Improving the odds in discriminating "drug-like" from "non drug-like" compounds

We have used a feed-forward neural network technique to classify chemical compounds into potentially "drug-like" and "non drug-like" candidates. The neural network was trained to distinguish between a set of "drug-like" and "non drug-like" chemical compounds taken from the MACCS-II Drug Data Report (MDDR) and the Available Chemicals Directory (ACD). The 2D atom types (of the full atomic representation) were assigned and applied as descriptors to encode numerically each compound. There are four main conclusions: First the method performs well, correctly assigning 88% of the compounds in both MDDR and ACD. Improved discrimination was achieved by a more critical selection of training sets. Second, the method gives much better prediction performance than the widely used "Rule of Five", which accepts as many as 74% of the ACD compounds but only 66% of those in MDDR, resulting in a correlation coefficient which is effectively zero, compared to a value of 0.63 for the neural network prediction. Third, based on a standard Tanimoto similarity search the selection of drug-like compounds in the evaluation set is not biased toward compounds similar to those in the training set. Fourth, the trained neural network was applied to evaluate the drug-likeness of 136 GABA uptake inhibitors with impressive results. The implications of applying a neural network to characterize chemical compounds are discussed.