Neural networks, used in deep learning, have become a popular tool for various prediction tasks. Whilst up to now traditional machine learning approaches are commonly used for activity predictions, the Merck Kaggle competition and the Tox21 Challenge were won by deep neural network architectures. The results of the two challenges indicate that deep learning could improve the predictivity, especially for complex endpoints such as hepatotoxicity. <br/> <br/>However, a major problem in this area is the small size and imbalancy of available toxicity datasets. In order to overcome overtraining and a bias towards the majority class, traditional machine learning approaches oversample the minority class (e.g. SMOTE) or apply cost sensitive learning. Other techniques used comprise bagging or boosting. For deep neural networks these techniques are rarely used as they increase the training time or lead to even smaller datasets. With the aim to solve both, imbalancy and insufficient size of the training set, we use conformational sampling along with 3D based description of the molecules. This allows to increase the size of the dataset without creating artificial samples. First results indicate that this is a versatile method for applying deep learning technologies on specific toxicity endpoints such as drug induced liver injury.

Prediction of toxicity: Deep learning with small and imbalanced datasets

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.