Using data mining to improve mutation in a tool for molecular evolution

We have developed an evolutionary algorithm-based program for drug design, the molecule evoluator. This program transforms known molecules into new molecules which may have improved properties relative to the parent molecule. Transforming the parent molecule into a derivative by mutation is necessary to find molecules with increased fitness. However, mutations that just randomly add and substitute atoms often result in molecules that contain undesirable chemical substructures, and can therefore not be used as drugs. We therefore want to add knowledge to the program about which mutations result in proper chemical structures and which ones do not. In this research we have mined a large chemical database, the World Drug Index, to obtain the frequencies of small substructures in drug-like molecules. Some of our mutation operators were subsequently modified to use these frequencies. Testing the new mutation frequencies on another large database of molecules, the NCI database, we found that the knowledge-based mutations more often produced existing molecules than the original mutations. This suggests that the modified mutations produce molecules that are easier to synthesize and more drug-like compared to the molecules generated using the original uninformed mutation operators.