Enhancing chemical reaction search through contrastive representation learning and human-in-the-loop.

Area of Science:

• Chemistry
• Computer Science
• Data Science

Background:

• Chemical reaction databases are crucial for synthesis planning.
• Manual query formulation for reaction databases is challenging and time-consuming.
• Existing search methods lack adaptability to user-specific needs.

Purpose of the Study:

• To develop an intelligent chemical reaction search system that simplifies result enhancement.
• To improve the alignment of search results with user preferences and requirements.
• To provide an alternative to complex query formulation for reaction databases.

Main Methods:

• Incorporation of contrastive representation learning to embed reaction records as numerical vectors.
• Application of dimensionality reduction for efficient vector compression and search.
• Integration of human-in-the-loop techniques to iteratively refine the representation model with user feedback.

Main Results:

• The proposed system effectively refines search results based on user binary ratings.
• Experimental investigations demonstrate improved chemical reaction search accuracy.
• The system enhances the discovery of relevant reaction records, especially for users with limited explicit knowledge.

Conclusions:

• The intelligent search system simplifies and enhances chemical reaction discovery.
• The method offers a user-centric approach inspired by recommender systems.
• This approach assists chemists in efficiently finding relevant reactions without complex query engineering.