An Adaptive, Modular Machine Learning System for Toxicity Prediction

We propose an innovative approach to the development of toxicity evaluation software. We will adapt a modular machine-learning software system that was originally designed for modeling inorganic and organometallic substances in catalysis and materialsscience problems to the challenge of predicting toxicity endpoints for a wide range of chemicals. The software's descriptor-based representation of substances allows us to generalize this modeling system and to simultaneously include many different typesof chemicals. The modular nature of the software allows us to create multiple models for different toxicity endpoints one by one, and then combine them seamlessly into a final product giving one prediction along with a degree of confidence in thatprediction. Due to the nature of the machine-learning approaches being applied, the reliability of the predictions will improve as more data are used to train each module. Phase I of this research will demonstrate the ability of this software to predictacute lethality as represented by oral rat LD50 values and to provide measures of confidence for those predictions. Phase II will involve creating and combining modules for other toxicity endpoints and will also involve training on human toxicologicaldata.The Air Force and other DoD components must perform risk-assessment determinations on all the new (and some existing) chemical entities that they develop and deploy. Current methods of toxicological studies are slow and expensive. The benefits toDoD of an alternative method such as the computational model proposed herein will be to speed the rate at which chemical substances can be evaluated and lower the cost of those evaluations. Other government agencies (e.g., NIST, EPA, FDA) also have a needfor toxicological evaluation software to streamline their risk-assessment procedures. In private industry, the need to perform toxicological evaluations is present and growing in the chemical manufacturing, agricultural, and pharmaceutical sectors, toname only a few. New substances are continually being developed, and stricter government regulations require that more of them be tested, at a cost in time, dollars, and animal lives that can reach staggering proportions. Effective toxicity evaluationsoftware will save enormous amounts of time and money. The market for alternative toxicity evaluation techniques is large and is expected to continue to grow.