AI Drug Discovery: How Artificial Intelligence Is Changing the Search for New Medicines

Can AI Invent a Drug? Inside the Race to Turn Algorithms Into Medicine

AI Drug Discovery in 2026

AI promises to speed up drug discovery, but the real journey still runs from biology and molecule design to testing, trials, and finally patients.

Can AI Invent a Drug? Inside the Race to Turn Algorithms Into Medicine

Every new drug begins with a biological guess. A protein may be driving disease. A signaling pathway may be stuck in the wrong state. Somewhere among millions of experiments, patient records and chemical structures, there may be a clue worth following. The hard part is finding the right clue before years of laboratory work and millions of dollars are spent on the wrong one. This is where artificial intelligence has started to matter.

AI drug discovery is not one technology. Some systems read scientific papers and connect genes to diseases. Others predict protein structures, rank compounds, estimate toxicity or generate molecules that have never been synthesized before. Used well, these tools can help researchers discard weak ideas earlier and spend laboratory time on candidates that have a better reason to succeed.

The results, however, are still easier to see in research workflows than in pharmacies. A 2026 perspective in Nature Reviews Drug Discovery found substantial technical progress but much thinner evidence that AI has already improved clinical success across the industry. Most AI-first programs remain preclinical or in early human testing, with relatively few in late-stage trials. AI is clearly changing discovery; it has not removed the hardest test of all - showing that a treatment is safe and genuinely useful in people.

That distinction is the heart of this story. Generating a plausible molecule is becoming fast. Turning one into a medicine is still slow, expensive and unforgiving. The real question is whether AI can improve enough decisions along that path to change the odds.

Why Finding a New Drug Is So Difficult

Drug discovery is often shown as a funnel, and the metaphor is useful. At the wide end are thousands or millions of biological ideas and possible compounds. At the narrow end is a medicine that has survived chemistry, toxicology, manufacturing, regulatory review and several phases of clinical testing.

A useful drug has to do far more than bind to a target in a computer model. It must reach the right tissue, remain active for long enough, avoid dangerous off-target effects, be manufacturable at scale and, most importantly, improve outcomes in real patients. A candidate can look excellent in a simulation and fail for a reason the model never captured.

So the biggest opportunity for AI is not an instant-drug machine. It is better triage. If a weak target can be abandoned earlier, a toxic chemical series can be rejected before months of synthesis, or a clinical trial can enroll a more informative group of patients, the savings can be enormous even when approval still takes years.

Where AI Enters the Drug Discovery Pipeline

There is no single 'AI for drug discovery.' Different models solve different parts of the problem: reading papers, finding disease targets, predicting molecular interactions, generating chemistry, estimating safety or analyzing clinical data. The easiest way to understand their value is to follow a drug from the first biological idea toward the clinic.

1. Finding the right biological target

Before chemists design a molecule, researchers have to decide what they want to change in the body. The target might be a protein, receptor, enzyme, gene or signaling pathway involved in disease. If that choice is wrong, elegant chemistry will not rescue the program.

AI can search across genomics, transcriptomics, proteomics, electronic health data and scientific literature for relationships that would be difficult to assemble manually. A 2026 Nature Reviews Drug Discovery review called target identification one of the most promising uses of AI, while making an essential distinction: a statistical association is a lead, not proof that the target causes the disease or will respond to treatment. 

Related on Next Horizon: Genetics and AI: How Algorithms Help Decode DNA

One of AI’s most valuable roles in drug discovery is helping researchers find the right biological target before a single drug candidate is tested in people.

2. Understanding the molecular machinery

Once a target looks promising, researchers need to understand how it is built and how it behaves. Proteins fold into complex three-dimensional shapes, interact with other molecules and can change conformation as they work. Those details often determine whether a drug can affect them at all.

AlphaFold changed this part of biology by making high-quality structural predictions available at a scale that had been difficult to imagine. AlphaFold 3, published in Nature in 2024, extended prediction beyond proteins to complexes involving DNA, RNA, small molecules, ions and modified residues. It cannot tell a researcher which drug will succeed in a patient, but it can provide a much better structural starting point for experiments. 

For a drug designer, that can turn a blind search into a more focused one. Instead of testing chemistry against a poorly understood target, the team may have a plausible picture of the binding pocket or molecular interaction it wants to influence.

3. Virtual screening: rejecting bad candidates before making them

A pharmaceutical company can physically screen enormous libraries of compounds, but every experiment costs time and money. Virtual screening moves part of that search into the computer so that fewer molecules have to be made and tested first.

Models can rank compounds by predicted binding, similarity to known active molecules, ability to cross biological barriers or risk of toxicity. None of those predictions is a verdict. Their value is in prioritization: reducing a huge list to a smaller set worth spending laboratory resources on.

That boundary matters. A 'hit' from a model is not a drug, and sometimes it is not even a useful chemical lead. Chemistry and biology still get the final vote.

4. Generative AI: designing molecules that have never existed

Generative AI takes the idea one step further. Instead of only ranking molecules that already exist in a database, a model can propose new structures designed around several goals at once: strong interaction with a target, acceptable solubility, lower predicted toxicity, suitable size and a realistic path to synthesis.

A rough analogy is an image generator working under strict scientific rules. The model explores many possible structures, but the output is not judged by whether it looks convincing. It has to obey chemistry, be synthesizable and eventually behave as predicted in experiments. That makes molecular generation far less forgiving than generating text or images.

Antibiotic research shows both the promise and the bottleneck. In 2025, researchers reported generative-AI approaches that produced structurally novel candidates against pathogens including Neisseria gonorrhoeae and Staphylococcus aureus. Some worked in laboratory or animal models. Many other generated structures were difficult to synthesize, a reminder that a clever molecular design is only the first step toward a usable antibiotic.

Even after AI helps design a promising molecule, the hardest part begins: proving safety and effectiveness through preclinical and clinical development.

5. Predicting ADMET before a drug reaches humans

Potency is only one part of a drug's job. A candidate may bind strongly to its target and still fail because the body absorbs it poorly, clears it too fast, sends it to the wrong tissue or converts it into something toxic. Researchers group many of these questions under ADMET: absorption, distribution, metabolism, excretion and toxicity.

Machine-learning models can estimate some of these properties before a compound reaches expensive stages of testing. That helps chemists compare candidates on more than raw potency. The estimates are useful because they reduce uncertainty - not because they make experiments optional.

6. Drug repurposing: finding a new job for an old medicine

Sometimes the best AI-assisted drug discovery project does not create a new molecule at all. Drug repurposing asks whether an existing medicine could treat a different disease. Because approved drugs already come with substantial safety, manufacturing and pharmacology data, a successful repurposing idea can start further down the development path.

During the COVID-19 pandemic, BenevolentAI used a biomedical knowledge graph linking drugs, genes, pathways and published evidence to flag baricitinib, an existing rheumatoid arthritis medicine, as a candidate worth testing. Clinical studies later supported its use in COVID-19. AI did not 'discover a cure'; it helped researchers connect existing evidence quickly enough to generate a testable clinical hypothesis. [5]

The Most Important Real-World Test: Rentosertib

For several years, AI drug discovery had an uncomfortable imbalance: impressive molecule-generation demos and large investment rounds on one side, but very little randomized human evidence on the other. Rentosertib matters because it pushed an AI-origin program further into the part of medicine where claims become harder to make.

Rentosertib, previously called ISM001-055, is an experimental small molecule for idiopathic pulmonary fibrosis (IPF), a progressive lung disease. AI methods were used both to identify TNIK as a target and to help design the molecule. In 2025, Nature Medicine published a multicenter, double-blind, randomized Phase 2a trial involving 71 patients.

Safety was the primary endpoint, and adverse-event rates were broadly similar across treatment groups and placebo. At the highest once-daily dose, the researchers also saw a signal of improvement in forced vital capacity, a measure of lung function, over 12 weeks. That is encouraging, but the trial was small and short; it was not designed to prove durable efficacy. Larger and longer studies are still needed.

That makes rentosertib more useful as evidence of feasibility than as a victory lap. An AI-assisted hypothesis produced a real molecule, the molecule reached patients, and a controlled trial generated data that can now be challenged at larger scale. That is a much higher bar than showing that a model can generate chemistry on a screen.

Modern drug discovery is becoming a connected system where AI models, biological data, and lab automation work together to narrow the search for better medicines.

Why AI Still Cannot Replace the Laboratory

AI models do not study biology directly. They study data that represent biology - structures, assays, sequences, images, papers and clinical records. The gap between representation and living systems is where many promising predictions break down.

A protein may behave differently inside a cell than in a purified assay. A mouse may not reproduce a human disease. A biomarker may track with illness without causing it. A compound may bind perfectly in a simulation and become unstable in blood. Even a treatment that looks safe in a small Phase 1 trial can behave differently when hundreds of patients bring different ages, genes, medications and other diseases into the picture.

This is why benchmark performance is not enough. The more useful question is whether a model changes a real R&D decision for the better: does it help select the right target, avoid a dead-end molecule, design a better experiment or improve the chance that a clinical program succeeds? The 2026 Nature Reviews Drug Discovery perspective argues that the field now needs exactly this kind of evidence. 

The main limitations researchers still face

·         Biological data are incomplete. We have enormous datasets, but many are biased toward well-studied diseases, proteins and patient populations.

·         Correlation is not causation. AI can find patterns that look convincing without identifying the true mechanism of disease.

·         Models can fail outside their training domain. A system trained on one chemical space or assay can become unreliable when researchers move into unfamiliar biology.

·         Synthetic feasibility matters. A generated molecule is useless if chemists cannot make it efficiently or reproducibly.

·         Preclinical success does not guarantee clinical success. Human physiology remains the final test.

·         Commercial secrecy makes evaluation difficult. Some of the most advanced industrial systems are proprietary, so independent researchers cannot fully verify claims.

·         Regulation and validation are evolving. In January 2026, FDA and EMA published joint principles for good AI practice in drug development, emphasizing context of use, data governance, performance assessment, documentation and human oversight. 

AI Is Also Moving Into Clinical Trials

AI does not stop being useful once a molecule is chosen. Clinical trials themselves contain difficult matching and prediction problems: finding eligible patients, defining meaningful subgroups, monitoring large streams of data and deciding when a program is strong enough to continue - or weak enough to stop.

A September 2026 review in Nature Reviews Bioengineering describes emerging uses including patient-trial matching, prognostic subgrouping, data monitoring, external comparator groups, surrogate-endpoint validation and support for go/no-go decisions. Digital twins and agentic systems are also being studied, but these are exactly the areas where attractive demonstrations need especially strict validation before they influence high-stakes decisions. 

The practical rule is simple: AI may help make a trial smarter, but it does not make the trial unnecessary. Randomized controlled evidence remains the most reliable way to learn whether a new treatment delivers more benefit than harm.

Will AI Make Drugs Cheaper and Faster?

In some parts of the pipeline, almost certainly. Across the entire journey from idea to approved medicine, the answer is still unknown.

AI can compress tasks that once took weeks of manual searching or screening into hours or days. It can generate candidates quickly and help teams choose experiments more selectively. But faster early discovery does not automatically shorten toxicology studies, manufacturing work, regulatory review or multi-phase clinical trials, all of which operate on biological and regulatory timelines that software cannot simply skip.

The biggest economic win may be less glamorous than a drug designed in record time: failing sooner. If a weak program can be abandoned after three months instead of three years, that is valuable even though no medicine was produced. In an industry where failure is common, avoiding the wrong experiment can matter as much as accelerating the right one.

What Changes in 2, 5 and 10 Years?

In about 2 years: AI becomes a routine research co-pilot

The most plausible near-term future is not a pharmaceutical company run by one autonomous agent. It is AI becoming ordinary infrastructure. Medicinal chemists will use generative models to suggest analogs; biologists will rank targets with AI; literature agents will track new evidence; and automated labs will increasingly run closed-loop experiments in which software proposes a test, robots perform it and the results shape the next prediction.

The important shift will be less visible than the headlines: teams will stop treating AI as a special initiative and start treating it like another research instrument.

In about 5 years: we get a much better clinical answer

By the early 2030s, many AI-origin programs being discussed today should have had enough time to reach Phase 2 or Phase 3. That will finally give the field something it still lacks in 2026: a sizeable set of clinical outcomes that can be compared with conventional drug-development programs.

If those programs show better success rates, AI will have evidence that it improves more than speed. If they fail at roughly the same rates, the industry's attention may move away from molecule generation and toward the harder problems that models cannot sidestep: disease biology, patient selection and translation from laboratory systems to humans.

In about 10 years: the distinction between AI drug discovery and drug discovery may disappear

If these tools mature, the label 'AI-designed drug' may eventually become unremarkable. Computational systems could sit inside almost every stage of pharmaceutical R&D, from target selection and medicinal chemistry to experiment planning, trial design and post-market safety monitoring.

A more ambitious version is a partly autonomous discovery loop: multimodal models generate hypotheses, robotic laboratories test them, software analyzes the results and the cycle repeats. Even then, medicine will remain a human and regulated enterprise. The hardest questions are not just computational: Is the benefit meaningful? Is the risk acceptable? Who should receive the treatment? Who is accountable when the system is wrong?

AI can evaluate vast numbers of possible molecules far faster than humans alone, helping researchers focus experimental work on the strongest candidates.

The Real Revolution Is Better Decisions, Not Instant Drugs

AI is already making parts of pharmaceutical research faster and more searchable. It can connect evidence across datasets no person could read in full, explore chemical spaces too large to test manually and help researchers decide which ideas deserve an experiment.

What it has not yet done is prove that those technical gains translate into a broad improvement in clinical success. That is the gap the next decade has to close.

The companies most likely to benefit may not be the ones that generate the largest number of molecules. They may be the ones that use AI to choose better biological questions, design better experiments and kill weak programs before they become expensive failures.

A computer can propose a chemical structure in seconds. A medicine still has to earn its way through evidence. If AI changes drug discovery, its biggest contribution may be helping scientists reach the right evidence sooner.

Continue reading on Next Horizon: AI in Pharmacogenomics 2026: Can Your DNA Predict the Right Drug and Dose?

FAQ

Can AI discover a new drug by itself?

Not by itself in the full medical sense. AI can identify targets, rank compounds and generate molecules, but laboratory validation, toxicology, manufacturing, clinical trials and regulatory review remain human-led, evidence-heavy processes.

Are any AI-designed drugs already in clinical trials?

Yes. Several AI-origin programs have entered human testing. Rentosertib is a prominent example because a randomized Phase 2a trial was published in Nature Medicine in 2025. Early clinical progress, however, is very different from regulatory approval.

Has an AI-discovered drug been approved by the FDA?

As of September 2026, peer-reviewed reviews still describe AI-first pipelines as concentrated mainly in preclinical, Phase 1 and Phase 2 development, with relatively few late-stage programs. The field has not yet shown a broad wave of approved medicines whose clinical success can clearly be attributed to AI-driven discovery.

How does generative AI design molecules?

Generative models learn patterns in chemical and biological data, then propose structures that are predicted to meet selected constraints such as target binding, solubility, size or toxicity. Chemists still have to decide which candidates are realistic, synthesize them and test whether the predictions survive contact with biology.

Will AI replace medicinal chemists and pharmaceutical researchers?

More likely, AI will change what these scientists spend time on. Human researchers are still needed to frame the biological problem, judge model outputs, design experiments, interpret conflicting evidence and take responsibility for decisions that affect safety.

What is the biggest limitation of AI drug discovery?

Translation from model performance to real clinical benefit. A system can excel on a benchmark and still fail to improve a drug program. Predictions have to survive chemistry, cells, preclinical models and, finally, diverse human patients.