Deep Learning for In Silico ADMET Prediction

Deep learning (DL) is transforming in silico ADMET (absorption, distribution, metabolism, elimination, and toxicity) prediction, overcoming key limitations of traditional computational models. This chapter examines the evolution of ADMET modeling, emphasizing how DL techniques such as deep neural networks (DNNs), convolutional neural networks (CNNs), and graph neural networks (GNNs) enhance predictive accuracy by identifying complex molecular patterns. Compared to machine learning (ML) approaches like random forests (RF) and support vector machines (SVMs), DL excels in processing high-dimensional data and capturing nonlinear relationships. Advances in multitask learning, transfer learning, and molecular representations, including equivariant GNNs, have further improved model performance. Despite these advancements, challenges persist in improving in vitro–in vivo correlations, managing data heterogeneity, and enhancing interpretability. Explainable AI techniques, such as Shapley additive explanations and attention-based mechanisms, are increasing model transparency. Additionally, hybrid modeling approaches and pretraining strategies are enhancing generalization while reducing reliance on extensive datasets. In summary, this chapter outlines recent advancements, discusses limitations, and suggests future directions, ultimately enabling researchers to prioritize safer drug candidates earlier in the discovery process and accelerate regulatory approval.