Has AI discovered a real drug?

Has AI actually discovered a drug that works in humans?

Yes, AI has helped discover real drug candidates that have already shown measurable effects in human patients. The clearest example is rentosertib, an experimental treatment for lung fibrosis that has now reached Phase III clinical trials.

Developed by Insilico Medicine, rentosertib targets idiopathic pulmonary fibrosis (IPF), a progressive disease that scars the lungs and makes breathing increasingly difficult. AI helped identify the biological target and design the molecule. In a randomized Phase IIa trial published in Nature Medicine in June 2025, patients receiving the highest once-daily dose experienced an average improvement in lung function, while the placebo group deteriorated slightly.

That finding is particularly interesting because the drug's development began with AI-assisted identification of a target that had not previously been established as a treatment route for fibrosis.

Other examples are emerging. Recursion Pharmaceuticals reported that its experimental drug REC-4881 reduced gastrointestinal polyps in patients with a hereditary condition that greatly increases colorectal cancer risk. Researchers at MIT and McMaster University have also used machine learning to discover compounds that kill dangerous antibiotic-resistant bacteria.

The remaining question is whether any of these treatments can survive the final stages of testing and become an approved medicine.

Has the FDA approved any AI-discovered drug yet?

No novel medicine clearly documented as having been discovered and designed through an end-to-end AI-driven process has yet been established as having received FDA approval.

That distinction matters because pharmaceutical companies have used computational tools for decades. Today, many approved drugs have benefited from machine learning somewhere in their research or development. Counting every medicine touched by an algorithm as an AI-discovered drug would make the claim almost meaningless.

For this analysis, we consider AI discovery meaningful when the technology makes a substantial contribution to finding a therapeutic target, designing a novel molecular structure or identifying a previously unrecognized treatment opportunity.

Rentosertib meets that definition through its AI-assisted target identification and molecular design. Exscientia's DSP-1181, which entered human trials in 2020, is another example, although development was later discontinued. Recursion's REC-4881 represents a different approach because AI helped identify a new medical application for an existing compound.

The FDA's published novel-drug approval records do not currently establish a completed approval for a clearly documented, end-to-end AI-discovered novel medicine. However, regulators do not classify approved treatments according to whether AI was used in their discovery, making a definitive count of every AI-assisted medicine impossible.

If you want more recent data on this point, please see our latest AI in drug discovery market report.

What makes a drug genuinely AI-discovered?

An AI-discovered drug is a treatment candidate whose discovery depended meaningfully on artificial intelligence, rather than a conventional compound that researchers happened to analyze using AI software.

There are several ways this can happen.

Insilico Medicine used AI to identify a potential biological target and design rentosertib around it. Exscientia used AI-assisted molecular optimization when developing DSP-1181. Recursion analyzes large collections of cellular images and experimental measurements to identify treatment opportunities that traditional research might overlook.

These approaches solve different problems. Finding a disease target is different from designing a molecule that interacts with it, and both differ from identifying a new use for an existing drug.

There is also a separate question about what counts as a successful discovery. A molecule that works in a laboratory is a genuine scientific finding, even if it never becomes a commercial medicine.

We can divide the evidence into seven levels:

Development stage	Evidence available	What researchers actually know
AI prediction	Computational model proposes a candidate	The molecule might work
Laboratory testing	Activity against cells or proteins	The molecule has biological activity
Animal studies	Effects demonstrated in living animals	The approach has preclinical potential
Phase I	First human safety and dosing studies	The compound can be tested in humans
Phase II	Early patient efficacy studies	The treatment may benefit patients
Phase III	Larger confirmatory clinical trials	Whether benefits hold up in broader testing
Regulatory approval	Positive benefit-risk assessment	The treatment can be marketed for an approved use

Is Insilico Medicine's rentosertib the first real AI drug?

Rentosertib is currently one of the strongest examples of a genuinely AI-discovered drug approaching the final stages of clinical testing, although earlier AI-designed compounds had already entered human trials.

Insilico Medicine developed rentosertib, previously known as INS018_055 or ISM001-055, to inhibit an enzyme called TNIK. Researchers identified a potential connection between this target and the biological processes driving lung fibrosis.

The company combined its PandaOmics platform for target identification with its Chemistry42 system for molecular design. Its research was published in Nature Biotechnology, followed by human clinical findings in Nature Medicine.

The development timeline was relatively short. Insilico reported approximately 18 months from target identification to preclinical candidate nomination, followed by early human testing.

In September 2026, Insilico announced that the first patient had received the drug in its GENESIS-IPF-3 Phase III study. The trial plans to enroll approximately 320 patients across 47 centers in China, with treatment lasting 52 weeks.

That is a major change from the earlier Phase IIa study, which involved only 71 participants over 12 weeks. The larger study will help determine whether rentosertib's initial benefits are clinically meaningful and lasting.

Did rentosertib really improve lung function in clinical trials?

Yes, rentosertib improved average lung-function measurements in its highest-dose Phase IIa group, but the benefit came from a small trial, and safety remains an important concern.

The Nature Medicine study followed 71 patients with idiopathic pulmonary fibrosis for 12 weeks. Researchers compared three dosing regimens with placebo and measured forced vital capacity (FVC), the amount of air patients can forcibly exhale.

Patients receiving 60 milligrams once daily experienced an average FVC improvement of 98.4 milliliters. The placebo group experienced a decline of 20.3 milliliters.

The difference between those mean changes was approximately 119 milliliters over three months.

That is an encouraging result in a disease where declining lung function is the central clinical problem. However, only 18 patients received the highest daily dose, and FVC was a secondary endpoint rather than the study's main safety objective.

The other two active-treatment regimens did not show equally convincing changes. Patients receiving 30 milligrams once daily experienced an average decline of 27 milliliters, while those receiving 30 milligrams twice daily improved by 19.7 milliliters.

There were also tolerability concerns. Sixteen of the 71 participants discontinued the study, including 12 because of adverse events. Seven discontinuations involved liver injury or liver-function abnormalities.

Current IPF treatments, including nintedanib and pirfenidone, already have evidence that they can slow disease progression. Rentosertib still has to establish that its benefits are durable, clinically relevant and worth the risks.

If you want more recent data on this point, please see our latest AI in drug discovery market report.

Has Recursion discovered an AI drug that actually helps patients?

Recursion Pharmaceuticals has reported encouraging human results from two AI-associated drug programs, including a 47% median reduction in gastrointestinal polyp burden with REC-4881.

REC-4881 is being developed for familial adenomatous polyposis (FAP), an inherited condition that causes numerous gastrointestinal polyps and dramatically increases colorectal cancer risk.

Recursion's computational platform helped identify MEK inhibition as a possible treatment approach. Unlike rentosertib, REC-4881 involves finding a new therapeutic use for a previously known compound.

The most recent Phase Ib/II results, disclosed on October 15, 2026, showed that 13 of 15 evaluable adult patients experienced reductions in total gastrointestinal polyp burden after 12 weeks of treatment.

The median reduction was 47%, and approximately one-third of patients achieved reductions of at least 65%.

The company also reported a 55% median reduction in upper gastrointestinal polyp burden, including two patients with reductions of at least 98%.

Some effects persisted after treatment stopped. Among patients evaluated 12 weeks after completing treatment, 85% still had total gastrointestinal polyp burden below baseline.

These findings are particularly relevant because removing gastrointestinal polyps and preventing their progression are central concerns for people with FAP. Still, the study was small and lacked a randomized placebo comparison.

Recursion has also reported early activity from REC-1245, an experimental cancer treatment designed to degrade the protein RBM39.

According to its October 2026 clinical update, two patients had confirmed partial tumor responses, while a third had an unconfirmed response. Among 12 efficacy-evaluable patients receiving doses of at least 45 milligrams, 11 experienced disease control after eight weeks.

The company reported synthesizing just 204 compounds over approximately 18 months during REC-1245's discovery process.

Has AI discovered an antibiotic that can kill resistant bacteria?

Yes, machine learning has helped scientists discover several compounds capable of killing dangerous bacteria, including pathogens that resist existing antibiotics.

One of the best examples is abaucin, discovered through research involving McMaster University, MIT and collaborating institutions.

In a 2023 study published in Nature Chemical Biology, scientists trained a machine-learning system using experimental results from approximately 7,500 compounds tested against Acinetobacter baumannii.

The model then helped researchers identify additional candidates worth testing. Abaucin emerged as a promising compound with relatively narrow antibacterial activity against A. baumannii, a pathogen associated with serious hospital infections.

Laboratory experiments showed activity against clinical bacterial isolates, including resistant strains. The compound also demonstrated antibacterial activity in a mouse wound-infection model.

Researchers identified a mechanism involving bacterial lipoprotein trafficking, linked to a protein called LolE.

Another example is halicin, identified in research led by MIT and published in Cell in 2020. Scientists used deep learning to recognize a compound with antibacterial properties that conventional researchers had not previously identified in that context.

Subsequent machine-learning studies also found promising compounds against antibiotic-resistant Staphylococcus aureus.

The discoveries are real, although none of these famous AI-identified experimental antibiotics has become a widely approved human treatment.

Abaucin's reported minimum inhibitory concentrations against tested A. baumannii strains ranged from 2 to 64 micrograms per milliliter, a 32-fold difference across the range. That variation shows why positive laboratory results still require extensive pharmacological and clinical investigation.

Are AI-discovered drugs passing clinical trials more often?

AI-discovered drug candidates appear to perform unusually well in early clinical testing, but we still have no convincing evidence that they succeed more often in the trials that determine whether a treatment actually works.

A 2024 analysis published in Drug Discovery Today by researchers associated with Boston Consulting Group examined clinical development outcomes among AI-native biotechnology companies.

The study estimated Phase I success rates of approximately 80–90% for AI-discovered molecules, considerably higher than the historical benchmarks considered by the researchers.

That suggests AI may be particularly useful for identifying compounds with suitable properties for human testing.

Phase II outcomes were less impressive. The same analysis estimated success at approximately 40%, although the sample was limited.

For comparison, a Biotechnology Innovation Organization analysis covering clinical development between 2011 and 2020 found that 28.9% of conventional drug programs successfully moved through Phase II.

The broader study examined 9,704 development programs and reported an overall 7.9% probability of advancing from Phase I to eventual regulatory approval.

These figures come from different populations and time periods, so we cannot use them to calculate a precise AI advantage. AI-native companies may also differ from the broader industry in the diseases they target, the candidates they select and the way they report development progress.

Clinical metric	AI-discovered candidates	Historical industry comparison
Phase I success	Approximately 80–90% in a limited 2024 study	Lower historical success
Phase II success	Approximately 40% in the same study	28.9% in BIO's 2011–2020 dataset
Phase III success	Insufficient comparable data	Established historical outcomes
Phase I-to-approval success	Not reliably established	7.9%
Early safety evidence	Encouraging	Extensive conventional baseline
Proven efficacy advantage	Not established	No reliable AI-specific comparison
Approved medicines per research dollar	Not yet measurable across mature AI cohorts	Historical industry benchmarks available

If you want more recent data on this point, please see our latest AI in drug discovery market report.

Have any famous AI-discovered drugs already failed?

Yes, several high-profile AI-associated drug candidates have failed to deliver the clinical results their developers expected, including programs from Exscientia and BenevolentAI. Recursion has also reported mixed Phase II findings.

DSP-1181 was one of the earliest examples to attract international attention.

Developed through a collaboration between Exscientia and Sumitomo Dainippon Pharma, the compound was designed as a potential treatment for obsessive-compulsive disorder. In January 2020, the partners announced its entry into Phase I testing after an AI-assisted discovery process lasting less than 12 months.

That was impressive speed, but DSP-1181 was later discontinued after failing to meet the requirements for further development.

BenevolentAI encountered a similar problem with BEN-2293, an experimental topical treatment for atopic dermatitis.

In 2023, the company reported that its Phase IIa trial met its primary safety and tolerability endpoint but failed to achieve secondary efficacy objectives related to reducing itching and inflammation. BenevolentAI subsequently stopped further spending on the program.

Recursion's REC-994, investigated for cerebral cavernous malformation, produced a more mixed picture in Phase II. The SYCAMORE study reported encouraging safety findings and some exploratory efficacy signals, but did not establish a convincing clinical benefit sufficient to validate the program's original expectations.

Across these three programs, we see different discovery methods and diseases, but a familiar problem: getting a candidate into human trials is much easier than proving it delivers meaningful benefits.

Is AI really making drug discovery faster and cheaper?

AI has already shortened the early stages of drug discovery in several documented programs, sometimes reducing years of research to roughly 12–18 months. The cost savings are promising, although the industry has not yet demonstrated a consistent reduction in the total cost of bringing an approved drug to market.

Exscientia and Sumitomo reported developing DSP-1181 to its clinical-candidate stage in less than 12 months. The companies compared that timeline with a conventional reference estimate of approximately 4.5 years for exploratory research.

Insilico Medicine reported approximately 18 months from identifying rentosertib's biological target to nominating its preclinical candidate.

Recursion reported a similar 18-month discovery process for REC-1245, involving the synthesis of 204 compounds before reaching its development candidate.

These programs did not all begin at the same stage or use identical development methods, so a simple average would be misleading. Nevertheless, several independent developers have demonstrated that computational discovery can considerably reduce the time spent selecting and optimizing molecules.

AI can also reduce experimental costs by predicting which compounds are more likely to have desirable properties before they are physically produced. That means fewer molecules may need to be synthesized and tested.

However, neither the 204-compound figure for REC-1245 nor the shorter discovery timelines tell us exactly how much a conventional team would have spent solving the same problem.

AI platforms also have costs of their own, including computing infrastructure, training data, automated laboratories and specialized researchers.

The problem becomes more expensive once candidates enter clinical testing.

Clinical trials need to establish whether a medicine is safe, what dose patients should receive and whether the benefits last long enough to matter.

Rentosertib's current Phase III study requires 52 weeks of treatment. Even if the molecular design phase had taken only a few days, researchers would still need substantial time to measure disease progression.

The industry also lacks a mature group of approved AI-discovered medicines from which to calculate the average total cost per successful treatment.

AI is already making specific research activities faster and potentially cheaper. The much bigger productivity gain will come if companies can also reduce failed experiments, avoid ineffective targets earlier and improve the proportion of candidates that survive clinical testing.

Has AlphaFold helped discover a real medicine?

AlphaFold has transformed how researchers study proteins, but no approved medicine has yet been clearly established as the direct result of AlphaFold independently discovering and designing it.

DeepMind's AlphaFold systems made large-scale protein structure prediction dramatically more accessible.

The AlphaFold Protein Structure Database, developed with EMBL-EBI, expanded to include predicted structures for more than 200 million proteins.

That matters because proteins are involved in nearly every biological process, and many medicines work by interacting with them.

AlphaFold 3, described in Nature in 2024, extended the technology toward predicting interactions between proteins, DNA, RNA, small molecules and other biological structures.

Researchers can use these predictions to investigate potential binding sites and understand how a drug candidate might interact with its target.

Isomorphic Labs is now trying to turn that scientific foundation into an industrial drug-design platform.

In February 2026, the company introduced its IsoDDE system and reported stronger performance on selected protein-ligand structure and binding-affinity prediction benchmarks.

Three months later, Isomorphic announced a $2.1 billion Series B financing round led by Thrive Capital, with participation from investors including Alphabet, GV, MGX, Temasek and CapitalG.

The money is intended to expand its drug-design engine and advance therapeutic programs toward clinical testing.

AlphaFold's contribution to biological research is already substantial. However, the ability to predict a protein's shape or a molecule's binding behavior does not establish that a treatment will improve patients' health.

Can ChatGPT or Claude discover drugs without professional scientists?

Advanced general-purpose AI models can now help design molecules that work in laboratory experiments, but they still depend heavily on specialist scientific tools and physical testing.

A particularly revealing experiment was published by Anthropic in October 2026.

Working with Citrini Research, Anthropic used an internal Claude research model to design synthetic protein binders against dangerous scorpion neurotoxins.

The project began with five toxins associated with scorpion species found in regions including Mexico, Brazil, North Africa and the Middle East.

Claude generated 318,484 computational protein designs, evaluated more than 553,000 predicted design-toxin interactions and helped select 211 candidates for laboratory testing.

Researchers at Adaptyv Bio and Twist Bioscience then tested the candidates experimentally.

Seventeen of the 211 designs met the researchers' binding criteria, a hit rate of approximately 8.1%. Successful binders were identified against four of the five toxins.

The computational campaign took approximately 64 hours and involved an estimated 10,000 GPU hours.

Those details put the idea of discovering drugs with a chatbot into perspective.

The researchers used a specialized internal model, multiple scientific tools, substantial computing resources and professional laboratories. This was far beyond typing a simple prompt into a consumer AI application.

The experiment also fell short of its medical objective. None of the designs targeting multiple toxins successfully met the criteria for binding more than one, and the researchers did not demonstrate that the proteins could neutralize venom.

If you want more recent data on this point, please see our latest AI in drug discovery market report.

Are pharmaceutical companies spending billions on AI drug discovery?

Major pharmaceutical companies are committing billions of dollars in potential AI drug-discovery partnerships, although the amounts actually paid upfront are usually much smaller than the headlines suggest.

AstraZeneca's latest collaboration with Insitro provides a clear example.

Announced on October 15, 2026, the agreement focuses on discovering treatments for metabolic dysfunction-associated steatohepatitis, a serious liver disease, and cardiomyopathies affecting the heart muscle.

The agreement has a potential value of approximately $1.5 billion.

However, AstraZeneca is paying only $40 million upfront. The remaining $1.46 billion depends on development and commercial milestones, with additional royalties potentially payable if resulting medicines reach the market.

That means the immediate commitment represents only about 2.7% of the headline value.

Recursion provides another useful comparison. Its partnership with Sanofi generated a new $10 million milestone in October 2026 after Sanofi selected a development candidate from their collaboration.

Recursion reported cumulative upfront and milestone payments of $144 million under that partnership.

Unlike the full potential value of AstraZeneca's agreement, these are payments associated with actual collaboration progress.

Isomorphic Labs represents a third approach: financing the AI discovery platform itself. Its $2.1 billion funding round provides capital to develop computational systems and advance a pipeline before the company has established an approved medicine.

These arrangements show that major pharmaceutical companies and investors are willing to finance AI discovery well before late-stage clinical validation.

However, the amounts have different meanings. A contingent licensing deal, a milestone payment and a venture financing round should never be added together as though they were all revenue generated from successful AI drugs.

Why can AI design good molecules but still fail to make effective drugs?

AI is getting better at designing molecules with promising properties, but predicting how an entire human disease will respond remains much harder than predicting a chemical interaction.

Consider what happens when scientists develop a drug to inhibit a specific protein.

A model may correctly predict that the molecule will bind strongly to its intended target. Laboratory experiments may confirm the interaction, and early human studies may show that the compound reaches the right tissues without causing unacceptable toxicity.

Yet patients may still experience little or no clinical improvement.

That happens because diseases involve many interacting biological processes.

Idiopathic pulmonary fibrosis includes complex mechanisms driving progressive scarring. Cancer cells can develop alternative signaling pathways that reduce treatment effectiveness. Neurological and psychiatric disorders depend on networks of biological processes that researchers still do not fully understand.

Blocking a protein successfully does not guarantee that the disease will respond as expected.

The problem also extends to the data used for training AI models. A compound may perform well in a specific cell line, animal model or laboratory environment without producing the same effect in human patients.

The next major breakthrough may therefore come as much from better experimental data and biological models as from more powerful AI systems.

Which AI-discovered drugs are closest to becoming approved medicines?

Rentosertib is currently among the furthest-advanced novel drugs developed through generative AI-driven target identification and molecular design, while Recursion's REC-4881 is another promising program with human efficacy data.

These candidates deserve more attention than the many AI-generated molecules that remain in laboratory testing.

They also illustrate different types of discovery. Rentosertib involves a new molecular design and an AI-identified biological target. REC-4881 involves an AI-enabled approach to finding a new therapeutic application for a known compound.

Among earlier programs, some have already been discontinued, providing useful reminders that clinical progress is uncertain.

Drug or program	Developer	AI contribution	Current evidence
Rentosertib	Insilico Medicine	Target identification and molecular design	Positive Phase IIa findings; Phase III underway
REC-4881	Recursion	Biological discovery and repurposing	Phase Ib/II polyp reductions
REC-1245	Recursion	Target discovery and molecular design	Early Phase I tumor responses
DSP-1181	Exscientia / Sumitomo	Molecular design	Phase I; discontinued
BEN-2293	BenevolentAI	AI-associated discovery	Phase IIa efficacy disappointment
REC-994	Recursion	AI-enabled drug repurposing	Mixed Phase II results; exploratory signals
Abaucin	McMaster / MIT collaborators	Machine-learning antibiotic discovery	Laboratory and animal evidence
Halicin	MIT-led researchers	Machine-learning antibiotic identification	Preclinical evidence
Scorpion-toxin protein binders	Anthropic / Citrini	AI protein design	17 experimentally validated binders
Isomorphic Labs programs	Isomorphic Labs	AI molecular and interaction modeling	Discovery-stage and development pipeline

Has AI discovered a real drug, or is the industry exaggerating?

Yes, AI has discovered real experimental drugs, including candidates that have already produced measurable effects in human patients. The achievement is genuine, although the industry's broader promises still run ahead of the clinical evidence.

Rentosertib has demonstrated an encouraging lung-function result and reached Phase III testing. Recursion's REC-4881 has reduced gastrointestinal polyp burden in patients, while REC-1245 has produced early tumor responses. AI-assisted antibiotic research has also identified compounds with experimentally confirmed antibacterial activity.

The more difficult claims are still unresolved. No clearly documented end-to-end AI-discovered novel medicine has yet established a complete FDA approval record. AI-associated candidates have disappointed in clinical trials, and the published evidence does not show that AI consistently improves the probability of proving efficacy in patients.

There is also a difference between discovering drugs faster and producing better medicines. Insilico, Exscientia and Recursion have demonstrated faster early research processes, but the industry still lacks reliable evidence that AI substantially reduces the total cost or time required to deliver an approved treatment.

AI drug discovery has already become a real scientific capability with promising clinical results.

But calling it a proven pharmaceutical revolution would be premature. That judgment requires more than one successful candidate. We need repeated approvals, stronger clinical outcomes and evidence that AI-based research produces more useful medicines for the resources invested.

For now, the answer is yes: AI has discovered real drugs. Whether it can reliably turn those discoveries into successful medicines is still the question that matters most.

If you want more recent data on this point, please see our latest AI in drug discovery market report.