Can Machine Learning Predict the Space Group Preference of Organic Molecules?

Area of Science:

• Crystallography
• Computational Chemistry
• Materials Science

Background:

• Crystal structure prediction (CSP) is vital for pharmaceuticals and materials discovery.
• High computational cost limits widespread CSP application.
• Current CSP methods often restrict search spaces by pre-selecting common space groups, risking exclusion of the true structure.

Purpose of the Study:

• To reduce the computational cost and ambiguity of selecting space groups for CSP.
• To investigate the use of machine learning models for predicting the most likely space group(s) of organic molecules.

Main Methods:

• Developed and evaluated machine learning models, including random forests and graph neural networks.
• Trained models using both 2D bonding information and 3D molecular information.
• Compared model performance against random prediction and selection based on overall space group frequencies.

Main Results:

• Both random forests and graph neural networks significantly outperformed random prediction.
• The best graph neural network model achieved 47.2% top-1 accuracy for space group prediction, an 8.2% improvement over the reference.
• Models trained with 3D molecular information showed higher accuracy than those trained with 2D information.
• Random forests performed best when incorporating both chemical and geometric molecular features.

Conclusions:

• Machine learning models can effectively predict likely space groups for organic molecules, aiding CSP.
• This approach offers a more accurate and less ambiguous alternative to traditional space group selection methods.
• The findings suggest that incorporating 3D structural and chemical/geometric features is crucial for accurate space group prediction in CSP.