Dr. Igor Tetko is visiting the chair
Uko Maran
On Thursday, September 4, 2025, Dr. habil. Igor Tetko will give a lecture at the University of Tartu Institute of Chemistry titled “Tox24 Challenge: An In-Depth Analysis of Modern Artificial Intelligence Methods for Accurate Prediction of Transthyretin Binding.”
Dr. Tetko is one of Europe’s leading cheminformatics researchers, whose work focuses on the application of machine learning and artificial intelligence in chemistry. He holds a master’s degree in chemistry from the Moscow Institute of Physics and Technology and a PhD from the Academy of Sciences of Ukraine. He has worked as a postdoctoral fellow in neuroinformatics at the University of Lausanne and obtained his habilitation at the University of Strasbourg. Since 2012, he has led the cheminformatics group at the Helmholtz Institute of Structural Biology in Munich. In addition, he is the founder and CEO of BigChem GmbH, coordinator of the European Union Horizon Europe Marie Skłodowska-Curie doctoral network “AiChemist” and program chair of the International Conference on Artificial Neural Networks (ICANN2025).
Abstract of lecture by Dr Igor Tetko:
The Tox24 Challenge was aimed at identifying cutting-edge computational methods for predicting in vitro chemical binding to transthyretin (TTR), a crucial serum protein involved in thyroid hormone transport. It attracted 78 teams from 27 countries, nearly doubling previous participants in Tox21 and highlighting growing interest in New Approach Methodologies (NAMs) for toxicity prediction. Key findings demonstrated the significant impact of advanced Machine Learning (ML)/Artificial Intelligence (AI) methods. Namely, representation learning methods, including Graph Neural Networks (GNNs) and Natural Language Processing (NLP) techniques (e.g., Transformer CNN, CNF2), were predominantly used by top performers as well as the winning approaches were based on consensus modelling strategies. A crucial post-challenge analysis revealed that "More data matters!": heavy data filtering detrimentally affected model accuracy. The models developed using a whole dataset of measurements significantly outperforming those with heavily cleaned data. These results underscore the robustness and high accuracy of the winning AI protocols, prompting a need to extend existing regulatory frameworks, such as the OECD QSAR principles, to fully embrace advanced ML approaches like consensus modelling, pre-training, and fine-tuning for chemical risk assessments.
