What Are Pharmacokinetic Models and How Do They Work?

Pharmacokinetic models are mathematical descriptions of how a drug moves through the body over time, from absorption into the bloodstream to distribution into tissues and eventual elimination. They link a dosing regimen to the concentration of drug a patient actually experiences, and when extended to include drug effects, they connect dose to concentration to therapeutic outcome.¹ The idea sounds abstract, but these models influence nearly every drug you have ever taken, shaping how doses were chosen in clinical trials, how the pill was formulated, and why your doctor might adjust the amount based on your kidney function or body weight.

What a Pharmacokinetic Model Actually Does

Imagine swallowing a tablet. The drug dissolves in your gut, enters the bloodstream, spreads into tissues, gets broken down (mostly by the liver), and is excreted (mostly by the kidneys). At every stage, the amount of drug present is changing. A pharmacokinetic model captures that change with equations so you can predict what the drug concentration in your blood will look like at any moment after a given dose.

The simplest version treats your entire body as a single, well-stirred container. Drug enters, mixes instantly, and leaves at a rate proportional to how much is there. That proportional-exit behavior is called first-order elimination: the higher the concentration, the faster drug disappears, and the drug’s half-life stays the same regardless of dose.² Most drugs follow this pattern at therapeutic doses. Some drugs, however, overwhelm their metabolic pathways and switch to zero-order elimination, where the body clears a fixed amount per hour no matter how much is circulating. Alcohol is the classic example. When elimination is zero-order, doubling the dose more than doubles the time the drug hangs around, making dosing predictions considerably trickier.²

Compartment Models and the Bateman Function

The one-compartment model is the workhorse of introductory pharmacokinetics. It assumes a drug distributes instantly throughout the body and then is eliminated at a constant fractional rate. For an oral dose, this setup produces a characteristic rise-then-fall concentration curve sometimes called the Bateman function: drug levels climb as the tablet is absorbed, peak, and then decline exponentially as the body clears the drug.³ It is a useful approximation for many drugs, though it rests on two unrealistic assumptions: that absorption rate is highest right at the start and that the drug distributes through the body instantly.³

When those assumptions are too crude, modelers add compartments. A two-compartment model splits the body into a central space (blood and highly perfused organs) and a peripheral space (muscle, fat, or other slow-to-equilibrate tissues). Drug shuttles between the two at rates that are estimated from the data. Recirculatory versions of two-compartment models go further: they separate the lungs from the rest of the body and use cardiac output as a driver of drug movement, which matters a great deal for anesthetic agents injected intravenously.⁴

There is also a model-free approach called non-compartmental analysis, or NCA. Rather than fitting an equation to a concentration curve, NCA uses simple geometry to measure the total drug exposure (the area under the concentration-time curve) directly from the raw data points.⁵ NCA is popular in early clinical trials because it requires fewer assumptions, though it cannot predict what would happen under a different dosing schedule the way a fitted compartment model can.

Physiologically Based Models

Conventional compartment models are defined by the drug data themselves: you give the drug, measure blood levels, and fit curves. Physiologically based pharmacokinetic (PBPK) models flip that approach. They start from the body’s own anatomy and physiology, building in organ volumes, blood flow rates, tissue composition, and enzyme activity, then layer in drug-specific properties like how readily the molecule crosses membranes or how quickly liver enzymes chew it up.⁶

Because the biology is baked in, PBPK models can extrapolate in ways compartment models cannot. You can change the “patient” in the simulation: swap in the organ sizes of a child, the reduced liver function of someone with cirrhosis, or even the physiology of a different species entirely.⁶ That flexibility is why PBPK modeling has become central to regulatory submissions. Both the FDA and EMA have issued guidelines encouraging its use, and pharmaceutical companies routinely submit PBPK analyses when seeking approval for dose adjustments or waivers for certain patient populations.⁷

One limitation worth noting: the more detailed the model, the more parameters it needs, and each parameter carries its own uncertainty. PBPK models can contain dozens of tissue-specific values, and when those values are poorly characterized the predictions can drift, particularly in severe disease states.

Why the Same Dose Does Not Work the Same Way in Everyone

Much of the practical value of pharmacokinetic models comes from their ability to account for the enormous variability among patients. Body weight, age, genetics, organ function, pregnancy, and even time of day all influence how quickly you absorb, distribute, and clear a drug.

Kidney disease is a straightforward example. When the kidneys are impaired, drug clearance drops and the volume of distribution may stay the same or even increase, pushing drug levels higher than expected for a given dose.⁸ Liver disease complicates things further. PBPK simulations of clozapine, for instance, predict that severe hepatic impairment can cut oral clearance by roughly 70%, leading to about a threefold increase in total drug exposure compared to healthy adults.⁹ Without modeling those changes, a standard dose could easily become a toxic one.

A large consortium evaluation of PBPK predictions across 29 drugs found that models predicted the change in drug exposure due to kidney impairment within twofold of observed values over 90% of the time. For liver impairment the success rate was lower, around 70%, with accuracy dropping in moderate and severe disease.¹⁰ The liver is simply harder to model: its functions are numerous, disease affects them unevenly, and good biomarkers of metabolic capacity are scarce.

Genetics adds another layer. The speed at which your liver enzymes process a drug can vary dramatically depending on which gene variants you carry. In children with tic disorders treated with aripiprazole, for example, both body weight and CYP2D6 genotype were found to markedly affect drug clearance, meaning two kids the same age could need very different doses.¹¹ Population pharmacokinetic models capture this kind of variability by estimating not just average parameter values but the spread around those averages and which patient characteristics (covariates) explain the spread.

Pregnancy is a particularly underappreciated source of pharmacokinetic change. Blood volume expands, kidney filtration rate rises, hormone-driven shifts alter enzyme activity, and a new organ, the placenta, enters the picture. Over two-thirds of pregnant women receive prescription drugs, yet dosing has historically been based on data from healthy male volunteers and non-pregnant women with little adjustment for these physiological upheavals.¹² PBPK models that incorporate pregnancy-specific physiology are beginning to fill that gap.

Linking Concentration to Effect

Knowing the drug level in your blood is only half the story. What clinicians and patients actually care about is the effect: pain relief, blood pressure reduction, infection clearance. Pharmacokinetic-pharmacodynamic (PK/PD) models add a second layer on top of the concentration predictions, connecting blood levels to a measured outcome.

One common challenge is that the peak effect of many drugs lags behind the peak blood concentration. Opioid pain relief, for instance, does not coincide with the moment the drug level in plasma is highest. Modelers handle this by introducing a hypothetical “effect compartment” that the drug enters and exits with a slight delay, then relating the concentration in that compartment to the observed effect using a standard response curve.¹³ For drugs that act indirectly, say by blocking synthesis of an inflammatory mediator rather than producing an immediate response, indirect-response models track the rise and fall of the mediator itself rather than just the drug.

Bayesian Dosing at the Bedside

Population models find their most immediate clinical use in a technique called Bayesian dose forecasting. The idea is to start with a population model’s average parameters, then update them with one or two blood-level measurements from the actual patient. The result is an individualized prediction of where drug concentrations are heading and, critically, what the next dose should be to keep them in the therapeutic range.¹⁴

In intensive-care settings, where kidney and liver function can change hour by hour, this approach makes a real difference. A prospective study of vancomycin dosing in patients with kidney problems found that pairing therapeutic drug monitoring with Bayesian forecasting improved the rate at which patients hit target trough concentrations while reducing kidney toxicity, and it was cost-effective to boot.¹⁵ Another study in critically ill patients given various antibiotics showed that 86% achieved target drug exposures after a single software-guided dose adjustment, and 94% reached clinical cure.¹⁶ The method works well precisely because it needs very few blood samples; sometimes a single level drawn at any convenient time is enough to meaningfully refine the prediction.

Getting From Animals to Humans

Before a new drug candidate enters its first human trial, researchers need to guess a safe starting dose. Pharmacokinetic models are indispensable here. One common technique, allometric scaling, uses the mathematical relationship between body size and metabolic rate across species to extrapolate clearance and volume from mice, rats, or monkeys to a predicted human value.¹⁷ For antibodies, several variants of this scaling exist, each applying slightly different corrections for the fact that large molecules are handled differently than small ones.¹⁸

PBPK models offer an alternative or complementary route. Because they model organs and enzymes explicitly, you can replace the rat liver with a human liver (in silico, of course) and simulate what happens. A recent study of a novel drug candidate used both allometric scaling and PBPK modeling to predict human pharmacokinetics before the first-in-human trial, then compared predictions to the actual clinical data once it was available.¹⁹ This kind of dual approach has become standard practice, helping to set first-in-human doses that are safe yet high enough to have a reasonable chance of producing a measurable effect.

When Drugs Bind Their Own Targets So Tightly It Changes Their Fate

Standard models assume that binding to a drug’s target does not measurably affect its blood levels. For most small-molecule drugs, that is a safe assumption. But some biologic drugs, particularly monoclonal antibodies and newer constructs like T-cell engagers, bind their pharmacological targets with such high affinity and in such quantity that the target itself becomes a route of elimination. This phenomenon is called target-mediated drug disposition, or TMDD.²⁰

In practical terms, TMDD means the drug’s half-life depends on how much target is available. At low doses, target-mediated clearance dominates, and the drug disappears fast. At higher doses, the targets get saturated, target-mediated clearance becomes negligible, and the drug hangs around much longer. This creates a nonlinear relationship between dose and exposure that a standard compartment model will miss entirely. Specialized TMDD models account for the binding, internalization, and turnover of the target, and they are increasingly important for predicting the pharmacokinetics of T-cell engagers, which interact with receptors on both T-cells and tumor cells simultaneously.²¹

PBPK models for large molecules face their own structural challenges. A PBPK model for an antibody, for example, may need to represent nine or more tissue compartments, each subdivided into vascular, interstitial, and intracellular spaces, with transport driven partly by convection through pores and partly by receptor-mediated recycling.²² The complexity is greater than for small molecules, but so is the payoff, since the behavior of biologics is almost impossible to predict intuitively.

Designing Long-Acting Formulations

Pharmacokinetic models are not just used to understand drugs that already exist. They also guide the design of new formulations, particularly long-acting injectables meant to release drug slowly over weeks or months. The challenge is matching the formulation’s release rate to the drug’s own clearance rate so that blood levels stay within the therapeutic window for the entire dosing interval.

PBPK simulations have been used to explore which HIV antiretrovirals could feasibly be given as monthly intramuscular injections. The models predicted that several drugs, including dolutegravir, rilpivirine, and efavirenz, could maintain adequate levels with monthly dosing, provided the formulation technology can deliver the right release kinetics.²³ Similar work in tuberculosis showed that monthly injections of delamanid and rifapentine could maintain effective drug concentrations, while other TB drugs like bedaquiline would need more frequent dosing.²⁴ These simulations save enormous amounts of time and money by narrowing the formulation options before anyone starts manufacturing batches for clinical testing.

A separate modeling framework focuses on matching a drug’s own dissolution characteristics to its therapeutic index. By plotting the release rate a formulation needs to achieve against the release rate the drug’s physical properties naturally produce, researchers can quickly identify whether a simple suspension will do, whether a controlled-release technology is necessary, or whether the drug actually dissolves too slowly and needs something to speed things up.²⁵

Models for Overdose and Poisoning

Some of the most consequential applications of pharmacokinetic models involve scenarios nobody wants to face: accidental or intentional overdose. Here, the model’s ability to simulate what happens at extreme doses, and to test the timing and effectiveness of interventions, can directly inform treatment guidelines.

Acetaminophen overdose has been modeled extensively using quantitative systems toxicology platforms that go beyond simple pharmacokinetics and track the cascade of liver injury downstream. These models can simulate both immediate-release and extended-release formulations, predict hepatic biomarker trajectories, and test scenarios involving repeated supratherapeutic dosing. Recent simulations found that the two formulations produced similar liver-injury profiles after overdose, supporting the clinical decision to manage both formulations with the same treatment protocol.²⁶ Further simulations extended these findings to adults with chronic alcohol use or depleted glutathione reserves, populations at higher risk of liver damage, and reached similar conclusions.²⁷

Ethylene glycol poisoning (antifreeze ingestion) illustrates how models can compare treatment strategies. A PBPK model incorporating hemodialysis, ethanol administration, and fomepizole as therapeutic interventions demonstrated that fomepizole, if given early enough, could be more effective than ethanol or dialysis at preventing the metabolism of ethylene glycol into its truly dangerous metabolites. Dialysis remained critical, though, once significant metabolism had already occurred or kidney damage had set in.²⁸ Models like these let toxicologists stress-test treatment algorithms on simulated patients before relying on them in emergency rooms.

Machine Learning Entering the Field

The newest frontier in pharmacokinetic modeling merges traditional mechanistic equations with machine learning. The appeal is clear: neural networks excel at finding patterns in complex, noisy data, while pharmacokinetic equations encode known physical laws. Combining them could produce models that learn from data more efficiently while still respecting the fact that, say, drug concentration cannot be negative or that elimination must be tied to clearance.

One recent approach, called distributional physics-informed neural networks, embeds pharmacokinetic differential equations directly into the neural network’s training process. This allows the network to recover population-level parameter distributions from published summary data rather than needing raw individual patient records, which are often unavailable.²⁹ Another system combined a type of recurrent neural network with pharmacokinetic constraints to predict individualized antibiotic concentrations in critically ill patients. The model accurately estimated population clearance and volume parameters while providing calibrated uncertainty bounds: about 92% of actual measurements fell within the model’s prediction intervals.³⁰

These hybrid approaches are still early-stage. They have not yet replaced the standard population modeling software that most clinical pharmacologists use daily. But they open possibilities for situations where traditional models struggle, like pulling useful pharmacokinetic information out of messy real-world datasets, electronic health records with irregular sampling, or published literature where only aggregate statistics are reported. The field is evolving rapidly, and the next decade will likely see machine-learning-augmented models move from proof-of-concept papers into bedside dosing tools.

How Interindividual Variability Gets Modeled Mechanistically

One area that deserves a closer look is how models handle the gap between “average patient” and “this patient.” Classic population pharmacokinetic models assign random variability terms to each parameter: some patients clear drug faster, some slower, and the model estimates how wide that spread is. This approach works, but it treats variability as statistical noise rather than something with a biological explanation.

Mechanistic covariate modeling tries to explain the variability instead of just measuring it. A novel approach proposed linking interindividual differences in drug distribution to lean body weight rather than total body weight, and deriving covariate relationships that account for a drug’s own physical-chemical properties. The result is a model where the predicted volume of distribution seamlessly adjusts for both body weight and body composition, with the balance between the two depending on the drug itself.³¹ This matters in practice because a lipophilic drug distributes preferentially into fat, so an obese patient needs a different dose adjustment than they would for a hydrophilic drug that stays mostly in lean tissue. Generic weight-based dosing misses that distinction; a mechanistic covariate model captures it.

The ambition, still being pursued across the field, is to make pharmacokinetic models precise enough that one day a patient’s genotype, organ function, body composition, and concurrent medications can all be fed into a model that spits out a personalized starting dose before the first pill is ever swallowed.