Xaira shows X-Design AI making drug-ready antibodies in weeks — including one GPCR traditional methods missed
Xaira Therapeutics on Friday published results for X-Design, its generative AI for antibodies, showing it produced a progressable binder for lead oncology program XA-1 in seven weeks and unlocked XA-4, a hard GPCR target where conventional discovery had already failed.
A billion-dollar AI drug shop just showed wet-lab receipts — weeks to a progressable oncology binder, and a GPCR antagonist where display and immunization already failed. That is the bar AI drug discovery has to clear if it wants to change what gets into the clinic.
On Friday, 2 October 2026, Xaira Therapeutics published “De Novo Design of Progressable Antibodies to Therapeutic Targets.” The byline is the Xaira Design Team. The page dates the post October 2, 2026, and does not print an hour. The post says the company’s model, X-Design, did two things on real pipeline programs: it made a drug lead faster, and it reached a target that had already resisted other methods. The programs are XA-1 and XA-4. An antibody is a protein the immune system uses to grab a target. Drug companies make antibody drugs to do that on purpose. De novo, the word in the title, means the model wrote the antibody from scratch instead of pulling it from a preexisting collection. Those lines are Xaira’s.
What Xaira says the model is aiming at. The Vega class of X-Design, the post says, optimizes at the same time for a binder that is human, potent, selective, cross-reactive, and developable. It is not a model that only tries to stick. Potent means a small amount has a strong effect. Selective means it prefers the intended target over close relatives. Cross-reactive, here, means it also grabs the same protein in the animals a company tests before any study in people. Developable means it can be made, stays stable, and does not stick to the wrong things. Xaira calls a molecule that clears those bars a progressable binder. The post says the designs below were generated zero-shot: no data from previous campaigns was used to train the model. Zero-shot, in that sentence, means the model had not been shown earlier lab results against these same targets. Xaira says a fuller assessment of X-Design is coming in a later post. Those lines are Xaira’s.
XA-1 is the speed case. Xaira calls it the lead oncology program. Oncology means cancer. The campaign needed a binder to a defined epitope, the exact patch on the target, that stays selective against closely related family members, also grabs the matching protein in a surrogate species, and lands inside a set window of affinity. Affinity is how tightly the antibody holds on. A surrogate species is the animal used to check the drug before people. Xaira used Vega to design 182 VHH sequences from scratch against that epitope. A VHH is the binding tip of a camel or llama antibody, one small chain instead of the usual pair. People sometimes call it a nanobody. The company then measured binding, against the human target and the surrogate, with surface plasmon resonance, a standard lab method for how tightly two molecules hold. Those lines are Xaira’s. The post does not name the protein.
What that screen found, as Xaira states it. The first pass identified 25 hits. A hit, here, is a design that showed binding. After full quality-control filtering, 18 were confirmed binders. The post calls that a 10 percent binder rate. Eighteen of 182 is about 1 in 10. Nine of the confirmed binders gave high-quality kinetic fits, meaning the binding curve was clean enough to trust the tightness number. The best binder had a KD of 17 nanomolar on the human target and 83 nanomolar on the cynomolgus monkey version. The post prints those as 17 nM and 83 nM. KD, the dissociation constant, is the lab’s number for how tightly two molecules hold. A lower number is a tighter hold. Nanomolar means billionths of a mole per liter: a very small amount of the antibody still grabs the target. Seventeen nanomolar is a tight early binder. Eighty-three is looser, and still in that same tight range, on the monkey protein. Cynomolgus, often shortened to cyno, is a monkey species used in safety work before human studies. The best molecule passed the company’s developability checks and scored as human-like on a score the post calls OASis. OASis compares the sequence with antibodies seen in people. One more optimization round improved the monkey cross-reactivity and produced a preclinical lead. A preclinical lead is a molecule chosen for the lab work that comes before any test in people. Those numbers are Xaira’s, from the company’s own screen. The post does not say an outside lab repeated it.
The clock, and the limit that sits next to it. Xaira says the design step produced a progressable molecule, human-like and developable at once, in three weeks of wet-lab work. Wet lab means the samples were real, not only a computer result. The full run from DNA synthesis to the lead molecule took seven weeks. DNA synthesis, here, means ordering the gene so cells can build the antibody. The figure caption splits the same seven weeks: three weeks to the progressable binder, then one round of optimization for the surrogate. A later paragraph says the lead came in an additional four weeks. Three plus four is seven. That sum is arithmetic on the post’s two spans. The post also says the speed was possible in part because prep was already done. The antigen, the target protein, and other reagents were ready. Most crucially, Xaira says, the desired epitope had been chosen ahead of time. An earlier version of X-Design could design binders to this target, the post says, but those needed more work before they were progressable. Vega’s lead arrived at roughly the same time those earlier hits were being turned into leads. Those lines are Xaira’s. The seven weeks is this campaign’s clock. It is not a promise that every future target will take seven weeks, and it is not a date for a study in people.
The comparison in the company’s Figure 1. Xaira says Vega triples the rate of human-like VHH sequences that meet every developability check, versus a baseline model trained without the company’s proprietary data. Triple means three times as many of the designed sequences cleared all of those checks, in that comparison. Proprietary data is lab data the company owns. The post says X-Design was built on public data plus proprietary data expanded over several years, and that the proprietary data helped Vega balance binding against humanness and developability. That triple is Xaira’s reading of its own figure. The zero-shot line is a different claim. The designs for these targets were not trained on earlier campaigns against those targets. The model still had the company’s broader data behind it.
XA-4 is the hard target. Xaira calls it a GPCR, a G-protein-coupled receptor. That is a protein in the cell’s outer membrane that passes a signal inward. Many medicines aim at this class. Antibodies often miss, because the part that sticks out of the cell is small. XA-4, the post says, does not have a well-understood epitope. It carries extensive post-translational modifications, chemical tags the cell adds after the protein is built. Those tags can hide or change the spot a drug would grab. The region outside the cell is small and flexible, and the link from epitope to function is unclear. Xaira says there is no preexisting antibody that broadly blocks it. An antagonist is a molecule that blocks the receptor instead of turning it on. The company calls XA-4 the most challenging GPCR in its current pipeline, and says earlier versions of X-Design had already failed on it. The goal was an antagonist to the human protein that also grabs the cynomolgus and mouse versions. Those lines are Xaira’s. The post does not name the receptor.
What ordinary discovery did, as Xaira states it. While the models advanced, the company also tried conventional methods, at least to get tools for studying the target. The post’s introduction says prior display and immunization efforts had already failed on XA-4. In the details: two campaigns of a conventional naive VHH library, with four rounds of selection, produced some putative hits. Those sequences grew more common in the mix, then failed to bind when each one was made on its own. A naive library is a large preexisting collection, not sequences written for this target. A llama immunization, injecting the target into a llama and harvesting the antibodies the animal made, also yielded no functional binders. Xaira says those results show XA-4 is not readily druggable by conventional means. Those lines are the company’s account of its own campaigns. The post does not name an outside lab that ran them.
What Vega designed instead. Because the working epitope and the structure were not clear, Xaira says it modeled the target bending across several candidate patches. After probing the target with designed minibinders, small designed proteins used as probes, Vega designed a library of 60,000 VHHs. One campaign could then cover multiple shapes and patches. The company screened that library against the target held in nanodiscs, small membrane-like discs that keep a membrane protein closer to the shape it has in a cell. An enrichment model ranked candidates across rounds of selection. Xaira took 92 of the enriched molecules and checked binding on cells that display the target, against control cells that do not. A hit had to show more than a threefold signal on the target cells. The figure caption says that check found 7 binders. In a beta-arrestin reporter assay, run against the receptor’s natural ligand, the company found 6 confirmed inhibitors. Beta-arrestin is one of the proteins a cell uses when this kind of receptor fires. The assay watches that response in cells in a dish. The post prints the assay with the Greek letter beta. It is not a result in a patient. The best candidate was made as a VHH-Fc, the small antibody fused to the stem of a normal antibody. It reached more than 96 percent maximum inhibition, with an IC50 of 183 nanomolar. IC50 is the concentration that cuts the measured response in half. The figure caption calls that complete antagonism. Those figures are Xaira’s.
What else that lead did, as the post states it. It binds human cells that display the antigen with an EC50 of 32 nanomolar. EC50 is the concentration that gets halfway to the strongest effect in that test. It also grabs the mouse and cynomolgus versions of the protein. It has a clean developability profile and a high OASis score. Xaira says it has entered lead optimization, the next lab stage of making the molecule a better drug candidate. Lead optimization is not a trial in people. The post says progressable antibodies to XA-1 and XA-4 do not guarantee that either program will succeed. Those lines are Xaira’s.
Two easier GPCRs, kept separate from the pipeline programs. Where the structure is known and the epitope is clear, Xaira says X-Design can design antibodies from scratch at a high hit rate. The examples are CXCR4, a chemokine receptor, and APJ, the apelin receptor. The company designed 91 structurally different VHHs against each, made them in a two-arm antibody format, and tested them on cells that overexpress the receptor. A hit was a greater than threefold signal over control cells. Xaira reports 32 of 91 on CXCR4, a 35 percent hit rate, and 26 of 91 on APJ, a 29 percent hit rate. The best EC50s sat next to the positive controls: 12.8 nanomolar for CXCR4 against a control of 8.6, and 8.4 nanomolar for APJ against a control of 8.8. Those are Xaira’s screens of better-characterized receptors. They are not XA-1 or XA-4, and the post does not present them as drugs.
What Fierce Biotech reported the same day. Darren Incorvaia’s article is dated Oct. 2, 2026, 11:00 a.m. The page does not print a time zone beside that clock. Chief executive Marc Tessier-Lavigne told Fierce that X-Design is “central,” and that designing antibodies that are “drug-ready from the start” has been a major focus. He said XA-1, a cancer candidate, was created from scratch in seven weeks. The article says the targets of XA-1 and XA-4 are not being disclosed. It calls XA-4 a tricky G-protein-coupled receptor that was otherwise “seemingly intractable.” Tessier-Lavigne said both programs are targets the company has “great conviction” in, “but where prior efforts to make antibodies have failed.” Xaira’s own post is more specific about which effort failed where. The display screens and the llama immunization are the documented misses on XA-4. XA-1 is the speed campaign on an epitope the company had already chosen. Fierce also says Xaira plans to advance both candidates into human studies, and that the chief executive declined to share any timing. The article names two other preclinical candidates, XA-2 and XA-3. Preclinical means before studies in people. The article does not say any of the four is in a human trial.
The money line, and whose number it is. Fierce says Xaira debuted in 2024 with $1 billion in funding. That $1 billion is the debut raise, as Fierce states it. It is not a figure in Friday’s Xaira post, and it is not a new round announced on Friday. Tessier-Lavigne was previously president of Stanford and chief scientific officer at Genentech, Fierce notes. He told Fierce the company wanted enough capital that it would not have to return to investors until it had molecules in the clinic. He also said he is open to raising sooner if investors are interested or a program can move faster. Fierce says he would not discuss how much cash is left. David Baker, a protein designer and an Xaira co-founder, said in a statement Fierce carries that some disease-driving proteins have stayed off-limits to antibody drugs because traditional discovery failed, and that “X-Design shows that AI is beginning to design those antibodies from scratch.” That quotation is his, in Fierce. The article also names two other Xaira models, X-Cell and X-Patient, beside X-Design. It does not attach Friday’s binding numbers to those models.
The picture is a navy card with a pink bar along the left edge. White type reads XAIRA. Pink type reads X-Design · Vega. Under that, the card lists an XA-1 oncology lead, then 7 weeks · 17 nM KD, then XA-4 hard GPCR, then traditional discovery failed, then AI antagonist progressed. The right side is the blue Xaira wordmark on a white panel, with a dot over the i. The frame does not print a calendar date. It is a summary card of the company’s Friday post. It is not a photograph of a lab, a gel, or a patient.
In plain terms, Xaira said on Friday that X-Design Vega designed a progressable binder for its lead cancer program, XA-1, in three weeks of lab work, and a preclinical lead in seven weeks from ordering the DNA. The epitope for that program was already chosen. The best binder held the human target at 17 nanomolar and the monkey version at 83. On XA-4, a hard cell-surface receptor where a naive antibody screen and a llama immunization had not produced a working binder, a library of 60,000 designs produced a blocking antibody. That molecule has entered lead optimization. Xaira says it cuts a cell signal in half at 183 nanomolar, binds human target cells halfway at 32 nanomolar, and also grabs the monkey and mouse proteins. On two better-known receptors, CXCR4 and APJ, the reported hit rates were 35 percent and 29 percent. Fierce’s same-day interview quotes the chief executive calling X-Design central, and saying XA-1 was made from scratch in seven weeks. The protein names are still undisclosed. The company told Fierce it plans studies in people and did not give a date. Neither source says either molecule is being given to patients. The $1 billion is Fierce’s figure for the 2024 debut, not a new raise.