On April 16, at WIRED Health in London, Isomorphic Labs president Max Jaderberg said the company is "gearing up to go into the clinic." Read that sentence again. The verb is "gearing up." Not "starting." Not "in." The most-quoted line out of the talk is also the most carefully-chosen one. It is a company moving its first AI-designed molecules toward the threshold where human biology gets to answer back.
That distinction is the whole story.
The threshold is narrower than the headlines
Isomorphic Labs is the UK biotech spun out of Google DeepMind in 2021, and its lead programs sit in oncology and immunology. The April 16 announcement is that those programs are nearing first-in-human dosing. That is a real milestone and a meaningful one. It is also not the same as "AI is treating patients."
Most coverage collapsed those two things into one. Several outlets ran headlines that read as if AI-designed drugs are already in trials. They are preparing to enter trials. The difference is months of regulatory choreography and is the whole reason Jaderberg used the phrase he used.
The credibility of the announcement comes from the lineage. AlphaFold predicted the structures of virtually all 200 million proteins known to researchers, and the AlphaFold Server crossed three million researchers across more than 190 countries in February. That is the platform Isomorphic Labs is building drug discovery on top of.
What actually changed
Isomorphic's proprietary engine is called IsoDDE, and according to its own technical report, it more than doubles AlphaFold 3's accuracy on the hardest protein-ligand pose-prediction tasks. On the Runs N Poses benchmark for novel pockets and ligands, IsoDDE hits 50% accuracy where AlphaFold 3 hits 23.3%. That gap is what gets a candidate molecule into pre-clinical with enough confidence to schedule a Phase 1.
Demis Hassabis, who runs both DeepMind and Isomorphic, has described the technical core this way: "You need to see how a small molecule is going to bind to a drug, how strongly, and also what else it might bind to." Strip that to its skeleton and you get the claim. The model has to predict not just whether a molecule fits, but how cleanly, and what side effects it might trigger. Get any of those wrong and a beautiful candidate fails in pre-clinical. Get all three approximately right and the candidate is worth testing in humans.
That is the bar Isomorphic says it has cleared.
Isomorphic is not first. It is catching up to a cohort.
Here is the part most coverage left out. Several AI-native drug discovery companies have had candidates in human trials for years.
Insilico Medicine's INS018_055, an AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis, reached Phase 2a in 2024. Recursion Pharmaceuticals, Exscientia, and AbCellera have also moved AI-derived candidates into the clinic. Reaching trials is no longer the news it was three years ago.
What remains true, and what makes the milestone matter, is that no AI-discovered drug has yet been approved by regulators. None has cleared the gate that turns a promising candidate into a product on a pharmacy shelf. The entire AI-pharma cohort is sitting in the same waiting room. Isomorphic is joining them, with the AlphaFold lineage as its calling card.
The bottleneck moves
If AI keeps producing more candidates faster, the scarce resource is no longer the molecule. It is everything that comes after the molecule.
Patients in the right disease populations. Regulatory bandwidth at the FDA, EMA, and MHRA. Trial sites. Data-monitoring committees. Cohort enrollment timelines. The actual machinery of human pharmacology, which has its own clock and does not care that compute got cheaper.
This is what reframes the announcement. Generative biology has been improving fast enough that it is approaching a wall it cannot move on its own. Compute can accelerate hypotheses. Human biology, regulators, and clinicians set the schedule for the test.
Why it matters
The conversation about AI drug discovery has been carried by promise. AlphaFold cracking protein folding. Recursion's high-throughput phenotypic screens. Isomorphic's pharma partnerships with Eli Lilly and Novartis. Each was a real win at a research milestone, and each got covered as if approval was the next natural step.
The clinical threshold is where that simplification breaks. Molecules that look brilliant in software fail in patients all the time, for reasons that have nothing to do with how they were discovered. A candidate that wins on every prediction can collapse on dose, on efficacy in a real disease cohort, on side effects that only appear in long enrollment, on adherence among real patients with real lives. Discovery and approval are connected. They are not the same problem.
That is the part Jaderberg's careful wording acknowledged. "Gearing up" is the honest tense. The model has reached the gate. Now biology, regulators, clinicians, and patients decide what the model was worth.
If AI can design candidates faster than medicine can test them, where does the real bottleneck move?
Originally published as an Instagram carousel on @recul.ai.