The World's Best-Funded AI Drug Designer Hasn't Treated a Single Patient Yet

IsoDDE doubles AlphaFold 3 accuracy in lab tests, but Isomorphic's first trial target is weeks away

On November 2, Isomorphic Labs President Max Jaderberg told an audience at Bloomberg Tech London that artificial intelligence will eventually design every medicine a human being takes. It was the company's most expansive public claim since CEO Demis Hassabis founded the Alphabet-backed laboratory five years ago on the proposition that AI could "solve all disease."

What Jaderberg did not say is that as of this week, Isomorphic Labs — backed by $2.7 billion in venture capital, partnered with Eli Lilly, Novartis, and Johnson & Johnson, and built on the Nobel Prize-winning AlphaFold lineage — has not yet dosed a single patient with any drug it has designed.

That gap between aspiration and clinical reality defines the most consequential question in AI drug discovery as 2026 runs out: whether a computational system that is demonstrably superior to anything available five years ago can produce evidence in a human being. The company's self-imposed deadline to begin its first Phase I trial is approximately eight weeks away.

How IsoDDE Extends AlphaFold Into a Drug Design Engine

The technological foundation of Isomorphic's clinical ambitions is IsoDDE — the Isomorphic Drug Design Engine — published as a technical report on February 10, 2026. The platform extends AlphaFold 3, published in Nature in May 2024, from protein structure prediction into an end-to-end drug design pipeline. The distinction is substantive: AlphaFold 3 tells you what a protein looks like in three dimensions; IsoDDE is intended to tell you what molecule will bind to it, how tightly, and where.

IsoDDE builds on four interlocking capabilities. The first is protein-ligand structure prediction — modeling how a small-molecule drug candidate physically fits into a target protein's binding pocket, including dynamic conformations that static structural models miss. The second is antibody-antigen interface modeling, extending the same geometric reasoning to biologics and immunology targets where the binding geometry involves much larger molecular structures. The third is binding affinity estimation: a numerical prediction of how strongly the molecule will capture its target, achieving a Pearson correlation of 0.89 against experimental binding free energy data, compared to 0.72 for AlphaFold 3 and 0.55 for traditional physics-based docking software, according to the company's technical report. The fourth capability is ligandable pocket identification — finding from amino acid sequence alone where on a protein's surface a drug might realistically bind, including so-called cryptic pockets that are not visible in static crystal structures.

The generalization benchmark result is what Isomorphic considers most significant. On a test measuring prediction accuracy for protein-ligand pairs with less than 20 percent sequence similarity to anything in training data — the hardest practical measure of whether the system understands chemistry rather than recognizing it — IsoDDE achieves approximately 50 percent accuracy against roughly 23 percent for AlphaFold 3. On FoldBench, an antibody-antigen structural prediction task, it reaches 75.58 percent versus AlphaFold 3's 47.90 percent. These numbers come from company-reported evaluations rather than independent peer review — IsoDDE has not been published in a journal — a distinction that matters when assessing what has been externally validated versus what has been internally benchmarked.

A Generative Design Loop That Runs in Days, Not Years

In September 2026, Isomorphic published an account of an automated drug design experiment intended to demonstrate the platform's generative loop in practice. Starting from a biological target, the system generated candidate molecules, scored them against predicted binding and selectivity criteria, selected the most promising, routed them for synthesis, and returned in-vitro activity data — all within two to four days per design iteration.

One candidate from that experiment showed measurable activity against the intended target. That is a meaningful demonstration: a fully automated pipeline completing a design cycle without human chemists making decisions at each step. What it is not is a drug, or even a preclinical candidate. In-vitro activity — a molecule doing something detectable in a cell culture — is the entry point to a development process that typically costs $2 billion, requires twelve to fifteen years, and fails over 90 percent of the time before reaching a patient.

What the September experiment also makes clear is what IsoDDE's public documentation does not cover: full ADMET profiling. ADMET refers to absorption, distribution, metabolism, excretion, and toxicity — the five pharmacokinetic dimensions that eliminate the vast majority of otherwise promising drug candidates between early lab work and clinical trials. The IsoDDE technical report addresses structure, binding affinity, pocket identification, and antibody interfaces. Whether Isomorphic's platform handles ADMET prediction systematically has not been disclosed.

A Deadline That Already Moved Once

Isomorphic's stated goal is to file its first Investigational New Drug application and begin first-in-human Phase I safety testing by the end of 2026. That deadline is now under two months away.

The original target was the end of 2025. At the World Economic Forum in Davos on January 20, 2026, Hassabis acknowledged the target had shifted by twelve months. Isomorphic has not publicly named a clinical candidate or confirmed when it expects to dose its first patient.

The pharmaceutical development timeline is dotted with announced milestones that slipped. Isomorphic has at least $2.7 billion in capital and three major pharmaceutical collaborations whose valuation rationale is tied directly to whether these programs reach the clinic. Bloomberg reported in October 2026 that the company was in talks for a new funding round at a valuation of $40 billion or more — a figure that would be entirely unsupported by clinical evidence if the year-end deadline is missed again.

Insilico Medicine Reached Phase III First

The company that currently sets the clinical benchmark for AI-designed drugs is not Isomorphic Labs. It is Insilico Medicine.

Insilico's rentosertib — a small-molecule TNIK inhibitor designed by its Pharma.AI platform for idiopathic pulmonary fibrosis — entered Phase III trials in July 2026, making it the first AI-designed drug to reach Phase III in any indication. The same company's ISM6331, a candidate for mesothelioma, received FDA Fast Track designation and is in Phase I. Insilico built its clinical position using a platform that integrates AI-driven target identification with molecule generation — a broader scope, pursued over a longer commercial runway, with less capital than Isomorphic has raised.

Recursion Pharmaceuticals, which absorbed Exscientia in 2024, has five to six programs in Phase I or II trials. Xaira Therapeutics, co-founded by Nobel laureate David Baker with $1.3 billion in initial capital, is building toward clinical programs in protein therapeutics. Industry-wide, the count of AI-assisted drug programs in clinical development now exceeds 170, though most represent AI contributing to target selection or early screening within traditional pharmaceutical workflows rather than generating the molecule entirely computationally.

Isomorphic's competitive distinction is not clinical seniority but the depth of its foundation model capabilities. No other drug discovery organization is building from the AlphaFold lineage at this scale, with Alphabet's research infrastructure and pharmaceutical partnerships of this magnitude feeding proprietary data back into the system.

The Gap Between Computational Excellence and Clinical Proof

Drug discovery has a biological complexity problem that no foundation model has yet resolved. A molecule can demonstrate ideal predicted binding geometry, strong computed affinity, and clean ADMET forecasts and still fail in a patient — because of off-target interactions that did not appear in training data distributions, immunogenic responses, metabolic conversion to toxic intermediates, or simply the irreducible difference between a cell assay and a living human system.

The same overconfidence risk documented in large language models applies to molecular AI: systems can produce confident, specific numerical outputs — predicted toxicity values, permeability estimates — that reflect training distribution patterns rather than the actual behavior of a novel compound. The pharmaceutical industry cannot discover these errors cheaply: each program that reaches Phase I represents tens of millions of dollars in development costs before efficacy data exists.

The large pharmaceutical companies partnering with Isomorphic are not betting that AI will replace medicinal chemistry. They are betting that AI narrows the chemical search space aggressively enough to reduce the number of compounds requiring synthesis before a tractable clinical candidate emerges. That is a meaningful value proposition. It is also not the same thing as designing every medicine a human will ever take.

Eight Weeks to Turn a Forecast Into Evidence

Whether Jaderberg's prediction about AI's long-run trajectory in drug design is correct — and over decades it may well be — the near-term test is narrower and specific. By January 2027, Isomorphic Labs will either have administered its first AI-designed molecule to a human patient or will have missed its own target for the second consecutive year. The outcome will determine whether the world's best-funded AI drug design laboratory has crossed the clinical threshold its competitors cleared first, and whether a $40 billion valuation narrative has any biological evidence to stand on.