Dr. Igor Tetko received his MSc from the Moscow Institute of Physics and Technology (summa cum laude - red diploma), one of the top-ranked Universities in the Soviet Union. He carried out his postdoctoral studies in neuroinformatics at the University of Lausanne, where he developed algorithms for the analysis of EEG data, synfire chains and theoretical modeling of thalamo-cortical organization of the brain. Since 2001, Dr. Tetko is a group leader in Chemical Informatics at Helmholtz Zentrum München, as well as CEO of BigChem GmbH. Dr. Tetko has co-authored >200 publications in chemoinformatics, bioinformatics, neuroinformatics and machine learning https://scholar.google.com/citations?user=eMe8DOkAAAAJ. He is a Program Chair of the International Conference on Artificial Neural Networks (https://e-nns.org/icann2025). He currently coordinates Horizon Europe Marie Skłodowska-Curie Innovative Training Network European Industrial Doctorates project Explainable AI for Molecules - AiChemist (https://aichemist.eu) and has coordinated three other such projects in the past. He is also an Associate Editor of the ACS ChemResTox (https://pubs.acs.org/journal/crtoec) journal.
Lessons from the Tox24 Challenge: Modern AI/ML contribute models with highest accuracy but we still need a better interpretation of them
Igor V. Tetko1,2,*
1Institute of Structural Biology, Helmholtz Munich, 86764 Neuherberg, Germany
2BIGCHEM GmbH, Valerystr. 49, 85716 Unterschleißheim, Germany
The Tox24 challenge1 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, which was organised by National Institutes of Health National Center for Advancing Translational Sciences (NCATS). The 78 teams representing 27 countries participants developed models to predict chemical binding to transthyretin (TTR), a serum binding protein, based on chemical structure.1
The winner #1, as well as runners-up #2 and #6, developed their models using the OCHEM (https://ochem.eu) platform thus further confirms its high accuracy. In 8 out of 11 models, the authors used at least one representation learning method. They included Graph Neural Networks, as well as Natural Language Processing (NLP) methods based on SMILES data processing or Foundation Chemistry Models. Many of the approaches used by top-ranked models are less than five years old and did not exist during the Tox21 Challenge. These observations clearly demonstrate the high impact that advanced ML/AI methods have made on the field.
The interpretation of models remains a significant barrier for adoptance of models by regulatory agencies. Although agnostic models for model interpretation are becoming popular, the use of noninterpretable descriptors does not allow it. In this respect functional groups such as Extended Functional Groups (EFG)2 provide a good compromise. Novel methods based on Explainable AI are expected to further advance the field. This would allow better hazard and risk assessment of chemical compounds, which is in particular important to develop safe by design chemicals.
Acknowledgements
The challenge was supported by Horizon Europe Marie Skłodowska-Curie Actions Doctoral Network grant agreement No. 101120466 “Explainable AI for Molecules” (AiChemist).
(1) Eytcheson, S. A.; Tetko, I. V. Which Modern AI Methods Provide Accurate Predictions of Toxicological End Points? Analysis of Tox24 Challenge Results. Chem. Res. Toxicol. 2025, 38 (9), 1443–1451. https://doi.org/10.1021/acs.chemrestox.5c00273.
(2) Salmina, E. S.; Haider, N.; Tetko, I. V. Extended Functional Groups (EFG): An Efficient Set for Chemical Characterization and Structure-Activity Relationship Studies of Chemical Compounds. Mol. Basel Switz. 2015, 21 (1), E1. https://doi.org/10.3390/molecules21010001.