We took a close look at Isomorphic Labs, the Alphabet-backed company building an AI drug design engine on the AlphaFold lineage, and what its pharma partnerships reveal about where computational biology actually creates leverage. Inside: why a general-purpose engine beats single-target AI, the ICP that makes partners pay nine-figure upfronts, and what the AI-for-drug-discovery category keeps getting wrong.

Drug discovery is the most expensive way to fail at something. A single new medicine costs roughly $2.6 billion and ten to fifteen years to develop, and about 90% of candidates that enter clinical trials never make it to market. Most of that attrition happens long before a human ever takes a pill. It happens at the bench, where medicinal chemists synthesize molecule after molecule, test them one at a time, and watch most of them do nothing useful against the target they care about. The targets themselves are often the problem. Pharma companies maintain long lists of proteins implicated in cancer, neurodegeneration, and autoimmune disease that have been studied for decades without yielding a drug. They are called "undruggable," which is really just a word for "we have run out of things to try at the lab bench."

The wedge: one engine, every target

Isomorphic Labs' core bet is that drug discovery is not a wet-lab problem that AI can accelerate at the margins. It is an information-processing problem that AI can reframe entirely. Demis Hassabis, the company's founder, has described biology as "an extraordinarily complex and dynamic" information processing system, and the company's name comes from the idea of an "isomorphic mapping" between biology and information science.

That framing is not just philosophical. It determines what they built. The Isomorphic Labs Drug Design Engine is a single, target-agnostic system. It does not need to be retrained from scratch for each new disease area or drug modality. The same engine works on small molecules, antibodies, peptides, and molecular glues. That generality is the wedge, and it is genuinely hard to copy because it requires you to have started from a foundation model lineage that most competitors do not have.

Iso sits on the AlphaFold family tree. AlphaFold 3, published in Nature, was 50% more accurate than the best traditional physics-based methods on the PoseBusters benchmark for predicting drug-like interactions, and it did not require any structural information as input. The IsoDDE system builds on that foundation and more than doubles AlphaFold 3's accuracy on challenging protein-ligand generalization benchmarks. It outperforms Boltz-2 by 19.8x on antibody-antigen structure prediction. AlphaFold itself is now used by over 3 million researchers across 190 countries and earned the 2024 Nobel Prize in Chemistry for Hassabis and John Jumper.

That heritage matters in a market full of AI-drug-discovery startups claiming to transform pharma. Most of them started with a narrow model trained on a specific target class or therapeutic area and are now trying to generalize outward. Iso started with a general-purpose biomolecular model and is pushing inward, toward drug-like accuracy and binding-affinity prediction at a level that beats gold-standard computational methods at a fraction of the time and cost.

The ICP they actually win

Iso's named partners are not mid-tier biotechs looking for a shortcut. They are Novartis, Eli Lilly, and Johnson & Johnson. Three of the largest pharmaceutical companies on earth.

The profile is specific. These are organizations with deep internal computational chemistry teams, massive compound libraries, and decades of institutional knowledge about which targets matter. They are not buying AI because it is fashionable. They are buying it because they have targets their own teams cannot crack. Novartis expanded its collaboration after one year to add up to three additional research programs. Fiona Marshall, President of Biomedical Research at Novartis, said the partnership had already explored "new chemical spaces that would be unavailable to probe through traditional methods."

That quote tells you who the ICP is. A partner who has exhausted conventional approaches on a difficult target and wants to access chemical space that physical experimentation cannot reach. The financial structure confirms it. Eli Lilly paid $45M upfront. Novartis paid $37.5M upfront. Together those deals are worth potentially nearly $3 billion to Iso, excluding royalties. That is milestone-heavy deal architecture. Pharma partners are paying for outcomes, specifically novel candidates against targets they could not drug on their own, and they are willing to structure enormous back-end payments because the unmet need is real and the alternative is years of trial and error that may yield nothing.

With $600 million in its first external round and $2.1 billion in Series B funding led by Thrive Capital with backing from Alphabet, GV, and sovereign AI funds, Iso has the runway to develop its own internal pipeline alongside these partnerships. That dual model matters. It signals to pharma partners that Iso has conviction in its own engine, not just a service to sell.

What the category still gets wrong

The AI-for-drug-discovery category is crowded and largely optimizing for the wrong number. Most players compete on the volume of molecules generated or the speed of virtual screening. Billions of candidate compounds. Millions of dockings per hour. Faster hit identification.

Speed and scale are the wrong scoreboard. The binding-affinity prediction gap is where drug discovery actually fails, and where most AI approaches still lose to careful physics-based methods. Generating a billion plausible molecules is useless if your model cannot accurately predict which of them will bind, how strongly, and whether the interaction will be therapeutically relevant in a living system. IsoDDE's claim to beat gold-standard physics-based methods on binding affinity at a fraction of the cost is aimed squarely at this weakness. The category also over-indexes on single-modality expertise. Many companies have built strong small-molecule engines or strong antibody engines, and then market that depth as a feature. Iso's position is the opposite. One engine spanning modalities is not a compromise. It is the point, because the hardest targets often require you to ask whether a small molecule, a biologic, or a molecular glue is even the right approach before you commit to a discovery program.

The deeper issue is that most AI drug-discovery companies are selling tools. Iso is selling a different epistemology for drug design. The distinction sounds abstract until you look at deal sizes. Tool vendors get software licenses. Companies that reframe the problem get $3 billion partnership structures and their own internal pipelines.

The takeaway

For operators watching this space, the lesson is not that AI will replace medicinal chemistry. It is that the boundary between computational prediction and experimental validation is shifting fast, and the companies that win will be the ones that can demonstrate accuracy on hard, novel, out-of-distribution targets rather than retrospective benchmarks on well-characterized systems. If you are evaluating any AI drug-discovery platform, ask one question: can it generalize to a target it has never seen, in a modality it was not built for, with predictive accuracy that changes what experiments you choose to run? If the answer is no, you are buying a faster version of the old bottleneck.