Aizen Therapeutics, which raised $13 million less than two years ago, has signed a collaboration worth up to $100 million per target with an unnamed public San Diego biotech to design oral peptide therapeutics with its DaX AI platform. The first commercial test of DaX-designed non-canonical amino…
Aizen Therapeutics, a seed-stage company that raised $13 million less than two years ago, disclosed this week a multi-program collaboration with an unnamed public San Diego biotech to design oral peptide therapeutics. The agreement carries several million dollars upfront and up to $100 million per nominated target in milestone-linked economics , a nine-figure per-target ceiling. It is the first disclosed commercial test of the DaX foundation model , Aizen's AI platform for designing peptides built from non-canonical amino acids .
The partner has nominated targets in immunology and neurology, though the specific proteins are undisclosed. Aizen's DaX-designed peptides have not yet been put in a pill, and none has advanced to an investigational new drug filing. The first job of the collaboration is narrower than its headline value suggests: demonstrate in vivo target engagement for DaX-designed oral peptides, with the partner's upfront payment funding that proof.
The clinical rationale is well established. Peptides offer excellent target specificity, but they almost universally require injection because the gut degrades them before systemic absorption. Aizen's approach uses non-canonical amino acids, chemical building blocks outside the standard 20, to create structures with improved stability and oral bioavailability. Whether that chemistry works in humans is precisely what the partner is paying to learn.
The financial structure shows how the partner is pricing platform risk. The per-target ceiling represents milestone-linked economics of up to $100 million for each nominated target, paid only if the collaboration hits successive development and commercialization benchmarks. The upfront payment is described only as "several million dollars," and the exact amount has not been disclosed. A partner writes a modest check to test a platform and reserves the large sums for results.
The gap between the upfront and the ceiling is the risk allocation in one number. For a seed-stage company with no clinical-stage assets, the value a licensee will commit before seeing data is small. Nine-figure per-target milestone economics, by contrast, are characteristic of deals in which the developer bears most of the technical risk and the licensee pays for success. Payments at that scale typically stand behind defined events such as IND clearance, proof-of-concept data, registration trial initiation, approval, and commercial launch, though Aizen's disclosure does not itemize them.
The most telling provision may be the option window:
A 12-month option clock puts early pressure on the first workstream. The partner will learn within roughly a year whether DaX can produce credible leads for its nominated targets, and it must decide on additional targets before any IND-quality data exists. That timing makes the first workstream decisive for the platform's commercial credibility, independent of the dollar figures attached to later milestones. The clock also disciplines the partner's own portfolio planning: the decision on additional targets will rest on preclinical evidence of engagement, potency, and oral exposure, not on clinical results.
The structure also reveals what each side is betting on. Aizen is betting that the first nominated targets validate the platform quickly enough to trigger a second nomination. The partner is betting that it can pay a small option price now and decide, with better information, whether to pay the larger sums later. The asymmetry is deliberate, and the 12-month window is the mechanism that forces the decision while the information is still fresh.
The regulatory frame is set by the FDA . The agency evaluates non-canonical amino acid-containing peptides, including DaX-designed oral peptides, under existing NDA frameworks, and there is no accelerated designation tied to the delivery innovation alone. Oral bioavailability is a drug property, not a regulatory category.
That means the standard requirements apply in full: pharmacokinetic characterization, toxicology, and manufacturing controls. A peptide that survives the gut and reaches its receptor still must show acceptable absorption, distribution, metabolism, excretion, and safety like any other new molecular entity. The agency will also want to understand the behavior of the non-canonical residues themselves, including their metabolic fate and any immunogenicity signal. None of these obligations is shortened by the fact that the molecule was designed by an AI model.
For an oral peptide, the pharmacokinetic package carries extra weight. The FDA will expect data on absorption under fed and fasted conditions, the effect of gastrointestinal pH, food interactions, and the variability of exposure between patients. The non-canonical residues add a further layer: each unnatural building block must be characterized for its own metabolic products, its potential to accumulate in tissues, and its contribution to the drug substance's impurity profile.
A further point follows from the legal structure. Synthetic peptides generally fall under the NDA pathway rather than the biologics license pathway , because they are chemically manufactured rather than derived from living systems. Nothing about non-canonical amino acid chemistry changes that classification. Delivery innovation alone will not change the regulatory path, and in vivo proof of target engagement remains the essential validation step before any IND.
The NDA route has a long history with synthetic peptides. Enfuvirtide , a 36-amino-acid HIV fusion inhibitor, was approved through an NDA in 2003, and its chemistry, manufacturing, and controls package established the pattern for synthetic peptide drug substances: defined impurity profiles, controlled synthesis, and full characterization of every building block. A non-canonical residue extends that obligation rather than changing it, because the agency will expect impurity, stability, and fate data for every monomer, whether or not that monomer has appeared in an approved drug before.
The standard genetic code specifies 20 amino acids . Non-canonical amino acids fall outside those 20, and they are the raw material of Aizen's approach. Substituting these building blocks into a peptide sequence can yield structures that resist the proteolytic enzymes of the gastrointestinal tract, resist rapid systemic clearance, and carry enough stability to reach circulation after oral dosing.
The oral barrier is a cascade, not a single wall. In the stomach, acidic conditions and pepsin begin the degradation. In the small intestine, pancreatic endopeptidases such as trypsin and chymotrypsin cleave internal peptide bonds, while exopeptidases on the brush border remove residues from the termini. The mucus layer traps large or charged molecules, the epithelial tight junctions exclude paracellular passage, and efflux transporters at the apical membrane pump out compounds that do cross. Any oral peptide must survive every layer in sequence, and then withstand first-pass hepatic metabolism after absorption.
Non-canonical residues can address several of these layers at once. Unnatural side chains and backbone modifications can evade peptidases, while reduced hydrogen-bonding capacity can improve passive permeability across the lipid bilayer. The tradeoff is that each substitution can also reduce potency or introduce metabolic liabilities, which is why the search space matters.
Specific edits illustrate the trade. N-methylating a backbone amide nitrogen removes a hydrogen-bond donor that proteases and membranes both recognize, which can improve stability and permeability in a single change, but it can also delete contacts needed for receptor binding. Replacing an L-amino acid with its D-enantiomer blocks the stereospecific peptidases of the gut, at the cost of a residue whose metabolic fate the FDA will expect to see characterized. Neither edit is free, and the design model exists to find substitutions whose penalties are smaller than their gains.
The scale of the problem favors a computational approach. With 20 canonical amino acids, a 10-residue sequence represents 20^10 possible combinations, more than 10^13 candidates, and the non-canonical vocabulary multiplies that figure by orders of magnitude. Brute-force screening of such a space is impractical. A model that can propose promising scaffolds and rank them before synthesis is not a convenience; it is a prerequisite.
Aizen describes DaX as a foundation model , a term borrowed from AI research for a system trained once on a broad corpus and then adapted to specific tasks. In this case the corpus is millions of annotated molecules and receptors. The annotation matters as much as the volume: the model is not merely cataloging structures. It is learning associations between chemical form and biological behavior, including the binding, stability, and permeability properties that determine whether a peptide can be swallowed and still act on its target.
For a design task, the model proposes sequences built from a vocabulary that extends beyond the standard 20 amino acids. Each non-canonical residue carries its own side chain, backbone geometry, and metabolic sensitivities. A proposed sequence therefore bundles several predictions at once: that the peptide will adopt a structure compatible with the target, that it will resist gut proteases, and that it will cross the intestinal epithelium. The output of the model is not a molecule but a ranked set of candidates, and the value of the platform depends on whether those rankings concentrate the scarce resources, synthesis and animal testing, on sequences that actually work.
DaX was incubated at Caltech, and Aizen claims it samples the non-canonical amino acid chemical space at ten times the scale of conventional discovery methods. That is a claim about reach, not about accuracy. Tenfold coverage of a larger search space only matters if the prioritization holds, and the only way to test prioritization is to make the molecules and put them in biological systems. Nothing in the training data substitutes for that. The collaboration's first workstream is built around exactly this test: the partner's upfront money funds the synthesis and in vivo evaluation of DaX-designed leads against the nominated targets.
There is a further constraint that no model can remove. De novo peptide design is data-hungry, and every new building block added to the non-canonical vocabulary increases the demand for training data that captures its behavior. The model's predictions for well-characterized residues will be more reliable than its predictions for exotic ones. That is why the first leads matter: their pharmacology will show whether the training corpus generalized to novel chemistry or only memorized familiar patterns.
For researchers, the deal is a live test of whether computational design can solve the two problems that keep peptides injectable: proteolytic degradation in the gastrointestinal tract and poor membrane permeability. The lessons will transfer broadly. If DaX-designed peptides show in vivo target engagement after oral dosing, the…
Peptides referenced: Enfuvirtide.
Related reading: Meant launches LegitScript-certified GLP-1 telehealth platform, NCI tests neoantigen vaccine in metastatic triple-negative breast cancer, GLP-1 Analogs for Psoriatic Arthritis With Obesity and Diabetes, Peptide Nanostructures Kill Resistant Bacteria and Restore Meropenem.