MLSolv-A: A Novel Machine Learning-Based Prediction of Solvation Free Energies from Pairwise Atomistic Interactions
share
Recent advances in machine learning technologies and their chemical applications lead to the developments of diverse structure-property relationship based prediction models for various chemical properties. The free energy of solvation is one of them and plays a dominant role as a fundamental measure of solvation chemistry. Here, we introduce a novel machine learning-based solvation model, which calculates the target solvation free energy from pairwise atomistic interactions. The novelty of our proposed solvation model involves rather a simple architecture: two encoding function extracts vector representations of the atomic and the molecular features from the given chemical structure, while the inner product between two atomistic features calculates their interactions, instead of black-boxed perceptron networks. The cross-validation results on 6,493 experimental measurements for 952 organic solutes and 147 organic solvents achieve an outstanding performance, which is 0.2 kcal/mol in mean-unsigned error. The scaffold-based split method exhibits 0.6 kcal/mol, which shows that the proposed model guarantees good accuracy even for extrapolated cases. Moreover, the proposed model shows an excellent transferability for enlarging training data due to its solvent-non-specific nature. Analysis of the atomistic interaction map shows there is a great potential that our proposed model reproduces group contributions on the solvation energy, which makes us believe that the proposed model not only provides the predicted target property, but also gives us more detailed physicochemical insights.
READ FULL TEXT