Uncertainty Quantification for Aquatic Ecotoxicity Prediction
7 affiliations
- 1National Institute for Public Health and the Environment
- 2Department of Mathematics, Computer Science and Statistic
- 3Ghent University
- 4Leiden University
- 5Leiden Institute of Advanced Computer Science
- 6Department of Environmental Science
- 7Radboud University
Abstract
Aquatic ecosystems are threatened by thousands of chemicals whose hazard is often unknown. Laboratory tests of individual chemicals on individual species establish, for example, the Lethal Concentration for 50% of test individuals (LC50), but for most chemicals in use such data exist for only a small fraction of species. A growing body of research applies machine learning to bridge these gaps. Before such models can be relied on in regulatory and environmental management contexts, they must not only be accurate in general, but also quantify how far individual predictions can be trusted. We address this by training a Bayesian Factorization Machine (BFM) on ecotoxicity test data. The BFM matches the accuracy of the current state of the art in predicting LC50 values, and additionally quantifies the uncertainty of every prediction, separating the part arising from a lack of data (epistemic) from the part that is irreducible, reflecting inter-test and biological variability (aleatoric). We then show how this uncertainty propagates into environmental quality assessment: Species Sensitivity Distributions (SSDs) become ensembles of plausible curves rather than single fits, and the hazardous concentrations derived from them carry credible intervals that are widest for the data-poor chemicals where conventional SSDs are least reliable.
Related
- High performance white blood cell classification using ensemble CNNs with Monte Carlo dropout and conformal prediction for clinical uncertainty quantification
- Nondimensionalization, Variational Inequalities and Uncertainty Quantification in a Morphoelastic Model for Burn Injury Healing
- Quantification of fat content in the prediction of lymph node metastasis in patients with head and neck squamous cell carcinoma: a study based on dixon-MRI
- Multiomic liquid biopsy: emerging applications in disease quantification, biomarker detection, molecular phenotyping, and cancer signal of origin prediction
- Screening Chemical Alternatives with Limited Ecotoxicity Data To Support Safe and Sustainable by Design Innovation.
BibTeX
@misc{price2026uncertainty,
title = {{Uncertainty Quantification for Aquatic Ecotoxicity Prediction}},
author = {Tadeusz Price and Oliver Urs Lenz and Markus Viljanen and Leo Posthuma},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15009892/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009892/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Uncertainty Quantification for Aquatic Ecotoxicity Prediction
AU - Tadeusz Price
AU - Oliver Urs Lenz
AU - Markus Viljanen
AU - Leo Posthuma
PY - 2026
DA - 2026/10/05
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009892/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009892/v1
AB - Aquatic ecosystems are threatened by thousands of chemicals whose hazard is often unknown. Laboratory tests of individual chemicals on individual species establish, for example, the Lethal Concentration for 50% of test individuals (LC50), but for most chemicals in use such data exist for only a small fraction of species. A growing body of research applies machine learning to bridge these gaps. Before such models can be relied on in regulatory and environmental management contexts, they must not only be accurate in general, but also quantify how far individual predictions can be trusted. We address this by training a Bayesian Factorization Machine (BFM) on ecotoxicity test data. The BFM matches the accuracy of the current state of the art in predicting LC50 values, and additionally quantifies the uncertainty of every prediction, separating the part arising from a lack of data (epistemic) from the part that is irreducible, reflecting inter-test and biological variability (aleatoric). We then show how this uncertainty propagates into environmental quality assessment: Species Sensitivity Distributions (SSDs) become ensembles of plausible curves rather than single fits, and the hazardous concentrations derived from them carry credible intervals that are widest for the data-poor chemicals where conventional SSDs are least reliable.
ER -