Machine Learning for Acute Oral System Toxicity Regression and Classification

In vivotoxicity testing remains a costly and time-consuming component of any pre-clinical drug development campaign. In particular, LD50 measurements require the loss of animal life but remain a critical component in preventing lethal compounds from entering the clinic. With advances in machine learning, in silicoLD50 prediction now has the potential to greatly reduce this burden. We study various types of machine learning models to predict acute oral LD50 measurements in rats as regression and classification problems. We demonstrate that transfer learning a ResNet34 model pretrained on ImageNet with test time augmentation generates the best performing regression model and that random forest augmented with conformal prediction provides a robust methodology to perform classification.