Link to paper

The full paper is available here.

You can also find the paper on PapersWithCode here.

Abstract

Atomic partial charges are important for molecular dynamics simulations.
Traditionally, partial charges are assigned using quantum chemical methods.
A hybrid physical/graph neural network-based approach is proposed to approximate the widely popular AM1-BCC charge model.
This hybrid approach is orders of magnitude faster and maintains accuracy comparable to differences in AM1-BCC implementations.
The hybrid approach scales linearly with the number of atoms.
Source code is available at a specified URL.

Espaloma uses graph neural networks to assign continuous latent representations of chemical environments to atoms
GNNs are used to assign symmetry-preserving parameters for atomic, bond, angle, torsion, and improper force terms
Initial features associated with nodes are determined based on resonance-independent atomic chemical features
Message-passing step of GNN consists of edge update, neighborhood aggregation, and node update
Hyperparameters are optimized during training to produce robust models
Espaloma framework is used to predict atomic parameters
Partial charges are predicted by predicting electronegativity and hardness of each atom
Runtime complexity of EspalomaCharge is ( )

EspalomaCharge is comparable to or better than AmberTools and OpenEye in terms of charge RMSE
EspalomaCharge is generalizable to other molecules of significance to chemical and biophysical modeling
EspalomaCharge is most accurate where there is abundant data in the training set
EspalomaCharge is 300 to 3000 times faster than AmberTools and 15 to 75 times faster than OpenEye
EspalomaCharge has ( ) complexity and is capable of parameterizing peptides of a few hundred residues within seconds
EspalomaCharge provides a seamless way to accelerate parameterization by distributing calculations on GPU hardware
EspalomaCharge provides statistically indistinguishable performance compared to AmberTools and the OpenEye toolkit on both RMSE and R2 metrics

EspalomaCharge assigns high-quality conformation-independent AM1-BCC charges
Uses modern machine learning infrastructure
Supports accelerated hardware
Runtime complexity is ( ) with respect to number of atoms
Small discrepancies to high-quality AM1-BCC reference implementations
Assigns charges to biopolymers with hundreds of residues
Python-based machine learning framework
One-hot element encoding prevents model from perceiving elemental similarities
Future work could aim to reduce error for uncommon chemistries
Incorporate ESP as a target in the training process
Multi-objective strategy to include multiple targets and charge regularization terms

Research was supported by the National Institute for General Medical Sciences of the National Institutes of Health
YW was funded by NIH grant R01GM132386 and the Sloan Kettering Institute
JDC was funded by NIH grants R01GM132386 and R01GM140090

Authors are responsible for the content
Content does not represent the official views of the National Institutes of Health

JDC is a member of the Scientific Advisory Board of four companies
The Chodera laboratory receives funding from multiple sources
Conformers were generated using RDKit and MMFF94 charges
ESPs were calculated using OpenFF Recharge
Induced solvent potential was calculated using OpenEye ZAP
Hydration free energies were calculated using a modified protocol
Charges were assigned using a hybrid physical/GNN model
EspalomaCharge showed smaller average charge RMSE than AmberTools
EspalomaCharge is fast, even for large systems
EspalomaCharge can be used in batch mode
EspalomaCharge introduced little error to explicit hydration free energy prediction