An Adaptive, Modular Machine-Learning System for Toxicity Prediction

We propose a novel, modular machine-learning based toxicity prediction software system that addresses in part the Air Force's request for an Integrated Toxicity Assessment System (ITAS). Reliable predictions of toxicity will fill the critical need ofprotecting human life and reducing risks, and will effect significant cost and time savings for several government and many industrial sectors. Our proposed software will address and extend the QSAR component of ITAS and fill the role of the ComputationalChemistry component. We will expand upon the progress we have made in Phase I, where we are adapting proprietary modeling techniques intended for catalysis and materials science problems to the challenges of predicting acute oral toxicity and of providingreliability estimates for every prediction. In Phase II, we will develop "living" predictive models for several toxicological endpoints: the software will automatically generate improved models when fed new data. The software is designed to takeadvantage of an anticipated wealth of new toxicity data expected from various sources, including high throughput screening and microarray techniques. We will also develop modules necessary for managing and analyzing large quantities of toxicological data.This software will be accessible both from within ITAS and as a standalone application.