Predicting aggregate morphology of sequence-defined macromolecules with recurrent neural networks

Area of Science:

• Polymer Science
• Computational Chemistry
• Materials Science

Background:

• Self-assembly of sequence-defined macromolecules creates complex structures.
• Predicting this behavior is difficult due to vast design spaces and computational costs.

Purpose of the Study:

• Accurately predict macromolecule aggregate morphology using machine learning.
• Develop a high-throughput screening method for sequence design.

Main Methods:

• Applied supervised machine learning, specifically regression models with implicit representation learning.
• Compared nine different model architectures, including recurrent neural networks and k-mer based methods.

Main Results:

• Implicit representation learning models significantly outperformed engineered feature models.
• A recurrent neural network-based regressor achieved the best performance.
• Successfully identified multiple sequences for self-assembly into desired morphologies via high-throughput screening.

Conclusions:

• Machine learning, particularly recurrent neural networks, offers an effective approach to predict macromolecule self-assembly.
• This strategy enables efficient design and screening of sequences for targeted aggregate structures.
• The methodology holds potential for complex design scenarios involving polydispersity and environmental variations.