Deep learning-based framework for toxicological assessment in early-stage drug discovery

Drug toxicity remains a major challenge in pharmaceutical development, contributing significantly to drug candidate failures and market withdrawals while raising growing concerns about pharmaceutical contamination of ecosystems. Following green chemistry principles, particularly the design of inherently safer chemicals and prevention of waste through early prediction of toxicity, this thesis investigates two critical aspects of drug safety: environmental Persistence, Bioaccumulation, and Toxicity (PBT), and Drug-Induced Liver Injury (DILI). In the first case study, we developed a message-passing neural network model for PBT prediction, that demonstrated robust performance even when tested on structurally diverse compounds. The model successfully identified potential PBT chemicals among pharmaceuticals and revealed specific structural features associated with PBT properties. These findings provide valuable guidance for designing environmentally safer drugs during early development stages, supporting the implementation of green chemistry principles in pharmaceutical research. The second case study advanced our understanding of DILI prediction through the development of an interpretable deep learning model. This work explored novel approaches in deep learning to understand and predict hepatotoxicity risk through analysis of molecular structures. Our research introduces computational approaches that can enhance the early assessment of drug toxicity, supporting the development of safer and more sustainable pharmaceuticals. These methodologies represent a significant step toward integrating green chemistry principles into drug discovery, potentially reducing both environmental impact and development costs through early identification of toxicity risks.