11.4 ADMET prediction

ADMET prediction estimates how a drug candidate will behave in the body and whether it's likely to cause harm. In computational medicinal chemistry, these predictions let you filter out problematic compounds early, before they fail expensively in clinical trials. This topic covers each component of ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity), the methods used to predict them, and how to integrate the results into drug design decisions.

Importance of ADMET prediction

ADMET stands for Absorption, Distribution, Metabolism, Excretion, and Toxicity. Together, these properties define a drug's pharmacokinetic and safety profile. Predicting them early in discovery lets you focus resources on compounds that are most likely to succeed.

Poor ADMET properties historically account for a large share of clinical trial failures. If a compound can't be absorbed orally, gets trapped by plasma proteins, is rapidly metabolized, or causes liver damage, no amount of target potency will save it. Early ADMET screening shifts these failures to the cheaper, earlier stages of the pipeline.

Role in drug discovery

ADMET prediction is applied during lead identification and optimization to rank and prioritize compounds. You evaluate candidates for oral absorption potential, distribution to the target tissue, metabolic stability, and safe elimination.

The predictions also guide structural modifications. If a lead compound has poor metabolic stability, for instance, a medicinal chemist can modify metabolically labile groups while monitoring the impact on other ADMET parameters.

Impact on clinical success

Compounds with favorable ADMET profiles have a substantially higher probability of demonstrating efficacy and safety in clinical studies. By identifying liabilities like rapid clearance, poor bioavailability, or hERG channel inhibition (a cardiotoxicity flag) before candidates enter the clinic, you reduce costly late-stage attrition and shorten development timelines.

Absorption prediction

Absorption prediction estimates how much of a drug reaches the systemic circulation from its site of administration. For oral drugs, this means crossing the intestinal epithelium and surviving first-pass metabolism in the gut wall and liver.

Physicochemical properties

Several physicochemical properties govern absorption:

• Molecular weight — Larger molecules generally have more difficulty crossing membranes. Lipinski's Rule of Five sets a cutoff at 500 Da.
• Lipophilicity (LogP) — Moderate lipophilicity favors membrane permeation, but too much increases metabolic liability and reduces aqueous solubility.
• Hydrogen bond donors/acceptors — Excessive H-bond donors (>5) or acceptors (>10) reduce passive permeability.
• Polar surface area (PSA) — A PSA above ~140 Ų typically correlates with poor oral absorption.

Computational quantitative structure-property relationship (QSPR) models predict these properties directly from chemical structure, allowing rapid screening of virtual compound libraries.

Permeability models

Permeability describes a drug's ability to cross biological membranes. Both experimental and computational approaches are used:

• In vitro assays: Caco-2 cells (human colon carcinoma) model intestinal permeability. MDCK cells offer faster throughput. PAMPA (Parallel Artificial Membrane Permeability Assay) uses a lipid-coated artificial membrane for passive permeability screening.
• In silico models: Machine learning classifiers trained on Caco-2 or PAMPA data can predict permeability from molecular descriptors. Structure-based approaches model interactions with lipid bilayers directly.

Bioavailability estimation

Bioavailability (F) is the fraction of an administered dose that reaches systemic circulation unchanged. For an oral drug:

$F = f_{abs} \times f_{gut} \times f_{hepatic}$

where $f_{abs}$ is the fraction absorbed, $f_{gut}$ is the fraction surviving gut-wall metabolism, and $f_{hepatic}$ is the fraction surviving hepatic first-pass metabolism.

Physiologically based pharmacokinetic (PBPK) models integrate solubility, permeability, and metabolic data to simulate bioavailability across different conditions (e.g., fed vs. fasted state).

Distribution prediction

Once absorbed, a drug distributes throughout the body. Distribution prediction estimates where the drug goes, how much reaches the target tissue, and how much remains free (unbound) to exert its effect.

Plasma protein binding

Drugs reversibly bind to plasma proteins, primarily albumin (for acidic and neutral drugs) and alpha-1-acid glycoprotein (AAG) (for basic drugs). Only the unbound fraction is pharmacologically active and available for distribution and elimination.

• In vitro methods: Equilibrium dialysis and ultrafiltration measure the bound/unbound ratio directly.
• In silico methods: QSAR and machine learning models predict binding from molecular descriptors like LogP, charge state, and aromatic ring count.

A drug with >99% protein binding may have very little free drug at the target, which can complicate dosing.

Volume of distribution

The volume of distribution ($V_d$) relates the total amount of drug in the body to its plasma concentration:

$V_d = \frac{\text{Amount of drug in body}}{\text{Plasma concentration}}$

A high $V_d$ (e.g., >1 L/kg) indicates extensive tissue distribution, while a low $V_d$ suggests the drug stays mostly in the plasma. $V_d$ is influenced by tissue binding affinity, lipophilicity, and blood flow. PBPK models and allometric scaling from preclinical species data are the primary prediction tools.

Blood-brain barrier penetration

The blood-brain barrier (BBB) is formed by tight junctions between brain endothelial cells and restricts most molecules from entering the CNS. For CNS-targeted drugs, adequate BBB penetration is essential. For peripherally acting drugs, low BBB penetration is desirable to avoid CNS side effects.

Predictive approaches include:

• In vitro: Cell-based BBB models (e.g., hCMEC/D3 cells) and PAMPA-BBB assays
• In silico: QSAR models using descriptors like PSA, LogP, and H-bond count. Generally, molecules with PSA <90 Ų and moderate lipophilicity cross the BBB more readily.

Metabolism prediction

Metabolism prediction identifies how a drug is biotransformed, which enzymes are involved, and how quickly the drug is cleared. This directly affects drug exposure, half-life, and the risk of drug-drug interactions.

Cytochrome P450 interactions

The cytochrome P450 (CYP) enzyme family handles the metabolism of roughly 75% of marketed drugs. The most important isoforms are CYP3A4, CYP2D6, CYP2C9, CYP2C19, and CYP1A2.

Predictions focus on three questions:

1. Is the compound a CYP substrate? This determines which enzymes clear the drug.
2. Is it a CYP inhibitor? Inhibiting a CYP enzyme can raise plasma levels of co-administered drugs, causing toxicity.
3. Is it a CYP inducer? Induction can lower plasma levels of co-administered drugs, reducing their efficacy.

In vitro assays use liver microsomes or hepatocytes. In silico approaches include docking into CYP crystal structures and ligand-based machine learning models trained on substrate/inhibitor datasets.

Phase I vs Phase II reactions

Drug metabolism occurs in two main phases:

• Phase I (functionalization): Oxidation, reduction, or hydrolysis reactions, often catalyzed by CYP enzymes. These introduce or expose a functional group, sometimes producing active or toxic metabolites.
• Phase II (conjugation): Attachment of polar endogenous molecules (glucuronic acid, sulfate, glutathione, acetyl groups) to the drug or its Phase I metabolites. This increases water solubility and facilitates excretion.

Some drugs undergo Phase II conjugation directly without a Phase I step. Predicting the balance between these pathways matters because Phase I metabolites can sometimes be more toxic than the parent drug.

Metabolic stability assessment

Metabolic stability measures how resistant a drug is to enzymatic breakdown. It's a key determinant of half-life and oral bioavailability.

• In vitro: Incubate the compound with liver microsomes or hepatocytes and measure the rate of disappearance over time. The intrinsic clearance ($CL_{int}$) is calculated from the elimination rate.
• In silico: Quantitative structure-metabolism relationship (QSMR) models and machine learning predict metabolic stability from molecular descriptors. These models can also identify specific sites of metabolism (soft spots) on the molecule.

Excretion prediction

Excretion prediction estimates how a drug and its metabolites are eliminated from the body. The two major routes are renal (kidneys) and biliary (liver to bile to feces).

Renal clearance estimation

Renal clearance involves three processes in the nephron:

1. Glomerular filtration — passive filtration of unbound drug from blood into the tubular fluid
2. Active tubular secretion — transporter-mediated secretion of drug into the tubule (e.g., via OAT, OCT transporters)
3. Tubular reabsorption — passive reabsorption of lipophilic drug back into the blood

Net renal clearance depends on the balance of these three processes. In silico prediction uses QSAR models incorporating molecular charge, size, lipophilicity, and transporter affinity, along with PBPK models.

Biliary excretion models

Biliary excretion is mediated by hepatic efflux transporters (e.g., P-gp, MRP2, BCRP) that pump drugs or conjugated metabolites into bile. This route is particularly relevant for larger, more polar molecules.

• In vitro: Sandwich-cultured hepatocytes maintain functional bile canaliculi and can measure biliary excretion. Vesicle-based transporter assays assess individual transporter contributions.
• In silico: QSAR and PBPK models predict biliary excretion based on molecular weight, polarity, and transporter substrate profiles.

Half-life prediction

Half-life ($t_{1/2}$) is the time for plasma drug concentration to decrease by 50%. It determines dosing frequency and the time to reach steady state.

$t_{1/2} = \frac{0.693 \times V_d}{CL}$

where $CL$ is total clearance. Half-life can be estimated by combining predicted $V_d$ and $CL$ values, or by scaling in vitro metabolic stability data using IVIVE approaches. PBPK models provide the most mechanistic predictions by integrating all ADMET processes simultaneously.

Toxicity prediction

Toxicity prediction identifies potential adverse effects before a compound reaches animal studies or clinical trials. Early toxicity screening is one of the most cost-effective steps in drug discovery.

In vitro toxicity assays

Common in vitro toxicity assays include:

• Cytotoxicity assays (MTT, LDH release) — measure general cell viability and membrane integrity
• Mitochondrial toxicity assays (glucose/galactose assay) — cells grown in galactose rely on mitochondrial oxidative phosphorylation, making them sensitive to mitochondrial toxins
• hERG channel assays — assess inhibition of the cardiac potassium channel, a major predictor of QT prolongation and arrhythmia risk
• High-content imaging — multi-parameter screening that simultaneously measures nuclear morphology, mitochondrial membrane potential, calcium levels, and cell membrane permeability

These assays generate rapid, cost-effective data that feeds into structure-activity relationship (SAR) studies.

In silico toxicity models

Computational toxicity prediction uses several approaches:

• QSAR models — correlate chemical structure with toxicity endpoints using statistical methods
• Rule-based expert systems (e.g., Derek Nexus, Toxtree) — flag known structural alerts (toxicophores) associated with specific toxicities
• Machine learning — deep learning and random forest models trained on large toxicity databases (e.g., ToxCast, Tox21) can predict mutagenicity, hepatotoxicity, and other endpoints

These models are most reliable for well-represented chemical classes and toxicity mechanisms. Novel scaffolds or idiosyncratic toxicities remain harder to predict.

Genotoxicity vs organ toxicity

• Genotoxicity is the ability of a substance to damage DNA. It's a critical concern because DNA damage can lead to mutations and carcinogenicity. The Ames test (bacterial reverse mutation assay) is the standard in vitro screen, and in silico models predict mutagenic structural alerts with reasonable accuracy.
• Organ toxicity refers to damage to specific organs. Hepatotoxicity (liver) is the most common cause of post-market drug withdrawal. Cardiotoxicity (heart), nephrotoxicity (kidney), and neurotoxicity (nervous system) are also major concerns.

Predicting both types is essential for a complete safety assessment. Structural alerts for genotoxicity (e.g., aromatic amines, nitro groups, epoxides) are relatively well characterized, while organ toxicity prediction often requires more complex, mechanism-based models.

ADMET prediction methods

Three broad categories of methods are used for ADMET prediction, often in combination.

In vitro techniques

In vitro assays provide direct experimental measurements of specific ADMET endpoints:

• Caco-2 and PAMPA for permeability
• Liver microsomes and hepatocytes for metabolic stability and CYP interactions
• Plasma protein binding assays (equilibrium dialysis)
• Cell-based toxicity screens

These assays are higher throughput than animal studies but still require synthesized compound, so they're typically applied to a narrower set of candidates than in silico screens.

In silico approaches

Computational methods predict ADMET properties from molecular structure alone, enabling screening of virtual libraries containing millions of compounds. The main approaches are:

• QSPR/QSAR models — statistical models relating molecular descriptors to ADMET endpoints
• PBPK models — mechanistic models simulating drug transit through physiological compartments
• Machine learning — algorithms (random forests, neural networks, gradient boosting) trained on experimental ADMET datasets
• Molecular docking — predicts binding to specific proteins (e.g., CYP enzymes, transporters)

The trade-off is speed vs. accuracy. In silico methods are fast and cheap but less reliable than well-designed in vitro assays.

In vitro-in vivo extrapolation

In vitro-in vivo extrapolation (IVIVE) translates in vitro measurements into predictions of in vivo pharmacokinetic parameters. For example, to predict hepatic clearance from microsomal data:

1. Measure intrinsic clearance ($CL_{int}$) in liver microsomes
2. Scale by microsomal protein content per gram of liver and liver weight per kg of body weight
3. Apply a liver model (e.g., the well-stirred model or parallel tube model) to account for hepatic blood flow and protein binding

$CL_{hepatic} = \frac{Q_H \times f_u \times CL_{int,scaled}}{Q_H + f_u \times CL_{int,scaled}}$

where $Q_H$ is hepatic blood flow and $f_u$ is the fraction unbound in plasma. IVIVE bridges the gap between in vitro data and clinically relevant predictions.

Integration of ADMET data

Optimizing a single ADMET property in isolation is rarely sufficient. Drug design requires balancing multiple properties simultaneously.

Multi-parameter optimization

Multi-parameter optimization (MPO) assigns scores or weights to each ADMET property and seeks compounds that perform well across all of them. A common approach is the CNS MPO score, which combines six physicochemical properties (LogP, LogD, PSA, HBD, MW, pKa) into a single desirability score for CNS drug candidates.

Computational tools like desirability functions and Pareto optimization algorithms help identify compounds in the "sweet spot" where no single property is unacceptable.

ADMET property trade-offs

ADMET properties frequently conflict with each other. Some common trade-offs:

• Increasing lipophilicity improves membrane permeability but worsens metabolic stability, solubility, and toxicity risk
• Reducing molecular weight improves oral absorption but may reduce target binding affinity
• Blocking metabolic soft spots can improve half-life but may shift metabolism to produce toxic metabolites

Recognizing these trade-offs is central to medicinal chemistry decision-making. There's rarely a "perfect" compound; the goal is an acceptable balance.

Decision-making in drug design

ADMET data feeds into go/no-go decisions at each stage of drug discovery. Medicinal chemists, DMPK (Drug Metabolism and Pharmacokinetics) scientists, and toxicologists collaborate to:

• Prioritize compounds for synthesis and testing
• Design focused libraries around scaffolds with favorable ADMET profiles
• Identify structural modifications that address specific liabilities
• Select appropriate preclinical species for further evaluation

Effective integration requires visualizing ADMET data across multiple dimensions (radar plots, traffic-light scoring systems) so that the team can quickly compare candidates.

Challenges in ADMET prediction

Species differences

ADMET properties can differ substantially between species. For example, CYP enzyme expression and substrate specificity vary between rodents and humans, and some metabolic pathways present in rodents are absent in humans (and vice versa). Allometric scaling and human-relevant PBPK models help bridge these gaps, but extrapolation always introduces uncertainty.

Inter-individual variability

Genetic polymorphisms in drug-metabolizing enzymes and transporters create significant variability in drug response across patient populations. CYP2D6, for instance, has poor metabolizer, intermediate, extensive, and ultra-rapid metabolizer phenotypes that can cause 10-fold or greater differences in drug exposure. Incorporating virtual populations and sensitivity analyses into ADMET predictions helps assess the range of expected responses.

Prediction accuracy limitations

Current ADMET models are constrained by the quality and diversity of their training data. Models trained primarily on drug-like molecules may perform poorly on novel chemical scaffolds or new modality therapeutics (e.g., PROTACs, macrocycles). Complex, multi-factorial toxicities like idiosyncratic hepatotoxicity remain particularly difficult to predict. Continuous model validation against new experimental data and expansion of training sets are necessary to improve reliability.

Future perspectives

Emerging technologies

Organ-on-a-chip systems and 3D cell culture models (spheroids, organoids) provide more physiologically relevant environments for ADMET testing. These platforms can recapitulate tissue architecture, cell-cell interactions, and dynamic flow conditions that monolayer cultures cannot. Coupling these advanced in vitro systems with computational modeling is expected to improve predictive accuracy and reduce reliance on animal studies.

Refinement of prediction models

Deep learning architectures, including graph neural networks that operate directly on molecular graphs, are showing improved performance on ADMET prediction benchmarks. Multi-task learning, where a single model predicts multiple ADMET endpoints simultaneously, can leverage shared information across related properties. Integration of multi-omics data (genomics, proteomics, metabolomics) into prediction frameworks promises a more mechanistic understanding of drug disposition.

Integration with systems biology

Systems pharmacology models combine ADMET predictions with target engagement, signaling pathway dynamics, and disease biology to simulate drug efficacy and safety at the whole-organism level. This holistic approach can capture feedback loops and compensatory mechanisms that single-endpoint predictions miss, ultimately supporting more informed decisions about which compounds to advance into clinical development.