Organic crystal structure prediction via coupled generative adversarial networks and graph convolutional networks

Area of Science:

• Crystallography
• Materials Science
• Computational Chemistry

Background:

• Organic crystal structures critically influence compound properties and biological activity.
• Experimental crystal structure investigations face limitations in comprehensive polymorphism studies.
• Quantum mechanics (QM)-based crystal structure prediction (CSP) is computationally expensive, hindering widespread adoption.

Purpose of the Study:

• To develop DeepCSP, a purely machine learning framework for rapid, minute-scale organic CSP.
• To overcome the computational cost limitations of traditional QM-based CSP methods.

Main Methods:

• A generative adversarial network was employed to generate trial crystal structures based on molecular features.
• A graph convolutional attention network was utilized to predict the density of stable crystal structures.
• Density-based ranking, comparing predicted and calculated densities, was implemented for screening and ordering crystal structures.

Main Results:

• DeepCSP demonstrated high performance in validating marketed drugs, with an accuracy exceeding 80% and a hit rate surpassing 85%.
• The framework achieves minute-scale prediction times, a significant speed improvement over traditional methods.

Conclusions:

• DeepCSP offers a feasible and efficient pure machine learning solution for organic crystal structure prediction.
• Artificial intelligence holds significant potential for advancing the field of CSP research through accelerated computation.