Eric March Vila (0000-0002-3161-1213) is a PhD student specializing in the integration of machine learning and cheminformatics in toxicology. With a background in bioinformatics and structural biology and a 5+ years research experience in machine learning and QSAR models at the Pharmaco Informatics laboratory (PhI lab) in Universitat Pompeu Fabra (UPF), Eric focuses on developing predictive models for chemical safety assessment and computational toxicology. His expertise spans the use of in silico methods for drug design, database management, and the application of artificial intelligence in toxicological sciences. He has been working lately on the applications of model combination in toxicology (https://doi.org/10.1016/j.toxlet.2023.10.013). Eric's current work involves collaborating on projects that leverage computational approaches to enhance the validation of new approach methodologies (NAMs) in toxicology by studying how to handle the uncertainty associated to these methods. He is passionate about bridging the gap between in vitro and in silico approaches by facilitating the adoption of machine learning tools to make toxicological data more accessible and interpretable. Eric's research aims to contribute to the evolving landscape of predictive toxicology, where computational tools play a crucial role in deriving new insights into toxic responses and their underlying biological mechanisms.
OpenTox Summer School 2025
Applications in Toxicology of predictive model combinations
In silico toxicology often benefits from combining multiple predictive models rather than relying on a single model. This seminar will explore practical approaches for integrating diverse models to address complex toxicological questions, such as predicting endpoints influenced by multiple mechanisms or spanning varied chemical spaces. These few examples show how predictive models
can be combined in different ways and for multiple purposes. The approaches will be illustrated using examples in which we will apply the open-source software Flame, developed in our group, which has been specifically designed to support the combination of multiple predictive models. For endpoints with binary outcomes, we will demonstrate how logical rules and machine learning algorithms (e.g., Random Forest or XGBoost) can be used to obtain combined predictions. For quantitative or more complex endpoints other alternatives will be presented A key focus will be on incorporating uncertainty estimates from individual models (e.g., conformal methods), either for quantitative or qualitative (classifiers) endpoints. The final aim of this seminar is to improve the usability of the results, facilitating their use for decision-making in different environments and promoting trust in in silico toxicology predictions.