Top

Can artificial intelligence discover new medicines faster?

Published 19 Nov 2026 • By Léonie Guerut

What if we could identify a molecule capable of slowing a disease, fighting a resistant bacterium or acting on a poorly understood mechanism more quickly? For people waiting for a new treatment, this prospect offers considerable hope.

Artificial intelligence now has a place in research laboratories. It helps analyse data, select molecules and explore potential treatments. But between a computer-generated proposal and a medicine available in a pharmacy, many essential steps remain.

So, what can it actually speed up? What results have already been achieved? And how can we distinguish a scientific advance from a promise that remains uncertain?

Can artificial intelligence discover new medicines faster?

Why does discovering a medicine take so long?

Developing a medicine is not simply about finding a substance that acts on a disease.

First, researchers need to understand enough about the mechanism they want to modify. For example, they may focus on a protein involved in inflammation or the multiplication of cancer cells. This protein then becomes a “therapeutic target”.

They then look for substances capable of acting on this target. Many will be abandoned: some do not produce the expected effect, while others are too toxic, poorly absorbed or difficult to manufacture.

The most promising candidates undergo preclinical studies, followed by trials in humans. These trials help determine doses, assess side effects and establish whether the treatment provides a benefit.

To understand these stages, you can read Carenity’s article on the different types of clinical trials.

AI can play a role at several points along this journey. One of its benefits is helping teams identify earlier which avenues deserve to be tested.

What does “discovering a medicine with AI” really mean?

This expression covers very different situations.

In some projects, AI helps identify a biological target. In others, it examines known molecules to identify those that might be useful. So-called generative systems can also propose new chemical structures.

AI used in pharmaceutical research is therefore not limited to a conversational assistant that is asked to invent a treatment. It often involves specialised models trained on biological, chemical or experimental data.

These tools can rank possibilities, estimate certain properties or suggest experiments. Researchers must then verify their results.

Saying that a drug candidate was “discovered by AI” does not, on its own, explain what the system did or what experiments have confirmed.

Finding a lead within a vast amount of data

Research generates considerable amounts of information: gene activity, protein structures, tissue images, experimental results and clinical data.

AI can help identify patterns within this information. For example, it might highlight a protein whose activity appears to differ between diseased and healthy cells.

This observation provides an avenue to explore. But an association is not enough: the protein may accompany the disease without being responsible for its progression. Researchers therefore need to verify that acting on it actually produces a useful effect.

The tool’s value lies in guiding research through a range of possibilities too vast to examine in full.

Understanding protein shapes better

A protein’s shape is very important to how it functions and which molecules can interact with it.

We can imagine a target as a lock: knowing its shape helps us look for a key. However, this comparison has its limits, because proteins move and biological interactions are more complex than simply fitting one object into another.

Tools such as AlphaFold have advanced protein structure prediction. Presented in a scientific publication in 2024, AlphaFold 3 extends this approach to the structures of complexes involving proteins, DNA, RNA and small molecules, among others.

In drug research, this can help researchers develop hypotheses about how a substance binds to a target.

However, a predicted structure remains a hypothesis. It does not demonstrate that a molecule will be effective in the body or well tolerated.

Designing molecules and choosing which ones to test

Once a target has been identified, researchers need to find a molecule with several qualities at once.

It must act on the target, but also have properties compatible with its use as a medicine. For example, a substance that looks promising in an experiment may be eliminated too quickly by the body.

AI models can help compare candidates or propose changes to their structure. Scientists then select some of these proposals, make the molecules and test them.

Experimental results allow researchers to revise their hypotheses and start again. AI can make this sequence of proposals, experiments and adjustments more efficient.

The expected benefit is therefore not simply producing more molecules: it also involves avoiding spending too much time on unpromising avenues.

A concrete example: a candidate studied in pulmonary fibrosis

Rentosertib illustrates the transition from AI-assisted discovery to clinical testing.

In this project, AI contributed to identifying a target, the TNIK protein, and then designing a molecule intended to inhibit it. The candidate was studied in idiopathic pulmonary fibrosis, a disease in which lung tissue gradually turns into scar tissue.

In 2025, results from a phase 2a trial were published in Nature Medicine. The study involved 71 patients followed for twelve weeks, randomly assigned to several groups receiving the treatment or a placebo.

The results provided information about tolerability and early signals regarding respiratory function. They represent a research milestone, but a trial of this size and duration is not enough to establish a lasting benefit or characterise all the risks.

This example shows that candidates arising from approaches using AI can be evaluated in patients. On its own, it does not establish that AI increases the success rate of pharmaceutical development as a whole.

Potential approaches are also being explored against resistant bacteria

The search for new antibiotics is another field in which AI has already contributed to experimental discoveries.

In a study published in 2023 in Nature Chemical Biology, researchers used machine learning to identify molecules that might act against Acinetobacter baumannii, a bacterium that can cause difficult-to-treat infections.

This approach led to the identification of abaucin, a compound showing targeted activity against this bacterium in the experiments conducted.

The example is interesting because the model helped guide the selection of molecules to test. Experiments then allowed researchers to examine their activity.

But discovering an antibacterial substance in the laboratory does not mean it is ready to treat an infection in humans. Questions about dosage, safety, distribution in the body and clinical effectiveness still need to be resolved.

Can AI also improve clinical trials?

AI can be used beyond the search for molecules. In particular, it can contribute to image analysis, processing data collected during trials or identifying patient profiles that match study criteria.

These applications could help organise certain research activities more effectively and make measurements more reproducible.

However, researchers must verify that the tools work under the conditions in which they will be used. A model that performs well on data from a single institution will not necessarily produce the same results elsewhere.

Participants’ needs also remain essential. Fatigue, travel or the frequency of tests can make a trial difficult to follow, regardless of how sophisticated the tools are.

Carenity explains how patients’ experiences can contribute to research in its article on studies, scientific publications and innovative projects carried out with patients.

Why is moving faster in the laboratory not enough?

An announcement may highlight how quickly a molecule was identified. This is useful information, provided we know which period was measured.

Does it refer to the time needed to select a target? To obtain a promising molecule? To start a first trial? Or to obtain marketing authorisation?

These stages are not equivalent.

Even if the initial discovery is accelerated, the treatment’s effects still need to be observed over time. An improvement measured after a few weeks does not answer every question about benefits and risks after several years.

To assess AI’s real contribution, researchers will therefore need to examine the results of many programmes: timelines, failures, tolerability and benefits for patients. Individual successes are not enough to describe the full picture.

The data used can also introduce bias

A model learns from the information it is given. If these data are incomplete, unrepresentative or of poor quality, its results may be too.

For example, a tool built using data mainly concerning certain patient groups may be less reliable for other populations.

There is also a risk of overstating performance when the data used to evaluate the model are too similar to those used to train it. Verification using independent data is therefore important.

Confidentiality raises further questions: when health data are used, their protection, conditions of access and research purposes must be clearly governed.

Who oversees the use of AI in drug research?

Health authorities are working to establish frameworks for these uses.

In January 2026, the European Medicines Agency (EMA) and the US Food and Drug Administration (FDA) published ten shared principles of good practice for using AI in drug development.

They cover data quality, a precise definition of the intended use, performance evaluation, risk management and human oversight, among other areas.

The level of scrutiny must reflect the role assigned to the system. A tool that helps a researcher explore hypotheses does not have the same impact as one whose results contribute to a regulatory decision about a treatment.

The aim is to be able to explain why the results are sufficiently reliable for the intended use.

What could this change for patients?

The hope is to generate more useful leads and improve certain stages of research.

For someone living with a chronic illness, the question remains practical: will a new treatment reduce symptoms, preserve independence, limit complications or help them live better with the disease?

The fact that a medicine was developed using AI does not answer these questions. The results of studies will provide the answers.

AI is also used in areas beyond drug discovery. To learn about applications related to monitoring a condition, you can read Carenity’s article on artificial intelligence in diabetes care. These uses should be distinguished from designing a new treatment.

How should you read an announcement about a “drug discovered by AI”?

A few questions can help clarify what has actually been demonstrated:

  • What was AI’s precise role: proposing a target, selecting a molecule or analysing results?
  • Do the data come from a simulation, laboratory experiments or a trial in humans?
  • How many people took part in the study, and for how long?
  • Was the treatment compared with a placebo or a standard treatment?
  • What benefits and side effects were observed?
  • Were the results published in a scientific journal or only announced by the company?

A discovery can be important while remaining a long way from routine use. Identifying its stage of development allows us to welcome progress with interest without turning a research lead into a promise of treatment.

Frequently asked questions

Can AI invent a medicine on its own?

It can propose new chemical structures or help identify targets. Research teams must then verify their potential in the laboratory and, when justified, in clinical trials.

Can it replace trials in patients?

A computer prediction does not, on its own, demonstrate a treatment’s effectiveness and safety in humans. AI can support certain stages of trials, but it does not automatically remove the need for clinical evaluation.

Is a medicine designed with AI necessarily better?

No. Its value depends on the benefits and risks observed in studies. The method used to discover it does not guarantee its success.

Can AI speed up research?

Yes, certain selection, prediction and analysis tasks can be accelerated. The impact on overall development time and on the number of genuinely useful treatments must be assessed case by case.

Did you find this article useful? Share your questions and your views on the use of artificial intelligence in medical research with the community in the comments.

Take care of yourself!

avatar Léonie Guerut

Author: Léonie Guerut, Health Content Writer

Léonie is a contributor to the Carenity community, where she helps create and share informative health content while engaging with members across the platform. She is particularly interested in health communication... >> Learn more

Comments

You will also like

Signs of stress and exhaustion: subtle changes your body may be showing

Signs of stress and exhaustion: subtle changes your body may be showing

Read the article
Medicines that are at risk of misuse and dependence

Medicines that are at risk of misuse and dependence

Read the article
Hot weather: risks related to taking certain medications

Hot weather: risks related to taking certain medications

Read the article
6 things you should know about your mental health!

6 things you should know about your mental health!

Read the article

Most commented discussions