A Simple Framework for Collaborative Development of Predictive Models Trained on Proprietary Data.

Area of Science:

• Computational chemistry
• Cheminformatics
• Toxicology

Background:

• Confidentiality of chemical structures is a major barrier in developing predictive models.
• Collaborative drug discovery and chemical safety assessment require robust predictive modeling.

Purpose of the Study:

• To present a methodology for building and sharing predictive models while preserving data confidentiality.
• To demonstrate the utility of ensemble models derived from shared predictive models.

Main Methods:

• A simple methodology enabling the construction and sharing of predictive models.
• Development of ensemble models using logical and machine learning algorithms from multiple shared models.
• Collaborative exercise involving four pharmaceutical and chemical companies.

Main Results:

• Ensemble models showed improved coverage of chemical space and enhanced prediction accuracy compared to individual models.
• Clear benefits in predictive quality were observed in the AMES mutagenicity endpoint prediction.
• The methodology ensures no confidential information is exported from company facilities.

Conclusions:

• The presented methodology offers a secure and effective way to build and share predictive models.
• Ensemble modeling significantly improves predictive performance in chemical and pharmaceutical research.
• The approach utilizes open-source software, is auditable, and maintains data privacy.