Although lack of efficacy is an important cause of late stage attrition in drug development the shortcomings in the translation of toxicities observed during the preclinical development to observations in clinical trials or post-approval is an ongoing topic of research. The concordance between preclinical and clinical safety observations has been analyzed only on relatively small data sets, mostly over short time periods of drug approvals. We therefore explored the feasibility of a big-data analysis on a set of 3,290 approved drugs and formulations for which 1,637,449 adverse events were reported for both humans animal species in regulatory submissions over a period of more than 70 years. The events reported in five species – rat, dog, mouse, rabbit, and cynomolgus monkey - were treated as diagnostic tests for human events and the diagnostic power was computed for each event/species pair using likelihood ratios.
The animal-human translation of many key observations is confirmed as being predictive, such as QT prolongation and arrhythmias in dog. Our study confirmed the general predictivity of animal safety observations for humans, but also identified issues of such automated analyses which are on the one hand related to data curation and controlled vocabularies, on the other hand to methodological changes over the course of time.