The tip of the iceberg: using AI to identify toxic chemicals

Abstract

This project was conducted to analyze the TRIDENT models described by Gustavsson et al. in the article Transformers enable accurate prediction of acute and chronic chemical toxicity in aquatic organisms [1]. The aims were to investigate the predictions made by the models, the relationship between the model’s chemical space, and the predictions they make. The methods used for the analyses were a combination of TRIDENT model predictions, modeling, and visualizations. The results of which were that there is a relationship between how accurately the TRIDENT models predict and the closeness the chemical has to the TRIDENT training data as well as the density of close neighbors in the training data. We also found that there are chemicals for which the TRIDENT models predict effective concentration values that are inconsistent with the measured value (label), possibly warranting further investigation of the chemical’s toxicity.