Data-Efficient Equivariant NNPs Enable DFT-Accurate Simulations and Implicit Solvation Free Energies.

Area of Science:

• Computational chemistry
• Machine learning in drug design
• Molecular modeling

Background:

• Machine learning (ML) potentials, including neural network potentials (NNPs), are increasingly vital in computational chemistry and computer-aided drug design.
• NNPs offer physics-informed estimations of binding and solvation, aiming to bridge the accuracy gap between classical force fields and quantum mechanical methods.
• Current universal NNPs lack the consistent chemical accuracy needed for reliable molecular dynamics simulations, highlighting the need for accurate and data-efficient potential energy surface representations.

Purpose of the Study:

• To develop data-efficient E(3)-equivariant graph neural network potentials for accurate solvation free energy (SFE) estimation.
• To achieve density functional theory (DFT)-level accuracy in SFE predictions with significantly reduced computational cost.
• To demonstrate the data efficiency of equivariant architectures for constructing ML potentials using a relatively small training dataset.

Main Methods:

• Utilized E(3)-equivariant graph neural network potentials for SFE calculations.
• Developed two distinct NNPs: one for gas-phase calculations and another for an implicit water model (Solvation Model based on Density - SMD).
• Trained and tested models on small compounds from the FreeSolv database, focusing on hydration free energy changes.

Main Results:

• The developed NNPs achieved chemical accuracy for SFE predictions.
• The implicit solvation NNP model demonstrated an accuracy of 89% with substantial computational speed-up compared to DFT.
• The DFT counterpart achieved 90% accuracy, validating the ML approach's performance.

Conclusions:

• Data-efficient E(3)-equivariant graph neural network potentials can accurately estimate SFEs at DFT level.
• These ML potentials significantly reduce computational cost, making them suitable for large-scale molecular simulations and drug-design workflows.
• The study highlights the potential of equivariant architectures for efficient and accurate ML potential development in computational chemistry.