From One in a Thousand to 15 Percent: Chai Discovery Treats Drug Design as a Scaling Problem

When Chai Discovery started, the state of the art for de novo antibody design was a 0.1 percent binding rate. Design a thousand molecules and about one of them would actually bind the target in the lab. Two years later the company’s Chai-2 model is at about 15 percent. Screen a thousand designs and you get roughly 150 binders back. That is the difference between a needle in a haystack and a sample size you can learn from.

The founders’ own budget makes the jump more striking. They told themselves that 20 percent in three or four years would be a win, and that the real goal was 1 percent — one in a hundred, which they thought would already shake the field. The model came in at 15 percent. Joshua Meier and Matthew McPartlon both say they did not expect it to work that fast.

The conversation is an episode of Sequoia Capital’s Training Data, recorded in August 2026, hosted by Pat Grady and Sonali Singh. Chai’s one-line pitch is that it engineers molecules with AI: a foundation-model lab for biology that wants drug research to look more like writing software. What follows is how they got the hit rate up, why they will not run their own clinical pipeline, and what they think a molecular CAD suite is for.

Meier’s résumé is a short history of modern AI. He went to a high school with a stem-cell lab, grew up as a programmer, and joined OpenAI when it was still a nonprofit, working on GPT-1, GPT-2, and scaling laws. The question that followed him out of that lab has not changed: if models can learn English, German, and French, why can’t they learn DNA and protein? He later became one of the original authors of the protein language model ESM. McPartlon came in from the other door. He studied pure math, started a PhD in theoretical computer science, and only in year three switched into deep learning and protein-structure prediction, in a lab that had built one of the first deep-learning folding systems. He is still, by his own account, a little terrified of biology. Once inside, he found the field simpler and more interconnected than it looks. An antibody and a mini-protein are both amino-acid sequences. They are different prompts, the way ChatGPT can both solve a math problem and help with English homework.

The name Chai is Chemistry plus AI, and they also drink a lot of spiced tea. Meier wanted something simple. Biotech names, he says, are one tongue-twister after another, and the company is trying to make the whole stack simpler, starting with what it is called. His diagnosis of the industry is older than the models. Early modern biotech found many medicines by luck, which is why the field is still called drug *discovery*: screen millions or billions of molecules and hope one works. Chai wants to swap that word for design. Put in the dream state of the molecule you want and let the model materialize it. Code gave software an abstraction layer that made iteration cheap. They want the same thing for biology.

Grady asks whether that means you stop doing experiments. Meier takes the other side. Software engineers got more productive and the market wanted *more* of them. If the return on a lab test goes up, people may run more tests, not fewer. The lab’s job is to verify. If you can evaluate a model, you can hill-climb it. That, he says, is a shared primitive of AI and of biology. They call the product vision a computer-aided design suite for molecules, biology’s version of CAD. Meier’s back-of-the-envelope is that going from an idea to a testable hypothesis still takes on the order of nine months. Compress that to nine weeks or nine days and each order of magnitude multiplies the number of ideas you can actually try. Medicines, in this telling, are what you grind out of that loop.

McPartlon walks the technical history without romance. Protein folding has a biennial contest, CASP, that holds out unpublished structures so academic groups can compete blind. Deep learning made a step change in 2018; AlphaFold 2 largely cracked the problem in 2020. The early systems predicted distances between amino acids and used a renderer to turn noisy distances into a fold, like the grainy first image models. Today you get high-definition structures in seconds. The inverse problem came next: given a shape, design a sequence that will fold into it. Around 2021 and 2022, diffusion models let people generate structure and sequence together and stuff real constraints into the prompt — bind here, look like this. Variational autoencoders tried to crush the whole distribution into a Gaussian, which was too hard. Diffusion breaks the protein a little, teaches the model to repair it a little, and asks again. Grady says that is the game he plays with his daughters: ask ChatGPT for a unicorn, then keep saying make it stronger, until you have the strongest unicorn in the world. McPartlon pauses. Yes. It is that easy.

They incorporated after Meier saw early antibody-antigen structure prediction at McPartlon’s house in 2024. For a long time the field treated antibody folding as unsolvable: not enough antibody structures in the Protein Data Bank, so antibody *design* was not even on the table. Early deep-learning protein design worked on mini-proteins, which are scientifically interesting and not what pharma actually buys. The holy grail was a computer-designed antibody with therapeutic function. Meier’s founding logic is a single sentence: if you cannot predict what an antibody looks like, how are you ever going to design one? McPartlon’s counter is the prompt view. Antibodies, mini-proteins, ordinary proteins — to the model they are kinds of sequence, not a hierarchy of sacred difficulty.

Go back to 0.1 percent and the scientific problem is not only waste. The gradient on the process is almost unusable. Many targets return no hits at all. The ones that do return too few molecules to see whether they look like drugs. At 15 percent, a thousand designs come back with 150 binders. For the first time the team can gather statistics on developability, stand up lab evaluations around those properties, and hill-climb. Other AI drug work at the time, Meier says, was mostly lab-in-the-loop tinkering: take a molecule and make it a little better. Chai’s bet was whether you could zero-shot a molecule that already had the properties you wanted.

The method sounds unoriginal on purpose. Scale data, scale models, scale compute, find the scaling laws, and keep stacking — the same recipe as self-driving cars and large language models. People treat biology as a bespoke exception. McPartlon says Chai is a “bitter lesson pill” company: general methods plus compute, over clever hand-built pipelines. The models are trained from scratch. They are not fine-tuning a general language model on protein data. Guiding principles matter in a blue-sky “neolab,” because the problem is open-ended and customers only show up when the science moves.

The first principle is simplicity. Chai-1 had 23 distinct submodules. Iterate on that and you have to understand each block’s behavior on its own, which does not scale. So they subtract until they can see the directions that actually matter. People on the team still propose a 24th module that would unlock one new target. Every time, the question is whether the change is an incremental patch or a compounding improvement that generalizes to a whole class. Glycosylation sites are the example. Someone wanted a special representation. The counter is that the sequence already carries motifs that hint at those sites, and a model that has to drive loss down further will have to learn them. If you believe the scaling laws, those targets unlock when the model gets stronger.

The second principle is rigor, because biology makes it easy to fool yourself. Wet-lab error bars are huge. Meier contrasts that with code generation: on SWE-Bench, a move from 16 to 17 to 18 percent is treated as real. In a biology lab, plus or minus five percent might be the same result, so the model has to show a step change, and it has to show it generally. A fancy architecture that looks good on one or two new tasks cannot carry a product. Singh says biology is famously hard to verify. Meier disagrees. He thinks it is one of the *more* verifiable domains. Code can compile, pass tests, and still be sloppy and unmaintainable — a taste problem. A molecule either binds, can be manufactured, or cannot. Those are hard numbers. The cost is time: a unit test returns in five seconds, a lab readout takes days or weeks. The numbers, at least, do not lie.

AI-for-drugs conversations usually run into Isomorphic Labs, which develops its own medicines. Chai took the other fork: build infrastructure for the existing industry. Meier gives three reasons. First, he and McPartlon had already worked at companies that owned a full-stack pipeline, models through the clinic. Because they are bitter-lesson people, they would rather spend the next dollar making the model stronger than standing up trials. Second, better models win better-paying partnerships, which buy more model work — a flywheel that a five-asset pipeline does not have. Third, and the one he likes most, partnering is an anti-self-deception device. Own a handful of targets and you might make exquisite molecules. You will never know whether the model generalizes to the other 500 molecules the industry makes that year. Partners force the truth.

The named partners include Eli Lilly, Novartis, argenx, and Pfizer. Those companies will not take a pitch on faith. The model has to work every time, not “this molecule is close, our chemists will clean it up.” Meier admits it is the harder business, which is why so few people run it, and also why the payoff is larger if it works. The stereotype going in was that pharma does not understand AI. The experience, he says, is the opposite. The customers are rigorous. They will test every claim. When the data holds, they go all in. Pharma is an innovation industry with a short exclusivity window. Lilly is a trillion-dollar company today; without new blockbusters it will not stay one. Once you are through the door, adoption moves fast, and partners start using the models in ways the lab did not script.

Hiring followed milestones, not a fantasy org chart. The founding team was almost all AI researchers. Antibody engineers arrived after the antibody design model existed. One of them is Andy Yeung, who was among the first people on yeast display at MIT two decades ago, spent twenty years at Pfizer and Genentech, and has an antibody approval to his name. When he joined, outsiders asked if Chai had pivoted into a full-stack drug company. Yeung then ran more antibody campaigns in a few months than he had in his whole career. Meier does not oversell the timing. When Chai-2 shipped, complicated formats such as multispecifics did not work. New hires waited weeks for the model to catch their case studies. Other names sketch the same software-company bet: Nathan Rollins, homeschooled, in David Baker’s protein-design lab at fourteen and in a PhD at eighteen; founding engineer Kevin Wu, who built what they believe was the first protein diffusion model; product people from Stripe, including co-founder Jack and Munaz, a top-ten Stripe code contributor; Neil, who had run a cybersecurity company. McPartlon’s summary is blunt. The deliverable is software, so the researchers have to be excellent engineers.

Scale has its own comedy. They started with 128 GPUs, and McPartlon felt rich — his PhD group had been four people on eight cards. The cluster is now large enough that a hyperscaler Slack-messaged them because the GPUs got too hot. Their reply was: isn’t that the point? “Hard target” means something different on every team, and they treat that as a feature. Science might mean an undruggable GPCR. ML research might mean a fold the model cannot produce. Product might mean a membrane protein with glycosylations that has to be shown to a user. There is rarely one bottleneck, which is why they refuse to be a company that ships one or two molecules. The goal is a CAD suite that changes how medicines are found, which means working every definition of hard at once.

You cannot scrape the web for protein structures the way you scrape it for English. McPartlon names two sources. The gold standard is the Protein Data Bank: experimental scientists have been depositing crystal structures since 1970, and without that archive, structure prediction and design would not exist. The second is sequence databases at something like a trillion tokens. Meier adds a jolt: there may be more biological sequence tokens on the internet than English tokens. Most of that data is noisy and redundant. Putting it together is a craft.

Early on the two founders stood in opposite camps. McPartlon was structure-pilled, obsessed with models that emit 3D coordinates, a class of problem that looks like neither an LLM nor an image model. Meier, coming from ESM, was sequence-pilled. Train a language model to predict the next amino acid and, to do that well, it has to represent the local microenvironment — which means 3D structure emerges inside the network. Today they refuse the religious war. Sequence or structure? Both, whichever solves the problem.

Compounding starts there. A stronger model run over the sequence libraries emits more predicted structures, which become training data for the next generation. The internal science team dogfoods the models, unlocks a use case, and then asks what new data that use case can write back. Exhaust from every lab experiment is collected too. Meier reaches for an OpenAI memory: they tried reinforcement learning on a GPT-1-class model and it did not work, because the base was too weak. Once the base was strong enough, those experiments started to move. Biology, he thinks, is entering the same phase. Partners generating binders in-house are part of the same flywheel. Chai never planned to ship one model that ended the field. It planned for iterations.

Asked to describe 2035, McPartlon flips an old industry slogan. You used to want first-in-class or best-in-class. Now you want last-in-class: the final answer, after which nobody needs another iteration. Those drugs should be specific to the disease and missing many of the off-target interactions that today’s medicines carry. Meier fills in the waste the loop is supposed to kill. The current habit is to find a molecule that is 80 percent of the way there and advance it to hit a timeline. It helps some patients and gets beaten a year later. Hard diseases such as Alzheimer’s become attackable if you can iterate hypotheses faster. Rare diseases and personalized medicines become economical if the cycle is cheap. A year ago, he says, nobody would zero-shot a molecule and green-light a program. Now someone might, and a year after that the first of those molecules can be in patients. McPartlon already has patents from a previous company on computer-generated molecules that are in people.

On competition, Meier’s answer is not another model lab. They are working against nature, and nature is a strong baseline. Wet-lab protocols have been tuned for decades. People have added the equivalent of module 240 to those protocols; that is twenty years of craft from someone like Yeung, and making a good antibody with existing tools is the bar AI has to clear. They did not start the company to build a ten-times-faster mouse. They started it for medicines that were not possible before. Inside the company, co-founder Jack has a rule: if you want to go fast in the long term, you sometimes have to go a little slower in the short term. Model releases cannot regress. The codebase has to let more people contribute. Partners the size of Lilly and Pfizer are now depending on the stack, which is a different responsibility from publishing a clever paper.

Meier and McPartlon have been on this problem for about ten years. A couple of years ago they were writing papers and privately wondering whether any of it would help a patient. The hit rate moved from one in a thousand to 15 percent, the demo moved from McPartlon’s living room to contracts with Lilly and Pfizer, and the next technical brink is control: a user who can ask for a binder at exactly 10 nanomolar affinity and have the model do it. The surface topic is pharma. The operating system is a research company trying to run the bitter lesson all the way through.