Instruction multi-constraint molecular generation using a teacher-student large language model.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Generating molecules with specific structures and properties is a significant challenge in computational chemistry.
• Existing models and tools often fall short in comprehensive molecular generation.

Purpose of the Study:

• To develop an advanced large language model for multi-constraint molecular generation.
• To enable the creation of novel molecules based on natural language descriptions of desired properties.

Main Methods:

• Introduced TSMMG (Text-based Structure and Multi-constraint Generation) model.
• Trained TSMMG on a large dataset of text-molecule pairs derived from expert knowledge ('teachers').
• Utilized natural language prompts to guide molecule generation.

Main Results:

• TSMMG achieved over 99% molecular validity across tasks.
• Demonstrated high success ratios: 82.58% (2-constraint), 68.03% (3-constraint), and 67.48% (4-constraint).
• Showcased zero-shot adaptability for novel property combinations and diverse language styles.

Conclusions:

• TSMMG is an effective framework for multi-constraint molecular generation via natural language.
• The model has significant implications for drug discovery and related scientific fields.