AI-Based Toxicity Prediction as an Alternative to Animal Testing

Abstract

In recent years, there has been a significant increase in the use of chemicals in our environment due to growing demand and consumption. Consequently, large-scale chemical regulation based on toxicological assays has been implemented to prevent exposure-related consequences for nature and human health. Historically, animal-based assays have been used for this purpose. However, there is now an increasing demand to replace these animal-based assessment methods with computer-based alternatives. Despite previous attempts to develop computer-based models, these models have proven to be unreliable and inaccurate, leading to a decrease in interest. Therefore, there is a pressing need to develop new computer-based models for toxicity assessment. Here, the introduction of deep learning models, particularly transformer architecture, has the potential to revolutionise the field. Deep neural networks have demonstrated the ability to handle complex and high-dimensional problems, surpassing older modelling techniques. Moreover, as the transformer has shown promise in handling chemical structure information, there is growing interest in its usage in the field of environmental toxicity assessment. The aim of this project was hence to explore the potential of transformer-based deep neural network models for the purpose of toxicity assessment. For this project, a subset of rat and mice in vivo toxicity assay data associated with EC50 and LOEC measurements, as well as different administration routes, were utilised. Here, three sets of data were analysed, each distinguished by the hazards: acute toxicity, carcinogenicity, or reproductive toxicity. The first type of model, the single-DNN model, was created for each data set separately. Subsequently, these models were expanded to the multiple-DNN model, able to handle all three data sets simultaneously. For all models, a pre-trained RoBERTa transformer was utilised to interpret canonicalised SMILES representation of chemical structures, with the performance then evaluated through repeated 10-fold crossvalidation. Principal Component Analysis demonstrated that the transformer could identify patterns in chemical structures related to toxicity. Moreover, the study found that the single-DNN model outperformed the multiple-DNN model in all trials, likely due to the latter’s increased complexity. All models exhibited leniency towards chemicals with low measured concentrations, and to mitigate this problem, a more stringent loss for lower concentrations was suggested. Overall, this project demonstrated the potential and effectiveness of transformer-based computer models for toxicity assessment, showcasing the versatility of this technology for addressing a broad range of toxic hazards