Dynamics-enhanced molecular property prediction guided by deep learning

Area of Science:

• Computational biology
• Drug discovery
• Molecular dynamics simulations

Background:

• Traditional molecular property prediction (MPP) relies on static molecular structures, neglecting crucial dynamic behaviors.
• Limited availability and high dimensionality of molecular dynamics (MD) data hinder its integration into MPP.
• There is a need for methods that leverage molecular dynamics for improved MPP.

Purpose of the Study:

• To develop and validate a novel approach for molecular property prediction (MPP) that incorporates molecular dynamics (MD) data.
• To create new, comprehensive MD datasets for MPP tasks.
• To introduce a dynamically enhanced molecular representation (DEMR) method.

Main Methods:

• Performed extensive molecular dynamics (MD) simulations to generate dynamics datasets for various molecules.
• Developed a dynamically enhanced molecular representation (DEMR) method incorporating multiple sampling strategies for MD frames.
• Utilized two deep learning pipelines to map DEMR to molecular properties across diverse MPP tasks.

Main Results:

• The proposed DEMR method and deep learning pipelines demonstrated superior performance in various MPP tasks compared to traditional methods.
• Achieved practical guidance for efficient frame selection in MD simulations for MPP.
• Highlighted the significant impact of integrating MD data into MPP.

Conclusions:

• Integrating molecular dynamics (MD) data significantly improves molecular property prediction (MPP) accuracy.
• The developed DEMR method offers a powerful tool for structure-based drug design.
• The generated MD datasets and code are publicly available to facilitate further research.