Evaluating TabPFN V3 for Multi-Label Toxicity Prediction: a Comparative Study on Tox21 and SIDER Datasets

• Henrique Abreu Piccolo USP
• Felipe Kenji Nakano UEL
• Ricardo Cerri USP

DOI: https://doi.org/10.5753/eniac.2026.31231

Resumo

Predicting molecular toxicity and side effects is a cornerstone of modern drug discovery, aimed at reducing the costs and ethical concerns associated with animal testing. This task is inherently multi-label, as a single chemical compound can interact with multiple biological targets and simultaneously trigger diverse toxicological pathways. However, tabular chemical datasets often face challenges such as extreme class imbalance and high sparsity. In such scenarios, TabPFN, a Transformer-based Prior-Data Fitted Network, emerges as a promising alternative, with studies showing its ability to generalize across a range of class imbalance ratios and its effectiveness on sparse and noisy data, often surpassing tuned XGBoost. However, none of the existing studies have investigated whether TabPFN maintains its performance in multi-label scenarios. Thus, this study investigates the performance of TabPFN V3 in predicting multi-label molecular toxicity and side effects across two benchmark datasets: Tox21 and SIDER. We represented molecular structures using 512-bit Morgan Fingerprints and investigated TabPFN V3 under two main multi-label strategies: Binary Relevance and Classifier Chains. Results on the Tox21 and SIDER datasets demonstrated that TabPFN V3 surpasses established frameworks XGBoost and Random Forest in PR-AUC scores across the evaluated multi-label spaces. However, under extreme class imbalance and positive label sparsity, these tree-based ensembles still retain a competitive advantage, yielding higher F1 scores compared to TabPFN V3. Our findings suggest that TabPFN V3 serves as a useful tool for chemical virtual screening, providing a tuning-free alternative predictive model that eliminates the operational overhead of manual hyperparameter optimization.

Referências

Breiman, L. (2001). Random forests. Machine Learning, 45:5–32.

Chen, T. and Guestrin, C. (2016). Xgboost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794. ACM.

Garcia de Lomana, M. et al. (2021). Multi-label classification models for predicting drug-induced liver injury targets. Toxicology in Vitro, 72:105094.

Grinsztajn, L., Oyallon, E., and Varoquaux, G. (2022). Why do tree-based models still outperform deep learning on tabular data? Advances in Neural Information Processing Systems, 35:507–523.

Hollmann, N., Müller, S., Eggensperger, K., and Hutter, F. (2022). TabPFN: A transformer that solves small tabular classification problems in a second. In NeurIPS 2022 First Table Representation Workshop.

Huang, R., Xia, M., Nguyen, D. T., et al. (2016). Tox21 challenge to predict the effects of toxic compounds on biological pathways. Frontiers in Environmental Science, 3:85.

Kuhn, M., Letunic, I., Jensen, L. J., and Bork, P. (2016). The sider database of drugs and side effects. Nucleic Acids Research, 44(D1):D1075–D1079.

Pedregosa, F., Varoquaux, G., Gramfort, A., et al. (2011). Scikit-learn: Machine learning in python. Journal of Machine Learning Research, 12:2825–2830.

Raies, S. and Bajic, V. B. (2018). In silico toxicology: comprehensive benchmarking of multi-label classification methods applied to chemical toxicity data. WIREs Computational Molecular Science, 8(3):e1352.

Read, J., Pfahringer, B., Holmes, G., and Frank, E. (2011). Classifier chains for multi-label classification. Machine Learning, 85(3):333–359.

Rogers, D. and Hahn, M. (2010). Extended-connectivity fingerprints. Journal of Chemical Information and Modeling, 50(5):742–754.

Sechidis, K., Tsoumakas, G., and Vlahavas, I. (2011). On the stratification of multi-label data. In European Conference on Machine Learning and Knowledge Discovery in Databases, pages 145–158, Berlin, Heidelberg. Springer-Verlag.

Szymanski, P. and Kajdanowicz, T. (2019). Scikit-multilearn: a scikit-based python environment for performing multi-label classification. J. Mach. Learn. Res., 20(1):209–230.

Tsoumakas, G., Katakis, I., and Vlahavas, I. (2009). Mining multi-label data. In Maimon, O. and Rokach, L., editors, Data Mining and Knowledge Discovery Handbook. Springer, Boston, MA.

Wu, Z. et al. (2018). Moleculenet: a benchmark for deep learning in molecular design. Chemical Science.

Yap, X. H. (2020). Multi-label model for toxicity prediction. Celebration of Research, Scholarship, and Creative Activities, Wright State University, 37.