DeepMind's Next Act Isn't AlphaFold — It's Designing the Drug Itself

Predicting a protein’s shape and actually shipping a drug that sells are separated by a deeper valley than most people assume. Isomorphic Labs, DeepMind’s drug-discovery spin-off, is betting it can cross that valley with AI. If AlphaFold was a game-changer for academia, this next round plays out on far colder turf: the pharmaceutical market.

One thing up front. This isn’t a story blowing up on Hacker News this week. It’s a long-dated bet the industry is watching quietly, the kind of thing that gets a polite nod at JPMorgan Healthcare and then goes back to grinding for a decade. So instead of narrating live reactions, this piece digs into the structure of the wager — why the picture matters, and where the real edges are.

AlphaFold Was the Starting Line, Not the Finish

A lot of people treat AlphaFold as the endpoint of drug discovery. It’s closer to the opening whistle.

What AlphaFold cracked was how proteins fold. The proteins in your body fold into three-dimensional shapes, and you have to know that shape before you can decide where to attack. Mapping a single structure used to take months or years and a mountain of lab spend. AlphaFold solved it in seconds, at something close to free.

But here’s where you have to stay cold-eyed. Knowing the target’s shape doesn’t hand you a drug. You still have to design a new molecule that slots into that shape precisely — then prove it’s safe in a body, prove it actually works, and prove you can manufacture it at scale. AlphaFold drew the map. Walking the road is an entirely different problem.

Isomorphic Is Going After the Design Itself

That gap is exactly why Isomorphic Labs was carved out of DeepMind. The pitch: move past the academic trophy of structure prediction and become a company that designs drugs that make money.

The core tech is the AlphaFold3 lineage. Where AlphaFold2 modeled proteins in isolation, AlphaFold3 predicts how a protein and a candidate drug molecule bind together. For a drug developer, that’s a different category of useful. Instead of testing millions of candidate compounds one by one at the bench, you can filter them inside a computer first.

The strategy runs on two tracks. One is partnering with Big Pharma to accelerate their existing pipelines with AI — Isomorphic has already locked in multi-billion-dollar collaboration deals with names like Eli Lilly and Novartis, which buys the ammunition. The other is eventually building its own drug end to end, from scratch. The first track is cash flow. The second is the actual ambition.

So Why Hasn’t Anyone Proven It Yet

This is where the skeptics get loud. And they have decent reasons.

First, the wall of clinical trials. No matter how good the candidate molecule the AI spits out, it still has to prove efficacy and safety in human bodies. That process usually takes years, and most candidates die there. Speeding up the front end with AI doesn’t fast-forward the back end. The old rule of thumb — north of a decade to bring one drug to market — still holds.

Second, the data diet problem. AI is only as smart as what it’s trained on. Existing drug data clusters around targets that are already well studied. The intractable diseases where we most desperately lack drugs are exactly where the data is thinnest. There’s a real risk of getting better and better at the places we already know well — and no better at the ones that matter.

Third, the clinical track record just isn’t there yet. There are dazzling announcements and huge contracts, but the knockout punch — “an AI-designed drug that actually saved a patient” — hasn’t landed anywhere in the industry. That’s not an Isomorphic problem. Every AI drug-discovery outfit in this arena is carrying the same homework.

Somewhere Between Hype and Real Change

So is this just another bubble? I’d be more careful than that.

What’s undeniable is that the cost and time of the discovery phase are genuinely dropping. Finding a promising candidate used to eat years; now that front end compresses dramatically. That’s not marketing — it’s something people feel at the bench. And nudging the base success rate of drug development up even a few percentage points moves tens of billions of dollars across the industry.

At the same time, the cold read is just as clear. AI can make drug development faster, but it can’t erase the fundamental uncertainty of the clinic. “AI replaces drug development” is an overstatement. “AI reshapes the front end of drug development” is close to reality. Telling those two apart is the first step to separating signal from noise.

In the end, whether Isomorphic Labs actually flipped the table won’t be settled by glossy papers. It’ll be settled by a headcount — how many of its molecules are sitting in clinical pipelines with real names a few years from now. So where do you land? The first blockbuster designed by AI — is that news closer than it looks, or still a long way off?