Stephen Quake on Genomics, AI Cells, and Medicine's Future

Stephen Quake has a physicist's instinct for finding the measurement that changes everything.

His PhD is in theoretical physics. His career landed in biology. And that gap — the deliberate, stubborn crossing of it — turns out to be responsible for at least three things that matter enormously to how medicine works right now. In a recent interview with ZME Science at the 2025 Falling Walls Conference in Berlin, the Stanford professor of bioengineering and applied physics walked through that career arc with the calm of someone who knew they were right before anyone else did.

The throughline isn't genomics, or diagnostics, or AI — it's the idea that biology needed a better philosophy of measurement. Physics had a whole field dedicated to it: precision measurement. Biology, Quake says, had nothing equivalent. So he went looking for ways to apply rigorous measurement to living systems, one breakthrough at a time.

---

A needle. Then a question.

The origin story of Quake's most publicly impactful work starts not in a lab but in a doctor's office. When he and his wife were expecting their first child, their doctor recommended amniocentesis — a procedure where a long needle is inserted through the abdomen into the amniotic sac to extract fetal DNA for genetic testing. Quake describes watching it happen:

"It was terrifying. A giant needle right into her belly and you know you're risking the life of the baby to ask a diagnostic question and I thought that's just ridiculous."

That question — why would you accept that risk when there might be another way? — stayed with him. The answer came when he encountered a phenomenon called cell-free DNA: the idea that when cells die, they shed fragments of their DNA into the bloodstream. Quake says this had been known as a phenomenon for decades but wasn't widely appreciated or in the textbooks. The insight was that if you're pregnant, some of that circulating DNA comes from the placenta and fetus — meaning you could read the fetal genome from a simple blood draw.

That became non-invasive prenatal testing (NIPT). Quake claims upwards of 10 million women a year now receive some version of it, and that the use of amniocentesis has dropped sharply as a result. That's not a minor clinical footnote. That's a procedure that carried real risk — to wanted pregnancies — becoming largely obsolete.

---

The $100 genome

The NIPT story is the one that tends to move people emotionally. But the sequencing story might matter more at scale.

When Quake co-founded a company to commercialize a new approach to single-molecule DNA sequencing, the reception was skeptical. People didn't think it would work. Instruments weren't selling. So in 2008, he sequenced his own genome on one — becoming, he confirms in the interview, the fifth person in the world to have their complete genome sequenced — and published the results to prove it could be done.

The proof of concept landed hard. At the time, he says, it was the cheapest complete genome by an order of magnitude, and required three authors instead of a hundred, one machine instead of a warehouse full.

The interviewer notes that the Human Genome Project cost close to $3 billion over its lifetime. Quake confirms the figure, then offers the contrast: sequencing a genome now runs to about a hundred dollars, maybe a few hundred. Twenty-odd years of technology development compressed the cost of reading your entire genetic blueprint by something close to seven orders of magnitude.

What that means practically: genome sequencing stopped being the exclusive province of massive research centers and became something any reasonably equipped lab can do. The question of who gets access to genomic tools shifted from "almost no one" to something far more open — though access and equity across healthcare systems remain genuinely unresolved. Cost falling doesn't automatically mean benefit distributed.

---

The cell problem nobody talks about enough 🧬

Here's where Quake's current work gets both fascinating and complicated.

Biology has long operated on the assumption that human cells come in a few hundred flavors. The interviewer cites the textbook figure of three to four hundred cell types. Quake's response is a laugh and then a correction: experimental work from his group's Tabula Sapiens project at the Chan Zuckerberg Biohub suggests there are at least an order of magnitude more than that. The human body is more cellular diverse than anyone thought, and — crucially — there's no way to predict these cell types from the genome alone. You have to go look. Experimentally. Cell by cell.

This isn't a minor detail. It blows up a tidy assumption that computational biology had been quietly relying on: that if you knew the genome well enough, you could work out the rest. You can't. The map requires the territory.

And that's what makes the next move interesting.

---

The 90/10 flip

Quake co-authored a paper laying out a vision for "virtual cells" — computational models of cells powered by AI, built on the massive experimental datasets that projects like Tabula Sapiens have generated. The ambition is specific: cell biology right now is roughly 90% experimental work and 10% computational. Quake thinks virtual cell models could invert that, letting researchers answer 90% of a question computationally and only needing experiments for the final 10%.

That's not modest. That's a claim that AI could restructure the fundamental workflow of biological research.

Whether it's visionary or wildly premature is genuinely contested — Quake mentions that the CZI workshop that produced the paper included "skeptics" alongside luminaries. The paper itself lays out a decade-long vision, not a present-day capability. The distance between "we believe this is possible" and "here it is" in AI biology is currently... significant. AlphaFold cracked protein structure in a way that stunned the field. Cell biology is messier, more dynamic, harder to pin down than protein folding. The bet Quake is making is that the data is now rich enough and the models are now good enough to start.

I find myself genuinely uncertain here — which is different from skeptical. The Tabula Sapiens data is real. The large language models being trained on transcriptomic cell atlases are real. What I can't yet evaluate is whether the emergent understanding those models produce will be biological understanding or very sophisticated pattern-matching that breaks the moment you push it outside training conditions. That distinction matters enormously for anyone hoping to use virtual cells to design treatments.

---

The 2050 doctor visit — and where I actually land

Asked to imagine a doctor's appointment in 2050, Quake doesn't hesitate long:

"I'm not sure you're seeing a human doctor at that point. It may be a full interaction with AI, and then the AI tells you what test to go get or what follow-ups to do before you're actually interacting with the human doctor."

He also predicts that cellular therapies — using cells themselves as treatment — will be as prominent by 2050 as small-molecule drugs are today. Medicine, in his telling, gets both more computational and more biological simultaneously.

I want to engage with the AI doctor claim honestly rather than just file it under "thought-provoking." My reaction is: the medical AI triage vision doesn't scare me on technical grounds — it scares me on distribution grounds. Who gets the good AI? We already live in a world where the quality of your healthcare correlates strongly with your zip code, your insurance, your language, your skin color. An AI that gives excellent diagnostic guidance to people with reliable internet, updated devices, and health data infrastructure, while people without those things get... what, exactly? That gap doesn't close itself. It tends to widen unless someone makes a deliberate choice to close it.

Quake isn't unaware of these tensions — he just wasn't asked about them in this conversation, and it's not his primary frame. His frame is capability: what becomes possible. The equity question is what becomes distributed. Those are different projects, and both matter.

---

The trajectory Quake describes — from precision measurement in physics, to a blood test that made a terrifying needle largely unnecessary, to a $100 genome, to AI models of cells — is a coherent argument that the tools for understanding biology are compressing faster than our institutions are built to handle. The science is moving. The frameworks for who benefits from it, and how, and whether the AI intermediary in your doctor's office speaks your language and knows your context, are moving slower.

That gap is the story underneath this one.

---

Mei Zhang covers biotechnology, genetics, and the future of medicine for Buzzrag.