On-the-fly selection of a training set for aqueous solubility prediction

Training sets are usually chosen so that they represent the database as a whole; random selection helps to maintain this integrity. In this study, the prediction of aqueous solubility was used as a specific example of using the individual molecule for which solubility is desired, the target molecule, as the basis for choosing a training set. Similarity of the training set to the target molecule rather than a random allocation was used as the selection criteria. The Tanimoto coefficients derived from Daylight's binary fingerprints were used as the molecular similarity selection tool. Prediction models derived from this type of customization will be designated as "on-the-fly local" models because a new model is generated for each target molecule which is necessarily local. Such models will be compared with "global" models which are derived from a one-time "preprocessed" partitioning of training and test sets which use fixed fitted parameters for each target molecule prediction. Although both fragment and molecular descriptors were examined, a minimum set of MOE (molecular operating environment) molecular descriptors were found to be more efficient and were use for both on-the-fly local and preprocessed global models. It was found that on-the-fly local predictions were more accurate (r2=0.87) than the preprocessed global predictions (r2=0.74) for the same test set. In addition, their precision was shown to increase as the degree of similarity increases. Correlation and distribution plots were used to visualize similarity cutoff groupings and their chemical structures. In summary, rapid "on-the-fly" similarity selection can enable the customization of a training set to each target molecule for which solubility is desired. In addition, the similarity information and the model's fitting statistics give the user criteria to judge the validity of the prediction since it is always possible that good prediction cannot be obtained because the database and the target molecule are too dissimilar. Although the rapid processing speed of binary fingerprints enable the "on-the-fly" real time prediction, slower but more feature rich similarity measures may improve follow-up predictions.