ADMET Prediction

The process by which a drug enters systemic circulation from its site of administration. Key determinants include passive permeability across biological membranes, active transport via influx and efflux transporters (e.g., P-glycoprotein), and solubility in gastrointestinal fluids.

• Caco-2 permeability: Gold-standard in vitro assay for intestinal absorption
• PAMPA: Artificial membrane assay isolating passive transcellular permeability
• Rule of Five: Lipinski's heuristic flagging compounds with poor oral absorption based on molecular weight (>500), logP (>5), H-bond donors (>5), and H-bond acceptors (>10)
• Fraction absorbed (Fa%): The percentage of an oral dose that traverses the gut wall

The reversible transfer of a drug from the bloodstream into tissues and organs. Governed by plasma protein binding (primarily to albumin and alpha-1 acid glycoprotein), tissue permeability, and volume of distribution (Vd).

• Blood-brain barrier (BBB) penetration: Critical for CNS drugs; predicted via logBB or brain/plasma ratio
• Fraction unbound (fu): The pharmacologically active portion not bound to plasma proteins
• Volume of distribution: A theoretical volume relating total drug in the body to plasma concentration; high Vd indicates extensive tissue binding
• Tissue-to-plasma partition coefficients (Kp): Used in physiologically-based pharmacokinetic (PBPK) models

The enzymatic biotransformation of a drug, primarily in the liver, into more polar metabolites for excretion. The cytochrome P450 (CYP) superfamily—particularly CYP3A4, CYP2D6, and CYP2C9—mediates the majority of Phase I oxidative reactions.

• Phase I reactions: Oxidation, reduction, and hydrolysis introducing or unmasking functional groups
• Phase II reactions: Conjugation with glucuronic acid, sulfate, or glutathione to increase water solubility
• Intrinsic clearance (CLint): The maximal enzymatic capacity of hepatocytes to metabolize a drug in the absence of blood flow limitations
• Metabolic stability: Often expressed as half-life (t1/2) in human liver microsomes or hepatocytes
• CYP inhibition/induction: A drug's potential to perpetrate drug-drug interactions by inhibiting or upregulating CYP enzymes

The irreversible removal of a drug and its metabolites from the body, primarily via renal excretion through glomerular filtration, tubular secretion, and reabsorption. Hepatobiliary excretion into bile and subsequent fecal elimination is the second major route.

• Total clearance (CL): The volume of plasma completely cleared of drug per unit time; the sum of renal and non-renal clearance
• Half-life (t1/2): The time required for plasma concentration to decrease by 50%; determines dosing frequency
• Renal clearance ratio: Relative to glomerular filtration rate, indicating net secretion or reabsorption
• Transporter-mediated excretion: Organic anion transporters (OATs) and organic cation transporters (OCTs) actively secrete drugs into urine

The prediction of adverse effects ranging from acute organ damage to chronic carcinogenicity. In silico toxicology employs structural alerts, quantitative structure-toxicity relationship (QSTR) models, and knowledge-based expert systems to flag hazardous substructures.

• hERG channel inhibition: Blockade of the cardiac potassium channel causing QT prolongation and risk of torsades de pointes arrhythmia—a leading cause of drug withdrawals
• Genotoxicity: DNA-reactive mutagenicity assessed via Ames test prediction (bacterial reverse mutation assay)
• Hepatotoxicity (DILI): Drug-induced liver injury, the most frequent cause of acute liver failure and market withdrawal
• Phototoxicity: Light-activated toxicity relevant for dermal and ophthalmic drugs
• Cytotoxicity panels: Multiparametric in vitro assays measuring mitochondrial integrity, membrane permeability, and cell viability

Machine learning and deep learning approaches that predict ADMET endpoints directly from molecular structure, bypassing costly and time-consuming in vitro assays. Modern architectures include graph neural networks (GNNs) operating on molecular graphs, transformer-based models trained on SMILES strings, and multitask deep learning that jointly predicts multiple endpoints.

• Random forest and XGBoost: Established baselines using molecular fingerprints and physicochemical descriptors
• Message-passing neural networks: Learn task-specific atomic representations directly from molecular topology
• Transfer learning: Pretraining on large unlabeled chemical corpora before fine-tuning on sparse ADMET data
• Applicability domain: The chemical space within which a model's predictions are reliable; critical for regulatory acceptance
• Conformal prediction: Produces prediction intervals with guaranteed coverage, quantifying uncertainty for each compound

A foundational heuristic for predicting oral bioavailability based on four physicochemical criteria:

• Molecular weight ≤ 500 Da
• Calculated logP ≤ 5
• Hydrogen bond donors ≤ 5
• Hydrogen bond acceptors ≤ 10

Compounds violating two or more rules are likely to exhibit poor absorption or permeation. While not an ADMET model itself, this rule set established the quantitative structure-property relationship paradigm that underpins modern in silico ADMET screening.

Computational models that forecast a compound's interaction with cytochrome P450 enzymes, the primary metabolic pathway for xenobiotics. Key prediction tasks include:

• Substrate identification: Will CYP3A4, 2D6, or 2C9 metabolize this compound?
• Inhibition liability: Does the compound act as a competitive or mechanism-based inhibitor?
• Site of metabolism: Which atomic position undergoes oxidative transformation?

Isoform-specific models use ligand-based QSAR, structure-based docking into CYP crystal structures, and deep learning on metabolic reaction databases.

In silico prediction of a compound's affinity for the human Ether-à-go-go-Related Gene (hERG) potassium ion channel. hERG blockade prolongs the cardiac QT interval, causing potentially fatal Torsades de Pointes arrhythmia.

Modern prediction approaches include:

• Pharmacophore models capturing the canonical positively charged nitrogen and surrounding hydrophobic features
• 3D-QSAR trained on patch-clamp electrophysiology IC50 data
• Graph neural networks learning directly from molecular topology

Regulatory agencies now expect hERG liability assessment early in lead optimization, making this a critical ADMET endpoint.

A mechanistic mathematical framework that integrates ADMET parameters into a whole-body compartmental model to simulate drug concentration-time profiles in plasma and tissues.

PBPK models incorporate:

• Absorption: Solubility, permeability, and intestinal transit
• Distribution: Tissue-to-plasma partition coefficients and protein binding
• Metabolism: Intrinsic clearance from hepatocyte assays
• Excretion: Renal clearance and biliary elimination

These models enable in vitro-to-in vivo extrapolation (IVIVE) and are increasingly accepted by the FDA for drug-drug interaction risk assessment and pediatric dose selection.

Computational models that predict the outcome of the Salmonella typhimurium reverse mutation assay, the standard in vitro test for genotoxic potential. A positive Ames result often halts development due to carcinogenicity risk.

Prediction methods leverage:

• Structural alerts: Expert-derived toxicophoric substructures (e.g., aromatic amines, nitro groups)
• Statistical models: Random forests and support vector machines trained on public mutagenicity databases
• Deep learning: Graph convolutional networks achieving >85% accuracy on benchmark sets

Regulatory frameworks including ICH M7 explicitly permit the use of in silico Ames predictions for impurity assessment.

Prediction of a compound's ability to cross the blood-brain barrier (BBB), a critical ADMET parameter distinguishing CNS-targeted therapeutics from peripherally restricted drugs.

Key molecular determinants include:

• Lipophilicity (logP/logD): Primary driver of passive diffusion
• Polar surface area (PSA): Values < 90 Ų favor CNS penetration
• Molecular weight: Lower MW correlates with higher BBB permeation
• P-glycoprotein (P-gp) efflux: Active transport out of the CNS

Binary classification models (BBB+ vs BBB-) and regression models predicting the logBB ratio (brain-to-blood concentration) guide CNS drug design.