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

The use of machine learning (ML) potentials has emerged
as a powerful
approach in computational chemistry, particularly in computer-aided
drug design studies. Neural network potentials (NNPs) provide a more
physics-informed estimation of binding and solvation events, bridging
the accuracy gap between classical force fields and quantum mechanical
methods. Current universal neural network potentials (NNPs) have not
yet achieved consistent chemical accuracy required for reliable molecular
dynamics simulations. Accurate yet data efficient representation of
potential energy surfaces and prediction of solvation free energies
are essential for large-scale molecular simulations and drug-design
workflows. Here, we utilize data efficient E(3)-equivariant graph
neural network potentials that are capable of estimating the solvation
free energies (SFEs) of small compounds with density functional theory
(DFT)-level accuracy and significantly reduced computational cost.
Leveraging the data efficiency of equivariant architectures, our models
achieve chemical accuracy with a relatively small training data set.
We demonstrate that the method is data-efficient in constructing ML
potentials. Our focus is on hydration free energy changes of small
compounds from the FreeSolv database. We develop and test two distinct
NNPsone for the gas phase and one for an implicit water modelthat
can be applied in molecular simulations and SFE calculations using
implicit solvation. The Solvation Model based on Density (SMD)-based
implicit NNP model achieves an accuracy of 89% while offering a substantial
computational speed-up compared to its DFT counterpart, which attains
90% accuracy.