Hassabis Left DeepMind's Day-to-Day to Build the Drug Engine Beyond AlphaFold

The biggest leadership shakeup in AI research this year was not really about AI research. When Google announced in August 2026 that Demis Hassabis would step back from running DeepMind day-to-day — becoming Chairman and Alphabet’s new Chief Scientist — the question was not whether DeepMind would be fine without him. It was: what does Hassabis think is more important than running the world’s most celebrated AI lab? The answer, it turns out, is drugs.

Why did a Nobel Prize-winning AI researcher leave the AI lab he built?

Because he thinks the bigger prize is on the other side of the research — not building smarter AI, but using smarter AI to build medicines that would otherwise take a decade and a billion dollars to develop through conventional methods.

Fortune reported that Hassabis had increasingly been focused on Isomorphic Labs even before the formal handover at DeepMind. The leadership change made official what had been happening in practice: the person who built one of the most important AI research organizations in history was spending most of his time on a drug-discovery company, and his move to Chairman was structured explicitly to make that possible.

This is not a pivot or a loss of ambition. It is a bet by someone who shared the 2024 Nobel Prize in Chemistry for his AlphaFold work that applying AI to drug discovery is a bigger problem than advancing AI itself. When a Nobel laureate decides to go all-in on turning that science into medicine, it is worth understanding what he is building.

What is Isomorphic Labs actually building?

The company’s core product is the Isomorphic Labs Drug Design Engine, known as IsoDDE. Detailed in a February 2026 technical whitepaper, it builds on AlphaFold 3 to do something AlphaFold alone was never designed for: predict how potential drug molecules interact with target proteins, including precise antibody structure predictions.

Here is why that distinction matters. AlphaFold answers the question “what shape is this protein?” IsoDDE attempts to answer the follow-up: “if I design this particular molecule, will it bind to that protein in a way that produces a therapeutic effect?” That second question is where drug discovery actually happens — and where it has traditionally consumed years of laboratory trial and error, billions of dollars, and extraordinary failure rates.

Columbia computational biologist Mohammed AlQuraishi told Scientific American, as reported by Forbes, that IsoDDE represents “a major advance, on the scale of an AlphaFold 4.” That is notable language from an independent academic — comparing a drug-interaction prediction engine not to an incremental improvement, but to a generational leap on the scale of the original breakthrough.

How much money is behind this bet?

Isomorphic Labs raised $2.1 billion in May 2026 — one of the largest funding rounds in AI drug discovery to date. That figure reflects something more than investor enthusiasm for AI: it reflects a specific thesis that the AlphaFold lineage of technology is ready to move from predicting protein structures (a scientific achievement) to designing drug candidates (a commercial and medical one).

Drug development is not a software problem where a talented team can ship in months. It requires years of computational modeling followed by years of clinical validation — and trials alone are enormously expensive. A $2.1 billion war chest gives Isomorphic Labs the runway to pursue that full pipeline without premature partnerships or compromised timelines.

Why is this harder than predicting protein shapes?

Because protein folding, as hard as it was, is essentially a one-body problem: given this sequence, what shape does it take? Drug interaction is a two-body problem with far more variables: given this protein shape and this candidate molecule, do they fit together? How tightly? At which binding sites? With what side effects on nearby biological processes? And does that interaction produce a therapeutic outcome in a living system, not just in a computational model?

1. Protein-shape prediction (AlphaFold's solved problem)

   Take a sequence of amino acids — essentially a string of letters — and predict the three-dimensional structure it folds into. AlphaFold achieved accuracy competitive with experimental methods, giving researchers a computational shortcut to a shape that previously required months of laboratory crystallography.

2. Drug-interaction prediction (IsoDDE's current problem)

   Take a known protein structure and a candidate drug molecule, then predict how they will interact — whether the molecule will bind to the protein, how strongly, at which sites, and whether that binding will produce a desired biological effect. This is what the Drug Design Engine is built to do.

3. Clinical validation (the next hurdle)

   Computational predictions, no matter how accurate, must be validated in real biological systems. 2026 marked the anticipated start of Phase I clinical trials for in-house oncology and immunology drug candidates designed using the engine — the moment where Isomorphic Labs’ computational work meets actual patients.

The progression from step one to step three is the entire reason Isomorphic Labs exists. Clinical trials are the only test that matters — a drug that works beautifully in a computational model and fails in a human body is not a drug.

What does this mean for how drugs get made?

If IsoDDE’s predictions hold up in clinical trials — and that is still an if, because Phase I trials test safety before efficacy — it would represent a fundamental compression of the drug-discovery timeline. The conventional process screens millions of candidate molecules through laboratory assays, taking years and eliminating the vast majority through brute-force experimentation. A computational engine that reliably predicts which candidates will interact productively with a target protein before any physical testing begins could reduce that phase from years to weeks.

Traditional drug discovery	AI-accelerated drug discovery
Screen millions of candidates in physical assays	Computationally predict promising candidates first
Years of laboratory trial and error for lead identification	Weeks of computational modeling for lead identification
High failure rates at every stage compound costs	Failures identified computationally before expensive trials
A single new drug costs over $1 billion on average	Potential to dramatically reduce pre-clinical costs

That compression matters not just for speed or cost, but for which diseases get treated at all. Many conditions with small patient populations are never pursued by pharmaceutical companies because the expected revenue does not justify the billion-dollar development cost. If AI can reduce that cost by an order of magnitude, diseases that were previously uneconomical to target become viable.

Why should anyone who does not care about AI pay attention to this?

Because this is one of the rare cases where an AI breakthrough might directly, materially affect how long you live or how well you recover from a serious illness. Most AI stories are about productivity, convenience, or economic disruption. This one is about whether a Nobel Prize-winning technology can be extended from understanding biological structures to designing medicines that change clinical outcomes.

Hassabis built his reputation by solving a problem biologists could not solve for fifty years. He won a Nobel Prize for it. And then he decided that solving it was not enough — that the point was never the prediction itself, but what you could build with it. Isomorphic Labs is that next build. The clinical trials anticipated in 2026 will begin to show whether the ambition matches the biology.