Comprehensive machine learning assessment of zebrafish behaviour and biochemical markers in response to caffeine exposure.

Abstract

Environmental exposure to caffeine (CAF) poses potential risks to aquatic ecosystems, affecting non-target species. This study investigated the chronic effects of environmentally relevant CAF concentrations, ranging from 0.16-50 µg/L, on zebrafish behaviour. A Kohonen-type artificial neural network classified zebrafish behaviour into nine behavioural classes based on a set of movement descriptors (mean meander, mean velocity, instantaneous velocity, distance to centre point, mean angular velocity and instantaneous acceleration), while a comprehensive analysis integrated behavioural classes previously defined and biochemical markers of oxidative stress, lipid peroxidation, reserve energy content, energetic pathways, and neurotoxicity. The discriminant analysis demonstrated that behaviour descriptors and biomarkers individually explained 38% and 67% of data variation, respectively, while the combination resulted in 19 models with 100% correct diagnosis. One of the models (Model A) seemed to suit the best dose-response relationship, incorporating key biomarkers including superoxide dismutase, catalase, glutathione peroxidase activities, and behavioural characteristics such as movement distance and velocity. This suggested methodology offers a different approach to evaluating CAF's ecological impact, highlighting behavioural analysis as a valuable complement to traditional ecotoxicological assessments. This study provides a novel framework for understanding organism-level responses to environmental stressors (e.g., several anthropogenic compounds), utilising Mahalanobis distance as an integrative response index. This approach shows promise for broader application in assessing the impact of various aquatic contaminants on aquatic organisms (from bacteria to fish), potentially extending to pharmaceuticals, pesticides, and industrial pollutants.