ADMET Prediction

Predicts a molecule's ability to cross biological membranes, primarily the intestinal epithelium. Key endpoints include Caco-2 permeability and PAMPA, which model passive diffusion. Models often integrate LogP and Polar Surface Area (PSA) to forecast oral absorption. A critical rule-of-thumb benchmark is Lipinski's Rule of Five, which flags compounds with poor absorption potential based on molecular weight, lipophilicity, and hydrogen bonding.

Estimates how a drug disperses throughout the body after entering the bloodstream. A primary endpoint is Plasma Protein Binding (PPB) , which predicts the fraction of drug sequestered by serum albumin and alpha-1-acid glycoprotein. Only the unbound fraction is pharmacologically active. Models also predict Blood-Brain Barrier (BBB) penetration to assess CNS exposure and the Volume of Distribution (Vd) to understand tissue affinity.

Forecasts the enzymatic biotransformation of a drug, primarily in the liver. Models identify Site of Metabolism (SOM) on the molecule and predict susceptibility to major Cytochrome P450 (CYP) isoforms like 3A4, 2D6, and 2C9. A critical safety endpoint is CYP450 Inhibition, where a drug blocks the metabolism of other co-administered drugs. Time-Dependent Inhibition (TDI) is a more complex, mechanism-based form of inhibition that leads to irreversible enzyme inactivation.

Predicts the rate at which a drug is eliminated from the body, a parameter that determines dosing frequency. Models estimate Intrinsic Clearance (CLint) in hepatocytes and predict whether excretion is renal or biliary. Transporters like P-glycoprotein (P-gp) can actively pump drugs out of cells, limiting absorption and facilitating excretion. Predicting whether a compound is a P-gp substrate is vital for understanding drug-drug interactions and bioavailability.

Evaluates the potential for a compound to cause harm. This is the most complex category, encompassing multiple distinct endpoints. hERG cardiotoxicity prediction models the blockade of a potassium channel linked to fatal arrhythmia. AMES mutagenicity predicts DNA damage potential. Drug-Induced Liver Injury (DILI) is a leading cause of drug failure, predicted by integrating multiple in vitro and chemical structure alerts. PAINS filters are used to weed out compounds that interfere with assay readouts rather than having true biological activity.

The foundational molecular properties that drive ADMET behavior. LogP (lipophilicity) and LogD (pH-dependent lipophilicity) govern membrane permeability and solubility. Aqueous Solubility is a prerequisite for absorption and bioavailability. pKa predicts ionization state at physiological pH, which directly impacts permeability and protein binding. These descriptors are not endpoints themselves but are the primary input features for predictive models and are often optimized directly in multi-parameter optimization (MPO) workflows.

Two critical concepts for deploying ADMET models with confidence. The Applicability Domain (AD) defines the chemical space where a model's predictions are reliable, typically bounded by the structural and physicochemical similarity to its training data. Uncertainty Quantification (UQ) assigns a confidence interval to each prediction, distinguishing between:

• Aleatoric uncertainty: Inherent noise in the data.
• Epistemic uncertainty: Model ignorance due to lack of data, reducible with more training examples. Methods like conformal prediction provide rigorous, finite-sample coverage guarantees, outputting a prediction set rather than a single point estimate.