Environmental safety assessments, as mandated by many regulations, require that toxicity data is generated for up to three trophic levels, algae, aquatic invertebrates, and fish. Conducting these testsin vivois resource-intensive, time-consuming, and causes undue suffering. Computational methods are fast and cost-efficient alternatives, however, their adaptation in regulatory settings has been slow, both due to low accuracy and narrow applicability domains. Here we present a new method for predicting chemical toxicity based on molecular structure. The method is based on a transformer, capturing structural features associated with toxicity, followed by a deep neural network that predicts the corresponding effect concentrations. After training on data from tens of thousands of exposure experiments, the model shows high predictive performance for each of the three trophic levels. Compared to commonly used QSAR methods, the model has both a larger applicability domain and a considerably lower error. In addition, training the model on data that combines multiple types of effect concentrations further improves the performance. We conclude that transformer-based models have the potential to significantly advance computational predictions of chemical toxicity and makein silicoapproaches a more attractive alternative when compared to animal-based exposure experiments.