Integrating bioassay and machine learning data for ecological risk assessments of herbicide use on Ulva australis

Herbicide contamination of aquatic ecosystems poses a critical risk to biodiversity. Bioassays provide useful ecological insights on responses to herbicides; however, they require a model organism. Ulva australis is an ideal candidate for herbicide toxicity evaluations. Conventional monitoring methods have certain limitations, necessitating innovative approaches for ecological risk assessment. We evaluated the toxicity of six herbicides (atrazine, chlorimuron-ethyl, diuron, hexazinone, simazine, and pendimethalin) to U. australis by integrating experimental bioassays with advanced machine learning models. Three key endpoints were measured-reproduction, relative growth rate, and photosynthetic efficiency. Species sensitivity distribution modelling was employed to determine the hazardous concentration values for 5 % of species (HC₅) and the predicted no-effect concentration (PNEC). The derived values aligned well with regulatory benchmarks. For diuron, the PNEC (0.37 ± 0.25 μg L^-1) closely matched the value of the European Chemicals Agency (0.32 μg L^-1). In contrast, the HC₅ for hexazinone (26.8 ± 28.7 μg L^-1) was lower than that specified by the Australian/New Zealand guideline (75 μg L^-1). Machine learning models showed high predictive accuracy, with gradient boosting outperforming random forest (R² = 0.933, RMSE = 0.0036 mg L^-1 vs R² = 0.878 and RMSE = 0.0048 mg L^-1). Sensitivity analysis confirmed the robustness of gradient boosting to input variability, highlighting its suitability for ecological risk assessment. This approach establishes a scalable framework for ecological risk evaluation by integrating experimental and computational methodologies. The resulting data can also generate adaptive strategies to mitigate herbicide impacts and protect aquatic ecosystems.