How well can we predict chemical toxicity using AI-based models?

Prediction biases in Quantitative Structure-Activity Relationship (QSAR) models 
The total number of chemicals on the world market continues to increase, and empirical toxicity testing is struggling to keep up. In silico tools, such as ecotoxicological quantitative structure-activity relationship (QSAR) models, are being used to fill data gaps when performing risk assessment, but to what extent can we trust these predictions? Three freely available ecotoxicity QSAR platforms (ECOSAR, Vega and T.E.S.T.) have been tested against empirical data from the United States Environmental Protection Agency (USEPA) ECOTOX database, to determine model predictivity for both acute and chronic toxicity. Predictivity for each platform was assessed with coverage (percent successful predictions across the list of chemicals), median absolute error and large deviations (percent predictions deviating from empirical data by more than a factor 10). Additionally, CLP-classification thresholds were used to examine predictivity for regulatory toxicity ranges, to find out if models are more or less accurate for different classes. Results indicate some differences in platform predictivity among the tested endpoints, and, more importantly, a problematic trend when stratifying with CLP-classes – toxicity of the most toxic compounds is underestimated.

TRIDENT-2: Predicting chemical toxicity across eukaryota
Chemical pollution is a major driver of biodiversity loss at a planetary scale and contributes substantially to the declining ecological status of ecosystems worldwide. Chemical risk assessments currently rely on animal exposure data from a limited number of model organisms to define environmentally safe boundaries, but generating such data is resource-intensive and time-consuming. Computational methods offer fast and cost-efficient alternatives; however, their limited accuracy, narrow applicability domains, and continued focus on a small set of model organisms make them ill-suited for addressing existing data gaps across both chemicals and species. Here, we present TRIDENT-2, a transformer-based model trained on over 550,000 exposure experiments, for predicting chemical toxicity towards 6674 eukaryotic species for a broad range of chemicals.