LAGNet: better electron density prediction for LCAO-based data and drug-like substances.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Drug Design

Background:

• Electron density is vital for drug design computations.
• Current deep learning models often use plane wave (PW) methods for training data.
• The drug design field predominantly utilizes Linear Combination of Atomic Orbitals (LCAO) for quantum property calculations.

Purpose of the Study:

• To develop deep learning models for predicting electron density using LCAO-based datasets for drug-like molecules.
• To address challenges in training neural networks with LCAO data, particularly handling large core orbital amplitudes.
• To optimize data storage and propose a novel neural network architecture for this task.

Main Methods:

• Training neural networks on LCAO-based electron density datasets.
• Implementing a novel storage method using standard grids instead of uniform grids.
• Developing a new deep learning architecture, LAGNet, based on the DeepDFT model.

Main Results:

• Proper handling of core orbital amplitudes is essential for successful LCAO-based training.
• The proposed grid storage method reduced probing points by 43x and storage by 8x.
• The LAGNet architecture demonstrated effectiveness for drug-like substances and DFT datasets.

Conclusions:

• Deep learning models can be effectively trained on LCAO-based electron density data for drug design applications.
• Optimized data representation and novel architectures like LAGNet significantly improve computational efficiency.
• This work bridges the gap between common computational chemistry practices (LCAO) and deep learning in drug discovery.