Predictive models of toxicity using deep learning: the case of mutagenicity

What is it about?

Deep neural networks can directly from low-level encoded data without the need of feature extraction. This paper shows how to take directly the molecular graphs to build predictive models for toxicity using available data sets of chemicals and property. Two Graph Convolutional Neural Networks (GCN) models are presented; they predict the yes/no output of the Ames test (a common mutagenicity test). These models take as input the molecular graphs and weight the role of their component subgraphs in producing the output. The net can also give an estimation of the uncertainty of the prediction. Those GCN models can be interpreted; the automatically extracted subgraphs can be compared with the chemical groups considered by the human experts in making the estimations.

Featured Image

Photo by Sangharsh Lohakare on Unsplash

Photo by Sangharsh Lohakare on Unsplash

Why is it important?

This paper exploits a large data set of mutagenicity data; besides a set of about six thousand Ames test data of public domain, a very large number of new data derives from the Ames/QSAR International Challenge Project of the National Institute of Health Sciences of Japan.

Perspectives