Deep neural networks represent an appealing solution to the problems of cost, time-investment and ethical concerns associated with the in vivo assessment of potential drug candidates from animal testing. However, from only the weight parameters of the network, it is nigh impossible for humans to comprehend why a trained network arrives at its predictions. There is also the problem of adversarial inputs in deep learning. Generally, these are inputs that are altered or perturbed in such a way that they exploit gaps in the deep neural network’s training and are misclassified, despite a potentially obvious correct classification.
Therefore, this project aims to train deep learning models to predict toxicological effects of molecules, such as mutagenicity, skin sensitisation or aquatic toxicity and apply methods for the interpretation of these models. These interpretations should provide explanations for the deep neural network decisions and allow the reasons for any adversarial inputs to be easily understood. Since chemistry and toxicology are widely studied disciplines, it should be possible to determine which interpretive methods produce explanations that are closest in line with our already established chemical understanding of the mechanisms behind toxicity. That is, this project aims to establish whether the toxicity models can reproduce our knowledge of chemistry, and furthermore, which of the interpretive methods are best able to reveal this.
Another limitation in the use of machine learning for predictive toxicology models is that datasets from in vivo experimental studies are often quite small. Thus, another aspect of this project will be to train other machine learning models that, empirically, seem to be able to make reliable predictions with smaller datasets (such as Gaussian process regression). The reduced-data models will be explained and interpreted along with the deep learning models, but also the reasons for the performance in low-data regimes will be investigated.
Deep learning and the interpretation of deep neural networks, application of explainable machine learning to molecular sciences.
MChem Chemistry, University of Bath.
Winner of the R W Bolland Prize For The Final Year Student With The Highest Marks In Master Of Chemistry, 2021.
Dr Matthew Grayson
Dr Pranav Singh