Comparison of logP and logD correction models trained with public and proprietary data sets