Virtual Second Species
The aim of this Challenge is to apply advanced computational and mathematical modelling approaches to develop a suite of virtual dog tissues and organs to model toxicological endpoints for New Chemical Entities (NCEs).
This is a Three Phase Challenge with Phase 3 awarded to winners of Phase 2 subject to successful delivery of Phase 2.
Challenge Partners collaborate with the NC3Rs to provide additional resources to successful applicants to help deliver the Challenge. The Challenge Partners for this Challenge are: eTransafe - Innovative Medicines Initiative (IMI) and Simomics Ltd.
Challenge briefing webinar
View the Challenges briefing webinar to find out more about this Challenge.
Apply for this Challenge
Please read the Full Challenge information below and the Guide for Participants. To apply for this Challenge, please use the Phase 1 application form for Three Phase Challenges.
Sponsored by Bayer AG, Eli Lilly and Company, Genentech Inc., Gilead Sciences Inc., GSK, Merck Healthcare KGaA and Roche, and in Partnership with eTransafe IMI and Simomics Litd., this Challenge aims to to apply advanced computational and mathematical modelling approaches to develop a suite of virtual dog tissues and organs to model toxicological endpoints for New Chemical Entities (NCEs).
Toxicology in the pharmaceutical industry
Animals are used in non-clinical studies to assess the efficacy and potential toxic effects of drugs before their first use in humans and alongside subsequent clinical studies. The safety studies aim to identify target organs of toxicity, assess reversibility of potential effects, and assess exposure-response relationships. The data generated informs the safe human starting dose and clinical monitoring. While first-in-human studies are usually regarded as very safe, not all relevant human toxicities are identified in animal studies (1).
Current regulatory guidelines for NCEs such as small molecule drugs usually require safety and tolerability data from two species, a rodent (i.e. rat or mouse) and a non-rodent (i.e. dog, minipig or non-human primate), before administration of potential new medicines in first-in-human clinical trials. Information typically available before these studies are started include early pharmacokinetic (PK) data, physiochemical information, in vitro data (e.g. secondary pharmacology off-target profiling and toxicological screening data), early safety pharmacology studies and any specific investigations to evaluate target-related risk or liabilities with high incidence (e.g. liver). Two different species are used for non-clinical safety assessment to account for species-specific differences in susceptibility. Unless an effect is known to be species specific, toxicities seen in one species are regarded as potential human toxicities to be closely monitored in clinical studies, if regarded as manageable. Longer term toxicity studies (including chronic treatment duration of up to 39 weeks) are also conducted in the same two species to support longer clinical studies as well as potential marketing authorisation of drugs for long-term use. There is good evidence that detection of toxicity in at least one of the toxicology study species is indicative of the potential detection of adverse events in clinical observations. However, the absence of toxicological findings in a non-clinical study is not always indicative that no adverse events will be detected in humans and can lead to late stage drug attrition (2,3). The variability in concordance between non-clinical species and adverse events detected in humans may be due to differences in the translation of effects for specific target organs of interest (e.g. good concordance for cardiovascular effects versus poor translation for central nervous system effects such as headache and nausea), the relatively small number of animals and humans studied during early trials and the lack of appropriate non-clinical models.
The research and development landscape has changed considerably since these regulatory requirements have been in place, with new in vitro and in silico technologies available to evaluate safety in addition to the standard in vivo approach. Opportunities to refine and reduce animal use within toxicology studies have been explored (4) and some flexibility in guideline testing requirements have been incorporated, for example, chronic toxicity studies are not required for drugs indicated for acute/short-term use, or in life-threatening indications. With emerging technologies, there is now the opportunity to consider the relevance of the animal models used in the development of NCEs.
Mathematical modelling and computational approaches to support a move to the use of a single rodent species in toxicology
Work by a NC3Rs-led consortium involving 30 pharmaceutical companies and regulatory bodies from the UK, Europe and North America identified opportunities to use only a single species for chronic toxicity studies (of 13 (sub-chronic) weeks to 26 or 39 (chronic) weeks duration) for a wider range of drug modalities than currently permitted by the regulations (5). This exercise evaluated data retrospectively and additional evidence is required to prospectively determine when a single species chronic toxicity approach may, or may not, be applicable. For NCEs such as small molecule drugs, these opportunities may not be widely adopted until convincing evidence on prediction of outcome of chronic studies is achieved followed by a change in international regulatory guidelines.
The pharmaceutical industry is exploiting advances in human-relevant in vitro sciences and mathematical modelling, data interrogation and analysis to improve approaches used in the assessment of the safety of new medicines. Organisations in the USA such as the National Toxicology Program and the Food and Drug Administration (6,7) have outlined commitments to support the development of these New Approach Methodologies (NAMs) that have the potential to revolutionise drug development. Advances in mathematical and computational modelling, accelerated by Machine Learning, are poised to revolutionise many aspects of daily life, (8,9) and there is a strong research base exploiting these capabilities to advance drug discovery and development. Integrating data from non-clinical studies with in silico tools to maximise the knowledge derived, has the potential to improve the safety assessment of candidate molecules and expedite the drug development process, delivering significant cost benefits and reducing animal use and drug attrition. For example, quantitative structure–activity relationship (QSARs) are widely used in the prediction of physicochemical properties during the selection and optimisation of lead candidates during drug discovery. QSARs have also been applied extensively in toxicology and when combined with read-across and advanced Machine Learning techniques, are increasingly powerful tools that have the potential to provide critical building blocks to parameterise complex biological models (10). However, QSARs are limited by their requirement for adequate knowledge of the chemical space and the extent of their validation (applicability domains), and while useful in screening of novel compounds based on their chemical similarity to historical data, they are not yet able to fully meet regulatory non-clinical testing requirements.
Other modelling approaches include integrated PK and pharmacodynamic (PD) modelling that aim to better characterise drug exposure and response relationships. These models can be further expanded into multi-scale models by building in layers of biological complexity from the sub-cellular to the organ system level and even across populations. Classical modelling approaches can be limited by inherently variable and often incomplete biological datasets that need to be accounted for when creating prediction models. Machine Learning can improve the ability of models to handle large, complex datasets, and when combined with the wealth of non-clinical data available and existing physiological models, has the potential to drive a step-change in predictive toxicology (11). The ability of some Machine Learning algorithms to learn and improve over time and their potential to address data gaps where there is adequate coverage from overlapping areas, further improves their potential predictivity and accuracy. Advances specifically looking to improve models in predictive toxicology include work using neural network approaches to model human organ toxicity and multi-species acute toxicity end points using novel prediction models (12,13). Additionally, agent-based morphometric models of dynamic signalling networks, parameterised by in vitro data on human molecular and cellular targets, can represent emergent behaviours and serve as in silico chemical screening platforms (14,15). Multi-scale mathematical models combining Machine Learning, agent-based, PK/PD and cell or tissue system models within a quantitative systems toxicology framework also offer the opportunity to represent the interplay between organs in a virtual environment (15,16). With the global bio-simulation market projected to grow to over $9 billion by 2028 (17), there are strong drivers to accelerate the development and accuracy of these models in assessing the safety of new medicines.
There is now the potential to apply advanced in silico tools and approaches to support the building of a more robust evidence base to facilitate moving towards using a single (rodent) species, without increasing risk to humans. Utilising the vast amount of historical and contemporary dog study data, a virtual model could be used for the assessment of potential target organ toxicities in the dog. Other non-rodents, such as non-human primates (NHPs) and minipigs, are also common non-rodent species used for small molecule testing. However, the historical use of the dog for small molecule development has produced considerable data in the literature and study reports that can be exploited. There may be potential to expand the model to develop other virtual non-rodent species in the future, and to other drug modalities.
A virtual human in which to assess potential toxicity would perhaps be the ideal model to develop, however, this is likely to be too big of a translational step (from rodent directly into human, without some level of non-rodent data) for acceptance by clinicians and regulators. The validation required for a model predicting human effects will also be much more difficult to achieve than acceptance of a model used to predict long-term dog study outcomes (where there may be the possibility to also validate predictions with shorter term dosing studies in dog and/or rat). Dogs used in toxicology studies are genetically less heterogeneous than the human population providing data more amenable for modelling purposes.
This Challenge aims to develop a model or suite of models that will ultimately replace the use of the dog in six to nine-month (chronic) toxicology studies. Proposals that aim to tackle the problem using a system-wide approach are welcomed. It is, however, acknowledged that the early and mid-term aims may focus on the development of a model that can identify the potential toxicity risk for a series of key target organs known to be affected in the dog, or that have been documented to be likely indicators of risk in humans.
The model is targeted towards the identification of toxicities that may develop upon long-term dosing in the dog, but it could also be used early within discovery as a screen for potential ‘show-stopper’ toxicities in the dog and focus investigations towards those target organs of concern (using other in vitro and in silico models) before testing in animals. Replacement of the dog in chronic toxicity testing will require significant validation for regulatory acceptance and there will likely be a period where results from this model would be run in parallel with the current regulatory testing requirements. The model developed through this Challenge will contribute to the growing evidence base and capability of NAMs to provide robust and predictive methods to assess toxicities and reduce the reliance on animals.
Rodent data will be available to support model development where required. Dog data is generally evaluated in conjunction with rodent data for small molecule safety programmes, and predictivity of the model in relation to associated rodent data is an important factor when developing the model.
In the UK in 2020, of the experimental procedures for repeat dose toxicity , there were 10,670 using mice, 27,432 using rats, 2,082 using dogs and 1,142 using non-human primates.
The typical design for a dog chronic toxicology study is to dose animals daily for six or nine months, dependent on the region where data will be submitted and reviewed (Europe or USA/Japan respectively (18)). It is common to include four main test groups (control and three dose levels) of three or four male and female animals per group, plus additional groups to assess recovery from any effects (usually restricted to control and high dose, using two male and female animals per group) resulting in 40 dogs typically being used for each chronic toxicity study, with each animal undergoing clinical assessment and repeated blood sampling over specific time periods.
The cost of a typical chronic toxicity study is often in excess of £500k, not including the additional costs associated with compound production and quality certification, formulation methodology support and data review. If this study could be replaced with a modelling tool, there would be a significant cost and time saving for industry, as well as the potential for the generation of more informative data.
With the time taken to develop new medicines ranging from 10 to 15 years, advances in NAMs such as in silico models have the potential to deliver improved decision-making tools, with better mechanistic understanding that result in more rapid discovery and development of medicines.
Over the past decade, collaborative industry projects have collected a significant amount of non-clinical data from toxicology and safety assessment studies, providing resources of curated databases of study data from partner pharmaceutical companies. These databases have then been used to generate novel tools for predictive toxicological modelling (19, 20, 21). One current project – Enhancing TRANslational SAFEty Assessment through Integrative Knowledge Management (eTRANSAFE), building on the previous project –eTox, is developing a technological architecture for data sharing of non-clinical and clinical data, with in depth data integration and data exploitation capabilities. This effort is creating a database of curated SEND (“standards for the exchange of non-clinical data”) reports provided by 12 partner pharmaceutical companies in addition to what was generated during the preceding eTox initiative and combining it with chemoinformatics, bioinformatics and clinical drug safety data to assess the predictive translation of non-clinical data to humans. The project has already developed the potential for a ‘rat virtual control group’ (22) which could reduce the need for concurrent controls in each study, instead making comparisons of test data with the historical database. The wealth of data that has been collected and curated through eTox and eTransafe provides a significant resource for delivery of this Challenge, and the activities developed through this Challenge will add additional outcomes and impacts to the core eTRANSAFE goals. eTRANSAFE are partnering with the NC3Rs to support data provision for the successful Challenge winners.
Software technologies from Simomics help support transparency for in silico models, embedded in their virtual laboratory approach (23). Components of their core technology were developed through previous CRACK IT funding and Simomics are keen to offer this to the successful teams, if required, to help increase transparency of the models developed, aiding uptake and acceptance.
Simomics will offer free, non-commercial, licences (for the purposes of this project) for a number of their software tools, along with reasonable technical support, depending on the needs of awardees. Tools include: the ability to annotate of software models with assumptions and the rationale for design, the ability to trace the reliance of models on key evidence and drawing tools that allow the creation of arguments to support the design and rationale for models and results. These tools will be offered to successful applicants as an option to help build their models but are not required.
1. Ponzano S., et al. (2018). Promoting Safe Early Clinical Research of Novel Drug Candidates: A European Union Regulatory Perspective. Clinical Pharmacology Therapeutics 103: 564-566. https://doi.org/10.1002/cpt.899.
2. Monticello TM., et al. (2017) Current non-clinical testing paradigm enables safe entry to First-In-Human clinical trials: The IQ consortium non-clinical to clinical translational database. Toxicology and Applied Pharmacology. 1;334:100-109. doi: 10.1016/j.taap.2017.09.006
3. Sanz, F., et al. (2017). Legacy data sharing to improve drug safety assessment: the eTOX project. Nature Reviews Drug Discovery 16(12): 811-812. doi.org/10.1038/nrd.2017.177
4. Sewell F., et al. (2016). Opportunities to apply the 3Rs in Safety Assessment programs. ILAR journal 57: 234-245. doi.org/10.1093/ilar/ilw024
5. Prior H., et al. (2020). Opportunities for use of one species for longer-term toxicology testing during drug development: a cross-industry evaluation. Regulatory Toxicology and Pharmacology 113: e104624. doi: 10.1016/j.yrtph.2020.104624
7. Avila A.M., et al. 2020 An FDA/CDER perspective on non-clinical testing strategies: Classical toxicology approaches and new approach methodologies (NAMs). Regulatory Toxicology and Pharmacology. Volume 114, 104662. doi.org/10.1016/j.yrtph.2020.104662.
10. Luechtefeld T., et al. (2018). Machine Learning of Toxicological Big Data Enables Read-Across Structure Activity Relationships (RASAR) Outperforming Animal Test Reproducibility. Toxicological Sciences 165(1), 198-212. doi.org/10.1093/toxsci/kfy152
11. Wenzel J., Matter, H., and Schmidt, F. (2019). Predictive Multitask Deep Neural Network Models for ADME-Tox Properties: Learning from Large Data Sets. Journal of Chemical Information and Modeling, 59 (3), 1253-1268 doi: 10.1021/acs.jcim.8b00785
12. Sosnin S., et al. (2019). Comparative Study of Multitask Toxicity Modeling on a Broad Chemical Space. Journal of Chemical Information and Modeling 59 (3), 1062-1072 doi: 10.1021/acs.jcim.8b00685
13. Jain S., et al. (2021) Large-Scale Modeling of Multispecies Acute Toxicity End Points Using Consensus of Multitask Deep Learning Methods. Journal of Chemical Information and Modeling. 61(2):653-663. doi: 10.1021/acs.jcim.0c01164
14. Saili KS., et al. (2019) Systems Modeling of Developmental Vascular Toxicity. Current Opinion in Toxicology. 15(1):55-63. doi:10.1016/j.cotox.2019.04.004
15. Kleinstreuer N., et al. (2013) A Computational Model Predicting Disruption of Blood Vessel Development. PLOS Computational Biology 9(4): e1002996. doi.org/10.1371/journal.pcbi.1002996
16. Ferreira S., et al. (200) Quantitative Systems Toxicology Modeling To Address Key Safety Questions in Drug Development: A Focus of the TransQST Consortium. Chemical Research in Toxicology 33 (1), 7-9. doi: 10.1021/acs.chemrestox.9b00499
19. Steger-Hartmann, T. et al. (2017). "The IMI eTOX initiative Data mining, read-across and predictive models for target evaluation and early drug candidate assessment." Toxicology Letters 280: S19-S19. https://doi.org/10.1016/j.toxlet.2017.07.036
20. Clark M, Steger-Hartmann T. (2018). A big data approach to the concordance of the toxicity of pharmaceuticals in animals and humans. Regulatory Toxicology and Pharmacology 96:94-105. doi: 10.1016/j.yrtph.2018.04.018.
21. Pognan F., et al. (2021). The eTRANSAFE Project on Translational Safety Assessment through Integrative Knowledge Management: Achievements and Perspectives. Pharmaceuticals 14, 237: 1-18. doi.org/10.3390/ph14030237
22. Steger-Hartmann T., et al. (2020). Introducing the concept of virtual control groups into preclinical toxicology testing. ALTEX 37(3):343-349 doi/10.14573/altex.2001311
 In the UK, the number of animals used in scientific procedures is provided annually by the Home Office. Although the UK figures specify the number and species of animals used, their use for repeated dose toxicity covers a broader range of toxicity tests than described for this Challenge.
 Integrating bioinformatics and chemoinformatics approaches for the development of expert systems allowing the in silico prediction of toxicities.