PhAI: A deep-learning approach to solve the crystallographic phase problem

Area of Science:

• Crystallography and Structural Biology
• Artificial Intelligence in Science
• Computational Chemistry

Background:

• X-ray crystallography is vital for determining 3D molecular structures.
• Reconstructing electron density maps requires complex structure factors, including amplitudes and phases.
• The loss of phase information during experiments is known as the crystallographic phase problem.

Purpose of the Study:

• To investigate the potential of neural networks in solving the crystallographic phase problem.
• To develop an AI-driven method for reconstructing crystal structures from X-ray diffraction data.
• To assess the efficiency and resolution achievable by a neural network approach.

Main Methods:

• Training a neural network on millions of artificial structure factor datasets.
• Utilizing the trained neural network to predict phase information from diffraction data.
• Evaluating the network's performance on solving the phase problem at a resolution of 2 angstroms.

Main Results:

• The neural network successfully solved the crystallographic phase problem at 2 angstrom resolution.
• The AI method required only 10-20% of the data typically needed for direct methods.
• The network demonstrated efficacy in common space groups and for modest unit-cell dimensions.

Conclusions:

• Neural networks offer a powerful new tool for addressing the crystallographic phase problem.
• This AI-driven method significantly reduces data requirements and computational time.
• The approach shows promise for analyzing weakly scattering crystals and advancing structural biology.