Leveraging tree-transformer VAE with fragment tokenization for high-performance large chemical model generation

Area of Science:

• Cheminformatics
• Computational Chemistry
• Artificial Intelligence in Drug Discovery

Background:

• Chemical Language Models (CLMs) using SMILES struggle with large, complex molecules and structural accuracy.
• Existing molecular generation techniques face scalability and precision challenges.

Purpose of the Study:

• To introduce the Fragment Tree-Transformer based VAE (FRATTVAE) for enhanced molecular generation.
• To address limitations in handling molecular complexity and ensuring structural integrity.
• To develop a scalable and accurate model for diverse chemical datasets.

Main Methods:

• Decomposition of molecules into fragments organized into tree structures.
• Implementation of Tree Positional Encoding to capture hierarchical relationships.
• Utilization of Transformer's self-attention mechanism for fragment dependencies.

Main Results:

• FRATTVAE demonstrates superior performance over existing methods in molecular generation tasks.
• Achieved high accuracy across various benchmark datasets, balancing reconstruction and generation quality.
• Successfully generated stable molecules with desired properties, avoiding structural alerts in optimization tasks.

Conclusions:

• FRATTVAE is a robust and versatile solution for molecular generation and optimization.
• The model's tree-based structure enhances scalability for large datasets and complex molecules.
• FRATTVAE shows significant promise for applications in cheminformatics and drug discovery.