ADMET Explained: Why Strong Binding Is Only the Beginning

A molecule looks promising in an experiment. It binds to the intended protein, and a small chemical change makes that binding stronger. The team has a reason to be excited. It also has a new set of questions: can the molecule reach that protein in the body, stay there long enough, and avoid unacceptable harm?

Those questions are the territory of ADMET: absorption, distribution, metabolism, excretion, and toxicity. They help explain why a good result in a binding assay is an early step in developing a medicine.

AI can help researchers estimate some of these properties and choose what to investigate next. Its value depends on connecting a prediction to the right experimental question. A favorable score, on its own, cannot tell the team how a candidate will behave in a person.

Five questions behind the acronym

The FDA’s overview of discovery and development describes the work needed to investigate how a compound behaves, what effects it produces, and what harm it might cause. ADMET organizes many of those questions.

The first four letters describe processes that influence exposure, meaning how much compound is present in a particular place over time. They are part of pharmacokinetics, often shortened to PK. Toxicity concerns harmful effects, which also depend on exposure and the biological setting.

This guide focuses on the questions faced when developing an oral small-molecule medicine. The intended route and target tissue matter: a compound meant to act inside the brain has different distribution requirements from one intended to act within the gut.

Absorption and oral bioavailability are different

Absorption asks how much compound gets across the intestinal barrier. Oral bioavailability, in the example here, asks how much of the swallowed compound reaches the systemic circulation unchanged. Systemic circulation is the blood supply carrying it around the body.

A compound can be absorbed and still be metabolized in the intestinal wall or liver before it reaches that circulation. This is called first-pass metabolism. Researchers therefore distinguish the fraction absorbed from the fraction that survives those processes. A primary study of human oral bioavailability examined these as separate contributors.

The invented example below follows 100 arbitrary units of a compound. It is conceptual accounting, with no real drug, dose, or patient data.

Here, 80% is absorbed, but only half of that survives first-pass metabolism. The resulting oral bioavailability is 80% × 50% = 40%. The numbers are assumptions chosen to illustrate the distinction, not typical values for medicines.

Even that 40% does not tell us how much reaches a target inside a particular tissue or cell. Distribution, binding to proteins, transport, and elimination still matter. “Present in blood” and “available at the target” are different research questions.

Why investigate these properties early?

Suppose a team spends several design rounds improving a molecule’s activity against a protein. It later discovers that too little compound dissolves under the conditions needed for an experiment. The promising activity result may be difficult to reproduce or translate into the next setting.

An earlier solubility check could have changed the design priorities. The team might have considered a different structure or investigated formulation options before committing to more work on the same series.

That is the practical case for early ADMET assessment. A model can help flag a question while alternatives remain open. An experiment then helps establish whether the concern is real and what it means for the program.

What AI can predict, and what a result means

ADMET models learn relationships between molecular features and measured properties. A team might use one model to estimate solubility and another to predict behavior in a particular metabolism assay. Each model has its own training data, output definition, and limitations.

The SwissADME paper , published in 2017, is a concrete example of computational methods for estimating properties relevant to small-molecule development. It describes several models and calculations, rather than a single universal assessment of whether a compound will succeed.

That distinction matters when reading a dashboard. A model may predict an experimental endpoint, classify a potential liability, or calculate a descriptor directly from a structure. Those outputs should be labeled so a scientist can tell what was calculated, what was predicted, and what was measured.

Before using a prediction, check three things:

1. What exactly is the output? A solubility value needs units and conditions. A toxicity label needs a defined endpoint. A score without that context can be hard to interpret.
2. Does the evaluation resemble this use? Performance on familiar chemical families may not carry over to a substantially different molecule. Ask how the model was tested and what evidence supports its use here.
3. What decision will it change? A prediction should help choose a candidate, a redesign, or an experiment. If two candidates are effectively tied within the model’s uncertainty, a precise ranking may be misleading.

Keeping the original input, model version, and result together also makes a later disagreement easier to investigate.

Why there is no single best ADMET profile

Drug design involves competing goals. A chemical change that improves one property can affect another. A team needs enough activity at the intended target, suitable exposure for its purpose, and an acceptable safety profile under the conditions being studied.

“More stable” is not automatically better in every setting. A compound that persists longer may create a different exposure pattern, and its metabolites may also need investigation. Similarly, higher blood concentrations do not settle whether enough active compound reaches the intended tissue.

The right comparison starts with the program’s goal. Which tissue matters? What evidence supports the proposed biological effect? Which property is currently limiting progress? Answering those questions is more useful than maximizing every score on a dashboard.

Turn a property concern into a testable question

A useful ADMET result narrows the next investigation. The examples below show how different findings can lead to different questions. They are illustrative situations, not reports of AgentCures candidates.

The hERG example deserves context. hERG is an ion channel involved in the heart’s electrical activity. FDA research describes how ion-channel measurements contribute to cardiac safety assessment . A prediction about one channel is one piece of evidence. Interpreting risk requires appropriate measurements and an integrated assessment, including exposure.

Metabolism also connects ADMET to drug interactions. One compound can affect an enzyme or transporter involved in another compound’s behavior, changing exposure. The FDA’s M12 guidance on drug interaction studies covers evaluation of these mechanisms using laboratory and clinical evidence. An early computational flag can identify a question for that work; it does not resolve the interaction on its own.

Can computational models reduce animal testing?

Using a suitable computational or laboratory method to answer a question can help avoid unnecessary experiments. Human cells, organoids, organ-on-chip systems, and models can contribute different kinds of evidence.

In April 2026, the FDA reported progress under its roadmap to reduce animal testing, launched in 2025. That work concerns the development and use of human-relevant alternatives. It should not be read as a claim that any ADMET prediction can replace every safety study.

The scientific question and the evidence supporting a method determine where it can help. A tool that is useful for prioritizing molecules may need different validation before being used for another purpose.

What earlier assessment changes for a team

Early assessment gives researchers more chances to change direction before a candidate becomes difficult or expensive to replace. It can also reveal that the next useful step is a focused measurement, rather than another round of molecular generation.

That benefit is worth evaluating directly. Did the prediction identify a problem that experiments confirmed? Did it help resolve a trade-off? Did it wrongly discourage work on a useful candidate? Following those outcomes helps the team learn where a model deserves trust.

At AgentCures, we compare ADMET predictions alongside target activity and selectivity, and keep analyses connected to research decisions. Our aim is to help scientists inspect the evidence, identify missing information, and decide what to test next.

Our guides to AI in drug discovery , research agents , and the economics of development explain how this work fits into the wider process.

Strong binding gives a team something worth investigating. ADMET helps it ask whether that molecular idea has a plausible path through the body to the effect the team wants to study.