AI in Drug Discovery: How It Works, With Real Examples

Imagine you have a promising idea for treating a disease and enough lab time to test only a handful of molecules. There are far more possibilities than you can afford to investigate. Which ones should you make first? What would you need to learn before moving forward?

That is where AI can help. It gives researchers ways to compare possibilities, propose new molecular structures, and choose experiments that could settle an important question. The result might be a better candidate. It might also be an early reason to stop pursuing a bad idea.

AI helps researchers decide what to test. Experiments reveal whether the prediction holds up. Understanding that relationship makes the field easier to follow, whether you are a scientist, an investor, or someone hoping for better treatments.

What does AI actually do in drug discovery?

An “AI drug” can sound as if a computer invented a medicine in one step. In practice, the phrase can describe several different contributions. A model might help identify a disease target, predict a molecular property, or suggest a chemical change. Another system might help organize the research around those predictions.

These jobs require different tools. A language model can help find and explain information. A model trained on molecular structures and experimental measurements can estimate a property. A generative model can propose new structures. An agent can connect tools into a sequence of tasks. Calling all of them “AI” hides the differences that matter when judging a result.

For a useful starting point, think of discovery as a repeating cycle:

The valuable output of a cycle is new knowledge. A failed experiment can still be useful if it rules out an explanation or shows which property needs attention.

From a disease question to a molecule worth testing

Consider a hypothetical team trying to reduce harmful inflammation. The team suspects that changing the activity of a particular protein could help. That protein is the target. A molecule that affects it is a possible starting point, rather than a treatment ready for patients.

1. Decide whether the target makes biological sense

Before designing a molecule, the team needs a reason to believe that changing this protein will improve the disease. It might examine genetic studies, patient samples, published experiments, and evidence that points in the opposite direction.

AI can help bring this information together and suggest connections to investigate. Scientists still need to ask whether the evidence supports a cause or merely an association. A protein that changes during illness is not automatically a useful target.

2. Find or design possible starting molecules

With a target in mind, researchers can screen existing compounds or propose new ones. Structural models can suggest how a molecule might fit into a protein’s binding site. Generative tools can suggest alternative structures to explore.

AlphaFold 3 is a concrete example of progress in structural prediction. Its 2024 paper describes a model that predicts complexes involving proteins, DNA, RNA, and small molecules. That can give researchers a structural hypothesis to work with. It does not, by itself, establish that a proposed molecule will treat a disease. Read the original AlphaFold 3 study.

3. Check more than binding

A molecule may look like a good fit for a target and still be a poor development candidate. Can it dissolve well enough? Reach the intended tissue? Stay in the body long enough? Avoid unwanted effects on other proteins? Can chemists make it reliably?

The FDA’s overview of discovery and development describes the experiments needed to investigate a compound’s behavior, effects, and potential toxicity. AI predictions can help prioritize these questions, but each prediction has limits.

This is why a single “best molecule” score can be misleading. The candidate with the strongest predicted binding may have the least useful overall profile.

In this example, there is no justified winner yet. Candidate B might deserve a test for unwanted binding. Candidate A might be worth a chemistry change to improve solubility. Candidate C might help the team understand how much activity it can trade for a different property.

The better question is: Which experiment would most change our decision?

4. Run an experiment that can challenge the prediction

Suppose the team chooses Candidate B. A useful first experiment might measure its activity against the intended target and a set of related proteins. If the prediction is wrong, the result should be able to show that clearly.

The team would record how the measurement was made, what controls were used, and how repeatable the result was. Those details help explain why a prediction and a measurement disagree. Was the model wrong? Was the compound unstable? Did the assay measure something different from what the team intended?

That is the work behind the simple phrase “learn from the results.” Feeding a number back into a model is only useful when the team understands what that number represents.

What has moved beyond a computer prediction?

There are published examples of AI-assisted work reaching human studies. One is rentosertib, a small molecule studied for idiopathic pulmonary fibrosis, a disease that scars the lungs.

A 2025 randomized phase 2a study in Nature Medicine reported results from 71 participants assigned to rentosertib or placebo for 12 weeks. AI had contributed to identifying the target and designing the molecule. The trial’s primary endpoint concerned adverse events; lung function was among the secondary endpoints.

That is meaningful clinical research, but the study’s size and duration matter. The authors called for larger, longer trials. The paper provides evidence about one candidate in one study, rather than a verdict on every AI discovery platform or proof of lasting patient benefit.

When reading an announcement, first locate it on the evidence map below. Each stage answers different questions.

The FDA describes clinical research as studies in people, with different phases addressing safety, dosing, and treatment benefit. A prediction, a laboratory result, and a trial result should be reported in language that makes their differences clear.

Where can AI save time, and what still takes time?

Some gains are easy to picture. A researcher who can compare analyses without copying results between tools has less administrative work. A team that spots a likely property problem before making a compound can reconsider its next experiment earlier. A model that suggests an unfamiliar but testable structure can broaden the search.

Those are useful possibilities. They are also different from showing that an entire drug program reached patients faster or cost less.

To evaluate a speed claim, ask what the clock includes. Does “discovered in weeks” mean generating structures, making a compound, completing preclinical studies, or finishing a clinical trial? Was the comparison made against a similar project with similar resources? Did the faster process produce candidates of comparable quality?

The FDA’s drug development overview separates discovery, preclinical research, clinical research, regulatory review, and monitoring after a drug reaches the market. Progress in an early step does not answer the questions posed by later ones.

Three reasons a convincing model can disappoint

The test was too familiar. If a model is evaluated on molecules very similar to those it learned from, its score may overstate how well it handles a new chemical family. Ask what was kept out of training and whether the evaluation resembles the intended use.

The biological setting changed. A result from a simplified assay may not carry over to a cell, a tissue, or a person. A team needs to establish which setting its evidence actually describes.

The team optimized the wrong thing. A higher predicted binding score is not always a better research outcome. If a program needs a molecule that reaches a particular tissue, improving binding while worsening exposure may move it away from its goal.

These are practical reasons to keep uncertainty visible. A useful recommendation explains what is known, what is inferred, and what observation would change the plan.

Five questions to ask about an AI drug discovery claim

You do not need to inspect a model’s code to ask useful questions:

1. What did AI contribute? Target selection, molecular design, a property prediction, or help running the workflow?
2. What was measured? A computer score, a laboratory effect, or an outcome in people?
3. What was the comparison? An existing method, a control compound, a placebo, or no comparison at all?
4. What remains uncertain? Look for failed predictions, safety findings, missing experiments, and limits on who or what was studied.
5. What happens next? A credible result should make the next research decision easier to explain.

These questions help distinguish an interesting demonstration from evidence that changes a development program.

How we approach this at AgentCures

At AgentCures, we are building tools that connect molecular design, property assessment, experiment planning, and research records. Our aim is to make the path from a scientific question to a reviewable proposal easier to follow.

Scientists need to see the inputs behind a recommendation, challenge its assumptions, and decide what evidence to collect next. Keeping analyses and decisions in version control helps preserve that context as the work changes.

If you want to go further, our guides explain how AI agents use scientific tools , what ADMET means for a candidate’s properties , and how protein structures support drug design .

The prospect of better medicines is what makes this work exciting. Its value will be established one useful prediction, well-designed experiment, and carefully interpreted result at a time.