ANN ARBOR, Mich.—Although scientists can produce vast arrays of compounds easily, most of these compounds will fail as drug candidates. Computational screening helps, but univariate analysis approaches don't necessarily accommodate compound characteristic codependencies. Researchers at Pfizer, University of Georgia, and Georgia Institute of Technology may have addressed this problem.
As they published in the Journal of Chemical Information and Modeling, the researchers took a two-step approach, developing a desirability function defined by individual compound characteristics (e.g., MW, rotatable bonds, cLogP) and a method known as sequential elimination of level combinations (SELC), which combines genetic algorithms (GAs) and "forbidden arrays" to identify and screen likely candidates.
Using the desirability function, the researchers calculate scores for each characteristic as a function of its position within limits based on a priori knowledge. They then use a multiplicative approach to set the overall compound score. To ensure they don't throw out the baby with the bath water, however, they incorporate a penalty term to prevent poor individual desirability characteristics from dominating the whole score.
With SELC, a series of compounds are synthesized and screened for their fitness. The researchers place poor performers in a "forbidden array" that prevents potentially poor compounds from being synthesized in later rounds and may be prepopulated from data known in advance. They then recombine and mutate the best performers over several cycles via GAs to generate "offspring" compounds for further testing. Unlike standard GAs, however, SELC incorporates prior data into the mutation probabilities, thereby hedging the bets of designing the most likely candidates.
In a proof-of-concept study, the researchers found that SELC was able to identify several strong drug candidates and significantly improved their understanding of the chemical space and yet only used 15% of the resources required by a typical screen.