Multipolar electrostatics based on the Kriging machine learning method: an application to serine.

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Machine Learning in Chemistry

Background:

• Developing accurate and efficient electrostatic models is crucial for molecular force fields.
• Current methods often struggle to capture polarization effects dynamically.
• A novel approach combining quantum chemical topology and machine learning is explored.

Purpose of the Study:

• To describe a multipolar, polarizable electrostatic method for future force field applications.
• To evaluate the performance of this method using the pilot system serine.
• To assess the impact of different data generation methods and levels of theory.

Main Methods:

• Utilizing Quantum Chemical Topology (QCT) to partition electron density.
• Employing Kriging, a machine learning technique, to model atomic multipole moments.
• Generating conformational data via Protein Data Bank sampling and normal mode distortion (Cartesian and redundant internal coordinates).
• Calculating wavefunctions at HF/6-31G(d,p), B3LYP/apc-1, and MP2/cc-pVDZ levels.

Main Results:

• The Kriging models demonstrated predictive accuracy for electrostatic interaction energy.
• Errors varied by data generation method and level of theory, with redundant internal distortion and higher theory levels yielding lower errors.
• At the MP2/cc-pVDZ level, errors ranged from 4.0 to 6.5 kJ mol⁻¹.
• B3LYP/apc-1 and MP2/cc-pVDZ levels showed similar, superior performance compared to HF/6-31G(d,p).

Conclusions:

• The developed multipolar, polarizable electrostatic method shows promise for novel force fields.
• Machine learning, specifically Kriging, effectively models atomic multipole moments for electrostatic predictions.
• The choice of data generation method and level of theory significantly impacts prediction accuracy.