Daphne Koller on Why Insitro Skipped Biotech’s Shortcut
Daphne Koller does not like the phrase “move fast and break things.” She says so almost immediately, and she says it the way people say things they’ve been asked before and have long since stopped softening. “Living in a sector where the products that we produce are drugs that have to go into a person, the notion of ‘break things’ is a very bad idea,” Koller tells Observer. “You do not want to break things when you’re going into a clinical trial and cause risk to your patients.”
It is an obvious point. Coming from Koller, it is also a rejection of the operating philosophy that built the industry she has spent the last eight years trying to out-execute. Koller trained as a Bayesian statistician, taught computer science at Stanford for 20 years, and won a MacArthur “genius” grant in 2004 for work bridging logic and probability before co-founding Coursera, the online education company she left at a valuation of roughly $1 billion. In 2018, Koller started over, founding Insitro on the premise that drug discovery—an industry famous for burning a decade and a billion dollars per approved therapy—could be re-engineered, as she’d once tried to re-engineer the university lecture. She has since built a company backed by more than $750 million in capital, including roughly $140 million in pharma collaboration revenue, toward turning algorithmic target discovery into clinical candidates.
Insitro raised its first $100 million in six months, from ARCH Venture Partners, Andreessen Horowitz, Foresite Capital, Alphabet’s GV and Third Rock, with Jeff Bezos joining later. Within a year, Koller had a deal with Gilead worth $15 million upfront against roughly $1 billion in potential milestones, for work on fatty liver disease. Seven years later, Insitro is running two live programs—one in ALS with Bristol Myers Squibb, one in metabolic disease with Eli Lilly—both tracing back to targets Koller’s platform surfaced from human genetics, not from a chemist’s hunch.
The Bristol Myers Squibb partnership is the clearest evidence that the model is compounding. BMS paid to expand its collaboration with Insitro twice in the past year: $20 million in October 2025 to extend work on an Insitro-discovered ALS target, then $10 million in March 2026 when BMS nominated two additional targets sourced from Insitro’s Virtual Human platform—bringing the portfolio to three ALS targets and more than $2 billion in potential milestones plus royalties. Separately, in September 2025, Eli Lilly signed on to train machine-learning models on its own proprietary preclinical data, a bet that Insitro’s modeling can shrink the animal-study cycles required before a candidate is chosen.
Following its January acquisition of the Israeli biologics-design startup CombinAbleAI—folded into a new platform Insitro calls TherML, built to extend the company’s target-discovery engine into molecule design and manufacturability across small molecules, oligonucleotides and antibodies—Insitro’s lead candidate has moved from genetics to mice. CTRO-1013, an A.I.-discovered siRNA therapy that silences the gene IRS1 to treat MASH, the metabolic liver disease formerly known as NASH, showed a 37 percent reduction in the fibrosis marker TIMP-1 and a 68 percent reduction in the liver-injury marker CK-18 in a standard preclinical mouse model, data the company presented at the American Diabetes Association’s June meeting. First-in-human trials are expected this year.
None of it, Koller insists, has moved faster than biology allows. “You cannot make a cell grow faster by throwing more GPUs at it,” she says, rebuking anyone assuming A.I. collapses biology’s timelines the way it’s collapsed, say, code review.
Koller’s aversion to “move fast and break things” is not just a rhetorical flourish about patient safety. It also describes a specific strategic decision she made years into building the company. Many platform biotechs hit an uncomfortable timing problem: building the discovery platform and then waiting for it to produce usable programs stretches out the road to becoming a “clinical-stage company.” The common workaround is to in-license an already-developed asset that never came off the company’s own platform, purely to have something to point to. Koller argues that shortcut is close to a trap. If the in-licensed asset fails, the decision was simply bad. But even if it succeeds, “it doesn’t read on your platform”; the capital spent developing it is capital that wasn’t spent advancing whatever the platform itself might eventually produce, and a success story about someone else’s molecule tells investors nothing about whether Insitro’s engine works.
“We made the decision to go the slow path,” she says. “We did not in-license anything. We waited for insights to come off of a platform that we first had to build, and then get the programs off that platform, and then move those into the clinic. That took longer, but it was the right decision.”
Koller’s founding logic for both of her companies is that real value gets paid for. “If you are providing something to people that is truly valuable, by and large, there will be a willingness to pay for that,” she says, pointing to high-quality education and transformative medicine as proof. She is equally direct about the alternative—”it’s also absolutely possible to go and sell people something that has modest, if any, value and charge them exorbitant amounts of money for that”—and says the discipline against doing so is exactly why she structured both Coursera and Insitro as for-profit companies rather than nonprofits. Philanthropic capital, in her telling, is unreliable by design.
“There is not really any reliable revenue stream, because you’re not charging for the value that you bring,” she says of donor-dependent models, where interest can shift from one cause to the next. Capitalism, for all its flaws, remains the only system she has seen reliably tie the money an organization receives to the value it actually produces—and building a for-profit company forces that correlation to hold. “That was the rationale for the decision at Coursera,” she says, “and it’s also the same rationale for the decision at Insitro.”
Koller does not pretend the CEO chair is a costless upgrade from the lab bench. Plenty of scientists prefer to stay scientists, she says, rather than take on selling the work, raising capital and “the nitty-gritty of HR and audits.” She has found the split stimulating, but is candid about what it costs her. “I do often miss having enough spare time to really sit down and think deeply about a scientific question. So, like everything in life, it’s about trade-offs.”
Asked to place Insitro within the sprawling category of “machine learning in healthcare,” Koller draws a sharp distinction. Traditional biotech companies, she argues, typically end up with only one or two precious programs to their name, forcing an early, defensive choice to sell a program off before it hits a real value inflection point, and forfeit most of the upside it could have generated. Pure-play tech companies pursuing the same market tend toward the opposite failure mode, avoiding the “heavy lifting” of pushing a drug through clinical proof of concept altogether, and defaulting instead to something closer to a service model for pharma, without taking on the cost and risk of clinical development themselves.
Insitro’s platform gives it a third option pharma companies don’t have: a discovery engine productive enough to supply partners with programs they are short on, while still retaining enough capacity to advance a subset of programs Insitro itself majority-owns through the clinic. “That allows us to also participate in the win-state economics that you get when a program actually proves itself out,” she says. It is, in effect, a hedge; collaboration revenue from partners like BMS and Lilly funds the company while it works, patiently, toward the far larger payout of owning a drug that clears human trials.
BMS’s decision to nominate two more ALS targets after the first—designated ALS-2 and ALS-3—validates the underlying biology. “These other targets that are related reinforce the fact that this wasn’t a mistake,” she says. More importantly to the business, it validates that the discovery platform itself is repeatable rather than a one-time hit. It shows that the platform is real, that it didn’t just happen upon something by accident, but that it is reliably reproducible. “It’s not a one-trick pony,” Koller says.
She frames the scarcity driving that value. High-confidence, novel drug targets are rare enough that pharma companies routinely crowd 30 or 40 programs onto a single known target rather than risk pursuing something new—a pattern Insitro’s platform is built to break. That “novel biology is the scarcest commodity in this industry” underlines Koller’s recurring worry—not about her own targets, but about the tools used to find them.
Pressed on A.I. reliability in drug discovery, Koller draws a distinction sharper than the industry’s usual hallucination anxiety. Fabricated citations are the easy failure mode to catch, she explains. “You’d have to be an idiot to just take what an A.I. said at face value,” since a hallucinated source can simply be checked against the primary literature. The harder problem is trusting an A.I.-generated insight built on today’s incomplete and imperfect data, and advancing a program based on conclusions that aren’t as grounded as they appear.
Koller shares a concrete instance. While reviewing an early A.I.-generated ranking of prioritized drug targets, her team pressure-tested the data and found that the rationale for prioritizing a particular target didn’t hold up. The team caught it before advancing the program, then built new checks into its A.I. pipeline to guard against the same failure mode going forward. Mistaking a statistical association for a causal relationship is one of the most consequential errors in the industry. “Causality is at the heart of what makes a drug work or not work, because you’re making an intervention in a person,” she says. “If that intervention is not causally upstream of the disease that you’re looking to modify, then the fact that they were statistically associated is meaningless.”
It is why Insitro built its platform around causality as “the fundamental underpinning” rather than around whatever associations happen to surface in public data, which Koller says is often riddled with spurious correlations traceable to nothing more than how different labs ran their experiments.
Insitro’s data stack breaks into three layers, and the proprietary share is bigger than a company built partly on public biobanks might suggest. Human data starts out largely non-proprietary—Insitro works with national biobanks and academic medical centers rather than owning a patient population outright—but the company layers A.I.-inferred data on top of what was directly measured, and those inferred layers are proprietary. The third layer, entirely proprietary, is data generated in Insitro’s own labs: causally anchored cellular data from cells deliberately perturbed, rather than simply observed.
By volume, Koller estimates roughly three-quarters of the company’s data now comes from its own lab-based “cell factory.” Of the remaining quarter drawn from human data, she says about half is publicly available, and the other half consists of Insitro’s proprietary inferences layered on top.
Koller argues that commoditized knowledge—first accelerated by search, now by conversational A.I.—has shifted what she recruits for in building Insitro. Koller hires people with traits no model can supply. Intellectual curiosity and a “non-trivial modicum of humility” top the list, the willingness to recognize what you don’t know and go learn it. Humility also governs how Insitro’s cross-functional teams interact. Koller warns against hiring people who believe deep expertise in one discipline makes them the smartest person in every room. It is a dynamic that produces siloed, bifurcated cultures. “Create together” is one of Insitro’s stated values. People from different disciplines, contributing distinct skills and ways of thinking, don’t just arrive at better solutions; they arrive at better problems, she says.
The other qualities Koller looks for are boldness balanced against rigor, a willingness to try things with no precedent, paired with a discipline to scrutinize new ground carefully rather than assume it’s working. The stakes of getting that balance wrong compound with every stage of drug development, she notes, since each downstream step is an order of magnitude more expensive than the one before it.
Trained as a Bayesian statistician, Koller describes Insitro’s confidence-building process in explicitly probabilistic terms: a prior that strengthens as data accumulates, drawn from human cohorts, lab work and existing literature, continuously pressure-tested against new evidence. The same logic applies to technology decisions. Starting small, creating benchmarks to test predictions against data the model wasn’t trained on, then testing against still another cohort. “The creation of the benchmarks is at least as much of an intellectual exercise as the creation of the solutions,” she says, calling the ultimate test of any hypothesis the same regardless of discipline. What happens when you put the drug in a human being?
That patience runs alongside a self-described impatience that she treats as productive rather than contradictory. Some timelines cannot be compressed. Cell growth, mouse experiments and clinical trials move at biology’s pace, not capital’s, but her impatience pushes her to keep asking whether an assumed timeline is actually fixed. Could a blood biomarker produce a confident readout faster than waiting for a clinical endpoint? Is a contract research organization’s standard turnaround a real constraint or just an unexamined habit? “When you actually pressure-test the question, [you find that] you could actually do it faster, and that’s great.”
Koller cites Jim Collins’s Good to Great as a longstanding reference point for building an enduring organization, but the book she says she has pushed on more people inside Insitro is Switch, by Chip and Dan Heath. Switch is an “incredibly illuminating” book about behavior change, which Koller says maps directly onto what both of her companies have had to do. Coursera and Insitro were each built to run counter to how their industries traditionally operate, requiring Koller to drive change both externally, in how the surrounding ecosystem behaves, and internally, in an organization unlike almost any other. Computational scientists learn to work alongside life scientists, and vice versa, inside a company she believes has no close precedent.
