Phase seeding may provide a gateway to structure solution by deep learning.

Area of Science:

• Crystallography
• Artificial Intelligence in Science
• Computational Chemistry

Background:

• Ab initio phasing is a cornerstone of X-ray crystallography for determining molecular structures.
• Solving complex crystal structures, especially large and non-centrosymmetric ones, remains computationally challenging.
• Integrating artificial intelligence (AI) into traditional crystallographic workflows offers potential for enhanced efficiency.

Purpose of the Study:

• To propose a novel phase-seeding method that combines AI with established ab initio phasing techniques.
• To demonstrate how AI-generated phase seeds can enhance traditional crystallographic structure solution.
• To reduce the computational burden of AI in phasing by transforming the problem into a classification task.

Main Methods:

• The proposed method utilizes a small subset of approximate phase values, termed 'phase seeds', potentially generated by machine learning models.
• Phase values are discretized into angular bins, converting the continuous phase problem into a classification task.
• This approach aims to assist, rather than replace, traditional ab initio phasing methods.

Main Results:

• The phase-seeding method significantly enhances traditional crystallographic phasing techniques.
• The hybrid approach shows promise for improving structure solution, particularly for challenging crystal samples.
• Discretization reduces the computational demands for AI training in crystallographic phasing.

Conclusions:

• The phase-seeding method represents a promising hybrid strategy for AI-assisted crystallographic workflows.
• This approach facilitates the integration of AI with established crystallographic tools for improved structure determination.
• Future research can explore further applications of AI in advancing crystallographic phasing.