Predicting Aquatic Species Sensitivity Distributions Using Machine Learning in a Regulatory Setting

The assessment of chemical risks is increasingly challenged by insufficient toxicological data available for the large number of marketed substances. Recent advancements in artificial intelligence (AI) and machine learning (ML) offer promising avenues to bridge these data gaps through model-based predictions. However, the integration of these methodologies into risk assessment practices remains limited, largely due to issues surrounding (regulatory) trust in the model outcomes. We developed an ML model to predict the ecotoxicity of chemicals across a broad range of aquatic species. These predictions are utilized to construct Species Sensitivity Distributions (SSDs), which constitute a key tool in environmental risk assessment. Trust of risk assessors in the SSDs is essential for use in chemical risk assessment, especially for modeled SSDs. Hence, besides quantitative performance evaluation, we also performed an extensive qualitative validation encompassing aspects such as interpretability and transparency. However, to truly advance the development of ML models for risk assessment, it is crucial to foster interdisciplinary collaboration to enhance the applicability of these technologies within regulatory frameworks. Therefore, this study emphasizes the importance of regulatory acceptance when developing new ML models for SSD predictions. We highlight future opportunities for ML in SSD prediction, while also addressing the challenges of qualitative model validation. As such, this work aims to stimulate the discussion on advancing in silico methodologies beyond the current state of the art and bridging the gap between the technological efforts made in the field of AI and the regulatory needs of chemical risk assessors.