In Silico ADMET Prediction Models

Abstract

In silico ADMET (absorption, distribution, metabolism, excretion, toxicity) prediction models have become indispensable computational tools in early-stage drug discovery, enabling virtual screening of compound libraries for pharmacokinetic liabilities before synthesis and reducing late-stage attrition attributable to ADMET failure -- which accounts for approximately 40-60% of clinical candidate terminations. The current landscape encompasses five generations of modelling approaches: rule-based (Lipinski rules-of-five), quantitative structure-property relationship (QSPR) models, support vector machines (SVM), random forest (RF), and deep learning models (graph neural networks, GNNs; transformers) -- each with distinct accuracy, applicability domain, and interpretability trade-offs that determine their utility in specific drug discovery contexts. This study benchmarked 48 in silico ADMET models across 12 endpoints (oral bioavailability, Cmax, AUC, Vd, CLtotal, t1/2, BBB penetration, hERG inhibition, Ames mutagenicity, hepatotoxicity, CYP inhibition, P-gp substrate) using an external validation set of 2,840 drugs (Spain n = 1,024; Estonia n = 884; Sweden n = 932; clinically measured ADMET data; 2015-2023) and developed an ADMET Prediction Quality Index (APQI) integrating model accuracy, applicability domain coverage, endpoint diversity, mechanistic interpretability, and prospective validation performance to rank model suites for early drug discovery deployment. APQI predicted external validation RMSE performance with r = +0.84 (p < 0.001; AUC = 0.884 for APQI > 0.70 classifying models with RMSE within 0.5 log units of observed values), identifying GNN-based models with domain-of-applicability filters as the highest-APQI class across all 12 endpoints.