Discovering a Transferable Charge Assignment Model Using Machine Learning.

Area of Science:

• Computational chemistry
• Machine learning applications in chemistry
• Molecular modeling

Background:

• Partial atomic charge assignment is crucial for force field parametrization, molecular docking, and cheminformatics.
• Machine learning offers high-speed chemical modeling but requires accurate reference data.
• Charge assignment lacks a unique, universally accepted solution.

Purpose of the Study:

• To develop a novel, computationally efficient machine learning model for atomic charge assignment.
• To accurately replicate molecular dipole moments across diverse molecular conformations.
• To assess the transferability and applicability of the new charge model.

Main Methods:

• Utilized a machine learning algorithm to learn charge assignment from molecular dipole moments.
• Trained the model on a large, diverse dataset of nonequilibrium molecular conformations (C, H, N, O atoms).
• Applied the developed Affordable Charge Assignment (ACA) model to predict dipoles and charges for out-of-sample molecules and biomolecular trajectories.

Main Results:

• The ACA model accurately predicts molecular dipole moments for new molecules.
• ACA charges demonstrate transferability to dipole and quadrupole moments of larger molecules than those in the training set.
• Application to biomolecular dynamics trajectories enabled infrared spectra prediction.
• ACA yields results comparable to Charge Model 5 at a significantly reduced computational cost.

Conclusions:

• ACA provides an accurate, computationally inexpensive method for atomic charge assignment.
• The model's ability to predict dipoles and charges accurately enhances its utility in various chemical applications.
• ACA's transferability and efficiency make it a valuable tool for large-scale molecular simulations and spectral analysis.