Zero. That is how many people have ever been dosed with a drug that an artificial intelligence designed atom by atom, from nothing, to fit a target it was told to hit. Not discovered by sifting existing compounds. Not optimised from a known molecule. Designed. As of June 2026 the count stands at zero, and the most credible projection in the field is that it will leave zero before the year is out. When it does, it will mark a threshold the previous decade of AI-in-medicine never quite crossed, and the distance between the hype and that threshold is worth measuring precisely.
End of 2026
On May 12, 2026, Isomorphic Labs, the drug-design company Alphabet spun out of Google DeepMind, closed a $2.1 billion Series B. Thrive Capital led it, with Alphabet, GV, CapitalG, Temasek and the UK's sovereign AI fund alongside. The money has one near-term purpose stated plainly by the company: to carry its first wholly AI-designed drug candidates into human clinical trials by the end of 2026.
The precision of that date matters, because it has slipped before. Demis Hassabis, who runs both DeepMind and Isomorphic, had earlier pointed to the end of 2025 for first-in-human trials. That target came and went. The honest reading of the new date is not "AI drugs are here" but "the people best positioned to know now believe they are roughly six months away, having been wrong about the timing once already." Isomorphic's pipeline is concentrated in oncology and immunology, with partnered programmes running alongside Novartis, Eli Lilly and Johnson & Johnson.
So the headline is a forecast, not a result. But a forecast from this vantage point carries weight, because the same field has already produced something the screens cannot dismiss: human data. To understand why end-of-2026 is a real inflection and not just another press release, you have to look at the number that already exists.
Plus 98.4 millilitres
In June 2025, the journal Nature Medicine published Phase IIa results for a drug called rentosertib, developed by Insilico Medicine. It was the first time a compound whose target and chemical structure were both generated by AI had returned controlled human efficacy data. The disease was idiopathic pulmonary fibrosis, a progressive scarring of the lungs that is usually fatal within a few years of diagnosis and has almost no effective treatments.
The headline figure: patients on the 60 mg daily dose showed a mean improvement in forced vital capacity, the standard measure of lung function, of +98.4 mL. The placebo group declined by 20.3 mL over the same window. In a disease defined by relentless loss of lung capacity, a group that gained nearly 100 mL while the untreated group lost ground is a signal worth its decimal places. The trial was small, roughly 71 patients, and Phase IIa is an early read, not a licence. But it established the proof of concept the whole field had been waiting on: a target picked by a machine, a molecule shaped by a machine, tested in human lungs, moving the right number in the right direction.
One distinction has to be held firmly here, because it is the axis the entire story turns on. Rentosertib was AI-discovered: the system identified a biological target (a kinase called TNIK) and generated chemistry to hit it. The molecules Isomorphic is preparing are AI-designed in a stricter sense, built de novo from a computational model of the target's three-dimensional structure rather than found by searching chemical space. The first proves AI can pick winners. The second would prove AI can author them.
AI-discovered means the system searched an existing or generated library and surfaced a promising compound, the way Insilico's platform did for rentosertib. The intelligence is in the selection.
AI-designed (de novo) means the system started from a target's atomic structure and built a molecule to fit it that need not resemble anything in any library, the way Isomorphic's AlphaFold-derived models aim to. The intelligence is in the construction. Crossing from the first to the second, in a human body, is the line 2026 is meant to cross.
From a folded protein to a finished molecule
The reason de novo design was out of reach until recently is that it requires two things at once: an accurate three-dimensional picture of the protein you want to act on, and a way to invent a molecule that grips it. AlphaFold solved the first half. In 2020 it cracked the fifty-year-old protein-structure-prediction problem, work that earned a share of the 2024 Nobel Prize in Chemistry for Demis Hassabis and John Jumper, with the other half going to David Baker for computational protein design. Knowing a protein's shape is the precondition for designing anything to bind it.
The second half, generating the binder, is where 2026 has moved fast. Isomorphic reported in early 2026 that an internal system it calls IsoDDE roughly doubled AlphaFold 3's accuracy on novel targets. The lineage Baker founded, the diffusion-based design tools RFdiffusion and ProteinMPNN, has made it possible to generate proteins and antibodies with functions that do not exist in nature. The field is no longer only reading the language of molecules. It is starting to write in it.
Antibodies are where written-from-scratch design has travelled furthest into the clinic, and the numbers there are no longer preclinical. Generate:Biomedicines, a Flagship company, took an AI-designed anti-TSLP antibody called GB-0895 into Phase 3 in December, the late-stage, large-population trials that sit one step from approval. Two studies, SOLAIRIA-1 and SOLAIRIA-2, are enrolling roughly 1,600 patients with severe asthma. Earlier data showed a single injection every six months lowering the asthma-driving protein without notable side effects. A six-month dosing interval is itself an engineering result: the AI was used to extend the antibody's half-life and sharpen its specificity, not just to find something that binds.
| Programme | Modality / Target | AI role | Stage, mid-2026 |
|---|---|---|---|
| Rentosertib Insilico |
Small molecule, TNIK inhibitor (lung fibrosis) | AI-discovered: target + chemistry | Phase IIa data published (Nature Medicine, 2025); +98.4 mL FVC |
| GB-0895 Generate:Biomedicines |
Antibody, anti-TSLP (severe asthma) | AI-designed antibody (affinity, half-life) | Phase 3 (SOLAIRIA-1/-2), ~1,600 patients |
| Undisclosed Isomorphic Labs |
Small molecules, oncology & immunology | De novo design from structure (AlphaFold lineage) | Pre-clinical; first-in-human targeted by end of 2026 |
Figure 1 — The gradient from discovered to designed, read by clinical stage.
"We want to reduce the time it takes to discover a drug from an average of five to ten years down to a matter of months."— Demis Hassabis, founder of Isomorphic Labs and DeepMind
The readouts that decide it
The next two years contain the only test that counts. Design accuracy, binding affinity, half-life: these are intermediate numbers, and the field has learned to move them. The number that decides whether AI-designed medicine is real is the efficacy readout in a human population, and the calendar between 2026 and 2028 is when the first of those arrive. Generate's Phase 3 asthma data will be among the earliest large-scale verdicts on an AI-designed molecule. Isomorphic's first-in-human results, if the end-of-2026 target holds, will be the earliest read on a de novo small molecule.
It is worth stating precisely what a success would and would not mean. It would mean that a molecule conceived inside a model, with no human intuition about its shape, can be safe and effective in people. It would not mean that drug development has been solved. The slow, expensive, failure-prone parts of the pipeline, the multi-year trials, the manufacturing, the regulatory review, the roughly nine-in-ten attrition rate between a clinical candidate and an approved drug, are largely untouched by better molecule design. AI has compressed the front of the pipeline, the discovery and design that used to take years. The clinical middle still takes about as long as it always has.
And the failure case has to be named with the same precision as the hope, because the field has a recent memory of it. For years the promise was that AI would slash failure rates, and the early clinical record of AI-discovered drugs has been mixed, with some high-profile candidates stumbling in trials exactly as conventionally discovered ones do. A machine can design a molecule that binds its target perfectly and is still toxic, or ineffective, or undone by the messy biology of a living person. The screen does not see the whole body. That is what the trial is for.
Which returns the story to its first number. Zero patients, today. The interesting fact about that zero is not that it is small. It is that, for the first time, a credible date is attached to the moment it changes, and a small number of efficacy readouts in the eighteen months after it will settle an argument the field has been having, on faith, for a decade. The drugs are no longer a thought experiment about chemical space. They are about to become a question that a few hundred patients will answer.
Buy me a coffee