Machine learning approaches to understand the uptake and elimination of anthropogenic stressors in animal health

Abstract

The detection of pharmaceuticals in aquatic ecosystems raises questions about their potential effects on non-target organisms. However, environmental concentrations do not directly reflect the internal body burden of organisms. As part of this work, an investigation of contaminants of emerging concern in a coastal ecosystem as well as in marine biota highlighted this, with internal and surface water concentrations differing substantially. The gap may be linked to both environmental conditions and compound-specific uptake and elimination processes, which affect their bioaccumulation potential. Laboratory-based studies to understand compound-dependent uptake and elimination kinetics have traditionally focused on fish. In line with the 3Rs principle and regulatory efforts to move away from vertebrate testing, and to broaden taxonomic coverage, this work aimed to assess the potential of the invertebrate model species Daphnia magna to be used in mixture-based exposures as a high-throughput alternative. This necessitated the development and validation of a broad targeted analytical method for the determination of multi-class pharmaceuticals in D. magna, which had not previously been developed. This work successfully developed a method applicable across multiple pharmaceutical classes, capable of quantifying more than 50 compounds to acceptance criteria defined by ICH method validation guidelines. Mixture toxicokinetic exposures were performed, ranging from a single compound to simultaneous exposure of up to 50 compounds, and toxicokinetic profiles were successfully derived for 49 pharmaceuticals. Pharmaceuticals generally exhibited low bioconcentration potential with the majority of compounds having bioconcentration factors (BCFs) of <60 L kg-1 dry weight (dw), yet values ranged from 2 to >10,000 L kg-1 dw. The broad range highlights the importance of considering internal exposure when assessing hazard and subsequent risk. The generated toxicokinetic data was applied to machine learning models alongside published fish BCF data. The models achieved good performance for fish but showed limited predictive ability for D. magna. Mechanistic evaluation confirmed that models identified established properties and relationships that are linked to bioaccumulation and membrane permeability (logP, TPSA), demonstrating their ability to capture relevant processes when sufficient representative data is available. Overall, this work highlighted the potential of mixture-based approaches in the non-vertebrate organism D. magna to enable higher-throughput bioconcentration assessment. Utilising this high-throughput testing could rapidly generate ecotoxicity data to fulfil regulatory requirements and enable in silico modelling approaches to potentially replace animal testing in the future.