95% of Drug Programs Fail. Daphne Koller Is Betting Machine Learning on the First Decision

Daphne Koller joined Sarah Guo and Elad Gil on *No Priors* as CEO and founder of Insitro, a company applying machine learning to pharmaceutical discovery. She did not start in a wet lab. In the mid-1990s, she said, the datasets available to machine-learning researchers were "kind of boring" — classifying text into 20 newsgroups, for example. Biology had more interesting data: scientists were beginning to measure gene activity across the entire genome in multiple samples. She followed the data, then fell for the science itself.

At Stanford she ran what she called a bifurcated lab. Half the group did core machine-learning work and published in traditional computer-science venues; the other half did core biology and published in *Nature*, *Cell*, and *Science*. Most of her computer-science colleagues had no idea she did biology. Most of her life-science colleagues had no idea she sat in a computer-science department. It was, she said, a bifurcated existence — and a lot of fun.

Gil had spent half an hour hunting for her textbook on probabilistic graphical models. Koller described the field as a pendulum. Early graphical models pulled artificial intelligence toward numerical data and machine learning, away from purely symbolic AI. Deep learning then shoved that work aside, because so much power came from pattern recognition on raw images and text without interpretable representations. The pendulum, she thinks, is swinging back. You need that pattern-recognition power, and you also need to reason about causality and to interpret the model so you can tell a clinician why it made the decision it did. The useful paradigm is a synthesis of both.

The path to Insitro ran through a "completely crazy" career decision. Through much of her Stanford years Koller felt an increasing urgency to have an impact on real people, not two steps removed by training students. When Stanford's first massive open online courses took off, she did not want to assume someone else would carry the flag. She left on what was supposed to be a two-year leave to co-found Coursera with Andrew Ng, fully intending to return to faculty life. Stanford's leave policy was strict. Two years later they asked her to come back; she said she needed another year to see the project through; they said that was not an option. She resigned an endowed chair and stayed in industry. Her mother thought she was nuts. Koller thinks her mother still does.

About five years in, in early 2016, she lifted her head. She had left Stanford for Coursera in late 2011, just before the machine-learning revolution took off in 2012. The ML world had totally transformed. It was not having much of an impact in the life sciences. Coursera, she said, is a wonderful company but not a deep-technology or science company. The place she could have a disproportionate impact was bringing the two disciplines together. There are not a lot of people who have spent roughly 20 years in machine learning and a decade in biology and can speak both languages.

Unsure where to go after five years away from science, she asked Art Levinson for advice — former CEO of Genentech, and chairman of Google and Apple, and in her view uniquely qualified to connect those worlds. She had been on the thesis committee of his son Jesse. Levinson's advice was, she said, "admittedly self-serving": come to Calico. She did not know much about the company beyond that it worked on aging, but it is not often you get to work with Levinson, or with Hal Barron, whom she also respects. She stayed 18 months and learned a ton. Then she realized she did not want a company focused on a particular biology. She wanted a platform for doing drug discovery differently — an "incredibly fraught, largely unsuccessful and very expensive endeavor." Calico was not the place to build that, so she left and founded Insitro.

What she took from Calico was not a single scientific trick. It was her first close look at how pharmaceuticals are actually developed. At Stanford she had already seen how much power machine learning could have even on small biological datasets; the datasets had since grown larger and richer. Two realizations pushed her out. One: the way insights are turned into therapeutic interventions was "so old-fashioned" and so unaccommodating of data that there had to be a better way. Two: data on human aging is hard to get, because aging is a long process and useful longitudinal cohorts would have needed to start 20 or 30 years ago. The intersection of ML and biology was huge. Aging, at least from her perspective, was not the first place to apply it.

Elad Gil laid out the industry's grim arithmetic: a successful drug still costs on the order of a billion to a billion and a half dollars and takes a decade-plus, and AI could plausibly help almost everywhere — small-molecule design, clinical-trial enrollment and data, diagnostics, endpoints. Koller agreed, and said the wealth of opportunities is itself one of the biggest challenges. "Everywhere you look there is a big opportunity for machine learning." A shrinking minority in biopharma still treats ML as a fad, or as something like X-ray crystallography: a tool that improves one narrow vertical. Her analogy is different. "It's not like x-ray crystallography. It's like computers. You're going to use it everywhere and it's going to be transformative everywhere." It will not be a silver bullet unless you figure out how to use it most effectively.

She divides drug development into three chunks. First is the original biology: which targets, in which indications, in which patient populations. Second is turning those targets into therapeutic matter — molecular design. Third is enabling clinical trials: patient selection, efficacy biomarkers, and the rest. All three are valuable. The number that makes discovery so expensive, though, is that 95 percent of drug programs fail. The biggest reason is not a poorly designed trial, though that still happens. It is not that the molecule fails to hit its target, though that happens too, in an increasingly smaller share of cases, because pharma companies have gotten better at making drugs.

Programs fail, she said, because "we're just not modulating the right thing" — the wrong target, the wrong indication, or the wrong patient population. If you want to bring down the two-and-a-half-billion-dollar number she cited, you have to bring down that 95 percent statistic so a successful program does not have to carry a long string of expensive failures on its back. Target identification, she figured, was maybe the hardest thing to do. It was also going to be the most impactful.

Target identification is a brutal machine-learning problem because you do not have the right training data. The question is: if I make this therapeutic intervention in this patient, what happens clinically? That label does not exist until the end of the process, in a clinical trial. So how do you train a model with no training data? Insitro's answer is two-pronged, and Koller thinks the synthesis is the point. One source is humans. Each of us is an experiment of nature: genetics have already modulated our genes to different activity levels. The mapping from genotype to phenotype can stand in as a surrogate for what a therapeutic intervention would do. That only gets you nature's experiments, which are not the same as a drug. In parallel the company built its own wet lab, makes interventions in cellular systems, and reads out phenotypic consequences at large scale with high-content modalities.

Machine learning is used three times. Once, to interrogate the phenotypic consequences of human genetic variation in high-content data — imaging, where ML already works well, plus transcriptomics, proteomics, and other omics. Once, to understand directed interventions: genome-edit a gene in a given cell background and watch how cell states respond. Third, to bring the two together and build cellular models that are predictive of human clinical outcomes. The goal is to replace "untranslatable animal models" with something driven from human biology.

For a computer-science audience, Guo asked how you even get a "Daphne neuron" in a dish. You take a white blood cell or a skin cell, run it through Nobel Prize-winning reprogramming, and turn it into an induced pluripotent stem cell — an iPSC — that can take any lineage: liver, heart, brain. From there you differentiate it into a neuron or a cardiomyocyte and recreate that person's genetics in a cellular system. You can go further and edit in a disease-causing variant across genetic backgrounds, then compare the same cell with and without the variant. Guo's analogy for a tech audience: an in vitro A/B test, isolating differences specifically associated with the disease-causing variant.

The platform is horizontal, so saying no is hard. Insitro looks for large unmet need plus a place where its own technology is a differentiated advantage. Neuroscience is the textbook case: very few effective therapies, in part because animal models do not translate. "Curing mice of schizophrenia, whatever the heck that means," she said, does not do much for human schizophrenia, because it is not even the same disease. iPSCs differentiate relatively easily into neurons with disease-relevant phenotypes, and human brain MRIs are becoming more abundant — more so as early Alzheimer's drugs make imaging part of who gets treated. Metabolism and oncology are the other bets, because high-content, unbiased, disease-informative data is already collected as standard of care. Complex diseases are not one cell type; Koller is pragmatic. You can push a hepatocyte out of its comfort zone with fatty acids and immune factors and get closer to a diseased liver without pretending it is the whole organ. Organoids and organs-on-chips, which mix cell types and even organ systems, are getting better by the day in other labs. Insitro goes first after diseases that manifest enough in a single cell lineage, and she estimates that in maybe three years another tranche of diseases will be unlocked by that technological wave.

Asked how far Insitro will take drugs in-house versus partnering, Koller was again pragmatic: maximize impact on patients. What the company has, over a lot of others, is an engine for novel insights and novel targets. Many biotechs have one or two programs; sell those off and the cupboard is empty. An engine can put some programs into partnerships, even fully out-license others, and keep producing — including as it expands into new indications and technologies. The company already had relationships with Bristol Myers Squibb and others on ALS and dementia. Sometimes the platform surfaces a target that already has a drug, because there are only about 20,000 genes, and someone developed a molecule for the wrong indication or the wrong patients. Insitro does not think new chemical matter is the only thing that makes its existence worthwhile. Going to the asset owner and bringing that drug to the right patients can shave two, three, maybe even five years off a program: the molecule exists, it may already have been in people, safety is shown, a biomarker may already say when it is working.

Gil asked about biomarkers, because so much of clinical development is waiting months or years for endpoints. Koller pointed to research: drugs that have a biomarker are about twice as likely to succeed in the clinic as those that do not. Drugs with support in human genetics are also twice as likely to succeed. Insitro believes both. Because the company is so focused on human data, clinical biomarkers for which patients will benefit tend to fall out of the work — not for free, but as a consequence — as long as you pay attention and do not do what a lot of companies do: find a target and apply it in all comers. If she is being cynical, that all-comers instinct is sometimes about maximizing revenue. It is also one of the big things that causes drugs to fail.

A geneticist colleague at Calico told her the Herceptin story. Herceptin, one of the earliest precision-oncology drugs, goes after HER2-positive breast cancer. If they had run that trial in an all-comers breast-cancer population, you would have needed about 10,000 patients — a very large trial — and even then you might not have seen a statistically significant signal, because side effects in non-responders can outweigh strong benefits in responders. Putting the right patients in the trial was "absolutely critical" to a successful, reasonably sized study. Gil added homologous-recombination-deficiency drugs: without selecting patients with specific mutations they did not work; focused on that population they worked extremely well. The medicine was there. The difference was who you targeted.

How much time can an ML-first approach actually cut? Gil noted that during Covid the industry did in six months what normally takes ten years. Koller said the answer has regulatory and biological halves. Talk to regulators early — not when you show up with an NDA package — about which biomarkers they would accept and how you would collect them robustly in the wild, not just in a research lab. That discussion can accelerate things. Covid also had a lot of patients and fast disease progression, so trials filled quickly. An Alzheimer's trial is different: disease progression is what it is, and you have to wait long enough to see a delta in the cognition curve. Amyloid beta is a proxy some people hoped would help; Koller’s guess is that it is the right proxy for some patients and not others. "There's only so much that you can speed up biology in certain cases, because biology takes as long as it takes."

People who have worked in both computer science and biology, Gil observed, have to unlearn the habit of manipulating data in the cloud and getting an answer. Koller said building the right culture at Insitro has been one of the most challenging things they had to do, and at the same time a big competitive advantage, because it is really not easy. You need computer scientists willing to learn a "fuzzy, ill-behaved" field where things do not do what they are supposed to do. Program a computer correctly and, bugs aside, the same thing happens. That is not true in biology. These are living systems; they do not respond the same way day after day. Scientists, conversely, get frustrated with the engineering instinct that you can take building blocks, put them together, and know what will happen.

A colleague's anecdote made the gap concrete. Some days an experiment went perfectly; other days the cells just died. The cells died on the days a particular technician, who had a fondness for onion sandwiches, was on shift. The onion on his breath made the cells "less happy." An engineer would not even think of that variable. The mindsets diverge on data, too. Show an engineer a bunch of dots and the instinct is to find the pattern that explains as many points as possible, because that is what you engineer a system around. Show a scientist the same dots and they look for the outliers, because exceptions are often the beginning of a discovery: why did this one behave differently?

Process is dual-track. Product managers run agile and scrums on platform development, including on the biology side. Drug-discovery projects that last years do not get scrums. A 45-day iPSC differentiation takes 45 days; there is no point holding an agile standup in the middle while you wait for the cells to do their thing. Project managers own those timelines. The two roles communicate, and each deploys its own discipline. Culture was written down at the start of the company. Koller's favorite behavioral norm is the last one: "we engage with each other openly, constructively, and with respect." Open means no silos, naive questions welcome, and naive suggestions from another discipline taken seriously — "why don't we do things this way" is often a good idea if you do not come in with because-that's-how-we've-always-done-it. Constructive means the goal is a better outcome, not being the smartest person in the room, which she called a big problem in companies full of smart people. Respect means valuing what everyone brings. Too many tech people, she said, walk into life sciences announcing that they have machine learning and are going to solve everything, without taking time to learn the other discipline's challenges. That raises hackles immediately, "and from there the conversation can only get worse."

If she were not at Insitro, what would she work on? This is the golden age of AI and machine learning, she said, and her personal compass has always been deploying it where it makes life better — improving life and health, not selling more ads. The interface between biology and technology is one of the richest areas that exist, because each field has been making huge progress on its own. Everyone hears about AI because of ChatGPT. The toolkit biologists now have — CRISPR, pluripotent stem cells, huge advances in microscopy — is not as visible to a lay person, but it is equally dramatic. Bring the two together and the spillover is not only drug discovery. She listed agriculture, environmental technology, energy, biomaterials that are less destructive and have better properties, and food tech.

There are other societal directions too. Energy applications of ML beyond biofuels are only starting, because the data have not been as readily available. Going back to her Coursera and Stanford days, she still sees a lot of room for machine learning in education: personalizing learning so the experience is more beneficial than letting ChatGPT write a student's essays. The list of useful things is almost unlimited. What it takes is a commitment to finding them, rather than falling into the comfortable place of going to a tech giant and doing something that earns a lot of money. Nice for you, she said. Maybe not as good for making the world better.

Guo asked for advice for founders who want to work in traditionally harder markets — edtech, healthtech — where budgets are thinner and regulation is thicker. Koller is hoping investors now see entire unpacked ecosystems where technology can make a difference and has not. Edtech had always been a backwater of investment, and yet Coursera attracted fairly significant funding even at the very early stages because the idea was compelling and differentiated. She is a believer, maybe an optimist, that a really good, differentiated idea whose impact you can make clear — as they could with those first MOOCs — will get people to go with it for a while. Will this be the place you make the most money with the greatest certainty? Maybe not. "We only have one life to live," she said, "and ultimately what you want to be able to do is look back on your life at some point [and say] I have done something that's really worthwhile and important."