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

We present a simple methodology that allows the building and sharing of predictive models without compromising the confidentiality of the structures of the training series. Multiple shared models can be used to obtain ensemble models, providing better coverage of the chemical space and better predictions than the original one. This approach is demonstrated in a collaborative exercise where four pharmaceutical and chemical companies developed predictive models for the AMES mutagenicity end point and shared them to build ensemble models using logical and machine learning algorithms. The results were systematically analyzed and compared, obtaining clear benefits in predictive quality. The method has the advantage of being very simple to execute, using only open-source software, and the possibility to audit the whole process and interchange files to guarantee that no confidential information is exported from the company facilities.