Talus Bio opens Ptarmigan-1, a structure-free AI model for proteome-wide drug screens

Talus Bioscience said Thursday it released Ptarmigan-1, an AI model that predicts small-molecule binding sites across the human proteome without modeling a 3D protein structure, and made access available through a public portal.

Most AI chemistry still assumes a pocket you can see. Talus is shipping a public model aimed at the proteins that refuse to hold still in a fold, which is where a lot of hard disease biology lives.

On Thursday, 1 October 2026, Talus Bioscience, Inc. said it released Ptarmigan-1. The company shortens its name to Talus Bio and calls itself an AI-native therapeutics company. Therapeutics, here, means drugs. The dateline is Seattle. Business Wire carries the release. BioSpace reprints it and dates the item October 1, 2026. MarketMinute’s reprint of the same wire stamps October 1, 2026, at 9:00 a.m. Eastern. The deck says Ptarmigan-1 is available now through a public portal, and calls it the first AI model to predict how small molecules bind to all proteins, including disordered and historically undruggable targets, without modeling a 3D structure. First, in that sentence, is Talus’s word. A small molecule is a drug-sized chemical, the kind a person can swallow, rather than a large protein drug. A binding site is the place on a protein where that chemical sticks. Disordered means the protein does not hold one fixed shape. Those lines are the release.

What the model is supposed to do, as the release states it. Ptarmigan-1 predicts small-molecule binding sites across the entire human proteome, including proteins too disordered for traditional 3D modeling to resolve. The proteome is the full set of proteins the human body can make. Traditional 3D modeling, here, means building a shape of the protein on a computer and then fitting the chemical into a pocket on that shape. The release says most AI drug-discovery tools still work that way: they fit a molecule onto a defined protein pocket. Ptarmigan-1 is designed to drop that requirement. Access, the release says, is available through a public portal. Those lines are Talus’s.

How large the gap is, on the release’s own figures, and how the chief executive puts it. Approximately 40 percent of the proteome, including many disease-relevant transcription factors and regulatory proteins, are flexible and lack a 3D structure for a drug to grab. Forty percent is two in five. A transcription factor is a protein that helps turn a gene on or off. Regulatory proteins are the ones that steer what a cell does. Alex Federation, PhD, chief executive and co-founder, is quoted on the release. He said that after the 2024 Nobel Prize in Chemistry, which recognized advances in protein structure prediction and design, scientists have focused more on using structure to drive discovery. “With Ptarmigan-1, we now have a complementary approach for roughly half of all human proteins that have evaded drug discovery because they can’t be folded in a computer,” he said. Complementary means an added route, beside the structure-based one. Roughly half is his phrase. The body of the release says approximately 40 percent. Those are two wordings of the same problem. They are not one number printed twice. The quotation is his, in the release.

The STAT6 example, as the release states it. On a recently disclosed STAT6 inhibitor series, Talus says Ptarmigan-1 outperformed structure-based methods at picking successful drug candidates, even though it had never seen that target or those molecules in training. STAT6 is a transcription factor. The release calls it a validated target for inflammatory disease, with a binding site that is disordered and hard to model. An inhibitor is a molecule meant to block the protein. Outperformed, in that sentence, is the company’s comparison. The release does not print the score, and it does not name the structure-based methods. The same paragraph says the model also found new small molecules that bound the flexible pocket on STAT6, and that those candidates were checked in a third-party lab. A third-party lab is a lab outside Talus. The release says that check shows the model can surface new chemistry from the protein’s sequence alone. Sequence is the order of building blocks in the protein, the letters, rather than a folded shape. The release does not name the lab. It does not describe a clinical trial, and it does not say a medicine from this work is approved. Those lines are Talus’s.

The speed claim on the release. By skipping protein folding, Talus says the model runs 5,000 times faster than structure-based methods and can screen 3.4 billion compounds against all human proteins in a day. Five thousand times means a job that takes a structure-based method a long stretch would, on this ratio, take one five-thousandth as long. 3.4 billion is the size of the chemical library in that sentence. A day is the window the release prints. The point of the speed, the release says, is to uncover cryptic pockets that a still picture of a protein cannot show. A cryptic pocket is a binding spot that is not obvious when the protein is drawn as one frozen shape. Those figures are Talus’s, in the wire. The release does not say a visitor on the public website can run that 3.4 billion screen.

The company site states the speed in a tighter way, and it points the big screen at a preprint. The site says Ptarmigan-1 is roughly 5,000 times higher throughput than Boltz-2. Throughput, here, is how many compounds the model can score in a given time. Boltz-2 is a structure-based model that builds a 3D fit. A footnote on the site says that comparison is the reported EGFR benchmark on one H100, measured over batched runs. EGFR is a protein often studied in cancer. An H100 is a high-end NVIDIA chip used to run the test. Batched means many compounds were scored together, not one at a time in a separate job. The same page says a search of 3.4 billion compounds across 20,431 proteins, with the top matches retrieved in 20 H100 GPU-hours after the library is embedded, is reported in the preprint, and that predictions still need experimental validation. Experimental validation means a lab has to test what the model suggests. 20,431 is the site’s count of human proteins in that screen. The wire says all human proteins. The site’s count is the more specific figure, and it is labeled as the preprint’s. Those lines are the company site’s.

What powers the model, as the release states it. Ptarmigan-1 is powered by MARMOT, Talus’s proprietary proteomics platform. Proteomics is the work of measuring many proteins at once. The release says MARMOT measures proteins at work inside living human cells, rather than in a test tube, and that this opens the door to complex biology across oncology, autoimmune disease, and more indications. Oncology is cancer. Autoimmune disease is illness in which the immune system attacks the body. An indication is a disease a drug might be aimed at. Opens the door is the company’s phrase for where the data might reach. It is not a list of approved drugs. The company site says Talus has tested 200 million compound–protein pairs in living human cells. Two hundred million is the site’s count of those measurements. It is not the 3.4 billion chemical library. Those lines stay with the page that printed them.

Lindsay Pino, PhD, chief technology officer and co-founder, is quoted on the release. “The data we’re building at Talus is structure-agnostic, meaning we can measure proteins whether or not they hold a fixed shape,” she said. Structure-agnostic means the measurement does not require the protein to sit still in one fold. “That means the model can learn just as well from flexible or intrinsically disordered proteins as it does from structured ones, which is what lets it generalize to targets nobody’s had a way to study before.” Intrinsically disordered means the protein’s own sequence does not settle into one stable shape. Generalize means the model can score a target it was not built around. That quotation is hers, in the release.

A scientific advisor is quoted on the same release, and the quotation is about chemistry, not a customer deal. Gavin Hirst, PhD, vice president of small molecule drug discovery at insitro and a scientific advisory board member at Talus Bio, said the point is new chemistry as well as new targets. “Structure-based methods are strongest on chemotypes that already have a solved 3D complex, which quietly biases every campaign toward the chemistry we have already explored,” he said. A chemotype is a family of related chemicals. A solved 3D complex is a picture, from an experiment, of a drug already sitting on a protein. Bias, in that sentence, means the method keeps favoring chemicals that already have such a picture. “Scoring without a pose breaks that coupling, and the speed of this model lets us test far more hypotheses, breaking us out of the cage of common chemistry.” A pose is the 3D fit of a molecule in a pocket. Scoring without a pose means ranking the chemical without building that fit. That quotation is his, in the release. The release lists his job at insitro and his seat on Talus’s advisory board. It does not say insitro is a customer.

What a person can actually run today, on the company blog and on the portal page. The blog, dated October 1, 2026, is by Will Fondrie, Alex Federation, and Lindsay Pino. It says Talus introduced Ptarmigan-1 in July, and that the public portal at ptarmigan.app opened on Thursday so a person can try predictions on a human protein. The portal screens one protein against up to 10,000 compounds, and it allows 10 screens a day. Ten thousand is the library cap. Ten a day is the screen cap. A visitor can use a default library of known drugs, probes, and decoy compounds, or upload a library. A probe is a chemical used to study a protein. A decoy is a chemical put in the list as a comparison, not as an expected hit. After the run, the blog says the page returns the top predicted hits. Each hit can be opened to see where the model thinks the compound may bind, down to a stretch of the protein chain, and those predictions can be drawn on a shape from AlphaFold DB. AlphaFold DB is a public set of predicted protein shapes. The blog says Ptarmigan-1 does not use those shapes. They are a picture for the person looking at the result. The blog also says larger screens are still to come, including through AI agents, and that a group that wants that scale now can write to the company. The release gives that address as bd@talus.bio, for expanded use or a larger campaign on one target. The pages do not print a price.

The about box on the release, and the limit on what Thursday’s news is. Talus says it enables AI-native drug discovery for the disordered proteome. It calls the flagship model Ptarmigan, and says that model opens the roughly 40 percent of the human proteome long treated as out of reach for structure-based methods. A lab-in-the-loop workflow, the about box says, discovers and optimizes small molecules against hard targets, including transcription factors. Lab-in-the-loop means the software and the lab experiments correct each other, rather than the model running alone. The about box calls the company’s protein-versus-small-molecule dataset the world’s most comprehensive. That superlative is Talus’s. The media contact is Thermal for Talus Bio, at press@talus.bio. Thursday’s announcement is the public portal. The July introduction of the model is the blog’s earlier date.

The July paper is a preprint, and its STAT6 test is a look backward, separate from the wire’s sentence about a third-party lab. bioRxiv posted it on July 30, 2026. The authors are at Talus Bioscience. A preprint is a paper the authors have shared before journal peer review. The page says it was not certified by peer review. The summary says a patent-only STAT6 inhibitor series, unseen in training, is a retrospective test: the model recovers the active molecules and points at the right place on the protein, where a structure-based baseline points at the wrong place. Retrospective means the authors looked back at molecules already disclosed. The figure caption says the actives are 40 inhibitors from two Pfizer patent disclosures. On one shared set, the authors report an AUC of 0.94 for Ptarmigan-1, against 0.58 for a docking method that was given the binding site and 0.58 for Boltz-2 run without a specified pocket. AUC is a ranking score from 0 to 1. A score of 0.5 is a coin flip. 0.94 means the model usually ranked a real inhibitor above the decoys, on the authors’ count. 0.58 is only a little above a coin flip. The caption says that shared set is the 40 actives plus 597 of 826 property-matched decoys, because Boltz-2 did not return a score for the other 229. On all 826 decoys, the caption says Ptarmigan-1 scores 0.92 and docking scores 0.62. The caption also says no public crystal structure shows this patent series bound to STAT6, so the exact spot the model names is a prediction. Those scores are the authors’, in the preprint. Pfizer, in that caption, is the company on the patents. The preprint does not say Pfizer hired Talus. The October 1 wire’s line about new molecules checked in a third-party lab is the wire’s. It is not this patent look-back.

The picture is the Ptarmigan-1 portal at ptarmigan.app. A dark navy page carries the Ptarmigan-1 wordmark and a small bird mark. The headline reads “Screen compounds against your protein target.” The line under it reads “Screen up to 10,000 compounds against a single protein target from the human proteome.” A counter reads “10 screens remaining today.” Three steps sit in a row: select a target, select compounds, and view results. The second step is the one filled in, and its text reads “Start with our default library or upload your own compounds.” The 10,000-compound line and the 10-screen counter match the limits on the October 1 blog. The page does not print a calendar date.

In plain terms, Talus said on Thursday that Ptarmigan-1 is open on a public portal. The model scores where a small molecule might stick on a human protein without first building a 3D fold, which is the company’s route into proteins that will not hold still. Talus says about 40 percent of the proteome is in that group, and Federation’s quotation puts the same idea at roughly half. The wire’s speed line is about 5,000 times faster than structure-based methods, and a screen of 3.4 billion compounds against human proteins in a day. The site ties that speed to a Boltz-2 comparison and says the giant screen is a preprint figure that still needs a lab check. The portal a person can open today is smaller: up to 10,000 compounds, and 10 screens a day. The STAT6 result in the wire is Talus’s account, including a third-party lab check the release does not name. The July preprint’s 0.94 score is the authors’ look back at 40 patent molecules. The pages do not print a price, and they do not say the Food and Drug Administration cleared Ptarmigan-1.