In silico ADMET: Supporting new drug discovery

Abstract

ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties are major determinants of drug success, accounting for nearly half of late-stage clinical attrition. In silico ADMET prediction tools offer a rapid, cost-effective, and animal-free strategy for early-stage screening, enabling the evaluation of hundreds to thousands of compounds prior to synthesis or in vivo testing. This review summarizes current in silico ADMET approaches, ranging from rule-based filters and QSAR models to advanced artificial intelligence (AI) and machine-learning platforms, and compares widely used tools such as SwissADME, pkCSM, and ADMETlab 3.0. AI-based ADMET models provide a clear advantage by supporting multi-endpoint prediction, high-throughput screening, and improved prioritization of lead compounds, thereby reducing experimental burden and development timelines. However, their predictive performance remains constrained by training-data bias, limited applicability domains, and reduced interpretability, necessitating experimental validation and cautious regulatory use. Overall, in silico ADMET prediction represents a transformative yet complementary component of modern drug discovery pipelines rather than a standalone replacement for experimental assessment.