Can we learn to distinguish between "drug-like" and "nondrug-like" molecules?

We have used a Bayesian neural network to distinguish between drugs and nondrugs. For this purpose, the CMC acts as a surrogate for drug-like molecules while the ACD is a surrogate for nondrug-like molecules. This task is performed by using two different set of 1D and 2D parameters. The 1D parameters contain information about the entire molecule like the molecular weight and the the 2D parameters contain information about specific functional groups within the molecule. Our best results predict correctly on over 90% of the compounds in the CMC while classifying about 10% of the molecules in the ACD as drug-like. Excellent generalization ability is shown by the models in that roughly 80% of the molecules in the MDDR are classified as drug-like. We propose to use the models to design combinatorial libraries. In a computer experiment on generating a drug-like library of size 100 from a set of 10 000 molecules we obtain at least a 3 or 4 order of magnitude improvement over random methods. The neighborhoods defined by our models are not similar to the ones generated by standard Tanimoto similarity calculations. Therefore, new and different information is being generated by our models, and so it can supplement standard diversity approaches to library design.