AI-Designed Drugs Are Entering Human Trials: Inside the Quiet Pharmaceutical Revolution of 2026

For most of pharmaceutical history, drug discovery has been a long, expensive, failure prone act of scientific detective work. A medicinal chemist would propose a molecule. A team would synthesize and test it. A decade later, with luck, one in ten candidates might survive the journey to patients. The average approved drug costs somewhere north of 2 billion dollars to bring to market, and around 90 percent of drug candidates that enter clinical trials still fail.

Presented By Our Partners

That long apprenticeship is beginning to end.

In the past eighteen months, a quiet shift has moved through the world’s drug development pipelines. Molecules designed almost entirely by artificial intelligence have entered mid stage human clinical trials. Some are being developed by publicly traded biotech firms with real oncology and fibrosis programs. Others are emerging from the research subsidiaries of Google DeepMind and from university labs that won the 2024 Nobel Prize in Chemistry. By the second quarter of 2026, the central question in pharmaceutical research is no longer whether AI can design a usable drug. It is whether AI designed drugs can survive the gauntlet of Phase 2 and Phase 3 trials, and whether the economics of discovery will finally bend.

What AI Actually Changes About Drug Discovery

To appreciate the shift, it helps to understand what AI is being asked to do.

Drug discovery breaks roughly into three stages. First, researchers identify a biological target, typically a protein whose activity they want to block, boost, or modulate. Second, they search for molecules that can bind to that target with high specificity and favorable drug like properties. Third, they optimize those molecules for safety, potency, bioavailability, and manufacturability, then run a long sequence of preclinical and clinical tests.

Traditional methods have relied on high throughput screening libraries of millions of compounds, trial and error medicinal chemistry, and an enormous amount of tacit human expertise. AI does not replace any of that. What it compresses is the search space. Modern generative models can propose molecules that are likely to bind to a given target and likely to behave like drugs, without having to physically synthesize and screen hundreds of thousands of failures first.

A 2023 analysis by the Boston Consulting Group, updated in 2025, found that the median time from target identification to preclinical candidate nomination has fallen from roughly 4.5 years using traditional methods to under 18 months at AI native firms. The early stage failure rate has also dropped, although later stage clinical failure rates remain almost unchanged. This pattern of faster early discovery followed by the same brutal clinical attrition is the key puzzle AI companies are now trying to solve.

AlphaFold and the Structural Biology Breakthrough

The transformation began with a problem that had defied molecular biology for fifty years: predicting how a protein folds from its amino acid sequence.

In 2020, DeepMind’s AlphaFold 2 solved that problem to near experimental accuracy at the Critical Assessment of Structure Prediction competition. By 2022, the AlphaFold Protein Structure Database contained predicted structures for more than 200 million proteins, essentially every cataloged protein across every known organism. The database is now used by more than two million researchers worldwide, and it has become the default starting point for hypothesis generation in structural biology.

In May 2024, DeepMind and Isomorphic Labs released AlphaFold 3 in Nature. The new model went beyond predicting single protein structures. It could model interactions between proteins and small molecules, proteins and DNA, proteins and RNA, and protein complexes with antibodies. John Jumper, the lead AlphaFold researcher, and Demis Hassabis, DeepMind’s CEO, shared the 2024 Nobel Prize in Chemistry for that body of work, along with University of Washington biochemist David Baker.

The significance of AlphaFold 3 for drug discovery is not subtle. If a model can predict, at atomic resolution, how a candidate molecule will sit inside the binding pocket of a disease relevant protein, then the initial screening phase of drug discovery becomes a computational problem rather than a synthesis and assay problem. It does not eliminate the need for wet lab validation. It does shrink the number of wet lab experiments required by orders of magnitude.

Insilico Medicine and the First End to End AI Designed Drug

The most closely watched real world test of AI driven drug discovery is unfolding at Insilico Medicine, a Hong Kong and New York based biotech founded by Alex Zhavoronkov.

Insilico’s lead asset, rentosertib, is a small molecule inhibitor of TNIK, a kinase implicated in fibrotic diseases. The company identified the target using its PandaOmics platform, then designed rentosertib using Chemistry42, its generative chemistry engine. The entire process from target discovery to preclinical candidate took under 18 months, a timeline that would have been unthinkable a decade ago.

In 2024, Insilico reported results from the Phase 2a trial of rentosertib in idiopathic pulmonary fibrosis, with full findings later published in Nature Medicine in 2025. The study enrolled patients with this progressive and often fatal lung scarring disease. The highest dose arm showed a clinically meaningful improvement in forced vital capacity over twelve weeks compared with placebo, on the order of 98 milliliters. Forced vital capacity is the standard regulatory endpoint for pulmonary fibrosis drugs. Rentosertib became the first fully AI designed and AI discovered small molecule to demonstrate benefit in a randomized Phase 2 trial.

Featured Partner

Invest in the Infrastructure Behind Modern Medicine

As healthcare expands beyond hospital walls, the buildings and campuses supporting that shift are generating compelling returns for investors who move early. The Healthcare Real Estate Fund offers qualified investors direct access to a curated portfolio of medical office, outpatient, and specialty care facilities.

Learn More →

Rentosertib is now progressing toward Phase 2b and Phase 3 studies. It is not yet approved, and as in any mid stage program the outcome is uncertain. What matters for the field is proof that a molecule conceived, optimized, and nominated by generative algorithms can clear a rigorous human efficacy threshold at all.

Isomorphic Labs and the Pharma Partnership Bet

If Insilico represents the biotech path, Isomorphic Labs represents the pharma alliance path. Spun out of Google DeepMind in 2021, Isomorphic Labs was founded specifically to apply AlphaFold class models to commercial drug discovery.

In January 2024, Isomorphic Labs signed research collaborations worth more than 2.9 billion dollars in potential milestone payments. Eli Lilly committed up to 1.7 billion dollars to co develop multiple small molecule therapeutics, and Novartis committed 1.2 billion dollars across three discovery programs. These are not the token AI partnerships of the 2010s. They are mid size acquisitions in structure, signaling that the world’s largest pharmaceutical companies now view generative structure prediction and generative chemistry as core inputs to their discovery operations.

The scientific bet underlying these deals is that models like AlphaFold 3 can meaningfully reduce attrition in lead optimization, a stage where tiny changes in molecular structure can completely change pharmacokinetics, off target binding, and toxicity. Traditional medicinal chemistry iterates this process manually, often for several years. If a well calibrated structural model can predict which modifications will preserve target engagement while improving drug like properties, the cycle can be compressed dramatically.

Isomorphic’s first wholly in house drug candidates are expected to enter clinical testing within the next year. The company has not publicly named indications, but its partnerships span oncology, immunology, and neuroscience targets.

David Baker and the Drugs That Do Not Exist in Nature

If AlphaFold answers the question of how existing proteins fold, the Baker Lab at the University of Washington’s Institute for Protein Design has been answering a different one. Can we design entirely new proteins from scratch that do things nature never attempted?

The 2024 Nobel Prize in Chemistry recognized exactly this. Baker’s group has spent two decades refining computational methods, including RoseTTAFold and the diffusion based RFdiffusion, to generate novel protein sequences with predicted structures optimized for specific functions. That includes binding pockets for disease targets, antibody like molecules that can be designed in hours rather than evolved over months, and vaccines whose components are computationally tuned to the exact epitopes the immune system needs to see.

Several Baker Lab spinoffs are now in or near the clinic. Icosavax, acquired by AstraZeneca in early 2024 for up to 1.1 billion dollars, is advancing a computationally designed RSV vaccine. Outpace Bio, Monod Bio, and others have moved into cell therapy and computational antibody programs. A 2024 paper in Nature from Nathaniel Bennett and colleagues at the Baker Lab demonstrated de novo antibody design against specific protein targets, a capability that had previously required animal immunization and months of laboratory optimization.

By 2026, protein design has moved from static binding proteins to dynamic conformational switches, enzymes with custom reaction chemistries, and therapeutic antibodies that outperform anything ever selected by natural evolution. The implications go beyond speed. These are molecules that do not exist anywhere in biology and would never be discovered by screening.

Generative Chemistry Beyond Structural Biology

Not every AI approach to drug discovery runs through protein structure. A parallel track, often called generative chemistry, focuses on the molecules themselves.

Recursion Pharmaceuticals, which merged with Exscientia in late 2024, combines high content cellular imaging with deep learning to identify compounds that produce desired phenotypes, even when the molecular target is unknown. Iambic Therapeutics uses physics informed neural networks to design selective kinase inhibitors, and its lead compound IAM1363, a HER2 selective inhibitor, is in Phase 1 testing. Generate Biomedicines, backed by Flagship Pioneering, has built generative models for therapeutic proteins and antibodies and has multiple clinical stage programs in oncology and immunology. Schrodinger, a long running computational chemistry platform, has combined physics based simulation with machine learning to advance internal and partner programs, including SGR 1505, a MALT1 inhibitor in hematologic cancers.

What ties these approaches together is a different view of the molecule design process. Instead of starting from a known scaffold and making incremental modifications, generative models propose entirely novel chemical structures that meet multiple constraints at once. Solubility, selectivity, off target avoidance, and synthetic accessibility are all scored simultaneously. The best candidates move forward not because a chemist believed in them, but because the model judged them as the highest probability winners across a multi dimensional design space.

The Hard Realities: Where AI Still Falls Short

Despite the progress, it would be a mistake to treat AI drug discovery as a solved problem.

The first difficulty is data. Generative models are only as good as the training data underneath them. For common drug targets such as kinases and G protein coupled receptors, there are decades of structural and pharmacological data. For newer target classes, including intrinsically disordered proteins, membrane complexes, and transient RNA structures, data are sparse and models tend to hallucinate plausible looking but invalid solutions.

The second difficulty is clinical validation. Compressing the discovery timeline is only half the pharmaceutical problem. The other half, clinical trial attrition, has not yet budged. A 2025 analysis by the biopharma research firm Evaluate Pharma found that AI originated drug candidates still experience roughly the same Phase 2 failure rates as traditionally discovered candidates, near 60 percent. Efficacy in humans remains far harder to predict than binding affinity in a test tube.

The third difficulty is the innovation versus incrementalism question. Critics such as Derek Lowe, the medicinal chemist who writes the long running In the Pipeline blog, have argued that many AI generated drug candidates are modest variations on known chemical series, not genuine leaps into new chemical space. The field’s response has been to push generative models further from training distributions, accepting higher synthesis failure rates in exchange for genuinely novel chemistry. Whether those novel molecules translate into better clinical outcomes is the open question of the next five years.

The fourth difficulty is regulatory. The United States Food and Drug Administration and the European Medicines Agency have begun publishing guidance on AI and machine learning in drug development, but frameworks for approving drugs discovered, optimized, or justified primarily by AI models are still evolving. Transparency, reproducibility, and explainability requirements will shape how AI derived evidence is weighted alongside wet lab data in regulatory submissions.

What Comes Next

The next two years will bring the first large scale readouts that tell the field whether AI drug discovery is a genuine productivity breakthrough or a more modest efficiency upgrade. Watch for three signals. First, the progression of rentosertib and other first wave AI designed small molecules through Phase 2b and Phase 3 in fibrosis, oncology, and immunology. Second, the first wholly Isomorphic Labs originated molecule to enter human trials and report binding and tolerability data. Third, the emergence of AI designed therapeutics for indications where traditional pharma has repeatedly failed, including Parkinson’s disease, amyotrophic lateral sclerosis, and complex metabolic disorders.

If even half of the current AI originated clinical programs produce approved drugs, the industry economics shift meaningfully. If the clinical failure curve matches historical norms, AI will still have proven useful in compressing early discovery, but the promise of radically cheaper medicines will take longer to materialize.

What This Means For You

For patients, the near term practical implications fall into three categories.

First, faster and broader discovery is likely to accelerate the number of new drug candidates entering trials for rare and poorly served diseases. Traditional pharma economics have favored blockbuster indications where a single approved drug can recoup development costs across millions of patients. AI driven discovery lowers the cost of the earliest stages enough that smaller indications may become economically viable, including rare cancers, pediatric conditions, and neglected tropical diseases.

Second, precision medicine and AI discovery are converging. When a drug is designed computationally against a precise target, it becomes easier to identify the subset of patients whose biology matches that target. This pairing of AI designed therapeutics with biomarker driven patient selection is already reshaping oncology trials, and within a few years it will touch cardiovascular and neurological conditions as well. If you have access to genomic testing or a family history of a complex condition, staying in touch with clinical trial registries such as ClinicalTrials.gov is becoming more valuable, not less.

Third, healthy skepticism remains appropriate. An AI designed drug that has completed Phase 1 testing is not automatically safer or more effective than a traditionally discovered one. The same biology, pharmacology, and clinical trial standards apply. The headlines in 2026 will trumpet more milestones, but only a few of those molecules will survive Phase 3 and reach pharmacy shelves. If you are hearing claims that a particular AI derived therapy will soon transform care, check the trial status, read the published data, and talk with your clinician about what the real evidence says.

The quiet revolution in pharmaceutical discovery is real. It is also still early. The first wave of AI designed drugs is now in human trials, and the next decade will reveal whether they deliver the outcomes that patients, clinicians, and investors have been promised.

Free Daily Briefing

The Latest Longevity Science.
Delivered Every Morning.

Join researchers, physicians, and health professionals getting daily breakthroughs in AI-driven medicine, epigenetics, and longevity research.

Support the research that powers this editorial

Subscribe Free

No spam. Unsubscribe anytime. We respect your inbox.