Bailey_masterpaper.qxd

ATLA 41, 335–350, 2013 335
An Analysis of the Use of Dogs in Predicting HumanToxicology and Drug Safety Jarrod Bailey,1 Michelle Thew1 and Michael Balls2 1British Union for the Abolition of Vivisection (BUAV), London, UK; 2c/o Fund for the Replacement ofAnimals in Medical Experiments (FRAME), Nottingham, UK Summary — Dogs remain the main non-rodent species in preclinical drug development. Despite the cur-
rent dearth of new drug approvals and meagre pipelines, this continues, with little supportive evidence of
its value or necessity. To estimate the evidential weight provided by canine data to the probability that a
new drug may be toxic to humans, we have calculated Likelihood Ratios (LRs) for an extensive dataset of
2,366 drugs with both animal and human data, including tissue-level effects and Medical Dictionary for
Regulatory Activities (MedDRA) Level 1–4 biomedical observations. The resulting LRs show that the absence
of toxicity in dogs provides virtually no evidence that adverse drug reactions (ADRs) will also be absent in
humans. While the LRs suggest that the presence of toxic effects in dogs can provide considerable eviden-
tial weight for a risk of potential ADRs in humans, this is highly inconsistent, varying by over two orders of
magnitude for different classes of compounds and their effects. Our results therefore have important impli-
cations for the value of the dog in predicting human toxicity, and suggest that alternative methods are
urgently required.
Key words: canine, dog, drug development, preclinical testing, toxicology.
Address for correspondence: Jarrod Bailey, British Union for the Abolition of Vivisection (BUAV), 16a
Crane Grove, London N7 8NN, UK.
E-mail: [email protected]
specific disorder (e.g. HIV infection), the insightsthey provide depend critically on the question It is generally assumed that testing new pharma- being asked of the diagnostic test. However, they ceuticals on animals helps to ensure human safety are not appropriate for assessing the salient ques- and efficacy. Regulatory agencies worldwide require tion at issue with animal models, which is whether preclinical trials (e.g. 1, 2), which involve at least or not they contribute significant weight to the evi- two species — typically one rodent and one non- dence for or against the toxicity of a given com- rodent species — to determine toxicity and pharma- pound in humans. Overcoming this key problem — cokinetics. The expectation is that additional data almost entirely overlooked by previous authors — from the non-rodent will detect adverse effects not requires a precise specification of the various detected by rodent tests. Despite the current dearth terms used (see Methods). Briefly, the appropriate of new drug approvals and meagre pipelines (e.g. 3, metrics are Likelihood Ratios (LRs; 10): the 4), this practice continues, with little supportive evi- Positive Likelihood Ratio (PLR) and the inverse Negative Likelihood Ratio (iNLR). Therefore, there Dogs are used in significant numbers in science is clearly a need for the kind of statistically-appro- ― approximately 90,000 are used per annum priate critical analysis that we provide here. The across the EU and the USA, according to the latest dataset we have used is unique, in that it is large available figures (6–8). About 80% of this use is as and allows the conditional probabilities required the non-rodent species in the evaluation of phar- for the LRs (PLR/iNLR) to be calculated.
maceutical safety and efficacy (6). However, onlylimited evaluations of the reliability of the caninemodel for this purpose have been conducted, chiefly due to the difficulty of accessing relevantdata, most of which are unpublished and propri- Animal models are widely used to assess the risk etary to pharmaceutical companies. Those evalua- that a given compound will prove toxic in humans.
tions that have been conducted have usually As with any diagnostic test, their reliability can employed ‘concordance’ metrics (e.g. 9), which var- only be assessed by performing tests in which the ious authors have interpreted as the true positive same compound is given to both animals and rate (‘sensitivity’) or the Positive Predictive Value humans, and the presence or absence of toxicity (PPV). While these metrics are appropriate for recorded. This leads to a 2 × 2 matrix of results, as assessing the reliability of a diagnostic test for a Compound toxic
Compound not
in humans
toxic in humans
Compound toxic in animal model
Compound not toxic in animal model
The basis of this matrix is that the human data makes PPVs dependent on the prevalence of toxic- are correct, and the dog data are true/false, if they ity in compounds, and thus an inappropriate meas- do/do not match them. The various cells in this ure of the reliability of the test with any specific matrix allow a variety of diagnostic metrics to be deduced, of which the most familiar and widely Thus, any appropriate metric of the evidential used are the true positive rate for the test (or ‘sen- value of animal models requires knowledge of both sitivity’ = a/[a + c]), and the true negative rate (or the sensitivity and the specificity of the model.
‘specificity’ = d/[d + b]). In previous research into This, in turn, implies that the appropriate metrics the reliability of animal models as predictors of for the evidential weight provided by an animal toxicity in humans, some authors (e.g. 9) have model are LRs (e.g. 13). In general, these are ratios focused on the sensitivity, expressed as the ‘true of functions of the sensitivity and specificity, which positive concordance rate’, or the so-called Positive can be extracted from the 2 × 2 matrix given above.
Predictive Value (PPV), given by a/(a+b), which In the specific case of animal models in general, reflects the probability that human toxicity was two LRs are relevant. The first is the so-called correctly identified by the animal model, given that toxicity was observed in the animal model(e.g. 12). However, neither of these metrics is suit- able for the role of assessing the evidential weight provided by any toxicity test. In the case of animalmodels, the sensitivity addresses only the ability of This LR captures the ability of an animal model to such models to detect toxicity that will subse- add evidential weight to the belief that a specific quently manifest itself in humans. This is a neces- compound is toxic. Any animal model that gives a sary, but not sufficient, measure of evidential PLR that is statistically significantly higher than weight. Suppose, for example, that the animal 1.0, can be regarded as contributing evidential model always indicates toxicity found in humans; weight to the probability that the compound under it would then have a sensitivity of 100%.
However, if, in addition, the model always indi- The other relevant LR is the so-called iNLR, cates toxicity, even in humans, its evidential value is no better than simply dismissing everycompound as toxic from the outset. Thus, a useful toxicity test must also be able to give insight into when toxicity seen in the animal model is notobserved in humans, which requires knowledge of This LR captures the ability of an animal model to add evidential weight to the belief that a specific There is, of course, an obvious reason for the compound is not toxic: any animal model that gives focus on sensitivity in animal model evaluation: if an iNLR that is statistically significantly higher a compound is found to be positive in an animal than 1.0, can be regarded as contributing eviden- model, it is unlikely to go into human evaluation.
tial weight to the probability that the compound Nevertheless, the fact remains that sensitivity alone cannot be an adequate guide to the value of It is worth noting at this point that the above definitions imply that a good animal model for The case of the PPV is more subtle. This metric detecting human toxicity is not necessarily also is a measure of the probability that human toxicity good for detecting an absence of toxicity. That is, a will be correctly identified, given that the animal high PLR does not guarantee a high iNLR; this will model detected toxicity. As such, PPVs are condi- tional probabilities, the condition being the pre- The above definitions also underscore the need existence of a positive animal test result. This for data on the human toxicity of compounds that The use of dogs in predicting human toxicology 337 fail initial animal tests. Again, a key feature of the mitigated by limiting the dataset to compounds current study is that this issue has been overcome reported in the FAERS database. Therefore, all via data mining methods. Data were obtained from the compounds are certain to have proceeded to a leading pharmaceutical safety consultancy, market, and animal preclinical data are available Instem Scientific Limited (Harston, Cambridge, for these compounds. Specific details of how the UK; http://www.instem-lss.com ‘Safety Intell FPs that were identified arose were not sought, igence Programme’), with funding provided by because they were not pertinent to this analysis, FRAME. All the information stemmed from pub- and this was not feasible, given the nature of the licly accessible sources, including: PubMed (http:// dataset. It must be assumed that the dog data www.ncbi.nlm.nih.gov/pubmed), the FDA Adverse were correlated with the human data retrospec- tively, and/or the human data arose from post- (http://www.drugbank.ca), and the National Tox marketing studies, and/or clinical trials were icology Program (http://ntp.niehs.nih.gov). Data applied for and approved, since the adverse were available for more than 2,300 drug com- effect(s) in dogs were minor and/or mitigated by pounds in humans and preclinical species. Inference of the good quality of the data used in this evaluation is outlined in the Discussion.
Compounds were selected that feature in the FAERS, FDA New Drug Applications (FDA NDAs)and DrugBank. Thus, the drugs selected for this The inappropriate nature of PPVs is demonstrated analysis are in clinical use, and have undergone in Figure 2, which shows a scatter plot of ‘ranked’ preclinical testing: human and animal data are PPVs against equivalent ranked PLRs. Each PPV therefore available for them. A non-redundant list and PLR was ranked according to its value for each of parent moieties was created, for example, by of the 436 classifications of effects, and these ranks normalising therapeutic products to their generic were plotted against each other. The disparity is names (e.g. Lipitor to Atorvastatin). This yielded evidenced by the scatter of points, few of which lie close to the y = x line that shows an ideal correla- A signature of the effects of each compound was tion. The misclassifications and misplaced ass created, focusing on tissue-level effects (e.g. brady- ump tions of the accuracy of canine data for the cardia and arrhythmic disorder would both be con- prediction of human adverse drug reactions sidered to be effects on heart tissues), as well as the (ADRs) are clear. For example, MedDRA ‘Level 4, individual observations, which were mapped to Vascular Disorder’ was ranked 20/436 with regard their MedDRA (Medical Dictionary for Regulatory to the most favourable classifications for human Activities; http://www.meddramsso.com) counter- predictivity based on PPV, but its cognate PLR parts. MedDRA observations are classified into four ranked 404/436 — one of the least predictive.
levels, Level 1 being the most specific and Level 4 Conversely, MedDRA classification ‘Level 2, providing a more generic ‘System Organ Class’.
Ventricular Conduction’ ranked 30/436 by PLR, These classifications help to eliminate false posi- tives that may arise from species-specific observa- Dog PLRs were generally high (median ~28), tions, and help the identification of concordant implying that compounds that are toxic in dogs are observations that might otherwise have been likely also to be toxic in humans. However, because missed, by their ‘rolling up’ into more-generic terms.
the PLRs vary considerably (range 4.7–548.7), LRs were derived for broad tissue-level effects (n with no obvious pattern regarding the form of tox- = 52), and more-specific biomedical observations icity, the reliability of this aspect of canine models (BMOs; n = 384), mapped to MedDRA classifica- cannot be generalised or regarded with confidence.
tions (Levels 1 [most specific] to Level 4 [more In contrast, the calculated inverse negative LRs generic ‘organ class’]). Fourteen BMO classifica- (iNLRs) are substantially more consistent, but tions not involving dogs were eliminated from the their median value of 1.11 (range 1.01–1.92) sup- study. A total of 3,275 comparisons were made ports the view that dogs provide essentially no evi- between the human and the dog, for 2,366 com- dential weight to this aspect of toxicity testing.
pounds, involving 436 (52 + 384) classifications of Specifically, the fact that a compound shows no effects. The Instem Scientific data on which our toxic effects in dogs provides essentially no insight analysis was based are shown in the Appendix, into whether the compound will also show no toxic and the full set of data, including 95% Confidence Intervals, are available on the FRAME website This lack of evidential weight has important implications for the role of dogs in toxicity testing, With regard to potential bias: FNs are more especially for the pharmaceutical industry. The common than FPs, since there is a bias resulting critical observation for deciding whether a candi- from a ‘precautionary principle’ not to progress date drug can proceed to testing in humans is the positives to human administration. This has been absence of toxicity in tests on animals. However, Figure 2: Scatter plot illustrating the lack of correlation of PPVs and PLRs of biomedical observations (BMOs) and tissue effects in humans and dogs PPVs and PLRs for all 436 results were ordered according to their value, with the highest ranking first and thelowest last. For each BMO and tissue effect, the corresponding PPV and PLR rank were plotted against each other.
If a perfect correlation exists, all points should lie on the line, where, for example, the 10th, 50th, and 100th highestPPV value would also be the 10th, 50th, and 100th highest PLR values. However, the significant scatter of the datapoints demonstrates that little correlation exists between PPV and PLR. For example: the 20th highest PPV ranksonly 404/436 for PLR, whereas the 30th highest PLR ranks only 406/436 for PPV. our findings show that the predictive value of the the dog. In 2012, a study that expressly set out to animal test in this regard is barely greater than minimise bias, showed that 63% of serious ADRs that that would be obtained by chance (see below).
had no counterparts in animals, and less than 20%of serious ADRs had a true positive corollary inanimal studies (15). Other similar examples exist for testing generally (e.g. 16–18) and more-specifi-cally, for example, in teratology (e.g. 19, 20) and The analysis presented here is urgently required, drug-induced liver injury (e.g. 5, 21). One notable to support informed debate about the worth of ani- study claimed a good concordance for dog and mal models in preclinical testing. It is acknowl- human toxicology (10), though neither the predic- edged among some stakeholders (if not universally tive nature of the animal data for humans, nor the among all stakeholders) that assessment of the sci- evidential weight provided by those data, were entific value of animal data in drug development is necessary, has been scarce, and has been thwarted We have, for the first time, addressed the salient for decades by the unavailability of relevant data question of contribution of evidential weight for or for analysis (e.g. 14). Nevertheless, primarily due against the toxicity of a given compound in humans to concerns over privacy and commercial interests, by data from dog tests, by using the appropriate data sharing and making data available continue metrics of LRs. Furthermore, we have applied the to be resisted, in spite of assurances to the contrary apposite LRs to a dataset of unprecedented scale, to critically question the value of the use of the dog Those few analyses that have been done, tend to as a preclinical species in the testing of new reflect unfavourably on animal models, including The use of dogs in predicting human toxicology 339 Substantiation of data quality is evidenced by: Our findings have practical implications for the the methods used to source the data and the use of animal models for toxicity testing, especially assured quality of the databases supplying them in the pharmaceutical industry. Reliance on flawed (listed above); the ways in which the data had been models of toxicity testing leads to two types of fail- used recently as a basis for scientific publications ure. If the models have poor PLRs, then there is a and presentations (e.g. 23–26); and the interna- risk that many potentially useful compounds will tional corporate and academic clients that have be wrongly discarded, because of ‘false positives’ used the consultancy and its data (e.g.
produced by the toxicity model. On the other hand, AstraZeneca; see 23–26). In addition, the impact of if the models have poor iNLRs, then many toxic ‘missing data’ (i.e. unpublished data held by phar- compounds will wrongly find their way into human maceutical companies) was mitigated by strictly tests, and will fail in clinical trials. The relatively limiting the dataset to drugs “with the greatest high PLRs found in this study show that animal chance of having been evaluated in all the species models may not be leading to the loss of many included in the study” (here, dogs and humans). In potentially valuable candidate drugs through false other words, “…lack of evidence for an association positives. However, our results do imply that many between a compound and a specific BMO demon- toxic drugs are not being detected by animal mod- strates a real absence of effect, and is not due to els, leading to the risk of unnecessary harm to missing data” (Instem Scientific Ltd. Analysis In this regard, our findings are entirely consis- Naturally, there must be caveats. Our analysis tent with the acknowledged failure of animal mod- was limited to data that are published and publicly els in general to provide guidance on likely toxicity available. It is widely acknowledged that many ahead of the entry of compounds into human trials.
animal experimental results/preclinical data Drug attrition has increased significantly over the remain unpublished and/or proprietary, for a vari- past two decades (e.g. 3, 4, 37–42): 92–94% of all ety of reasons (e.g. 15, 27–30). Such publication drugs that pass preclinical tests fail in clinical tri- bias is a major problem (e.g. 31–34), and, com- als, mostly due to unforeseen toxicities (43–45), pounded by other factors such as size and quality and half of those that succeed may be subsequently of the animal studies, variability in the require- withdrawn or re-labelled due to ADRs not detected ments for reporting animal studies, ‘optimism in animal tests (46). ADRs are a major cause of bias’, and lack of randomisation and blinding (28, premature death in developed countries (47). A 35), it means that gauging the true contribution of major contributing factor is the inadequacy of pre- animal data to human toxicology is impossible — clinical animal tests: one recent study showed that at least for third parties without access to phar- 63% of ADRs had no counterpart in animals, and maceutical company files. All datasets are imper- less than 20% had a positive corollary in animal fect to varying degrees. However, it is only possible to use data which are available, and to ensure that, With specific regard to the dog, the most exten- as far as feasible, those data are of good quality sive study prior to the report we present here, con- and as free from biases as possible, and that their cluded that 92% of dog toxicity studies did not analysis and derived conclusions are as objective provide relevant information in addition to that provided by the rat, and that the other 8% did not It must be made abundantly clear that we, the result in the immediate withdrawal of drugs from authors of this report, did not make decisions development, indicating that dog studies are not regarding the toxicity/non-toxicity of drugs, or decide required for the prediction of safe doses for upon or apply any criteria to such decisions. The humans (17). There is a scientific basis for this: mining of the data, and the decisions on toxicity of among several notable species differences which the drugs, were independent of the authors of this confound the extrapolation of data from dogs to paper, and were made by one or both of the authors humans, significant differences between humans of the drug/toxicity papers and/or database submis- and dogs in their cytochrome P450 enzymes sions used, and the data-mining consultancy/cura- (CYPs) — the major enzymes involved in drug tors of the Safety Intelligence Pro gramme, Instem metabolism — have been acknowledged for some Scientific Limited. Therefore, if any pharmaceutical time, compelling the conclusion that, “…it is read- industry stakeholders have issues or concerns with ily seen that the dog is frequently not a good meta- our conclusions, we would encourage them to con- bolic model for man and is poorly comparable to duct further analyses by using their own proprietary the rat and mouse” (for references, see 46). The data, and/or to facilitate such investigations by mak- lack of knowledge of canine CYPs has been high- ing available anonymised data, in accordance with lighted, which is surprising, considering the extent the promotion of transparency encouraged by EU of the use of dogs in preclinical testing. This prob- Directive 2010/63/EU (36), as well as to engage fully lem is likely to be amplified by intra-species differ- in constructive discussion and debate with us and ences, as well as by inter-species differences (49).
our colleagues in animal protection organisations.
It may therefore be argued that, if many differ- ences exist between different breeds or strains of median iNLR figure found by our study, if the com- the same species, then extrapolating pharmaco pound shows no sign of toxicity in the dog, the kinetic data from that highly variable species to probability that the compound will also show no humans must not only be difficult, but must also toxic effects in humans will have been increased by the animal testing from 70% to 72%. The testingthus contributes essentially no additional confi-dence in the outcome, but at considerable extra cost, both in monetary terms and in terms of ani-mal welfare. This also has obvious practical rele- This analysis of the most comprehensive quantita- vance to the issue of high attrition rates in clinical tive database of publicly-available animal toxicity studies yet compiled, suggests that dogs are highly It is argued that a comprehensive suite of more inconsistent predictors of toxic responses in humans, reliable alternative methods is now available (14, and that the predictions they can provide are little 51, 52). Combined with considerable public con- better than those that could be obtained by chance ― cern over the use of dogs in science (53), the high or tossing a coin ― when considering whether or not ethical costs of doing so, given the sensitive nature a compound should proceed to testing in humans. In of dogs (e.g. 15, 54), and the expressed desire for other words: “…for any putative source of evidential the use of dogs as a second species in drug testing weight to be deemed useful, its specificity and sensi- to have a scientific, rather than a habitual, basis tivity must be such that LR+ [PLR] >1. Tossing a (14), we conclude that the preclinical testing of coin contributes no evidential weight to a given pharmaceuticals in dogs cannot currently be justi- hypothesis, as the sensitivity and specificity are the same ― 50% ― and thus the LR+ [PLR] is equal to 1”(22).
Dog PLRs were generally high, showing that a drug which is toxic in the dog is likely to be toxic inhumans. However, they were extremely variable The authors are grateful to the British Union for the and with no obvious pattern, suggesting this aspect Abolition of Vivisection (BUAV), the Fund for the of dog tests cannot be considered particularly reli- Replacement of Animals in Medical Experi able or helpful. Further, though not within the scope (FRAME), and The Kennel Club (via FRAME), for of this analysis, it is of great interest whether the funding. They thank Robert Matthews for advice on dog revealed any significant toxicities, that were also inferential issues, Bob Coleman for his help and present in humans, that other species such as the encouragement during the inception of this rat did not. In other words, did the dog ‘catch’ any undertaking, and Instem Scientific (previously true human toxicities not caught by the rat? It has BioWisdom; Harston, Cambridge, UK) for scientific been previously argued that such toxicities are rela- consultancy and for data analysis on integrated data tively low in number (e.g. the development of just relating to adverse events in model animal species.
11% of new compounds was terminated due to The research described in this article is based on the effects uniquely seen in dogs, though the human sig- analysis and conclusions of the authors: it has not nificance of these could not be determined), which been subjected to each agency’s peer review and pol- would further diminish any value the canine model icy review; therefore, it does not necessarily reflect the views of the organisations, and no official More importantly, while iNLRs were much more consistent, they revealed that dogs provide essen-tially no evidential weight to this aspect of toxicitytesting. Specifically, if a compound shows no toxic Received 23.05.13; received in final form 11.09.13;accepted for publication 19.09.13. effects in dogs, this provides essentially no insightinto whether the compound will also show no toxiceffects in humans. This is crucial: the critical observation for deciding whether a candidate drugcan proceed to testing in humans is the absence of Anon. (2004). Directive 2004/27/EEC of the toxicity in tests on animals, and our findings show European Parliament and the Council of 31 March that the predictive value of the dog test in this 2004, amending Directive 2001/83/EC on the regard is barely greater than by chance. Community code relating to medicinal products for A quantitative example illustrates this. Suppose human use. Official Journal of the European Union researchers wish to investigate a candidate com- L136, 30.04.2004, 34–57.
Anon. (2010). Federal Food, Drug and Cosmetics pound belonging to a family which prior experience Act. Silver Spring, MD, USA: US Food and Drug indicates has a 70% probability of freedom from Administration. Available at: http://www.fda.gov/ ADRs in humans. Before conducting tests in RegulatoryInformation/Legislation/FederalFood humans, the drug is tested in dogs. By using the The use of dogs in predicting human toxicology 341 of Animals in Medical Experiments (FRAME).
Duyk, G. (2003). Attrition and translation. Science, 18. Litchfield, J.T.J. (1962). Symposium on clinical New York 302, 603–605.
drug evaluation and human pharmacology. XVI.
Kola, I. & Landis, J. (2004). Can the pharmaceuti- Evaluation of the safety of new drugs by means of cal industry reduce attrition rates? Nature Reviews tests in animals. Clinical Pharmacology & Thera - Drug Discovery 3, 711–715.
peutics 3, 665–672.
Aithal, G.P. (2010). Mind the gap. ATLA 38, Suppl.
19. Bailey, J. (2008). Developmental toxicity testing: Protecting future generations? ATLA 36, 718–721.
UK Home Office (2013). Statistics of Scientific Pro - 20. Schardein, J. (2000). Chemically Induced Birth cedures on Living Animals — Great Britain 2012. Defects, 3rd edn, 1019pp. Boca Raton, FL, USA: HC 549, 60pp. London, UK: The Stationery Office.
USDA (2011). Annual Report. Animal Usage by 21. Spanhaak, S., Cook, D., Barnes, J. & Reynolds, J.
Fiscal Year — Fiscal Year 2010, 2pp. Riverdale, (2008). Species Concordance for Liver Injury, 6pp.
MD, USA: United States Department of Agriculture Cambridge, UK: Biowisdom Ltd. Available at: (USDA), Animal and Plant Health Inspection http://www.biowisdom.com/files/SIP_Board_Species Service (APHIS). Available at: http://www. aphis.
_Concordance.pdf (Accessed 10.10.13).
usda.gov/animal_welfare/efoia/downloads/2010_ 22. Matthews, R.A. (2008). Medical progress depends Animals_Used_In_Research.pdf (Accessed 05.09.
on animal models — doesn’t it? Journal of the Royal Society of Medicine 101, 95–98.
Anon. (2010). Sixth Report on the Statistics on the 23. Barnes, J.C., Matis, S., Kenna, G., Swinton, J., Number of Animals Used for Experimental and Bradley, P.M., Day, N.C., Reed, J.Z., Reynolds, J. & Other Scientific Purposes in the Member States of the Cook, D. (2008). The Safety Intelligence Program: European Union. SEC(2010) 1107, 14pp. Brussels, An Intelligence Network for Drug-induced Liver Belgium: European Commission. Available at: Injury, 2pp. Cambridge, UK: Biowisdom Ltd.
http://eurlex.europa.eu/LexUriServ/LexUriServ.do? Available at: http://bioblog.instem.com/downloads/ uri=COM:2010:0511:REV1:EN:PDF (Accessed 10.
Carboxylic_acids_A4.pdf (Accessed 10.10.13).
24. Sidaway, J.R.M., Roberts, S., Huby, R., Nicholson, Olson, H., Betton, G., Robinson, D., Thomas, K., A., Pemberton, J., South, M., Noeske, T., Engkvist, Monro, A., Kolaja, G., Lilly, P., Sanders, J., Sipes, O., Bradley, P. & Reed, J. (2012). Drug Toxicities G., Bracken, W., Dorato, M., Van Deun, K., Smith, Associated With Pharmacological Activity: Using P., Berger, B. & Heller, A. (2000). Concordance of Harmonised Data to Make the ‘Known’ Visible, 2pp.
the toxicity of pharmaceuticals in humans and in Macclesfield, UK: Safety Assessment, AstraZeneca.
animals. Regulatory Toxicology & Pharmacology Available at: http://bioblog.instem.com/wp-content/ 32, 56–67.
uploads/downloads/2012/03/SOT2012_make-the- 10. Altman, D.G. & Bland, J.M. (1994). Diagnostic tests known-visible_poster.pdf (Accessed 10.10.13).
2: Predictive values. British Medical Journal 309,
25. Fourches, D., Barnes, J.C., Day, N.C., Bradley, P., Reed, J.Z. & Tropsha, A. (2010). Cheminformatics 11. Anon. (2012). Likelihood Ratios. Oxford, UK: analysis of assertions mined from literature that Centre for Evidence Based Medicine (CEBM).
describe drug-induced liver injury in different Available at: http://www.cebm.net/index.aspx?o= species. Chemical Research in Toxicology 23,
12. Greek, R. & Menache, A. (2013). Systematic reviews 26. Greco, I., Day, N., Riddoch-Contreras, J., Reed, J., of animal models: Methodology versus epistemol- Soininen, H., Kloszewska, I., Tsolaki, M., Vellas, B., ogy. International Journal of Medical Sciences 10,
Spenger, C., Mecocci, P., Wahlund, L.O., Simmons, A., Barnes, J. & Lovestone, S. (2012). Alzheimer’s 13. Grimes, D.A. & Schulz, K.F. (2005). Refining clini- disease biomarker discovery using in silico litera- cal diagnosis with likelihood ratios. Lancet 365,
ture mining and clinical validation. Journal of Translational Medicine 10, 217.
14. Hasiwa, N., Bailey, J., Clausing, P., Daneshian, M., 27. Wandall, B., Hansson, S.O. & Ruden, C. (2007).
Eileraas, M., Farkas, S., Gyertyan, I., Hubrecht, R., Bias in toxicology. Archives of Toxicology 81,
Kobel, W., Krummenacher, G., Leist, M., Lohi, H., Miklosi, A., Ohl, F., Olejniczak, K., Schmitt, G., 28. Hackam, D.G. (2007). Translating animal research Sinnett-Smith, P., Smith, D., Wagner, K., Yager, into clinical benefit. British Medical Journal 334,
J.D., Zurlo, J. & Hartung, T. (2011). Critical evalu- ation of the use of dogs in biomedical research and 29. ter Riet, G., Korevaar, D.A., Leenaars, M., Sterk, testing in Europe. ALTEX 28, 326–340.
P.J., Van Noorden, C.J., Bouter, L.M., Lutter, R., 15. van Meer, P.J., Kooijman, M., Gispen-de Wied, Elferink, R.P. & Hooft, L. (2012). Publication bias in C.C., Moors, E.H. & Schellekens, H. (2012). The laboratory animal research: A survey on magni- ability of animal studies to detect serious post mar- tude, drivers, consequences and potential solutions.
keting adverse events is limited. Regulatory Tox PLoS One 7, e43404.
icology & Pharmacology 64, 345–349.
30. Briel, M., Muller, K.F., Meerpohl, J.J., von Elm, E., 16. Igarashi, T., Nakane, S. & Kitagawa, T. (1995).
Lang, B., Motschall, E., Gloy, V., Lamontagne, F., Predictability of clinical adverse reactions of drugs Schwarzer, G. & Bassler, D. (2013). Publication bias by general pharmacology studies. Journal of Tox - in animal research: A systematic review protocol.
icological Sciences 20, 77–92.
Systematic Reviews 2, 23.
17. Broadhead, C.L., Jennings, M. & Combes, R. (1999).
31. van der Worp, H.B., Howells, D.W., Sena, E.S., A Critical Evaluation of the Use of Dogs in the Porritt, M.J., Rewell, S., O’Collins, V. & Macleod, Regulatory Toxicity Testing of Pharmaceuticals, M.R. (2010). Can animal models of disease reliably 106pp. Nottingham, UK: Fund for the Replacement inform human studies? PLoS Medicine 7, e1000245.
32. Sena, E.S., van der Worp, H.B., Bath, P.M., 23003/title/More-compounds-failing-Phase-I/ Howells, D.W. & Macleod, M.R. (2010). Publication bias in reports of animal stroke studies leads to 44. Okie, S. (2006). Access before approval — a right to major overstatement of efficacy. PLoS Biology 8,
take experimental drugs? New England Journal of Medicine 355, 437–440.
33. Perel, P., Roberts, I., Sena, E., Wheble, P., Briscoe, 45. Aurup, P. (2012). Er Danmark et Attraktivt Land C., Sandercock, P., Macleod, M., Mignini, L.E., for Klinisk Forskning? (Is Denmark an Attractive Jayaram, P. & Khan, K.S. (2007). Comparison of Country for Clinical research?), 23pp. Ballerup, treatment effects between animal experiments and Denmark: MSD Laboratories. Available at: http:// clinical trials: Systematic review. British Medical di.dk/SiteCollectionDocuments/Opinion/Sundhed/ Journal 334, 197.
Høring/Præsentation%20-%20Peter%20Aurup, 34. Schott, G., Pachl, H., Limbach, U., Gundert-Remy, U., Ludwig, W.D. & Lieb, K. (2010). The financing 46. Anon. (1990). FDA Drug Review: Post Approval Risks of drug trials by pharmaceutical companies and its 1976–1985. GAO/PEMD-90-15, 132pp. Washington, consequences. Part 1: A qualitative, systematic DC, USA: US General Accounting Office. Available review of the literature on possible influences on at: http://161.203.16.4/d24t8/141456.pdf (Accessed the findings, protocols, and quality of drug trials.
Deutsches Arzteblatt International 107, 279–285.
47. Lazarou, J., Pomeranz, B.H. & Corey, P.N. (1998).
35. Kilkenny, C., Parsons, N., Kadyszewski, E., Festing, Incidence of adverse drug reactions in hospitalized M.F., Cuthill, I.C., Fry, D., Hutton, J. & Altman, D.G.
patients: A meta-analysis of prospective studies.
(2009). Survey of the quality of experimental des - Journal of the American Medical Association 279,
ign, statistical analysis and reporting of research using animals. PLoS One 4, e7824.
48. Gad, S.C. (2006). Animal Models in Toxicology, 36. Anon. (2010). Directive 2010/63/EU of the Euro 952pp. Boca Raton, FL, USA: CRC Press.
pean Parliament and of the Council of 22 September 49. Martinez, M.N., Antonovic, L., Court, M., Dacasto, 2010 on the protection of animals used for scientific M., Fink-Gremmels, J., Kukanich, B., Locuson, C., purposes. Official Journal of the European Union Mealey, K., Myers, M.J. & Trepanier, L. (2013).
L276, 20.10.2010, 33–79.
Challenges in exploring the cytochrome P450 sys- 37. US FDA (2004). Innovation or Stagnation: Challenge tem as a source of variation in canine drug phar- and Opportunity on the Critical Path to New Med- macokinetics. Drug Metabolism Reviews 45, 218–
ical Products, 12pp. Silver Spring, MD, USA: US Department of Health and Human Services, Food 50. Broadhead, C.L., Betton, G., Combes, R., Damment, and Drug Administration. Available at: http://www.
S., Everett, D., Garner, C., Godsafe, Z., Healing, G., fda.gov/ScienceResearch/Special Topics/CriticalPathInitiative/CriticalPathOpportunitiesReports/ucm077 Heywood, R., Jennings, M., Lumley, C., Oliver, G., Smith, D., Straughan, D., Topham, J., Wallis, R., 38. Issa, A.M., Phillips, K.A., Van Bebber, S., Nida Wilson, S. & Buckley, P. (2000). Prospects for reduc- marthy, H.G., Lasser, K.E., Haas, J.S., Alldredge, ing and refining the use of dogs in the regulatory B.K., Wachter, R.M. & Bates, D.W. (2007). Drug toxicity testing of pharmaceuticals. Human & withdrawals in the United States: A systematic Experimental Toxicology 19, 440–447.
review of the evidence and analysis of trends. Curr - 51. Spielmann, H., Kral, V., Schäfer-Korting, M., Seidle, ent Drug Safety 2, 177–185.
T., McIvor, E., Rowan, A. & Schoeters, G. (2011).
39. Bennani, Y.L. (2011). Drug discovery in the next The AXLR8 Consortium. Alternative Testing Strat - decade: Innovation needed ASAP. Drug Discovery egies, Progress Report 2011, 364pp. Berlin, Germany: Today 16, 779–792.
Institute of Pharmacy, Free University of Berlin.
40. Eichler, H.G., Aronsson, B., Abadie, E. & Salmon - Available at: http://scrtox.eu/~scrtox/images/pdf/stories/ son, T. (2010). New drug approval success rate in Europe in 2009. Nature Reviews Drug Discovery 9,
52. Anon. (2007). Toxicity Testing for the 21st Century: A Vision and a Strategy, 216pp. Washington, DC, 41. Hughes, B. (2008). 2007 FDA drug approvals: A year of flux. Nature Reviews Drug Discovery 7, 107–109.
53. Anon. (2009). Public Opinion. London, UK: Euro - 42. Hartung, T. (2009). Toxicology for the twenty-first pean Coalition to End Animal Experiments.
century. Nature, London 460, 208–212.
Available at: http://www.eceae.org/en/what-we-do/ 43. Harding, A. (2004). More compounds failing phase campaigns/12-million-reasons/public-opinion (Acc I. FDA chief warns that high drug attrition rate is pushing up the cost of drug development. The 54. Hare, B., Brown, M., Williamson, C. & Tomasello, Scientist, 6 August 2004. Available at: http:// M. (2002). The domestication of social cognition in www.the-scientist.com/?articles.view/articleNo/ dogs. Science, New York 298, 1634–1636.
The use of dogs in predicting human toxicology 343 Table A1: Raw data from Instem Scientific’s ‘Safety Intelligence Programme’, showing the number of drugs associated with ADRs in humans and dogs Parameters
Number of drugs
Human/Dog
Human Neither Dog Human
Adverse effect: tissue-level or BMO (MedDRA Level 1–4)
Level 2 — glomerulonephritis and nephrotic syndrome Level 3 — glaucoma and ocular hypertension Level 2 — glaucomas (excluding congenital) Level 3 — renal and urinary tract neoplasms (malignant and Level 2 — urinary tract neoplasms (unspecified malignancy not Level 2 — hepatic peroxisome proliferation Level 4 — blood and lymphatic system disorders Level 3 — hepatic and hepatobiliary disorders Level 3 — renal disorders (excluding nephropathies) Level 4 — skin and subcutaneous tissue disease All entries are numbered for identification only (column 1). The second column (parameters) indicates the specific biomedical
observation (BMO) in question (e.g. ‘bradycardia’ or ‘arrhythmic disorder’), or tissue-level effects (e.g. ‘heart’, which would encom-
pass these two BMOs). The BMOs were mapped to their MedDRA (Medical Dictionary for Regulatory Activities) counterpart, which
are classified into four levels, level 1 being the most specific and level 4 providing a more generic ‘System Organ Class’. The number
of drugs for which ADRs were observed in each species is shown in columns 3–8. Human/Dog represents drugs for which an ADR
was reported in both humans and dogs: these are True Positives (TPs), and correspond to cell ‘a’ in the 2 × 2 matrix (see Methods,
Figure 1). Dog represents drugs for which an ADR was reported in dogs, but not in humans: these are False Positives (FPs), and
correspond to cell ‘b’ in the 2 × 2 matrix. Human represents drugs for which an ADR was reported in humans, but not in dogs:
these are False Negatives (FNs), and correspond to cell ‘c’ in the 2 × 2 matrix. Neither represents drugs for which an absence of
ADRs was evident in both humans and dogs: these are True Negatives (TNs), and correspond to cell ‘d’ in the 2 × 2 matrix. Notably,
lack of an association between a compound and a specific BMO was assumed (by the data provider) to demonstrate a real absence of
effect, and not be due to missing data. To minimise the impact of missing data, the group of compounds in the dataset were chosen
with the greatest chance of having been evaluated in all the species included in the study (see Methods). The total number of drugs
exhibiting ADRs in each species, regardless of the presence or absence of ADRs in the other species, is given in the final two
columns: Dog = a + b (TP + FP); Human = a + c (TP + FN).
Parameters
Number of drugs
Human/Dog
Human Neither Dog Human
Adverse effect: tissue-level or BMO (MedDRA Level 1–4)
Level 4 — injury, poisoning and procedural complications Level 3 — cardiac and vascular investigations (excluding Level 2 — hepatic microsomal lipid peroxidation Level 4 — respiratory, thoracic and mediastinal disorders Level 2 — non-site specific vascular disorders nec Level 2 — nephropathies and tubular disorders nec Level 3 — chemical injury and poisoning Level 2 — electrocardiogram observation Level 2 — rate and rhythm disorders nec Level 3 — epidermal and dermal conditions Level 3 — decreased and non-specific blood pressure disorders Level 3 — central nervous system vascular disorders Level 2 — dermal and epidermal conditions nec Level 3 — respiratory and mediastinal neoplasms (malignant Level 2 — lower respiratory tract neoplasms Level 3 — respiratory tract neoplastic disorder Level 2 — hepatocellular damage and hepatitis nec Level 2 — heart rate and pulse investigations Level 3 — arteriosclerosis, stenosis, vascular insufficiency and Level 2 — vascular hypotensive disorders Level 2 — cerebrovascular and spinal necrosis and vascular The use of dogs in predicting human toxicology 345 Parameters
Number of drugs
Human/Dog
Human Neither Dog Human
Adverse effect: tissue-level or BMO (MedDRA Level 1–4)
Level 3 — cardiac disorder signs and symptoms Level 2 — hepatic failure and associated disorders Level 3 — bronchial disorders (excluding neoplasms) Level 2 — central nervous system vascular disorders Level 3 — haemolyses and related conditions Level 3 — haematology investigations (including blood groups) Level 2 — encephalopathies (toxic and metabolic) Level 2 — respiratory tract and pleural neoplasms (malignancy Level 2 — diabetic complications (renal) Level 2 — hepatic enzymes and function abnormalities Level 3 — cardiac physiological observation Level 4 — general disorders and administration site conditions Level 2 — ventricular arrhythmias and cardiac arrest Level 2 — peripheral vasoconstriction, necrosis and vascular Level 2 — ischaemic coronary artery disease Level 3 — hepatic physiological phenomenon Level 2 — atrial natriuretic factor secretion Level 3 — renal and urinary tract disorders (congenital) Level 2 — non-site specific gastrointestinal haemorrhages Level 2 — hepatic and hepatobiliary disorders Parameters
Number of drugs
Human/Dog
Human Neither Dog Human
Adverse effect: tissue-level or BMO (MedDRA Level 1–4)
Level 2 — renal and urinary tract injuries nec Level 2 — renal structural abnormalities and trauma Level 2 — coronary necrosis and vascular insufficiency Level 4 — congenital, familial and genetic disorders Level 2 — central nervous system haemorrhages and Level 3 — lower respiratory tract disorders Level 2 — non-site specific embolism and thrombosis Level 2 — non-site specific injuries nec Level 2 — cardiac conduction abnormality Level 4 — metabolism and nutrition disorders Level 3 — cardiac and vascular disorders congenital Level 2 — conditions associated with abnormal gas exchange Level 2 — vascular smooth muscle cell proliferation The use of dogs in predicting human toxicology 347 Parameters
Number of drugs
Human/Dog
Human Neither Dog Human
Adverse effect: tissue-level or BMO (MedDRA Level 1–4)
Level 2 — vascular anomalies congenital nec Level 2 — purpuras (excluding thrombocytopenic) Level 3 — coagulopathies and bleeding diatheses Level 3 — infections (pathogen unspecified) Level 4 — neoplasms benign, malignant and unspecified Level 2 — pulmonary vascular resistance Level 2 — musculoskeletal and connective tissue signs and Level 2 — vascular malformations and acquired anomalies Level 3 — vascular physiological observation Level 2 — non-site specific necrosis and vascular insufficiency nec Level 2 — cell metabolism disorders nec Level 3 — miscellaneous and site unspecified neoplasms Level 2 — neurological signs and symptoms nec Level 3 — renal physiological observation Parameters
Number of drugs
Human/Dog
Human Neither Dog Human
Adverse effect: tissue-level or BMO (MedDRA Level 1–4)
Level 2 — hepatobiliary neoplasms (malignancy unspecified) Level 2 — inflammatory disorders following infection Level 2 — lower respiratory tract inflammatory and immunologic Level 3 — hepatobiliary neoplasms (malignant and unspecified) Level 2 — neoplasms unspecified (malignancy and site Level 1 — right ventricular hypertrophy Level 1 — hepatic mitochondrial swelling Level 2 — hepatic cytochrome p450 level Level 2 — lipid metabolism and deposit disorders nec Level 3 — electrolyte and fluid balance conditions The use of dogs in predicting human toxicology 349 Parameters
Number of drugs
Human/Dog
Human Neither Dog Human
Adverse effect: tissue-level or BMO (MedDRA Level 1–4)
Level 1 — centrilobular hepatic necrosis Level 3 — procedural related injuries and complications nec Level 2 — cardiac function diagnostic procedures Level 2 — renal vascular and ischaemic conditions Level 2 — circulatory collapse and shock Level 1 — complete atrioventricular block Level 2 — metabolic acidoses (excluding diabetic acidoses) Level 3 — increased intracranial pressure and hydrocephalus Level 2 — increased intracranial pressure disorders Level 2 — cerebrospinal fluid tests (excluding microbiology) Level 3 — neurological, special senses and psychiatric investigations Level 2 — nervous system haemorrhagic disorders Level 2 — cardiac and vascular procedural complications Level 2 — hepatobiliary signs and symptoms Level 2 — site specific embolism and thrombosis nec Level 2 — ventricular refractory period Level 2 — hepatic cytochrome p450 function Parameters
Number of drugs
Human/Dog
Human Neither Dog Human
Adverse effect: tissue-level or BMO (MedDRA Level 1–4)
Level 4 — pregnancy, puerperium and perinatal conditions Level 2 — site-specific vascular disorders nec Level 2 — bile duct infections and inflammations Level 1 — ventricular premature complex Level 3 — hepatic and biliary neoplasms (benign) Level 2 — hepatobiliary neoplasms benign Level 3 — neurological physiological observation Level 2 — peripheral vascular resistance Level 2 — left ventricular systolic blood pressure

Source: http://frame.org.uk/dynamic_files/bailey_final.pdf