COSMOtherm wins SAMPL6 blind challenge
Within the SAMPL6 blind challenge featuring octanol-water partition coefficients (logP) for a set of drug-like molecules COSMOtherm predictions proofed to be more accurate than all competing methods. The original data and results can be found here.
The objective of this challenge was to predict the partitioning of 11 protein kinase inhibitor fragments between a water and an octanol phase. This molecular property is of high relevance for the development of new drugs as it is strongly related to the distribution of active substances within the body. The experimental data was measured by the organizers beforehand but unknown to the participants in order to evaluate the true predictivity of the computational methods applied. Out of 91 contributions using all kinds of different approaches ranging from molecular dynamics to deep learning, COSMOtherm using the FINE parameterization scored the lowest RMSE (root mean squared error) and MAE (mean absolute deviation). COSMOquick based predictions using an empirical correction term based on machine learning on top of lower level COSMOtherm calculations came in third. The results of the SAMPL6 challenge supports the claim that COSMOtherm currently computes the most accurate free energies in solution.