Chemical knowledge-informed framework for privacy-aware retrosynthesis learning.

Area of Science:

• Computational chemistry
• Machine learning
• Chemical informatics

Background:

• Chemical reaction data is crucial for advancements in pharmaceuticals, materials science, and industrial chemistry.
• Proprietary reaction data requires strict confidentiality, posing challenges for current machine learning training paradigms.
• Existing methods for training retrosynthesis models involve aggregating data from multiple sources, creating significant privacy risks due to data transmission and potential exposure.

Purpose of the Study:

• To develop a privacy-preserving approach for training machine learning-based retrosynthesis models.
• To enable distributed training of retrosynthesis models across multiple organizations without compromising proprietary data confidentiality.
• To introduce the chemical knowledge-informed framework (CKIF) for secure and effective model training.

Main Methods:

• Developed the chemical knowledge-informed framework (CKIF) for privacy-preserving distributed training.
• CKIF aggregates model parameters iteratively, avoiding the need to share raw reaction data.
• Leveraged chemical properties of predicted reactants to assess model behaviors and determine adaptive aggregation weights.

Main Results:

• CKIF successfully enables distributed training of retrosynthesis models across organizations.
• The framework maintains the confidentiality of proprietary chemical reaction data.
• CKIF demonstrated superior performance compared to several strong baseline methods on various reaction datasets.

Conclusions:

• The chemical knowledge-informed framework (CKIF) offers a secure and effective solution for privacy-preserving retrosynthesis model training.
• CKIF addresses the critical need for confidentiality in handling sensitive chemical reaction data.
• This approach facilitates collaborative model development in chemical research without data breaches.