Re-imagining Toxicology in the age of AI

Abstract

The science of toxicology forms the cornerstone of modern medicine and healthcare. In toxicology, computational methods from omics-level data analysis to multi-level (cell, tissue, organ, and whole body) modeling and simulation techniques have been increasingly applied. New approaches in Analytical AI based on deep neural networks and representation learning hold further promise to enhance the field, providing the ability to query multiple modalities of data (quantitative, imaging, text processing) and higher predictive quality of results.

However, gaps exist in the adoption of such cutting-edge methods in modern machine learning by practicing toxicologists, where descriptor-based quantitative structure–activity relationship (QSAR) methods are still widely used in this field. With the exponential progress witnessed in the field of AI, and more recently the rise of Generative AI and large language models (LLMs), it is critical to Review, Re-assess, and Re-imagine the role of AI in toxicology.

In this talk, we highlight, through different case studies, the various applications of traditional and modern machine learning techniques in toxicology - ranging from large-scale risk assessment of specific compounds to the prediction of the impact of combination trial ingredients in cosmetology. Modern methods are typically data-hungry and require training on many examples. We also explore how emerging methods in generative AI can be developed to fine-tune existing large, pre-trained models on a large corpus of multi-modal data and applied to focused problems in toxicology.

Novel modeling architectures which leverage existing computational methods, together with modern neural networks, representational learning methods, and large language models hold significant promise to re-imagine the science of toxicology - from specific drug ADMEtox properties to whole body risk assessment for human health and ecotoxicity.