Graph-Based Approaches for Predicting Solvation Energy in Multiple Solvents: Open Datasets and Machine Learning...

Area of Science:

• Computational chemistry
• Materials science
• Drug discovery

Background:

• Solvation properties are crucial for designing materials and chemicals.
• Quantum chemistry simulations are accurate but computationally expensive.
• Efficient methods are needed to predict solvation energies for large molecular datasets.

Purpose of the Study:

• To develop accurate and efficient machine learning models for predicting molecular solvation energies.
• To create a comprehensive dataset of solvation energies for a wide range of molecules across multiple solvents.
• To provide accessible tools for researchers to utilize these models in chemical space exploration.

Main Methods:

• Utilized message-passing neural networks (MPNNs) trained on a large dataset of computed solvation energies.
• Generated a new dataset of solvation energies for 130,258 molecules in five common solvents using an implicit solvent model.
• Developed models requiring only the molecular graph as input, enabling rapid computation.

Main Results:

• Achieved high accuracy with mean absolute errors of 0.5 kcal/mol (9 or fewer non-hydrogen atoms) and 1 kcal/mol (10-14 non-hydrogen atoms).
• Created an openly available dataset containing 651,290 computed entries.
• Provided user-friendly interfaces for model access and application, including SMILES string input and domain of applicability estimation.

Conclusions:

• Machine learning models offer a computationally inexpensive alternative to quantum chemistry for solvation energy prediction.
• The developed dataset and models facilitate rapid screening of chemical spaces for novel molecule discovery.
• This work supports advancements in energy storage, pharmaceuticals, and industrial chemical synthesis.