MRlogP: Transfer Learning Enables Accurate logP Prediction Using Small Experimental Training Datasets

Small molecule lipophilicity is often included in generalized rules for medicinal chemistry. These aim to reduce time, effort, costs, and attrition rates drug discovery, allowing the rejection or prioritization of compounds without need synthesis testing. The availability high quality, abundant training data machine learning methods can be a major limiting factor building effective property predictors. We utilize transfer techniques get around this problem, first on large amount low accuracy predicted logP values before finally tuning our model using small, accurate dataset 244 druglike create MRlogP, neural network-based predictor capable outperforming state art freely available prediction small molecules. MRlogP achieves an average root mean squared error 0.988 0.715 against molecules from Reaxys PHYSPROP. have made trained network all associated code descriptor generation available. In addition, may used online via web interface.