BackgroundThe pharmaceutical industry is in the midst of a productivity crisis and rates of translation from bench to bedside are dismal. Patients are being let down by the current system of drug discovery; of the several 1000 diseases that affect humans, only a minority have any approved treatments and many of these cause adverse reactions in humans. A predominant reason for the poor rate of translation from bench to bedside is generally held to be the failure of preclinical animal models to predict clinical efficacy and safety. Attempts to explain this failure have focused on problems of internal validity in preclinical animal studies (e.g. poor study design, lack of measures to control bias). However there has been less discussion of another key factor that influences translation, namely the external validity of preclinical animal models.Review of problems of external validityExternal validity is the extent to which research findings derived in one setting, population or species can be reliably applied to other settings, populations and species. This paper argues that the reliable translation of findings from animals to humans will only occur if preclinical animal studies are both internally and externally valid. We review several key aspects that impact external validity in preclinical animal research, including unrepresentative animal samples, the inability of animal models to mimic the complexity of human conditions, the poor applicability of animal models to clinical settings and animal–human species differences. We suggest that while some problems of external validity can be overcome by improving animal models, the problem of species differences can never be overcome and will always undermine external validity and the reliable translation of preclinical findings to humans.ConclusionWe conclude that preclinical animal models can never be fully valid due to the uncertainties introduced by species differences. We suggest that even if the next several decades were spent improving the internal and external validity of animal models, the clinical relevance of those models would, in the end, only improve to some extent. This is because species differences would continue to make extrapolation from animals to humans unreliable. We suggest that to improve clinical translation and ultimately benefit patients, research should focus instead on human-relevant research methods and technologies.