compound behaviour. In a recent publication, we demonstrated that MSI could be used to show how drug-drug interactions influence blood-brain barrier permeability. Traditional imaging technology would not be able to show these interactions in all their complexity, so MSI bridges this gap (10).
A joint project between AstraZeneca and Cancer Research UK (CRUK) aims to map cancer topology in unprecedented molecular detail to help in PK/PD modelling, determining phenotypes and predicting responses to therapy. Massive datasets of spatially-resolved molecular information will be combined to create Google Earth-style multi-level rendered maps of tumours. Users will be able to use these to explore tumour architecture from a whole-tumour overview, down to individual molecules.
The amount of detail captured using MSI results in immense datasets. We are working with leaders in this field including the National Centre of Excellence in Mass Spectrometry Imaging (NiCEMSI) at the National Physical Laboratory (NPL), to explore how to mine this high-dimensionality data to predict how drugs will behave in a patient. As with other high-content techniques, the MSI development must be coupled with advances in data storage, artificial intelligence (AI) algorithms and processing power. As improvements in computational power expand to accommodate and make sense of these datasets, MSI can be used to build more informative clinical packages than ever before.
CRISPR-Cas9 genome editing
In 2013, two seminal studies published in Science described how CRISPR-Cas9 system could be harnessed for gene editing in mammalian cells (11,12). Since then, it has revolutionised gene editing for biomedical research. Thanks to its ease of use and implementation, CRISPR is a powerful tool in target selection and validation. It can be used to create precise genetic disease models, often involving the change of a single nucleotide.
We have applied CRISPR to create more than 120 cellular disease models in which we have deleted or introduced single nucleotide changes to genes to examine the effect of specific genes on disease pathways. For example, we used CRISPR to reveal a new drug target in chronic obstructive pulmonary disease (COPD). Previous research had shown that salt-inducible kinase (SIK) was a factor in inflammatory response in some COPD patients. SIK exists in three isoforms, and it was not clear which particular isoform was responsible for the pathological effect.
Using CRISPR-Cas9, we created iPSC-derived macrophages expressing mutants of the three isoforms (13). By creating kinase-dead versions of each of the three isoforms, we identified which isoform was responsible for disease. This has led to initiation of a programme to develop a specific inhibitor for the target isoform. While a broad-spectrum SIK inhibitor might have been effective, we decided to hone in on the precise target, maximising the specificity of the potential therapy and chance of success.
We have also used CRISPR to rule out what otherwise appeared to be promising targets before investing in screening programmes. In 2014, two Nature papers suggested that MTH-1 inhibitors killed cancer cells, making it a ‘hot topic’ at the time (14,15). We subsequently used CRISPR to delete MTH-1 – precisely knocking it down was a clean, definitive means of assessing its role. Surprisingly, this had a much smaller effect on cancer cell survival than had been predicted, invalidating it as a target and saving on further clinical development only for it to fail later on (16).
A core application at AstraZeneca is in the creation of libraries of CRISPR reagents to delete, upregulate or downregulate expression of every gene in the genome (17). Alongside partners at the Sanger Institute and the Innovative Genomics Institute, we have tested these reagents to identify new disease targets – a major priority in creating preclinical validation packages and, arguably, the most important decision made during research. In terms of target identification, CRISPR offers a precise, efficient new paradigm compared to a small molecule screen.
To put this into context, a genome-wide CRISPR screen may use 20,000 assay wells when a small molecule screen may use hundreds of thousands of compounds. Identifying the actual target of a small molecule screen post-hit can take years, whereas CRISPR screening directly identifies genes of interest. Target identification and validation are arguably some of the most crucial elements of drug discovery – at AstraZeneca, 83% of projects fail for reasons related to target validation, so adopting CRISPR to optimise this early on will deliver on driving pipeline efficiency and therapeutic prediction of successful candidates.
Summary - Improving Drug Development Productivity With Better Predictivity
Our 5R framework has been a major factor in transforming R&D productivity, and supporting AstraZeneca’s return to growth. We now set ourselves the challenge to continue our journey of R&D productivity improvement. By testing our hypotheses in preclinical models that more closely resemble the human conditions we are aiming to treat, we believe we will select better drug targets and make better predictions for which candidates to test clinically.
Looking to the future, there is great interest in the potential of machine learning and artificial intelligence algorithms to make the most of data. Constructing ‘knowledge graphs’ using all our preclinical and clinical datasets, including data from humanised models, MSI and CRISPR combined with public-domain data, will layer information in a way that will better inform and identify putative therapeutic targets.
The 5R framework has helped refine the selection of molecules we feed into the clinical pipeline, and has led to an improvement in the quality of the drug candidates we take forward into clinical trials. Continuous review and evolution of our drug discovery and development approaches will hopefully uncover further opportunities for continued improvement in our productivity, and our ability to bring innovative medicines to patients faster. DDW
Dr Lorna C. Ewart heads AstraZeneca’s Centre of Excellence for Microphyisological Systems. Here she has established strategic collaborations with leading external academics as well as led technology transfer into AstraZeneca. She is a trained pharmacologist with 20 years’ experience in the pharmaceutical industry spanning both the efficacy and safety domains.
Dr Richard J.A. Goodwin is a Principal Scientist at AstraZeneca and leads the mass spectrometry imaging group within Drug Safety and Metabolism. He joined AstraZeneca in Sweden as a post doc in 2011. Since then he has helped build the multimodal imaging capability through internal expansion and external collaborations.
Dr Stephen E. Fawell is Vice President, Head Oncology iScience, for the IMED Biotech Unit at AstraZeneca, with responsibility for target selection, drug discovery and optimisation and overseeing biology, pharmacology, DMPK and chemistry resources. Steve obtained his PhD at the University of Leeds, UK, and completed post-doctoral fellowships at Rutgers Medical School NJ and the Imperial Cancer Research Fund.
Dr Pernille B.L. Hansen is head of Bioscience Chronic Kidney Disease of AstraZeneca’s IMED Biotech Unit, responsible for delivering new targets for CKD and lead the bioscience kidney discipline. She holds a Professorship in Cardiovascular and Renal Research at University of Southern Denmark.
Dr Catherine C. Priestley leads Science Communications and Relations for IMED Biotech Unit at AstraZeneca and is a member of the ABPI Innovation Board where she advises on areas of focus for UK life sciences. As a toxicologist, she has published more than 30 manuscripts assessing the safety profile of new modalities.
Steve Rees is Vice President of Discovery Biology at AstraZeneca. He is Chair of the European Laboratory Research and Innovation group (ELRIG) and is a member of the Scientific Advisory Board for LifeArc and the Centre for Membrane Protein and Receptor research at the Universities of Nottingham and Birmingham.
Dr Menelas (Mene) Pangalos is Executive Vice President of AstraZeneca’s IMED Biotech Unit and AstraZeneca’s Global Business Development. Mene has overall responsibility for AstraZeneca’s research and early development activities spanning three continents and 2,500 people. As one of AstraZeneca’s leading scientists, Mene has published more than 150 peer-reviewed articles and served as editor of numerous books and journals.
1 DiMasi, J et al (2016). Innovation in the pharmaceutical industry: New estimates of R&D costs. Journal of Health Economics, 47: 20-33.
2 Morgan, P et al (2018). Impact of a five-dimensional framework on R&D productivity at AstraZeneca. Nature Reviews Drug Discovery, 17: 167-181.
3 O’Hagan, P et al (2009, March 30). Bringing Pharma R&D back to health. Retrieved from Bain & Company: http://www.bain.com/publications/articles/bringing-pharma-r-and-d-back-to-health.aspx.
4 Greene, J et al (2017). The novel ATM inhibitor (AZ31) enhances antitumor activity in patient derived xenografts that are resistant to irinotecan monotherapy. Oncotarget, 8 (67): 110904-110913.
5 Bruna, A et al (2016). A Biobank of Breast Cancer Explants with Preserved Intratumor Heterogeneity to Screen Anticancer Compounds. Cell, 167 (1): 260-274.
6 Huh, D et al (2010). Reconstituting organ-level lung functions on a chip. Science, 328(5986): 1662-1668.
7 Bauer, S et al (2017). Functional coupling of human pancreatic islets and liver spheroids on-a-chip: Towards a novel human ex vivo type 2 diabetes model. Scientific Reports, doi: 10.1038/s41598- 017-14815-w.
8 Nguyen, DO-H (2018). Humanizing Miniature Hearts through 4-Flow Cannulation Perfusion Decellurazition and Recellurization. Scientific Reports, DOI:10.1038/s41598- 018-25883-x.
9 Homan, K et al (2016). Bioprinting of 3D Convoluted Renal Proximal Tubules on Perfusable Chips. Scientific Reports, 6, 34845.
10 Vallianatou T et al (2018). A mass spectrometry imaging approach for investigating how drug-drug interactions influence drug blood-brain barrier permeability. Neuroimage, 172: 808-816.
11 Cong L et al (2013). Multiplex genome engineering using CRISPR/Cas systems. Science, 339(6121): 819-823.
12 Mali P et al (2013). RNAguided human genome engineering via Cas9. Science, 339 (6121): 823-826.
13 Goransson, M et al (2016). SIK Inhibition: A Novel Opportunity to Modulate Disease Phenotype in COPD. ATS 2016. san Francisco: American Journal of Respiratory and Critical Care Medicine.
14 Gad, H et al (2014). MTH1 inhibition eradicates cancer by preventing sanitation of the dNTP pool. Nature, 508 (7495): 215-221.
15 Huber, K et al. (2014). Stereospecific targeting of MTH1 by (S)-crizotinib as anticancer strategy. Nature, 508(7495): 222-227.
16 Kettle, J et al (2016). Potent and Selective Inhibitors of MTH1 Probe Its Role in Cancer Cell Survival. Journal of medicinal Chemistry, 59 (6): 2346-2361.
17 Canver, M et al (2018). Integrated design, execution, and analysis of arrayed and pooled CRISPR genome editing experiments. Nature Protocols, 13: 946-986.