AI Drug Discovery Enters Clinical Validation: After AlphaFold, How Are the First AI-Designed Drugs Progressing?

AI drug discovery is undergoing a key identity shift: from technology demonstrations in the lab to drug pipelines that have actually entered human trials. Just over two years after AlphaFold won the Nobel Prize, the first batch of drug candidates designed with AI involvement—or even led by AI—have advanced to the clinical stage, and the industry's yardstick has changed accordingly: no longer model benchmarks, but clinical trial readouts.

AlphaFold Lays the Foundation: From Protein Structure to Drug Design

In October 2024, the Nobel Prize in Chemistry was awarded to Demis Hassabis and John Jumper (Google DeepMind) for AlphaFold's breakthrough in protein structure prediction, alongside David Baker of the University of Washington. It was the first time the Nobel Prize treated an AI system as a core tool of discovery—a symbolic significance far beyond the prize itself.

AlphaFold2 solved a major biological problem in 2020: predicting a protein's three-dimensional structure from its amino acid sequence. Before this, resolving a single structure with X-ray crystallography or cryo-electron microscopy could take months or even years; AlphaFold needed only minutes, with accuracy competitive with laboratory methods. DeepMind subsequently partnered with EMBL's European Bioinformatics Institute to open more than 200 million predicted structures to researchers worldwide, covering the vast majority of the known protein catalog, and they have been used by over 2 million researchers from 190 countries.

Structure matters for drug design because most drugs work by binding to a target protein, like a key fitting a lock. Knowing the precise shape of the lock lets chemists design better keys instead of guessing. AlphaFold 3, released in 2024, further extended predictions to interactions between proteins and DNA, RNA, and small molecules.

Isomorphic Labs and an End-to-End Drug Design Engine

In 2021, DeepMind spun out its drug discovery business as Isomorphic Labs, led by Hassabis, with the goal of having AI carry out the entire drug design process in silico. In February 2026, the company released the Isomorphic Labs Drug Design Engine (IsoDDE); its technical report claims performance on protein-ligand structure prediction is twice that of AlphaFold 3, with binding affinity prediction accuracy higher than physics-based methods, at a fraction of their time and cost.

IsoDDE also tackles problems that older methods struggled to reach, such as discovering "cryptic pockets" on proteins—hidden binding sites that only open in the presence of a drug molecule—and simulating protein conformational changes during drug binding. The company's pipeline focuses on oncology and immunology, with its first AI-designed candidates approaching or entering first-in-human trials, and discovery collaborations with Novartis, Eli Lilly, and Johnson & Johnson. In May 2026, the company raised $2.1 billion, one of the largest private financings ever recorded in biotech.

PDB and Traditional Structure Determination: The Protein Data Bank (PDB), established in 1971, has long been the global archive for experimentally determined structures. By the time AlphaFold appeared, the PDB held roughly 180,000 experimental structures, a substantial portion from X-ray crystallography, while cryo-EM rose rapidly around 2017 thanks to detector technology breakthroughs. The bottleneck of experimental methods lies not only in time and cost but also in the fact that many proteins—especially membrane proteins and intrinsically disordered proteins—are extremely difficult to crystallize or stably image. AlphaFold's 200 million-plus predicted structures therefore do not simply replace experimental data; they fill the structural space that experimental methods could never cover, and their value lies in providing computable starting models—experimental validation remains a necessary step to confirm the conformation of key binding sites. Users should note the confidence metrics attached to predicted structures (such as pLDDT); low-confidence regions often correspond to flexible loops, and using them directly for docking easily produces false positives.

Technical Background on Binding Affinity and Cryptic Pockets: Binding affinity is typically measured by the dissociation constant (Kd) or half-maximal inhibitory concentration (IC50), with smaller values indicating tighter binding between drug and target. Traditionally, physics-based methods such as free energy perturbation (FEP) predict affinity fairly accurately, but each molecule requires hours to days of computation, making large-scale screening difficult. IsoDDE claims to surpass such methods in accuracy using deep learning; its significance lies in turning affinity prediction into a module that can be called at high throughput. A "cryptic pocket" refers to a binding site masked by flexible side chains in the ligand-free state and exposed only in a specific conformation—a classic example being the pocket bound by KRAS G12C inhibitors. Such sites are traditionally overlooked by structure-based screening because they simply aren't visible in static crystal structures. Simulating protein conformational changes involves molecular dynamics, which is extremely computationally expensive; AI surrogate models attempt to approximate this process at lower cost. If these capabilities hold, they directly expand the range of druggable targets.

A Clinical Pipeline Inventory: Who Is Actually Testing AI-Designed Drugs

No matter how elegant the structures and models, the ultimate test of AI drug discovery is whether AI-designed drugs work in humans. By 2026, a real pipeline has taken shape.

Company/Program	AI Focus	Status
Isomorphic Labs (IsoDDE)	Drug design engine, oncology and immunology	Approaching first-in-human trials; partners include Novartis, Eli Lilly, J&J
Insilico Medicine / rentosertib	AI-discovered target (TNIK) and molecule design	Phase IIa completed, Phase III ongoing
Recursion / REC-4881	Imaging-based biology platform	Phase 2 in familial adenomatous polyposis
Recursion / REC-7735	Structure-based AI design, mutation-selective oncology drug	IND cleared, Phase 1/2 expected to start in 2026
Generate:Biomedicines / GB-0895	Generative protein design	Clinical-stage antibody program
Iambic Therapeutics / IAM1363	Physics-informed deep learning	Phase 1, brain-penetrant HER2 inhibitor

Note that these statuses come from company disclosures and peer-reviewed literature as of September 2026, and will change as programs advance or are terminated.

Insilico Medicine: Validating the First Fully AI-Generated Drug Candidate

The most-cited proof of concept comes from Insilico Medicine's rentosertib. Its target, TNIK, is associated with fibrosis and was identified by the company's PandaOmics target discovery platform, while the small molecule itself was designed by the generative chemistry system Chemistry42. The company describes it as the first candidate in which AI led both target discovery and molecule design, rather than assisting at a single step.

The Phase IIa trial in idiopathic pulmonary fibrosis (an incurable progressive scarring disease of the lung) enrolled 71 patients, with results published in Nature Medicine in June 2025. The trial met its primary safety endpoint and showed a dose-dependent efficacy signal: patients in the 60 mg once-daily group saw forced vital capacity increase by an average of 98.4 mL over 12 weeks, while the placebo group declined by 20.3 mL. The drug has since advanced to Phase III. The company also says the program took about 18 months from target identification to preclinical candidate, far shorter than the traditional multi-year cycle.

A September 2026 analysis in Nature Biotechnology offered an interesting footnote: reanalysis of blood samples retained from the trial found that, while improving lung function, the drug shifted six independent proteomic "aging clocks" toward a younger biological age. This is an exploratory biomarker signal and does not prove an anti-aging effect, but it suggests that an AI-discovered target may involve multiple disease biologies at once.

Interpreting Trial Phases and Endpoints: Phase IIa trials are typically small and exploratory, with safety as the primary goal and a preliminary look at efficacy signals; sample sizes are often in the tens, with limited statistical power. Rentosertib's 71-patient trial fits this design, so the "dose-dependent efficacy signal" is hypothesis-generating rather than confirmatory. Phase III is the confirmatory trial, requiring hundreds to thousands of patients, prespecified primary endpoints, and statistical significance thresholds. Forced vital capacity (FVC), mentioned above, is a common measure in idiopathic pulmonary fibrosis trials, but regulatory approval usually also requires assessment of the rate of FVC decline or time to disease progression; an absolute change at a single time point has limited persuasive power. Understanding this helps judge: positive Phase IIa readouts reduce downstream risk, but there is still considerable distance to the conclusion that "the drug works"—which is why the industry emphasizes "look at approval rates, not press releases."

Recursion: Industrial-Scale Imaging and M&A Integration

Recursion Pharmaceuticals took a different path. Its Recursion OS platform is built on automated microscopy: robots capture millions of images of cells under genetic and chemical perturbations, and machine learning models learn the visual signatures of disease biology from them. In November 2024, Recursion acquired its closest AI drug discovery competitor, Exscientia, for $688 million in stock; the combined company says it has more than 10 clinical and preclinical programs.

The merged pipeline shows AI applied at different stages. REC-4881 is a MEK1/2 inhibitor for familial adenomatous polyposis, with median polyp burden reduced by 43% at week 13 and 53% at week 25. REC-7735 is an inhibitor targeting mutant forms of PI3K-alpha, going from first chemical hit to IND clearance in just 10 months, with more than 100-fold selectivity for the mutant protein over the healthy version. By mid-2026, the company reported 5 clinical-stage programs and more than $500 million in realized collaboration payments, with partners including Bayer, Roche/Genentech, Sanofi, and Merck.

Generative Design: How the Underlying Engine Works

These programs share a technical core: generative AI for molecules. Just as image generators produce pictures from text prompts, chemistry models generate drug candidate structures that simultaneously optimize multiple properties—binding affinity, selectivity for the disease target, metabolic stability, and the practical question of whether the compound can actually be synthesized in the lab. Human chemists still validate, test, and optimize, but the chemical space explored in weeks would previously have taken years of manual design.

Cost, Risk, and Regulation: The Hard Questions AI Drug Discovery Must Answer

To understand why AI matters here, look at the economics it is trying to fix. The Tufts Center for the Study of Drug Development estimates the average cost of bringing a new drug to market at about $2.6 billion, including failure costs and the time value of money, with a clinical approval success rate of only 11.83%—roughly 1 in 8.5 compounds entering clinical trials succeeds. Development timelines typically run 10 to 15 years, most of the money is spent on failed candidates, and failures often occur in expensive late-stage trials.

The economic case for AI has two layers: compress early discovery from years to months, and eliminate bad candidates earlier before expensive trials. Better target selection, more precise molecules, and AI-assisted trial design (selecting the right patients, predicting responders) all attack the failure rates that make drug development so expensive. But the industry has yet to prove this at scale, and the 2026 clinical readouts are only the first real test. Setbacks are expected: even AI-guided programs will fail in the clinic, and the field's credibility ultimately depends on approval rates, not press releases.

Regulators are adapting in parallel. In May 2023, the FDA published a discussion paper on the use of AI/ML in drug and biological product development, noting that more than 100 submissions filed in 2021 alone already included AI or ML components, and committed to a risk-based approach. In January 2025, the FDA issued draft guidance on using AI to support regulatory decision-making, structured around "context of use": the more heavily a regulatory decision relies on AI output, the more evidence the model must carry. In January 2026, the FDA and the European Medicines Agency jointly published Guiding Principles for Good AI Practice in Drug Development, with ten principles covering data governance, transparency, lifecycle monitoring, and keeping humans accountable for regulatory decisions.

The thread running through the regulatory stance is consistent: agencies do not prohibit AI in drug development, but shift the burden from "permission" to "proof." Sponsors must document what a model does, demonstrate that it is fit for that purpose, and monitor it over time.

Context of Use and Model Lifecycle Management: The FDA's concept of "context of use" draws on the biomarker qualification framework and means clearly defining the specific role of an AI model's output in a regulatory decision—whether it is used to screen compounds, support dose selection, or form part of the primary efficacy evidence. The more critical the role, the more validation evidence sponsors must submit, including the representativeness of training data, robustness testing of model performance, and performance on external validation sets. Model lifecycle monitoring requires sponsors to continuously track performance drift after deployment, because changes in the distribution of input data can render a model ineffective. The practical impact of this framework is that AI is no longer treated as a black-box tool but as a regulatory object that must be validated, documented, and audited like an analytical method—significantly increasing the documentation burden, but providing a clear path for the acceptability of AI outputs.

Where Things Stand in 2026: The Infrastructure Era Ends, the Clinical Era Begins

AI drug discovery has crossed a watershed. Protein structures have become a solved public good, generative design platforms are in operation in the industry's largest collaborations, and the infrastructure-building phase is essentially over. The clinical phase has begun: the first AI-designed drug candidates are being given to patients, and trial readouts over the next two to three years will determine whether AI has genuinely improved the industry's success rates or merely accelerated the early steps. Either way, that $2.6 billion question now has a new answer being tested in human bodies.