J Lane1; L Hosseini Gerami1; L Granston1; M Wilkinson1; S Windsor1;
1 AbsoluteAi, UK
AbstractAbsoluteAi is a start-up specialising in in silico drug toxicity prediction using machine learning, and we are excited to present our work on the development of TOXSCAPE, a suite of world-leading toxicity prediction tools.Our machine learning platform ATLAS will incorporate a range of featurisers, including fingerprints, images, graphs, and proprietary features, along with classic ML and deep learning algorithms, including proprietary architectures. The platform is designed to utilise a wide range of data sources, including public data, proprietary data, custom-generated data, partner data, and high-content image data. All data is cleaned, standardised, and augmented to form our proprietary mix of datasets. ATLAS then sweeps this large computational space of features, algorithms and hyperparameters to optimise the combination. It then uses hybridisation and model ensembling methods to find the best of the best model. We have also developed interpretability tools to help users better understand the predictions generated by our models. These include concentration analysis of the toxicity prediction, R-group contribution and mechanistic understanding of toxicity. We have exemplified the capabilities of ATLAS with hERG prediction, providing world-leading solutions for toxicity prediction. While hERG prediction is just one of the many prediction tools that make up TOXSCAPE, it exemplifies the state-of-the-art metrics we strive for in all our models.TOXSCAPE is providing valuable insights and predictions for a range of toxicity endpoints and contributing to safer and more effective drug discovery. We look forward to continuing to develop TOXSCAPE and collaborating with other researchers and companies to improve the accuracy and specificity of our models.