What works and what does not: Lessons from experience in a pharmaceutical company

Computational chemists make two types of contributions in drug research. The traditional contribution is to develop models that predict molecular properties and so provide experimentalists with information that helps them rule out certain compounds before detailed investigation - for example, the investigation can be following up a high-throughput screening hit or synthesizing a new molecule. Our experience suggests that such models are useful, but a corollary utility is the addition of alternative perspective of a computational chemist to the usual composition of a discovery team. This alternative perspective can lead the team to novel discoveries, independent of quantitative models. Although we, computational chemists, show the inadequacy of global models to predict octanol-water log P, pK(a), permeability, and bioavailability, scientists continue to use these models. Apparently, they are accurate enough for practical purposes. Computational chemists make a second type of contribution when they devise expert systems that encode knowledge into a computer system. An example of the latter is our development of red and yellow structure alerts that suggest that a compound might be reactive or unstable (red alert) or contain unattractive functional groups (yellow alert). These expert systems are not derived from traditional quantitative modeling techniques, nor can they be. Nonetheless, our experience had been that scientists make heavy use of such computer programs and take a responsibility for suggesting improvements in the rules. Experience suggests that computational chemists make the largest impact when they focus on solving the specific problems faced by their experimentalist colleagues.