Quantum Chemical Calculations with Machine Learning for Multipolar Electrostatics Prediction in RNA: An Application...

Area of Science:

• Computational Chemistry
• Biophysics
• Machine Learning

Background:

• Accurate electrostatic modeling of RNA requires accounting for anisotropic atomic electron density.
• High-rank atomic multipole moments are crucial for this, but quantum chemical calculations are computationally expensive for large systems like RNA.

Purpose of the Study:

• To evaluate five machine learning methods for predicting atomic multipole moments.
• To assess the accuracy of predicted multipole moments and subsequent electrostatic interaction energies in RNA pentose models.

Main Methods:

• Employed Gaussian process regression with automatic relevance determination (ARDGPR), Kriging, radial basis function neural networks, Bagging, and generalized regression neural networks.
• Predicted atomic multipole moments and calculated atom-atom electrostatic interaction energies using these methods.
• Compared prediction errors for multipole moments and electrostatic energies.

Main Results:

• ARDGPR and Kriging outperformed other methods in predicting high-rank multipole moments for O, C, N, and H elements.
• ARDGPR achieved the lowest absolute average energy error (1.83 kJ mol⁻¹), significantly outperforming Kriging (4.33 kJ mol⁻¹).
• ARDGPR demonstrated a 58% reduction in absolute average energy error compared to Kriging.

Conclusions:

• ARDGPR is a reliable and accurate method for predicting electrostatic interaction energy in RNA pentose systems.
• The strong feature extraction capability of ARDGPR enables correct and dependable electrostatic modeling for large biomolecules.