An equivariant generative framework for molecular graph-structure Co-design.

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Designing molecules with specific properties is a key challenge in chemistry and drug discovery.
• Existing machine learning methods often fail to unify 2D and 3D molecular information, limiting their effectiveness.
• There is a need for advanced generative models that can learn structure-property relationships for molecule design.

Purpose of the Study:

• To introduce MolCode, a novel generative framework for molecular graph-structure co-design.
• To leverage 3D geometric information to enhance 2D molecular graph generation and guide 3D structure prediction.
• To demonstrate the framework's ability to design molecules with desirable properties and high target affinity.

Main Methods:

• Developed MolCode, a roto-translation equivariant generative framework.
• Integrated 3D geometric information into 2D molecular graph generation.
• Utilized the generated 2D graph to guide 3D molecular structure prediction.
• Evaluated performance on de novo molecule design, targeted molecule discovery, and structure-based drug design tasks.

Main Results:

• MolCode achieved high validity (99.95%) and uniqueness (98.75%) in generating molecular graphs and structures.
• The framework successfully generated drug-like molecules with significant affinity to target proteins (61.8% high affinity ratio).
• MolCode outperformed existing methods in various molecule design and discovery tasks.

Conclusions:

• The integration of 2D topology and 3D geometry provides complementary information crucial for effective molecule design.
• MolCode offers a powerful and versatile approach for de novo molecule design, material science, and drug discovery.
• The study provides new insights into machine learning-based molecular representation and generation.