Although inappropriate pharmacokinetic properties were a major cause of attrition in the 1990s, safety issues are recognized as today’s single largest cause of drug candidate failure. It is expected that the right balance of in vivo, in vitro and computational toxicology predictions applied as early as possible in the discovery process will help to reduce the number of safety issues. This review focuses on recent developments in computational toxicology. Direct modeling of toxic endpoints has been deceiving and hampered the wide acceptance of computer predictions. The current trend is to make simpler predictions, closer to the mechanism of action, and to follow them up with in vitro or in vivo assays as appropriate.