Comparison of Different Models for Toxicity Prediction using AI

Abstract

Artificial intelligence (AI) has truly transformed the way we predict toxicity, offering a scalable, accurate, and scientifically sound alternative to traditional experimental methods. Early studies highlighted the effectiveness of rule-based and QSAR models, but recent advancements in machine learning (ML), deep learning (DL), and large language models (LLMs) have significantly improved predictive accuracy for various toxicological endpoints. Today, these applications span drug discovery, nanotoxicology, environmental safety, and human organ toxicity, integrating molecular descriptors, omics, and proteomics data to enhance both interpretability and accuracy. Innovative architectures like multitask learning, TabNet, TabTransformer, and explainable AI platforms such as TIRESIA and ToxNet are paving the way for more transparent predictions of acute, chronic, and developmental toxicity. Additionally, ensemble and hybrid models that combine QSAR, deep neural networks, and spatio-temporal representations are improving generalizability across different chemical spaces. Despite these advancements, challenges remain, including data heterogeneity, interpretability, regulatory acceptance, and the ability to apply findings in real-world scenarios. Future directions are focusing on mechanistic modeling, multi-omics integration, and the development of AI-based decision-support systems to accelerate safe drug development, chemical risk assessment, and environmental monitoring. Collectively, these innovations position AI as a cornerstone technology in predictive toxicology, bridging the gap between experimental and computational approaches for safer and more sustainable innovations.

Ähnliche Arbeiten