A Two‐Stage Machine Learning Framework With Physically Inspired Descriptors for Efficient Prediction of Hydration Free Energies

Hydration free energy is essential to understanding stability and reactivity in aqueous environments. Recently, we reported a machine learning model Mole2Solv to predict experimental hydration free energies of molecules ( J. Phys. Chem. Lett. , 2023, 14, 1877). Physically inspired descriptors such as electrostatic potential distribution around solute molecules were proposed instead of chemistry‐specific input features, thus improving transferability of the prediction model when it meets new compounds. However, most of these descriptors require ab initio calculations on solute molecules in vacuum, which seriously hinders its application on large‐scale molecular solvation modeling and high‐throughput screening. To address this limitation, we propose a two‐stage framework that combines Mole2Solv with DeepMoleNet, a multilevel attention‐based graph neural network developed by Ma and coworkers ( J. Chem. Inf. Model. , 2021, 61, 1066). Using the DeepMoleNet, 13 physically inspired descriptors can be predicted based on chemical structures of solutes and then applied to the subsequent Mole2Solv regression model. Expensive electronic structure calculation is no longer necessary for generating descriptors. In comparison with our previous model, the proposed two‐stage framework keeps the prediction accuracy while significantly reduces the computational cost. Additionally, the transferability of physically inspired descriptors to unseen chemical structures can be well preserved within the two‐stage framework.