Collective intelligence for AI-assisted chemical synthesis

Area of Science:

• Chemistry
• Artificial Intelligence
• Computational Science

Background:

• Exponential growth in scientific literature, particularly chemical reactions, poses challenges for experimental implementation.
• Existing large language models (LLMs) show promise but lack reliability for diverse de novo compound synthesis.
• Bridging the gap between reported reactions and practical, executable experiments is crucial for scientific advancement.

Purpose of the Study:

• To introduce MOSAIC (Multiple Optimized Specialists for AI-assisted Chemical Prediction), a novel computational framework.
• To enable chemists to leverage collective knowledge from millions of reaction protocols for experimental design.
• To provide reproducible and executable experimental protocols with confidence metrics for complex chemical syntheses.

Main Methods:

• MOSAIC utilizes the Llama-3.1-8B-instruct architecture.
• The framework trains 2,498 specialized chemical experts within Voronoi-clustered spaces.
• This approach partitions vast chemical domains into searchable expert regions for efficient knowledge access.

Main Results:

• MOSAIC achieved an overall 71% success rate in experimental validation.
• Over 35 novel compounds were synthesized across pharmaceuticals, materials, agrochemicals, and cosmetics.
• The framework demonstrated the ability to discover new reaction methodologies not present in its training data.

Conclusions:

• MOSAIC offers a scalable and generalizable strategy for AI-assisted chemical discovery.
• The framework effectively translates vast amounts of scientific literature into actionable experimental protocols.
• This approach accelerates knowledge application in fields facing information overload and facilitates the creation of novel chemical entities.