Summary

Chai-2’s headline is a roughly 20% antibody-design hit rate in up to 20 attempts, versus about 0.1% or lower for earlier computational methods and millions or billions of candidates in traditional screens. Across more than 50 held-out targets, it produced hits for about half, recasting open-ended search as “an engineering problem” with a roughly two-week lab-validation cycle.
The benchmark is unusually persuasive because Chai optimized its evaluation for breadth, not a therapeutically curated showcase. The team scraped in-stock vendor catalogs, excluded targets with more than 70% sequence identity to anything in SAbDab, and saw comparable success in a subset down to 25%; Joshua Meier’s summary: “the model didn’t care.” He still calls the reported target success potentially a lower bound because the hurried experimental setup may have introduced mistakes.
Chai’s technical leap is from an “atomic level microscope” to a molecular generator: structure prediction sees the lock, while Chai-2 designs a key atom by atom. Errors can average less than one atom’s width, and out-of-distribution success suggests the model learned something fundamental about molecular interactions rather than memorizing protein families—though the founders say they still do not fully understand why it works.
The thesis is market expansion, not merely cheaper discovery: AI’s different failure modes could unlock targets traditional screening cannot reach. Yet Chai-2 worked on only about half its benchmark targets, and Meier stresses that discovery is “just the tip of the iceberg”; manufacturing, stability, capital markets and clinical risk remain.
Wet labs become a scaling complement, not an obvious casualty. Because model sampling is probabilistic, moving from 20 designs to 100,000 might surface better drugs, potentially changing injection versus subcutaneous dosing. Unlike a person reviewing 10,000 chatbot answers, a laboratory can test all 10,000 candidates, preserving a role for CROs and high-throughput screening.
A partner case makes the economics concrete: a five-to-10-person effort costing an estimated $5 million-$10 million over several years had struggled to find a molecule binding both human and monkey versions of a target. Chai ordered 14 sequences; four hit human, one hit monkey, and one overlapping design hit both—enough, the founders said, to move the program forward.
Defensibility is shifting toward an integrated design product—multi-objective prompts, therapeutic-property optimization, lab feedback and expert workflows—not merely a model. Chai-1 was open-sourced; Chai-2 is a larger pipeline whose users must specify epitopes, multiple species and off-target avoidance. Sarah Guo’s implication is sharp: leading antibody engineers become expert prompters, while Chai argues domain expertise becomes more valuable, not less.
The founders’ “bullish on biotech” case rests on observed velocity rather than current market conditions: XBI has struggled for five years amid what they call one of biotech’s worst markets in decades. They contrast a move from below 0.1% to near 20% in a year with mini-protein results near 70% across all five tested targets, then speculate—rather than promise—that some molecular classes might reach 50%-plus or nearly 100% success in another year.

Deep dive

1. Chai was founded at the prediction-to-design inflection

Meier’s timing test: AI drug discovery remained a research idea until the team saw a “one, two years” window—early enough to build before proof, but late enough that a company would not wait a decade. Their bet was that diffusion and language-model advances would carry protein folding into molecular interaction and design.
AlphaFold 2’s around-2020 breakthrough predicted individual proteins from sequence with experimental accuracy, but drugs work by modulating other molecules. Earlier predictors returned “one view on a protein,” like image systems before diffusion; Chai wanted to generate diverse antibody-antigen and small-molecule-protein interactions.
Meier’s critique of earlier AI-bio companies is that laboratory integration was often “almost too tight.” Chai still runs experiments, but its goal is a “portable AI platform” deployable across hundreds or thousands of projects; open-source Chai-1 was the first proof of that generality.
Jack Dent’s career decision followed the same logic: once engineering molecules “with atomic precision” looked plausible, the future became “impossible to unsee.” For him, it was “almost hard to work on anything else with your life.”

2. Twenty attempts replace an open-ended antibody search

Antibodies account for close to 50% of recent drug approvals and seven of the 10 bestselling drugs, making Chai-2’s design result commercially consequential rather than a narrow scientific demonstration.
The workflow starts with a specified target, generates up to 20 antibodies for a 24-well plate, and sends them for a roughly two-week validation cycle. Close to 20% of ordered designs bound as intended, while at least one hit emerged for roughly half the tested targets.
Chai had set 1% as its company-wide annual objective because previous computational attempts succeeded around 0.1% or less. The result therefore cleared the internal goal by roughly an order of magnitude and prior approaches by multiple orders.
Traditional screening may search millions or billions of candidates, “panning for gold” through yeast or phage libraries. Alternatives include immunizing mice or llamas and extracting antibodies; during COVID, researchers instead searched antibodies taken from infected humans for one that neutralized the virus.

3. The benchmark was built to defeat the one-target demo

Most AI drug-design papers test one, two or three targets; Chai tested more than 50. Meier compares a one-problem result to claiming an LLM solved an entire contest: “You need a real benchmark,” especially when randomness can make one or two trials misleading.
The target-selection method was deliberately unglamorous: Chai scraped vendor catalogs for proteins currently in stock so everything could be ordered together. The team retrieved their sequences, compared them with SAbDab—a collection of antibody structures in the Protein Data Bank—and removed targets exceeding 70% sequence identity.
The resulting panel was not selected for therapeutic importance; some proteins already had drug programs. Meier therefore treats the exercise as model assessment, not a showcase portfolio, and says the 50% target-level result may be a lower bound because the hurried setup could have introduced mistakes.
An even harder subset pushed similarity to training data down to 25%, yet the success rate was essentially unchanged. Meier’s conclusion—“the model didn’t care”—challenges biological categories that humans consider dissimilar.

4. Chai-2 generates molecular keys with atomic precision

Structure prediction supplies an “atomic level microscope”: sequence goes in, and predicted atom locations in three-dimensional space come out. Chai-2 then moves from observation to generation, producing both a new sequence and a structure intended to bind a prompted location.
Dent’s analogy distinguishes an “ImageNet moment” from “Midjourney for molecules.” If the target is a lock, Chai-2 reasons by placing individual atoms to construct its key, with error across the entire structure sometimes below the width of one atom: “How can you hope to design the key if you can’t see the lock?”
Generalization to unfamiliar targets suggests the data contains a learnable signature of how proteins interact. Meier calls that profound but unresolved: the team “still don’t fully understand” why the models work, and does not claim to have extracted the biological principles the model may encode.

5. Adoption will pair generative models with more laboratory sampling

Meier separates two sources of value: screening existing programs faster on computers, and attacking problems unreachable by traditional methods. Because Chai-2 worked on only about half its targets, the near-term sweet spot may be cases where its failure mode differs from the laboratory’s.
Dent opened access to academic groups and industry partners because drug discovery is too resource-intensive and the opportunity too broad for one company to pursue every target. Hundreds of requests arrived within hours, forcing the roughly dozen-person team to prioritize.
Industry pushback is straightforward: companies can already discover drugs, so does faster discovery change which molecules are possible? The founders’ answer is to revisit stalled programs; the stronger response circulating in the community is that this is “another tool in the toolkit” that users may need to avoid being left behind.
Meier expects more samples to improve probabilistic outputs: 10 times, 100 times or orders of magnitude more designs than 20 should explore better regions. CROs are already asking about 100,000-design runs, where a superior molecule might determine whether patients need an antibody requiring an injection or can receive subcutaneous dosing—the “best of AI” married to “the best of biology.”

6. The biotech bull case expands beyond faster monoclonal antibodies

Dent contrasts five weak years for XBI and long investment cycles with Chai’s one-year jump from below 0.1% to nearly 20%. Mini-protein experiments reached close to 70% design success, picomolar affinities and hits on all five tested targets, supporting his proposed “bullish on biotech” baseball caps.
The long-term analogy is a computer-aided design suite for biology—SolidWorks for molecular engineering or Photoshop for creatives. Beyond binding, models must optimize manufacturability and stability, while easier design could make biparatopic antibodies that combine two paratopes easier to pursue than conventional single-site monoclonals.
One partner had spent several years, five to 10 people and an estimated $5 million-$10 million seeking a molecule cross-reactive with human and cynomolgus-monkey forms of a protein. Among only 14 Chai designs, four hit human, one hit monkey, and one of those overlapping hits hit both.
Guo’s pushback is worth keeping: many industry observers see pharma’s expense and bottleneck as clinical rather than discovery. Guo argues that if discovery risk fell, the industry could become more efficient and effective. Meier responds that capital markets, clinical failures and everything required to make a drug remain “the tip of the iceberg” beyond Chai’s current result.

7. Chai’s product must turn hits into multi-objective drug candidates

Dent’s boundary condition is explicit: “These just aren’t drugs yet. They’re hits.” Chai must characterize therapeutic properties and eventually design entire drug candidates zero-shot; after seeing antibodies emerge in 20 attempts, that possibility feels less futuristic, but it remains a future investment rather than a demonstrated outcome.
Chai-1 could be treated as a model: enter sequences and receive a structure. Chai-2 is a larger pipeline and product whose interface must express sophisticated design intent, then incorporate laboratory results so the system can act as a copilot for subsequent rounds.
Prompting may specify two targets, simultaneous human and animal binding, or affinity for one protein while avoiding another. Meier’s memorable framing—“it’s a good time to be a sick mouse”—captures how multi-species design could eliminate separate surrogate antibodies and reduce the risk that animal evidence reflects a subtly different molecule.
The advice to antibody engineers is to get access, learn to prompt and “start dreaming about the new possibilities.” Choosing an epitope shared between species—or deliberately choosing a dissimilar region for selectivity—becomes high-leverage judgment; Dent expects specialists to “fall off their chair” once a hypothetical capability becomes usable.

8. A small interdisciplinary team treats research code as infrastructure

Chai’s team is around a dozen people spanning chemistry, physics, biology, AI and software engineering; surprisingly few have conventional computer-science degrees. The operating constraint is shared focus: everyone attacks the same company problem rather than pursuing a favorite research project.
Dent describes reaching a new field’s frontier as “a total fight,” with “waves of excitement and misery.” His faster route was surrounding himself with specialists and learning across disciplines, while ensuring researchers were also strong engineers capable of shipping discoveries into partners’ hands.
His architectural rule is that someone must keep the whole system in mind, preserving simplicity and modularity before software entropy stops progress. Deep-learning bugs may hide for weeks inside expensive runs; Chai has bisected Git history with repeated training jobs to locate regressions costing tens of thousands of dollars, which is why it writes unit tests even for research code.
Compute follows a scrappy but aggressive threshold: Chai began on free cloud credits, scales when it sees “signs of life” or a scaling law, and avoids scaling unproven directions. With Chai-2 launched, hiring includes product engineering, antibody engineering, business development and account-executive roles.

Why I don’t think AGI is right around the corner Figma’s IPO: The Full Breakdown & Why Melio’s $2.5BN Acquisition is “Discouraging”