Profiling mechanisms that drive acute oral toxicity in mammals and its prediction via machine learning

Abstract

Abstract We present a mechanistic machine-learning quantitative structure-activity relationship (QSAR) model to predict mammalian acute oral toxicity. We trained our model using a rat acute toxicity database compiled by the US National Toxicology Program. We profiled the database using new and published profilers and identified the most plausible mechanisms that drive high acute toxicity (LD50 ≤ 50 mg/kg; GHS categories 1 or 2). Our QSAR model assigns primary mechanisms to compounds, followed by predicting their acute oral LD50 using a random-forest machine-learning model. These predictions were further refined based on structural and mechanistic read-across to substances within the training set. Our model is optimized for sensitivity and aims to minimize the likelihood of underpredicting the toxicity of assessed compounds. It displays high sensitivity (76.1% or 76.6% for compounds in GHS 1–2 or GHS 1–3 categories, respectively), coupled with ≥73.7% balanced accuracy. We further demonstrate the utility of undertaking a mechanistic approach when predicting the toxicity of compounds acting via a rare mode of action (MOA) (aconitase inhibition). The mechanistic profilers and framework of our QSAR model are route- and toxicity endpoint-agnostic, allowing for future applications to other endpoints and routes of administration. Furthermore, we present a preliminary exploration of the potential role of metabolic clearance in acute toxicity. To the best of our knowledge, this effort represents the first accurate mechanistic QSAR model for acute oral toxicity that combines machine learning with MOA assignment, while also seeking to minimize underprediction of more highly potent substances.