Which modern AI methods provide accurate predictions of toxicological endpoints? Analysis of Tox24 challenge results.

The Tox24 challenge was designed to evaluate the progress that has been made in computational method development for the prediction of in vitro activity since the Tox21 challenge. In this challenge, participants were tasked with developing models to predict chemical binding to transthyretin (TTR), a serum binding protein, based on chemical structure. The analyzed dataset included chemicals that were screened in a competitive binding assay designed to measure the reduction in fluorescence due to displacement of 8-anilino-1-naphthalenesulfonic acid ammonium salt (ANSA) from TTR. The data were randomly split into a training set of 1012 compounds, a leaderboard set of 200, and a blind set of 300. This article provides an overview of the Tox24 Challenge and some of the models developed by the participating teams. Some of the approaches taken by winning teams included use of mixtures, enumerating tautomers, data cleaning. Many of the teams used consensus models. Overall, there has been significant progress in the development of machine learning tools since the Tox21 Challenge.