Molecular Generation with Reduced Labeling through Constraint Architecture.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Molecular generation

Background:

• Machine learning (ML) models for molecular generation often require extensive labeled data, which is costly to obtain, especially for bioactivity.
• Existing multi-property molecule generation models show a dependence on the availability of such labeled datasets.

Purpose of the Study:

• To analyze the data requirements of multi-property molecule generation models.
• To develop a novel framework, Frag-G/M, for efficient molecular generation with reduced reliance on labeled data.

Main Methods:

• Proposed Frag-G/M, a fragment-based molecular generation framework.
• Utilized conditional transformer, recurrent neural networks (RNNs), and reinforcement learning (RL) within the framework.
• Evaluated model performance based on the generation of molecules with desired properties and scaffold diversity.

Main Results:

• Frag-G/M generated significantly more desired molecules than baseline models using the same amount of biological activity label data.
• Molecules generated by Frag-G/M demonstrated higher scaffold diversity compared to known active compounds and baseline-generated molecules.

Conclusions:

• Frag-G/M offers a more data-efficient approach to generating molecules with desirable properties.
• The framework's ability to produce diverse molecular scaffolds shows promise for real-world drug discovery applications.