Transferring Knowledge from MM to QM: A Graph Neural Network-Based Implicit Solvent Model for Small Organic...

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Machine Learning

Background:

• Molecular environment significantly impacts conformational ensembles.
• Accurate environmental modeling is crucial for computational studies.
• Machine learning (ML) shows promise for modeling molecular interactions, particularly in classical molecular dynamics.

Purpose of the Study:

• To develop a general machine-learned implicit solvent model for quantum-mechanical (QM) calculations.
• To overcome the computational cost limitations of existing ML implicit solvent models for QM.
• To create a model transferable across different QM functionals and basis sets.

Main Methods:

• Developed a novel approach by transferring knowledge from classical interactions to QM.
• Emulated a QM/MM (quantum mechanics/molecular mechanics) setup with electrostatic embedding and a nonpolarizable MM solvent.
• Utilized a graph neural network (GNN) to create the implicit solvent model (QM-GNNIS).

Main Results:

• The QM-GNNIS model requires neither QM/MM reference calculations nor experimental data for training.
• The GNN-based model is compatible with any QM functional and basis set.
• QM-GNNIS accurately describes 39 organic solvents for small organic molecules and reproduces experimental trends.

Conclusions:

• The developed QM-GNNIS model offers a computationally efficient and broadly applicable solution for implicit solvent modeling in QM calculations.
• This approach successfully bridges the gap between classical and quantum mechanical simulations for solvent effects.
• The model demonstrates superior performance compared to existing state-of-the-art methods in reproducing experimental observations.