Patrick Hsu Wants an App Store for Biology, Not Just Better Drugs

Patrick Hsu, a genome-editing pioneer and co-founder of the Arc Institute, joins Josephine Chen and Pat Grady on Sequoia Capital’s *Training Data*. The first question is the sore one. Computer science plus biology has been a slogan for years; now the slogan is AI plus biology. Where are the drugs? Hsu’s answer is time, then a bottleneck. Even if models poured a trillion or ten trillion perfect molecules out of the pipeline, you would still have to test them in animals and then in people.

That regulatory crawl, he says, is the real constraint. Packing the top of the funnel does not compress the years it takes to move through the apparatus. Then he climbs onto what he calls a soapbox: ML for bio is not just drug design. Designing molecules matters, but it is a narrow slice of biology as a STEM field and as something that shapes human lives. Chen asks whether he simply means understanding the human body. His academic motivation is larger. Physicists have hunted a unifying theory for a century. Biologists already have one, so obvious that people treat it as background weather: evolution.

Evolution, Hsu says, acts on every length scale of biology, from terraforming planets down through ecosystems, populations, individuals, tissues, cells, and single molecules. Arc has been training a family of models called Evo, named for those forces, that try to map biological sequence to function with modern sequence modeling. The bet is blunt. Natural selection passes its effects down the generations through DNA mutations, so you should train on genomes.

The 2024 Nobels went to David Baker, Demis Hassabis, and John Jumper for protein design and structure prediction. Hsu is happy to praise proteins as fundamental molecular machines. He also notes that both citations say, explicitly, “for proteins.” As a genome biologist he looks one layer down. Proteins, RNA, and regulatory DNA are all encoded in genomes. The question the lab asked was whether a long-context model could reason over the bases and molecules embedded there and learn how molecular interactions become function.

Evo 2 is an autoregressive, multi-convolutional hybrid: an efficient long-context model doing next-base prediction. The same machine-learning recipe that learns grammar from next-word prediction, he argues, finds higher-order patterns in biology. Predict the next base and you might be looking at the difference between health and a sickle-cell mutation. Predict the next amino acid and you might be looking at whether an enzyme’s catalytic pocket still works. Predict the next gene and you might complete, or lose, a biosynthetic pathway.

Asked for uses a grandmother could understand, Hsu starts with predicting breast-cancer mutations, where he says the model is best in class, and with designing new CRISPR gene-editing systems, often zero-shot. People are building “an app store for biology” on these foundation layers. He jokes that labs are in a contest over whose model is more foundational: DNA versus protein versus all-atom diffusion. He does not care who wins the purity fight. Capabilities that feel useful are what matter, and the weights are open source.

Sequence a person, via 23andMe or any other test, and you will find mutations. Some cause muscular dystrophy, cystic fibrosis, or breast cancer. Most are variants of unknown significance: scientist-speak for nobody knows. The model has an opinion, and Hsu says it is state of the art at the job. The paper’s example is BRCA1. Known causal mutations lead many women to elect double mastectomies, a family-scale decision. Known benign variants mean annual mammograms. The huge middle is VUS. The team uses ClinVar as ground truth and spent real effort on evals, the gory work nobody likes and that has to be rebuilt as models improve. The bar is a task that would make a cell biologist feel something, the way NLP people felt when models started solving math.

They trained on DNA with no lab in the loop and no reinforcement learning. DNA is the fundamental information layer of life, and it is also where the data is. Published sequencing from 25-plus years sits in the government-funded Sequence Read Archive: bacteria, phage, viruses, humans, monkeys, fish, flies — “the entire Noah’s ark.” The experiment already ran. Mutations that distinguish people from one another, and from chimpanzees, worms, and bacteria, live in a dataset Hsu describes as trillions of tokens. The model just looks.

Using Evo is DNA in, DNA out, and humans do not speak DNA well. Imagine ChatGPT in Russian with 1% of the words in English. Teams have to build annotators and interpretability tools just to read the internals. Prompting biological language models, he says, is in the very early innings; we speak DNA “with an extremely heavy accent.” He wants everybody to do that work. Tools get useful when they lower the energy barrier of adoption, the way everyone uses BLAST, AlphaFold, CRISPR, and sequencing. If 2025 is the year of AI agents, he expects a zoo of models for every step of the scientific method, not just mapping sequence to function, and Arc has started releasing agent work.

Arc’s job, Hsu says, is to be a mothership for long-term research-capability breakthroughs, with the multidisciplinary expertise to execute them. In biology, many of the biggest mechanistic discoveries still happen in universities, unlike AI, where industry leads. Thirty years ago basic science and industry asked different questions; today the distributions overlap heavily. Universities still optimize for grants and first-author papers. Companies put dozens or hundreds of people behind a molecule. He thinks people overemphasize the split, but the incentives still drive different behavior. Arc blends an academic side with technical staff built more like industry, hoping others copy the model.

OpenAI co-founder Greg Brockman visited during his sabbatical, the first vacation he had taken since starting the company — which, of course, meant more work. Hsu’s story: Brockman was excited about biological language modeling but kept saying he could not promise much time; he owed his wife Anna a vacation. Mid-meeting she got up for the bathroom. He said, here is my email, “get me on the repo.” Arc’s CTO Dave Burke, who ran engineering for Android and Pixel and has an older biomedical-engineering PhD, now leads computational modeling of virtual cells. The ops stack, from finance and legal to lab ops and university relations, is meant to look like a tech company, spanning Stanford, Berkeley, and UCSF.

Success is not a Nature and Science paper count. Give Berkeley or Stanford professors millions of dollars and papers are the default output. Hsu wants tangible artifacts: technical blogs, code repos, protocols, and platforms people can actually use, aimed at mechanistic insight and, over time, curing some diseases. If Arc can be “an Edison shop” inventing things that find real-world value, partners who specialize in the last mile can take them out.

Even perfect drug design still faces a long chain after the molecule exists. Hsu describes pharma as a large, decentralized bureaucracy, with some parts functional and some not. Models can help discrete steps: target identification, data analysis, literature review, better binders, half-life, selectivity, manufacturability. What he actually sees inside companies is compression of massive regulatory documents, then AI helping write more of them. Talk to the budget people rather than the discovery heads and they will say they do not spend that much on discovery. Most of the dollars are in development. Tell me how AI helps there.

Industry probability of success, he says, is about 10%. A lot of the complaints about discovery and development fall out of that number. Why is the FDA so focused on safety? Because 90% of the time the drug does not work. The promise of AI is moving that rate from 10% toward 20, 30, or 50. Wet-lab biology is still close to guess and check. The scientific literature is missing the reasoning trace of what did not work. Papers are narrativized as inexorable logic; everyone who makes the sausage knows that is not how it happens. That high-fidelity trace is not written down, and it is exactly what reasoning models would need to close the loop.

Why do things that work in cells and mice still fail in humans? The models are imperfect. Discovery is a stack of filters: cell lines, then primary cells or organoids, then a mouse, then a monkey, then people. By then, five years and a hundred million dollars have gone by. The steps run in series because there is no predictive power; you have to do it in the arena, and growing cells and animals takes months to years. A trusted oracle — the part of Dario Amodei’s *Machines of Loving Grace* Hsu thinks Amodei got right — would let you run 10,000 agents at once. We do not have that model.

Today’s science agents close small gaps from step X to X+1. Companies hunt the most commercially valuable, shortest set of steps. End-to-end closed-loop discovery is far away; tools that traverse a few steps efficiently will take off first. Arc’s example is the Arc Virtual Cell Atlas, which Hsu calls the world’s largest single-cell dataset. An agent crawls the Sequence Read Archive, cleans messy metadata, and reprocesses data in the cloud. Two lead researchers plus one agent reached community-wide scale on grind work a talented computational biologist would not want to do. That leverage was a mental unlock. He cares about using agents to make breakthroughs, not charting the full closed loop yet.

The best scientists Hsu has worked with do two things: they invent creative ideas and they execute. Creative ideas come from connections other people miss. In a room of ten smart people, a small subset hears everything and names the conceptual bridge — out-of-distribution generalization from reading and reasoning across chemical biology, molecular biology, AI, and physics. Truly bilingual ML-and-bio people are rare; those translators power everyone else. What he looks for, from technician to executive, is simple. Do you think about science outside the lab? Have you done something end to end? Do you have the grit to walk the path? Going from step 1 through 15 winnows the population. “The last 20% of a project is actually 80% of the work.” Finishing things, more than once, hones a killer instinct.

Beyond gene editing and machine learning, Arc is building high-throughput tools for multi-system interactions, especially at the neuroimmune interface. Two scientists hired from Penn study interoception: the body talking to the brain, the “my tummy feels funny” sense that turns out to have deep mechanism. Programming biology is usually imagined as a binder to a protein or a CRISPR edit to a gene. Hormones and drugs like Ozempic program how you think, feel, and behave — satiety, energy, mood, muscle synthesis, focus. Hsu has been thinking about how to program physiology, not just a binding site.

Christoph Thaiss, one of Arc’s principal investigators, showed a gut bacterium that produces a molecule, connects via the enteric nervous system to the brain, and releases dopamine: the circuit for runner’s high. Delete the bacteria or cut any step and the reward is gone. The pathway also runs in reverse. Deep psychological stress signals from the brain to astrocytes that innervate the gut, releasing pro-inflammatory cytokines, causing inflammation and ulcers. Stress causes ulcers; this is the mechanism, and then you can talk about treatment. That is a different drug paradigm: less binder-to-biomarker, more a holistic, almost Eastern-medicine question of how do I feel healthier.

Future care, he thinks, looks like AI doctors integrating data multimodally: continuous glucose monitors, an Oura ring or WHOOP for recovery, Quest or Function Health for blood. Today you only know up, down, or in reference range, not what to do. Personalized genetics has been missing a way to combine genome sequence with health biomarkers — the high-school equation genotype plus environment equals phenotype. 23andMe, which had recently filed Chapter 11, was a pioneer at putting genetics in millions of hands. Hsu wants something that ties body measurements to diet and sleep and gives longitudinal personalized advice. He does not want to be an unhinged big-data person; Grady says, but it will. Chen asks about gambling addiction and drugs like Mounjaro and Ozempic. Hsu calls that interoception and tells the Sequoia hosts they should start that company.

Asked for 2025, 2030, and 2050, Hsu starts near-term. By year-end he hopes the field can design full IgG antibodies — the real medicines, not nanobody fragments — in one shot, with CDRs that bind where you point on a protein surface. De novo enzyme design should mature over the next couple of years. Most of the field is protein-coded; Arc wants to zoom out to cells. Building “the PDB of virtual cells” will take years. PDB is the gold-standard bank of experimentally solved structures that DeepMind used to train AlphaFold. His 2030 bet is accurate, useful virtual-cell models that make cell biologists feel emotion and raise therapeutic probability of success.

For 2050 he wants scientific superintelligence: end-to-end recursion of the scientific method, lab in the loop, a fully automated wet lab. Speeding drug discovery still hinges on predictive power, and the obvious way to miss it is building the wrong data. You can model the mouse in all its glory; it still will not be the human. His other soapbox is that the field needs far more experiments in humans, which takes regulatory innovation plus creativity. One example: lungs from brain-dead donors can be perfused and kept alive for a week, and a recent paper showed experiments in that tissue.

In the lightning round he admits he is a daily user of OpenAI Deep Research, the only AI app that has changed how he works. It was the first time he felt real emotion that someday he might be automated. Killer app for scientists by year-end? Deep Research again. Arc’s virtual-cell models will be useful, he says, but they still need time to ripen. The most important thing he has learned at Arc is a Scottish proverb he paraphrases: be happy while you are alive, for you are a long time dead. Lab training teaches you to hate on everything. Optimism is not a path to mediocrity. It is the emotional capacity for persistence toward long-term goals.