The inability to understand and control the ADMET properties of molecules is an important reason why many candidate drugs fail late in the development pathway. Unfavorable pharmacokinetics, metabolism or toxicity, for example, can cause development candidates to be dropped. These failures are expensive and they contribute to the diminishing efficiency of the pharmaceutical industry. In silico models of ADMET properties allow these properties to be considered at an early, less costly stage, and should reduce the number of late-stage development candidates which fail. ADMET properties are multifactorial and complex, requiring very flexible methods to build predictive in silico models. This review summarizes the contribution neural networks are making to the development of useful ADMET models.