Large Language Models for Controllable Multi-property Multi-objective Molecule Optimization

Area of Science:

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

Background:

• Real-world drug design necessitates optimizing multiple molecular properties to meet pharmaceutical standards.
• Existing computational methods and instruction-tuned Large Language Models (LLMs) struggle with nuanced, property-specific optimization objectives.
• This limitation hinders the practical application of AI in complex drug development scenarios.

Purpose of the Study:

• To introduce C-MuMOInstruct, the first dataset for instruction tuning focused on multi-property optimization with explicit, property-specific goals.
• To develop GeLLM4O-Cs, a series of instruction-tuned LLMs capable of targeted, property-specific molecular optimization.
• To enhance the practical applicability of AI in realistic drug design workflows.

Main Methods:

• Creation of the C-MuMOInstruct dataset, featuring property-specific objectives for multi-property optimization.
• Development of GeLLM4O-Cs, LLMs fine-tuned using the C-MuMOInstruct dataset.
• Experimental evaluation across 5 in-distribution and 5 out-of-distribution tasks to assess performance against baselines.

Main Results:

• GeLLM4O-Cs demonstrated superior performance compared to strong baselines, achieving up to a 126% higher success rate in molecular optimization tasks.
• The models exhibited strong 0-shot generalization capabilities, successfully handling novel optimization tasks and unseen instructions.
• Consistent outperformance across diverse in-distribution and out-of-distribution test cases.

Conclusions:

• GeLLM4O-Cs represent a significant advancement in AI-driven molecular optimization for drug design.
• The developed models effectively address the challenge of property-specific multi-objective optimization.
• This work paves the way for foundational LLMs capable of supporting diverse and realistic optimization objectives in pharmaceutical research.