Enhancing Unconditional Molecule Generation via Online Knowledge Distillation of Scaffolds

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Machine learning for molecular generation

Background:

• Deep learning, particularly language models, accelerates drug discovery by generating novel molecular structures.
• Molecular scaffolds are crucial for creating chemically feasible and biologically relevant molecules.
• Directly using scaffolds can introduce bias, limiting the exploration of novel chemical space.

Purpose of the Study:

• To develop a novel framework, OMG, that integrates scaffold information into unconditional molecule generation.
• To enhance the validity and novelty of generated molecules using deep learning.
• To improve the discovery of drug candidates with balanced chemical and biological properties.

Main Methods:

• Implementation of an Online knowledge distillation framework (OMG) for unconditional molecule generation.
• Utilizing a GPT model for de novo SMILES string generation and a Transformer model for scaffold-based generation.
• Deep integration of scaffold and complete molecular structure knowledge through mutual learning between models.

Main Results:

• The OMG framework significantly improves the validity and novelty of GPT-based unconditional molecule generation.
• Experimental results on benchmark datasets demonstrate enhanced performance.
• Generated molecules exhibit a favorable balance across multiple chemical properties and biological activity.

Conclusions:

• The OMG framework effectively combines the benefits of scaffold-guided generation and unconditional generation while mitigating bias.
• This approach enhances the potential for discovering viable drug candidates through improved molecular generation.
• OMG represents a promising advancement in applying deep learning to drug discovery and molecular design.