Compressed graph representation for scalable molecular graph generation.

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Deep learning models are increasingly used for molecular graph generation.
• Computational complexity increases with molecular size, limiting deep learning applications to large molecules.
• Existing methods struggle with scalability for molecules containing many heavy atoms.

Purpose of the Study:

• To develop a molecular graph compression method to reduce computational complexity.
• To enable efficient and scalable deep learning-based generation of large molecules.
• To maintain the capability of generating chemically valid and diverse molecular graphs.

Main Methods:

• A molecular graph compression technique is proposed.
• Six prevalent small substructural patterns between atoms are identified.
• These substructures are converted to additional edge features, compressing the graph representation without information loss.

Main Results:

• The method significantly reduces the number of nodes in molecular graphs.
• Generative models can be built more efficiently and scalably on compressed representations.
• Effectiveness demonstrated for molecules with up to 88 heavy atoms using the GuacaMol benchmark.

Conclusions:

• The proposed molecular graph compression method effectively addresses computational complexity in deep learning-based molecular generation.
• This approach facilitates the application of deep learning to larger and more complex molecules.
• The method ensures the generation of chemically valid and diverse molecular structures.