Has AI discovered a real drug?
In our AI in drug discovery market deck, you will find everything you need to understand the market
SUMMARY
Yes, AI has discovered a real experimental drug, but it has not yet produced a de novo medicine that regulators have approved and doctors routinely prescribe.
Rentosertib is the clearest test so far because AI contributed to both sides of the discovery problem: choosing TNIK as the biological target and generating a new molecule designed to inhibit it.
The program has moved beyond a clever laboratory result. Rentosertib has a defined structure, a manufacturing process, a dosing schedule, human safety data and an encouraging patient-level efficacy result.
That efficacy result is promising but fragile. The strongest lung-capacity improvement came from only 18 patients treated for 12 weeks, while dropouts and liver-related events became more common at higher exposure.
The new Phase III trial is the real dividing line. It plans to follow 320 patients for 52 weeks, creating roughly 19 times the planned patient-treatment exposure of the earlier study.
AI appears strongest at the part of drug discovery that can be measured and optimized: selecting compounds with workable chemical, safety and pharmacokinetic properties. Early analyses suggest unusually high Phase I survival rates for AI-originated molecules.
The advantage becomes much less clear in Phase II, where a drug must change a complex human disease rather than merely behave like a usable molecule. So far, AI-discovered drugs look roughly ordinary at that stage.
Baricitinib remains an important success, but it proves a narrower point. AI helped identify a valuable new use for an existing medicine; it did not originate the molecule or its original target.
Several early AI-designed candidates reached human testing quickly and then failed for familiar reasons. AI can accelerate the search, but patients still expose weak biology, inadequate efficacy and unacceptable side effects.
Our conclusion is that AI drug discovery is real, not hypothetical. The unresolved question is whether it can repeatedly produce approved medicines faster, more cheaply or with better clinical success rates than conventional discovery.
This market map, featured in our AI in drug discovery market deck, highlights top companies and startups in the AI in drug discovery market
Why can we answer this more confidently now?
Rentosertib has moved AI drug discovery into a real late-stage clinical test.
Insilico Medicine recently started a Phase III trial of rentosertib in people with idiopathic pulmonary fibrosis, a progressive disease that scars the lungs. The study plans to recruit 320 patients across 47 centres in China and treat them for 52 weeks. Researchers will measure how quickly each patient’s lung capacity declines.
That is a major step up from the evidence available before. The drug’s earlier Phase IIa study included only 71 people and lasted 12 weeks. It was built to look for safety problems and early signs of benefit rather than provide a final answer.
Phase III forces the drug to perform over a much longer period, across more hospitals and in hundreds of patients. It also creates enough exposure to reveal safety issues that a small trial might miss.
Rentosertib remains experimental, and Insilico naturally presents its own program in the best possible light. Even so, the trial is real, registered and built around a standard clinical endpoint. AI drug discovery is now facing the same test that decides the future of any serious pharmaceutical candidate.
What should count as an AI-discovered drug?
The label fits when AI makes a decisive discovery choice, such as selecting the target, creating the molecule or uncovering a new medical use.
At the broadest level, almost every large pharmaceutical company now uses machine learning somewhere. An algorithm might predict toxicity, rank compounds, analyze images or help recruit patients. Calling every drug touched by one of these tools “AI-discovered” would make the term useless.
A stronger example is AI-assisted repurposing. The system searches existing medicines and finds one that could treat a different disease. The molecule already exists, but the new medical use can still be a valuable discovery.
AI-designed drugs go a step further. Here, the system proposes or optimizes the actual molecular structure. Scientists then synthesize the suggested compounds and test them in the laboratory.
Rentosertib belongs to the most ambitious category. AI did two central jobs: it ranked TNIK as a target for pulmonary fibrosis and helped generate a new molecule against it. Human researchers still chose the experiments, made the molecule and ran every clinical trial, but the AI systems shaped the original target-and-drug combination.
| Type of AI contribution | What AI actually does | Best current interpretation |
|---|---|---|
| Research assistance | Predicts properties, analyzes data or supports trials | AI-assisted pharmaceutical research |
| Drug repurposing | Finds a new use for an existing medicine | A real therapeutic discovery using AI |
| Molecular design | Creates or optimizes a new compound | An AI-designed drug candidate |
| Target and molecule discovery | Helps choose the biology and design the compound | The strongest form of AI drug discovery |
| Regulatory approval | Produces evidence strong enough for routine medical use | Still unachieved for a de novo AI-originated drug |
If you want more recent data on this point, please see our latest AI in drug discovery market report.
As this chart shows, and as featured in our AI in drug discovery market deck, search interest in AI drug discovery has grown rapidly
Is rentosertib genuinely AI-discovered?
Rentosertib meets the strongest practical definition because AI helped choose both TNIK and the molecule built to block it.
Insilico’s PandaOmics system analyzed biological networks, tissue data, scientific literature, patents and disease pathways. It ranked TNIK as a promising target connected to fibrosis and inflammation. Researchers then tested that prediction in cells and animals.
A second platform, Chemistry42, generated and optimized molecules capable of inhibiting TNIK. Medicinal chemists assessed the proposals, synthesized selected compounds and improved them through repeated experiments. The final candidate became rentosertib, previously known as ISM001-055 or INS018_055.
That story is more convincing than the usual claim that “AI found a drug.” We can point to a specific biological target, a new chemical structure and a documented path from the original computational work to human testing.
AI never worked alone. Scientists decided what data to use, which suggestions deserved laboratory testing and whether each result justified another experiment. The description “AI-discovered” identifies the method that guided the search. It does not imply that a machine independently invented, manufactured and tested the treatment.
Did rentosertib actually improve patients’ lungs?
The highest rentosertib dose improved average lung capacity over 12 weeks, giving the drug its first credible efficacy result.
The Phase IIa study randomly assigned 71 people with idiopathic pulmonary fibrosis to three rentosertib dosing groups or a placebo group. Researchers measured forced vital capacity, or FVC, which tracks how much air a person can forcibly breathe out.
Patients taking 60 milligrams once daily gained an average of 98.4 millilitres of FVC. The placebo group lost 20.3 millilitres. The observed gap between the two groups was therefore close to 119 millilitres.
The dose pattern adds credibility. Patients receiving 30 milligrams twice daily improved by 19.7 millilitres, while those taking 30 milligrams once daily declined by 27 millilitres. Higher exposure generally produced a better result.
The evidence becomes less impressive when we look at the scale. Each treatment group started with only 17 or 18 people. The trial was also too short to show whether patients remained stable, deteriorated more slowly or lived longer. Several other measures, including walking ability and quality of life, failed to show the same clear improvement.
It is a real reason to keep testing, not proof that the treatment works.
| Trial group | Patients at the start | Average FVC change after 12 weeks |
|---|---|---|
| Placebo | 17 | −20.3 mL |
| Rentosertib 30 mg once daily | 18 | −27.0 mL |
| Rentosertib 30 mg twice daily | 18 | +19.7 mL |
| Rentosertib 60 mg once daily | 18 | +98.4 mL |
This chart, featured in our AI in drug discovery market deck, shows annual venture capital investment in AI drug discovery startups
How convincing was the Phase II trial?
The Phase II result was good enough to justify a larger trial. The small sample and treatment dropouts still leave plenty of room for failure.
The study used a strong basic design: patients were randomly assigned, neither they nor their doctors knew who received the drug, and one group received a placebo. Nature Medicine published the full results, allowing outside researchers to inspect the data and limitations.
Only 55 of the 71 participants completed the 12-week treatment period. Sixteen stopped early, including six people in the highest-dose group and six in the twice-daily group. Treatment-related adverse events also became more frequent as the dose increased, reaching 77.8% in the highest-dose group compared with 29.4% for placebo.
Liver toxicity or abnormal liver results contributed to several withdrawals. Some affected patients were also taking nintedanib, an existing pulmonary fibrosis treatment, so the interaction between the two medicines needs closer study.
The trial population creates another limitation. All participants lived in China, the treatment period was short and each dose group was tiny. The authors themselves said that the sample was too homogeneous to settle long-term safety or effectiveness.
The lung-capacity improvement deserves attention. The safety and dropout data deserve equal scrutiny. Phase III now has to show that the benefit survives when the study becomes larger and longer.
If you want more recent data on this point, please see our latest AI in drug discovery market report.
What will Phase III actually decide?
Phase III will decide whether rentosertib can slow lung decline for a full year without creating unacceptable safety problems.
The new study will randomly assign patients to once-daily rentosertib or placebo. Its main endpoint is the annual rate of FVC decline over 52 weeks. Researchers will also track disease progression, gas exchange in the lungs and patient-reported quality of life.
The trial increases planned enrollment from 71 to 320 people. The treatment period grows from 12 weeks to 52. Combining those two changes gives the study roughly 19 times as much planned patient-treatment exposure as the earlier trial.
That larger exposure could produce three very different outcomes. A sustained reduction in lung decline would validate both the TNIK target and the molecule. A weak or inconsistent result would suggest that the early improvement overstated the true benefit. Safety problems could stop the program even if lung function improves.
The geographical scope remains limited because all 47 planned centres are in China. A positive result may therefore require confirmation in a broader international population before the drug gains wide regulatory acceptance.
This chart, featured in our AI in drug discovery market deck, shows how Shrödinger is positioned in AI drug discovery
Has any AI-discovered drug been approved?
No de novo AI-originated drug has received regulatory approval as of now.
Baricitinib, discussed below, gained approval for a use identified with help from AI, but the medicine itself existed long before that discovery. No new molecule widely recognized as having been originated through a modern AI platform has completed the full journey from discovery to approval.
Rentosertib is currently the most advanced clear example. Insilico explicitly describes it as investigational, and no regulator has authorized doctors to prescribe it for pulmonary fibrosis.
Approval sets a much higher bar than reaching Phase III. Regulators need convincing evidence that the treatment benefits patients, that the risks are manageable and that every manufactured batch meets the required quality standards. They also review dosing, labeling, drug interactions and the company’s production process.
The lack of an approval should temper the grandest claims around AI. It does not erase the progress already made. Drug development takes years, and the current generation of AI-originated candidates is only beginning to reach the stages where approvals become possible.
Does baricitinib count as an AI drug discovery success?
Baricitinib is a major AI-assisted success because an algorithm helped uncover a useful new treatment for an existing drug.
BenevolentAI used its knowledge graph to search for medicines that might interrupt the biological processes involved in COVID-19. It highlighted baricitinib, a drug already approved for rheumatoid arthritis, as a possible treatment early in the pandemic.
Randomized trials later showed that baricitinib could help certain hospitalized patients. The US Food and Drug Administration approved it for hospitalized adults requiring supplemental oxygen, ventilation or extracorporeal membrane oxygenation.
This was a meaningful medical result. Reusing an established medicine meant that researchers already understood much of its dosing, manufacturing and safety profile. That allowed large clinical trials to begin far faster than they could have for a newly invented molecule.
Baricitinib proves that AI can find valuable links between existing drugs and new diseases. It tells us much less about AI’s ability to design original chemistry or identify a completely new biological target. Rentosertib is attempting those harder tasks.
If you want more recent data on this point, please see our latest AI in drug discovery market report.
This chart, featured in our AI in drug discovery market deck, shows annual funding in AI drug discovery startups
What happened to the first AI-designed drug candidates?
The first AI-designed candidates reached human trials quickly, and several then ran into the same efficacy problems that stop conventional drugs.
DSP-1181 became one of the earliest widely publicized AI-designed molecules to enter a clinical trial. Exscientia and Sumitomo Pharma developed it for obsessive-compulsive disorder in less than 12 months, compared with the roughly four and a half years they cited for a conventional exploratory program.
The compound entered Phase I, disappeared from Sumitomo’s active pipeline and never became an approved treatment. The program still proved that AI could help generate a clinically testable molecule at unusual speed.
BenevolentAI’s BEN-2293 later entered a Phase IIa trial for atopic dermatitis. The drug met its main safety and tolerability goals, but the secondary efficacy measures failed. BenevolentAI ended further investment in the program.
These results are a useful reality check. AI can shorten the search for a plausible molecule. Patients then expose every weakness in the biological hypothesis.
| Candidate | Intended use | Furthest clear stage | Outcome |
|---|---|---|---|
| DSP-1181 | Obsessive-compulsive disorder | Phase I | Development ended |
| BEN-2293 | Atopic dermatitis | Phase IIa | Safety acceptable, efficacy insufficient |
| Rentosertib | Idiopathic pulmonary fibrosis | Phase III | Pivotal trial underway |
Are AI-discovered drugs better at passing Phase I?
Early evidence suggests that AI-discovered drugs survive Phase I at an unusually high rate.
A Boston Consulting Group analysis published in Drug Discovery Today examined the clinical pipelines of AI-native biotechnology companies. It estimated that AI-discovered molecules had an 80% to 90% Phase I success rate. Historical industry averages are considerably lower, although the exact comparison changes depending on the drugs and years included.
Phase I mainly tests whether a molecule behaves like a usable drug. Researchers look at safety, tolerability, absorption, metabolism and how long it remains in the body. Modern AI systems are well suited to optimizing these measurable chemical properties.
The finding also matches what companies are reporting. Insilico says its platform usually selects a preclinical candidate after synthesizing and testing roughly 60 to 200 molecules. Since 2021, it has nominated 31 preclinical candidates and secured permission to begin human studies for 13 of them. Those are company figures, so they are reported performance rather than independent validation.
The published success estimate still rests on a fairly small sample. Failed or quietly abandoned programs may be harder to identify than successful ones. Even with those caveats, the evidence points toward a genuine strength in early molecular design.
This chart, featured in our AI in drug discovery market deck, compares the main business model options for AI drug discovery biotech companies
Are AI drugs also winning in Phase II?
So far, AI-discovered drugs have shown no clear Phase II advantage.
The same published analysis estimated a Phase II success rate of around 40%, broadly comparable with historical industry performance. The authors warned that the sample was limited, but the gap between Phase I and Phase II was striking.
Phase I asks whether the molecule is safe enough to test. Phase II asks whether it changes the disease. That second question depends on far messier biology.
A molecule can bind perfectly to its target and still fail. The target may play only a minor role in patients. Another biological pathway may compensate. The treatment may reach the wrong tissue, work only in a subgroup or require a dose that causes unacceptable side effects.
BEN-2293 showed this problem clearly. It was safe enough, but patients failed to improve sufficiently. Rentosertib has produced a better result so far, although its small trial leaves the outcome open.
Right now, AI looks better at improving the molecules entering the clinic than at improving our understanding of human disease.
If you want more recent data on this point, please see our latest AI in drug discovery market report.
Is AI making drug discovery faster and cheaper?
AI is clearly shortening target selection and molecular design. Evidence of lower total development costs remains thin.
Rentosertib’s preclinical candidate was selected about 18 months after target discovery began. Phase 0 and Phase I testing were completed within 30 months of the project’s start. The Nature Medicine paper presents this as a major compression of the usual early discovery process.
DSP-1181 reached candidate selection in less than a year. Insilico currently reports an average of 12 to 18 months from starting a discovery program to choosing a preclinical candidate. These examples come from the companies developing the drugs, but their timelines are specific enough to show that early discovery can move faster.
The cost claim is harder to verify. Companies rarely publish complete figures covering computing, proprietary datasets, automated laboratories, failed programs, staff and experiments performed before the official project clock began.
Clinical development also dominates much of the eventual spending. Recruiting hundreds or thousands of patients, operating hospitals, manufacturing trial supplies and satisfying regulators remain expensive. AI cannot compress a 52-week clinical endpoint into a few computer simulations.
For now, the clearest economic gain comes from running fewer weak laboratory experiments and reaching candidate selection sooner. Whether those savings lower the final cost of an approved medicine will depend on how often the candidates succeed later.
This chart, featured in our AI in drug discovery market deck, shows revenue breakdown by customer segment in the AI in drug discovery market
Has AI discovered a real antibiotic?
AI has found powerful antibiotic candidates in laboratory and animal studies, with human testing still ahead.
In 2020, researchers trained a deep-learning model to predict which molecules could kill bacteria. The system identified halicin, a compound previously investigated for diabetes, as a broad-spectrum antibiotic candidate with a structure unlike common antibiotics. Halicin killed several resistant bacteria and worked in mouse infection models.
Researchers later used machine learning to discover abaucin, a narrow-spectrum compound active against Acinetobacter baumannii. This pathogen causes difficult hospital infections and often resists existing medicines. Abaucin reduced infection in a mouse wound model and appeared to disrupt bacterial lipoprotein transport.
Both discoveries were scientifically useful. AI searched chemical spaces too large for researchers to inspect manually and found compounds that conventional screening had overlooked.
Neither example has established safe, effective treatment in people. Halicin was also an existing molecule assigned a new purpose, while abaucin emerged from the screening of known compounds. Their strongest lesson concerns search efficiency: AI can uncover unusual antibacterial activity. Clinical usefulness remains a separate challenge.
Is rentosertib one lucky case or part of a real pipeline?
A growing clinical pipeline now sits behind rentosertib, although most programs remain in their earliest human tests.
Insilico alone reports 13 candidates cleared for clinical trials. Recursion, which combined with Exscientia, continues to run AI-enabled programs in oncology and rare diseases. Isomorphic Labs is building its own internal oncology and immunology pipeline while working with companies including Novartis, Eli Lilly and Johnson & Johnson.
Exact industry totals are unreliable. Some trackers count any molecule substantially shaped by AI. Others include repurposed drugs, vaccines, antibodies, trial-design software or programs where AI played a relatively small role. Depending on the definition, published figures range from several dozen clinical molecules to well over one hundred.
We do not need an inflated total to see the trend. Multiple AI-native companies have produced clinical candidates, and several large pharmaceutical groups are now paying to use these platforms. The pipeline is broad enough that success rates should become easier to measure over the next few years.
Rentosertib remains unusual because it has reached a confirmatory trial after showing a patient-level efficacy result. Most of the wider pipeline has only demonstrated that a molecule can enter the body safely.
This chart, featured in our AI in drug discovery market deck, shows how AI drug discovery platform technology has evolved over time
What would prove that AI drug discovery really works?
One approval would prove that AI can originate a medicine. Repeated approvals with better timelines or success rates would prove an enduring advantage.
The first test is straightforward. An AI-originated candidate must complete Phase III, receive regulatory approval and provide enough clinical value for doctors to use it. Rentosertib could reach that threshold, although its pivotal result is still unknown.
A single approval would carry enormous symbolic importance. Pharmaceutical history also contains many one-off successes produced by methods that never became consistently superior.
The broader test requires comparison. We need to know whether AI-originated programs reach candidate selection faster, synthesize fewer molecules, cost less and survive clinical trials more often than similar conventional programs. Disease difficulty, company quality and drug type must be comparable, or the analysis becomes misleading.
Phase I data already hint at better molecular quality. Phase II data currently look ordinary. Phase III remains almost empty territory for this generation of drugs.
The field will deserve its strongest claims when those three stages tell the same story.
Has AI discovered a real drug?
Yes, AI has discovered a real experimental drug.
Rentosertib has a novel structure, a defined target, a manufacturing process, a dosing schedule and human clinical data. AI materially influenced the discovery of both the target and the molecule. That clears any reasonable threshold for calling it a real drug candidate.
The evidence also sets a firm limit. Its encouraging patient result came from 18 people receiving the highest dose for 12 weeks. Dropouts and liver-related events need more investigation, and several secondary measures failed to confirm the lung-capacity improvement.
Rentosertib is now undergoing the late-stage test that will determine whether its early result holds up. No AI-originated molecule has yet completed that test and become an approved medicine.
Our final judgment is mostly true, with one essential qualification. AI has discovered a real drug in the scientific and clinical-development sense. It has yet to discover a medicine that regulators approve and doctors routinely prescribe.
That remaining step is the hardest one in drug development. It is also the only one that ultimately counts for patients.
If you want more recent data on this point, please see our latest AI in drug discovery market report.
In our AI in drug discovery market deck, we identify pain points entrepreneurs should prioritize
OUR METHODOLOGY
This analysis tests whether AI has discovered a real drug by separating the claim into the milestones that actually determine the answer: target identification, molecular design, preclinical validation, human safety, clinical efficacy and regulatory approval.
We use rentosertib as the central case because AI contributed to both the selection of TNIK and the generation of a new molecule designed to inhibit it. We treat that as a stronger form of AI discovery than research assistance or the repurposing of an existing medicine.
We keep different types of evidence separate. A fast discovery timeline does not prove clinical efficacy, a successful Phase I trial does not prove that a drug changes disease, and entry into Phase III does not equal regulatory approval.
The Phase IIa evidence is assessed using the published randomized trial design, FVC results, patient numbers, treatment discontinuations and adverse-event data. The Phase III program is assessed using its registered enrollment, duration, locations and primary endpoint.
Industry-level claims about clinical success rates are treated cautiously because the number of AI-originated drugs remains small and unsuccessful programs may be less visible. Company pipeline figures are identified as reported performance rather than independent validation.
We also compare rentosertib with other kinds of AI-enabled discovery. Baricitinib represents AI-assisted drug repurposing, while halicin and abaucin show how machine learning can uncover unusual antibacterial activity before human testing.
Key sources include the ClinicalTrials.gov registration for the rentosertib study, the Nature Medicine paper reporting the Phase IIa results, Insilico Medicine’s rentosertib disclosures, the PandaOmics platform description, the Chemistry42 platform description, the US Food and Drug Administration’s drug approval framework, and the European Medicines Agency’s medicines-development guidance.
Additional evidence comes from Eli Lilly’s prescribing information for baricitinib, BenevolentAI’s program disclosures, Exscientia’s program disclosures, Sumitomo Pharma’s pipeline materials, the AI clinical-success analysis published in Drug Discovery Today, the MIT summary of the halicin discovery, the underlying Cell paper on halicin, and the Nature Chemical Biology paper on abaucin.
For the broader pipeline, we use disclosures from Recursion, Isomorphic Labs, Novartis, Eli Lilly, and Johnson & Johnson Innovative Medicine. We prioritize sources that provide specific trial, regulatory, molecular or pipeline information over general claims about AI transforming pharmaceuticals.
This chart, featured in our AI in drug discovery market deck, shows revenue breakdown by region across Europe, Asia, North America, Africa, and South America in the AI in drug discovery market
Related blog posts
- Has any longevity drug worked yet?
- Who is closest to a real anti-aging drug today?
- Who will deliver the first in-body CRISPR cure?
