Predicting accurate ab initio DNA electron densities with equivariant neural networks

Area of Science:

• Computational chemistry
• Biomolecular modeling
• Machine learning

Background:

• Accurate modeling of large biomolecules like DNA is hindered by the computational cost of quantum chemistry calculations.
• Existing methods struggle with the scale required for comprehensive DNA simulations.

Purpose of the Study:

• To develop a machine learning model capable of calculating accurate ab initio electron densities for large DNA structures.
• To overcome the limitations of conventional quantum methods in biomolecular modeling.

Main Methods:

• An equivariant Euclidean neural network framework was employed.
• The model was trained on B-DNA basepair steps, learning base pairing and stacking interactions.
• The model's performance was evaluated on its accuracy, scalability, and generalization capabilities.

Main Results:

• The machine learning model achieved high accuracy (typically <1% error) for electron densities of arbitrary B-DNA structures.
• The model demonstrated negligible error increase with system size, indicating strong extrapolation capabilities.
• Computational scaling was found to be essentially linear, and the model generalized to A- and Z-DNA forms.
• Accurate electrostatic potentials for DNA were calculated, outperforming classical force fields.

Conclusions:

• The developed machine learning model offers a computationally efficient and accurate solution for electron density and electrostatic potential calculations in large DNA structures.
• This approach significantly advances the field of biomolecular modeling, enabling more precise studies of DNA behavior and interactions.