From 0.1% to 20%: Chai-2 Turns Antibody Design Into Prompt Engineering

Sarah Guo opens No Priors with the question that has hung over the field for a decade. Joshua Meier and several scientists on the Chai Discovery team have worked on AI drug discovery in different labs for about ten years. Guo has watched the same area. Computationally designed drugs have not yet reached the market. Why start a company now? Meier's answer is timing, not a sudden proof. For a long stretch the work was a research idea with "signs of life" that it would someday work, but not on a startup's clock. You cannot found a company that needs ten years to function, and you also do not want to wait until it is already working and miss the boat. The sweet spot, he says, was a one-to-two-year window to get the thing off the ground.

They placed two technical bets. Protein structure prediction had been treated as solved around 2020, when AlphaFold 2 reached experimental accuracy — for a single protein, one snapshot at a time. Drug discovery needs the interactions: how a small molecule modulates a protein, how an antibody binds an antigen. Early signs suggested that was becoming possible, and they bet that diffusion models and language models would do for proteins what they did for images: not one conformation, but a diversity of generations. The previous structure models, Meier says, were like early image generators before diffusion.

Almost every AI-bio company before them, he adds, glued itself too tightly to a wet lab. Chai runs plenty of experiments, but they wanted a portable platform that could be dropped onto hundreds or thousands of programs. Their earlier open-source model, Chai-1, is already used across the industry in ways the team cannot even enumerate. The company's one-line goal is to turn biology from a science into an engineering discipline. Jack Dent, who spent years as an engineering and product lead at Stripe, had stayed in touch with Meier since they were problem-set partners at Harvard — Dent maxing out CS, Meier stacking chemistry, physics, and the rest of the science curriculum, plus later BioAI work at Meta and OpenAI. Once the research stopped looking like a toy, Dent says, the future became impossible to unsee. Knowing that humans would soon engineer molecules at atomic precision made it "almost hard to work on anything else."

Chai-2 is a family of models that the company calls state of the art on several tasks. The result they are most excited about is design. Antibodies are among the most therapeutically interesting molecules: they account for close to 50% of recent drug approvals, and seven of the top ten bestselling drugs are antibodies. Give Chai-2 a target, ask it for an antibody, ship the designs to the lab, and wait about two weeks. In a 24-well plate — at most 20 attempts — roughly 20% of the antibodies bind their targets in the intended way.

When the project started, the company-wide goal for the entire year was a 1% success rate. Prior computational attempts sat around 0.1% or lower. Traditional high-throughput screening is a different universe: people screen millions to billions of compounds to find one that sticks. "There's a reason we call it drug discovery," Meier says. "It's a discovery problem. It's a search problem." Teams pan for gold in massive yeast or phage libraries, or they inject a mouse or a llama, wait weeks for the animal to get sick, bleed it, and isolate antibodies from plasma. COVID neutralizing antibodies were found that way from recovered patients. Chai-2, they argue, lifts computational hit rates by multiple orders of magnitude over the prior computational state of the art, and by many more over that lab search.

Faster, cheaper turnaround is the obvious reading. What Dent and Meier say they care about more is the class of targets previous methods simply could not reach. Biotech, Dent notes, is glum: the XBI has had a rough five years, one of the worst markets in decades. Chai-2 looks to them like an early flash of a platform shift of the kind the field last saw in the 1970s — new molecule classes, new targets, markets that were closed, diseases that had no therapy.

The Chai-2 paper looks at more than 50 targets. Most papers in AI drug discovery, Meier says, test one, two, or three. If you treat the work as engineering, that is not enough to know the system generalizes. He compares it to an LLM paper that brags about solving one IMO contest problem: cool, not a benchmark. Sometimes they try one or two targets just to check they have not shipped a huge bug. Even a 50% hit rate on two targets could be luck. So they ran a large benchmark to convince themselves, then the community.

How they picked the 50 problems is the part biologists would laugh at and engineers would love. They wrote a scraper against vendor catalogs, ordered whatever protein was in stock, and shipped every design in one batch so the experiment would turn around quickly. Each sequence was checked against SAbDab, the antibody-structure collection in the Protein Data Bank. Anything with more than 70% sequence identity to training data was removed. They did not think about therapeutic usefulness. Some of the proteins already have drug programs. This was model assessment, not a pipeline of future products. Meier even thinks the headline numbers may be an underbound, because the slapdash experimental setup could have hurt them.

The point of the exercise, Meier says, was to convince themselves and then the community that Chai-2 works, and only then to apply it. Hundreds of people already want to try the model on live drug programs. Treating antibody design as an engineering problem, on this telling, means a benchmark this size rather than a one-off demo on a famous antigen.

For listeners who know structure prediction, Meier starts there, because it is the piece that makes design work. The first thing the company did was sprint to a state-of-the-art structure engine and open-source it as Chai-1. Structure prediction, he says, is an atomic-level microscope: it shows where atoms sit in 3D. Once you can see the lock, you can start moving atoms around, changing sequence, and watching the structure respond. You prompt the design model with a target and a location; it places atoms until the constraint is satisfied and emits both a sequence and a structure that fits.

Dent supplies the computer-vision mapping. Going from sequence to a predicted structure is a classification-style task, the ImageNet moment for the field. Designing a binder is generative — Midjourney for molecules. The lock-and-key picture is literal: given a placement of atoms, craft a complementary set. The models reason by placing individual atoms in 3D, often with error across the whole structure less than the width of one atom. If you cannot see the lock at that resolution, he asks, how can you hope to cut the key?

Guo asks whether anything like an LLM's emergent grammar shows up. Meier points right back at the microscope. They still do not fully understand why the models work. They did not know this was possible; they only thought there was a chance. The signature of how proteins interact seems to be embedded in the data, and the model is generalizing into a new setting, not interpolating binders it has already seen for these targets. The test proteins look dissimilar by a biologist's family rules. In a supplement plot they push sequence similarity all the way down to 25%; the success rate is basically unchanged. The model, Meier says, does not care. Guo's follow-on is that if the network has learned principles the research community has not written down, people will start learning the rules from the model.

Hoped-for impact splits in two. First, screening moves onto the computer, so months or years of discovery compress. Second, and this is what they say they are more excited about, you go after problems traditional methods could not reach. The model is not perfect: it worked on about 50% of the targets they tried, maybe more given the experimental caveats. Its failure modes differ from the lab's, and that mismatch is the sweet spot. Chai is opening access to academic groups and industry partners. Drug discovery is too resource-intensive, Dent says, for one company to chase every target; assuming otherwise would be conceit. The access page drew hundreds of requests within hours of launch. A small team is still triaging who gets in.

Usual objections when papers like this appear: does it really work, will it work where training data is thin, are the molecules high quality? Benchmarking at 50-plus targets, Meier thinks, disarmed a lot of that. The sharper question is whether faster discovery changes which molecules you can even attempt. People who already know how to find drugs need a reason to pick up a new tool; the answer they are hearing from others is that the projects that were stuck now have another instrument, and sitting it out is how you get left behind. AI observers have joked that artificial antibodies arrived before some of the other breakthroughs they expected first.

Will large-scale wet-lab screening die? Meier compares the model to ChatGPT on a math problem: the top answer is not the top 10,000, and in a chat product you will never read 10,000 samples. In a lab you can actually test them. Companies are already asking Chai to run not 20 times but 100,000, even if 20 would have worked, because a better binder can be the difference between an IV antibody and subcutaneous dosing. The marriage of generative models and CRO-scale screening is, they say, still an open design question. Dent's 25-year sketch is deliberately reckless by AI standards: from under 0.1% to near 20% in a year, why not 50% or higher the year after? Mini-protein results already sit near 70% success on all five targets tested, with picomolar affinities. That is the beginning of a CAD suite for molecules — SolidWorks for mechanical engineers, Photoshop for creatives, something equivalent for biology. They have been kidding about ordering "bullish on biotech" baseball caps.

Success rate is only one axis. A drug is a needle in a haystack of properties: manufacturability, stability, and more, which Chai is putting on the list next. Faster design also unlocks formats beyond monoclonal antibodies aimed at a single epitope. Meier describes biparatopic constructs: pick two sites on the same target, make two primitive binders, and join them. Trade-offs that used to pit target risk against discovery risk against manufacturability start to move together. AI, he thinks, raises the bar across the board and opens an arbitrage of programs that were too crowded or too hard. There is still a lot to validate.

The anecdote that makes them optimistic predates even their first confirmatory batches. A partner had spent a few years, a team of five to ten people, and an estimated fully loaded $5–10 million, trying to build a molecule that cross-reacted with human and cynomolgus-monkey versions of a protein so animal studies would not fail on a slightly different sequence. Chai prompted the unfinished model for both targets at once and ordered 14 sequences. Four hit human, one hit cyno, one hit both. The program moved, with a pocket of diversity around that molecule. Guo notes the bear case on biotech: the bottleneck and the bill are clinical, not discovery. If discovery risk really falls, she says, the whole industry's odds improve. Meier will not oversimplify. Capital markets and clinical risk are real. Chai-2 is the tip of the iceberg. He still thinks the progress is worth re-running the industry's math.

Where does a model company invest next? Two places, Dent says. First, what comes out today is a hit, not a drug. Early preprint data suggest many of the properties you would want, but they need more characterization before anything is a viable human candidate. The more aggressive step is generating entire drug candidates zero-shot from the model — a sentence nobody at the company was really saying a few months ago, and that now follows from getting antibody hits in 20 attempts. Chai-1 felt like a model; Chai-2 is already a pipeline. A scientist starting a program will not write prompts from a script. They will think hard about how to describe the problem, including multi-species and multi-property constraints. Meier's industry aside: it is a good time to be a sick mouse, because a human drug usually has to work in animals, and many programs stall with a human binder or a mouse binder but not both, then invent a "surrogate antibody" just to convince the FDA the mechanism is real. They showed human-plus-monkey in the paper; the opposite prompt, hit this protein and miss that one, has early evidence too. The software around those prompts, Dent says, will get as advanced as Photoshop.

Guo asks what a working antibody engineer or biologist should learn if they believe CAD-for-biology is arriving. Meier's list is short. Get access to Chai-2. Get the prompts right — which epitope, which stretch of sequence, is itself experimental design. Then start dreaming about experiments that used to be unthinkable. They used to ask engineers, hypothetically, what they would do with a 1% success rate, and brains would not quite turn over. Now the number is 10% and up, and the conversations change. It is the LLM story again: five or ten years ago, "what would you do if we could predict the next word perfectly" was almost unanswerable until people had the product in their hands.

Strategy hides in the prompt. Want to hit human and miss monkey? You can lean on the engine, or you can simply bind a region that already differs between the two species. Those moves are obvious only after you have used the tool. Dent was still on calls a few weeks earlier with people who said this was three to five years out; after a demo they fell off their chairs. Antibody engineers, he says, are often the most excited, because they hate slow loops and intractable tasks. Give a software engineer a tool that means writing less code and they spend the surplus on architecture. The same raise-the-bar effect is coming for biologists, which is also why Chai wants partners: biology is too vast for one company to hold all the specialist knowledge.

Building that company is itself an engineering story. The team is about a dozen people, including co-founder Matt McPartland; Meier has a chemistry degree, another founder, Alex, a physics PhD, and surprisingly few computer-science diplomas. Everyone teaches everyone. Dent's Stripe years — Stripe Link, Stripe Capital scaling from zero toward 50 engineers — taught him that someone has to keep the whole system in their head or software entropy eats the schedule. Each module owner should load as little context as possible; simplicity and modularity have to be culture, with no shortcuts, because the next person is the user. Deep-learning repos make that harder: a bug can sit for weeks. Chai has already bisected git history with training runs to find a bad commit, burning tens of thousands of dollars of compute. Unit tests for everything became non-negotiable after that, a practice Meier and Dent say most research codebases still lack. They started scrappy, even unsure they would stay in San Francisco, on cloud free credits, and they only pour on compute when they see signs of life. They are now hiring product engineering, antibody engineering, and business development. The cold water at the end is theirs: Chai-2 produces hits. Developability and the clinic still sit between a binder and a medicine. The direction on the wall is that the most effective antibody engineers will soon be expert prompt engineers.