Insights from the SAMPL8 physical properties blind prediction challenge.

Abstract

The SAMPL community challenges help draw attention to major challenges impairing accuracy of biomolecular and physical models for drug discovery. The eighth iteration of SAMPL physical properties focused on prediction of pKa and logD. The pKa predictions were focused on compounds with multiple titratable functional groups. The logD portion of the challenge focused on making logD predictions for the same compounds in 7 different organic solvents such as between water and ethyl acetate or water and heptane. Here we report on the work we did for the SAMPL8 physical properties challenge which included the development of an automated high throughput to measure the pKa and the logD. We then used the acquired dataset to stage a blind challenge we used to benchmark different empirical and physical models. We analyzed the performance of various methods using error metrics such as RMSE, MAE and ME as well as correlational metrics such Rsquared, slope and Kendall Tau. Here, we discuss the SAMPL8 dataset and the results of the SAMPL8 challenge, including the pKa prediction and logD components. We will also discuss insights gained from the challenge, including which methods performed well and which performed poorly, and give suggestions for the design of future SAMPL challenges.