Generating diversity and securing completeness in algorithmic retrosynthesis.

Area of Science:

• Artificial Intelligence in Chemistry
• Computational Chemistry
• Machine Learning for Chemical Synthesis

Background:

• Machine learning, particularly neural networks, has advanced chemical synthesis planning by accurately predicting reactions.
• Retrosynthesis, the process of planning synthesis from simple building blocks to a target molecule, requires effective algorithms for assembling predicted reactions.
• Defining objective functions for retrosynthesis is challenging due to context-specific requirements.

Purpose of the Study:

• To develop an algorithm for generating diverse retrosynthesis plans.
• To introduce a novel chemical diversity score (CDS) to quantify the diversity of synthesis plans.
• To compare the performance of the new algorithm against established methods like Monte-Carlo Tree Search.

Main Methods:

• Adaptation of Depth-First Proof-Number Search (DFPN) and its variants for retrosynthesis.
• Implementation of a novel chemical diversity score (CDS) to guide the search for diverse solutions.
• Investigation of DFPN's completeness properties, including conditions for guaranteed solution finding.

Main Results:

• The developed algorithm significantly outperforms Monte-Carlo Tree Search in generating diverse synthesis plans, as measured by the novel CDS.
• The algorithm demonstrates improved time efficiency compared to Monte-Carlo Tree Search.
• Progress was made in understanding and improving the completeness of DFPN for retrosynthesis.

Conclusions:

• The novel DFPN-based algorithm offers a superior approach for generating diverse and efficient chemical synthesis plans.
• The chemical diversity score (CDS) provides a valuable metric for evaluating the diversity of retrosynthesis outcomes.
• Further understanding of DFPN's completeness, especially when enhanced, contributes to more reliable AI-driven synthesis planning.