DaX

Every co-crystal structure in the Protein Data Bank shows an L-amino acid ligand meeting an L-amino acid receptor. That is not a sampling gap that more deposition would close — it is the entire record, because that is the chirality biology uses. A generative model that wants to build the mirror image of a peptide, a chain assembled entirely from the right-handed enantiomers of the natural amino acids, therefore has no training examples of the thing it is being asked to make, and the design tools trained on that record emit L-chemistry because L-chemistry is all they were shown. Mirror peptides are worth the trouble anyway: proteases evolved against left-handed substrates and cut them poorly, and the fragments that would normally be presented to the immune system never appear. The established route to one, reverse mirror-image phage display, gets around the data problem by synthesizing the mirror image of the target and screening against that, and it takes months to years per target.

DaX — originally expanded as D-amino acid eXplorer — is Aizen Therapeutics' generative model for that chemistry. Aizen describes it as a foundation model trained to learn the underlying physics and geometry of receptor-ligand interaction, an abstraction of the binding interface general enough that handedness becomes something the model can be steered on rather than a fixed property inherited from the training set. That is the mechanism by which an all-L structural corpus can inform D-chirality design. The model came out of David Van Valen's lab at Caltech, was spun out into Aizen in 2022, and was named publicly in November 2024 when the company emerged from stealth with $13 million in seed funding.

The scope Aizen describes has widened since. The 2024 framing was specifically mirror peptides — full D-amino acid chains. By 2026 the company positions DaX over non-canonical amino acid (ncAA) chemical space generally, in service of oral biologics, with D-amino acids as one example of the building blocks involved.

#Key Features

• Chirality transfer: The model is trained to design in a chemistry its structural training data does not contain, learning binding at the interface rather than memorizing left-handed sequence-structure pairs.
• Guided diffusion: Generation is a guided diffusion process, the same class of method that made de novo backbone design tractable, steered here toward interfaces rather than toward isolated folds.
• Drugs, not just binders: Aizen frames the objective as designing peptide drugs — half-life, solubility and oral delivery optimized alongside affinity — rather than binders that then need medicinal chemistry to become tractable.
• Non-canonical building blocks: Design ranges over the ncAA alphabet that modern peptide therapeutics already depend on, which the company states covers ten times the scale of conventional ncAA discovery.
• Hard receptor classes: Targets include GPCRs, transmembrane proteins and ion channels, the classes that soluble-protein binder design and display screening handle least well.

#Technical Details

Aizen states that DaX was trained on millions of uniquely annotated molecules and receptors, and that its generative core is guided diffusion. The company's 2026 AI white paper is more specific about the corpus: the featurization was applied to natural protein structures and their paired ligands from the Protein Data Bank, and the training set contained no D-peptides bound to the three benchmark receptors — CD98, TROP-2 and GLP-1R — which is the stated basis for calling the D-peptide results zero-shot generalization. That document also diagrams the binding-interface featurization and reports hit-confirmation rates and potencies for each of those three targets, including 12% hit confirmation and ~635 nM EC50 cAMP activation for de novo GLP-1R agonists. No parameter count and no training-corpus manifest have been published, and no preprint or peer-reviewed article describes the model — every technical claim about it traces to the company or to trade press relaying the company. Neither code nor weights are available, and DaX is reachable only through a partnership with Aizen.

The reported validation is likewise company-sourced: Aizen's scientists synthesized D-peptides generated by the model and tested them against known receptors, and report demonstrated binding activity against multiple clinically relevant target receptors. The company puts a full design-build-test cycle at three to four weeks, against the six-to-twelve months or longer that mirror-image display campaigns require.

#Applications

Aizen's own pipeline is oral peptide therapeutics for chronic immune disorders, both gut-restricted and systemic, and the company has also described mirror peptides as targeting moieties for radioligand delivery into solid tumors and for transport across the blood-brain barrier. The commercial pattern is partner-nominated targets: a multi-program collaboration announced in August 2026 with an unnamed San Diego public biotech points DaX at immunology and neurology targets chosen by the partner, with milestone payments of up to $100 million per target.

#Impact

DaX is one of the clearest industrial bets on chirality as a design axis rather than a chemistry curiosity, and it arrived alongside academic work attacking the same transfer problem — PepMirror learns from homo-chiral binding data and generalizes to hetero-chiral design through explicitly handedness-aware geometric features. The contrast is instructive in a second way: PepMirror publishes its method and its wet-lab affinities in the literature, while DaX's disclosure stops at a company white paper. That paper does name targets, hit rates and potencies, so the claims are at least specific — but with no code, no weights, no released sequences and no peer review, none of them can be checked from outside, and DaX's place in the catalog rests on what a venture-backed company says about a system it has not described in the literature.