The significance of meaning: why do over 90% of behavioral neuroscience results fail to translate to humans, and what can we do to fix it?

The vast majority of drugs entering human trials fail. This problem (called "attrition") is widely recognized as a public health crisis, and has been discussed openly for the last two decades. Multiple recent reviews argue that animals may be just too different physiologically, anatomically, and psychologically from humans to be able to predict human outcomes, essentially questioning the justification of basic biomedical research in animals. This review argues instead that the philosophy and practice of experimental design and analysis is so different in basic animal work and human clinical trials that an animal experiment (as currently conducted) cannot reasonably predict the outcome of a human trial. Thus, attrition does reflect a lack of predictive validity of animal experiments, but it would be a tragic mistake to conclude that animal models cannot show predictive validity. A variety of contributing factors to poor validity are reviewed. The need to adopt methods and models that are highly specific (i.e., which can identify true negative results) in order to complement the current preponderance of highly sensitive methods (which are prone to false positive results) is emphasized. Concepts in biomarker-based medicine are offered as a potential solution, and changes in the use of animal models required to embrace a translational biomarker-based approach are outlined. In essence, this review advocates a fundamental shift, where we treat every aspect of an animal experiment that we can as if it was a clinical trial in a human population. However, it is unrealistic to expect researchers to adopt a new methodology that cannot be empirically justified until a successful human trial. "Validation with known failures" is proposed as a solution. Thus new methods or models can be compared against existing ones using a drug that has translated (a known positive) and one that has failed (a known negative). Current methods should incorrectly identify both as effective, but a more specific method should identify the negative compound correctly. By using a library of known failures we can thereby empirically test the impact of suggested solutions such as enrichment, controlled heterogenization, biomarker-based models, or reverse-translated measures.