Deep Learning Neural Networks for Quantitative Toxicity Predictions

The project will build on a continuing collaboration with Unilever on toxicity prediction and analysis. It will being with the construction, evaluation and delivery of regression Bayesian neural networks for implementation in the existing MIE atlas. These models will be evaluated across several biological targets and provide quantitative predictions and standard deviations in each prediction. Together with other computational approaches (Random Forests, Support Vector Machines, etc), these will be used to build up combination models, further extending our earlier discoveries on the power of consensus models in this area. Applicability domains and confidence in predictions will be considered throughout to ensure maximum model impact. The best performing algorithms will be compared to the accuracy and reliability of in vitro assays. These results should also be useful in finding the relative weaknesses and gaps in the experimental data. The knowledge built up from these models will be used to aid in the development of quantitative AOPs using literature available data.