Llamol: a dynamic multi-conditional generative transformer for de novo molecular design

Area of Science:

• Computational Chemistry
• Artificial Intelligence
• Organic Chemistry

Background:

• Generative models, such as General Pretrained Transformer (GPT), show promise in Natural Language Processing (NLP) and molecule design.
• Exploring organic chemical space for novel compounds, particularly electro-active ones, requires advanced computational tools.

Purpose of the Study:

• To develop a novel generative transformer model for exploring organic chemical space.
• To create a flexible and robust tool for de novo molecule design, capable of handling multiple conditions.

Main Methods:

• Developed Llamol, a generative transformer model based on the Llama 2 architecture.
• Trained Llamol on a 12.5 million superset of diverse organic compounds.
• Introduced Stochastic Context Learning (SCL) for flexible and robust training, accommodating potentially incomplete data.

Main Results:

• Llamol adeptly handles single- and multi-conditional organic molecule generation (up to four conditions, with potential for more).
• The model generates valid molecular structures in SMILES notation.
• Successfully incorporated numerical properties and token sequences for conditioning, individually or in combination.

Conclusions:

• Llamol is a potent and expandable tool for de novo molecule design.
• The Stochastic Context Learning (SCL) procedure enhances flexibility and robustness in molecular generation.
• The model's ability to utilize diverse conditioning makes it valuable for discovering novel electro-active compounds.