Transferring chemical and energetic knowledge between molecular systems with machine learning.

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning

Background:

• Predicting molecular properties is crucial for chemistry, biology, and medicine.
• Machine learning has increasingly impacted molecular simulations for property prediction.
• Transferring knowledge from simple to complex molecular systems remains a challenge.

Purpose of the Study:

• To develop a novel methodology for knowledge transfer in molecular simulations.
• To predict high and low free-energy conformations in complex molecular systems.
• To apply machine learning for property prediction in atomistic systems.

Main Methods:

• Utilized a novel hypergraph representation of molecules to encode multi-atom interactions.
• Developed new message passing and pooling layers for hypergraph-structured data.
• Applied transfer learning from tri-alanine to deca-alanine systems for free-energy classification.

Main Results:

• Achieved a remarkable Area Under the Curve (AUC) of 0.92 for transfer learning.
• Demonstrated successful classification of high and low free-energy conformations.
• Showcased unsupervised grouping of deca-alanine secondary structures based on free-energy values.

Conclusions:

• The proposed methodology enables reliable transfer learning for molecular systems.
• This approach paves the way for predicting structural and energetic properties of complex biological systems.
• The study validates the potential of hypergraph neural networks in molecular simulations.