TriaX is a tensor-trained foundation model that represents proteins in full conformational flexibility across their entire interaction surface. One generalized model for protein–protein interaction modulation and induced-proximity design — the molecular glues, degraders, and modulators that work by controlling which proteins come into contact.
Three axes — all flexibility, all spatial coverage, and induced proximity — trained into one tensor model, so a single system speaks the grammar of protein interaction across targets it has never seen.
Protein–protein interactions drive most of biology, and most of them remain undruggable. Their interfaces are broad and flat, their binding sites appear only when the protein moves, and the newest therapeutics don't block a target at all — they hold two proteins together. Conventional tools, built on static structures and known pockets, miss all three of these realities. TriaX was built to model them directly. The name is literal: three axes of the same problem — flexibility, spatial coverage, and proximity — learned together in one model rather than bolted on one at a time.
Most models pick one simplification — a rigid structure, a known pocket, a single target. TriaX refuses all three. Each axis feeds the same shared representation, so gains on one carry to the others.
Proteins are ensembles, not snapshots. TriaX represents the full conformational landscape — including transient and cryptic states — so interactions are scored against how a protein actually moves, not one frozen structure.
Interaction interfaces are large and historically undruggable. TriaX models the entire spatial surface of a protein rather than a handful of known pockets, surfacing modulation sites across the whole interface.
Proximity drugs work by holding two proteins together. TriaX models the full ternary complex — target, effector, and the molecule between them — to design glues and degraders, not just single-target binders.
TriaX learns a shared language of protein interaction that generalizes across targets. The same model proposes modulators and proximity-inducers for interactions it has never encountered, instead of being retrained from scratch for every new program — the difference between a tool per target and a foundation for the whole field.
A generative model that cannot be wrong cannot be useful. Every prediction TriaX makes is checked against complexes whose interfaces are already experimentally characterized, under the same settings that produced the prediction — because a method that can't recognize a known interface hasn't earned the authority to rule one out.
We screen for physical possibility before spending compute, calibrate every score against reference complexes, and say plainly where a model's assumptions end. Stating the limits of a method is what makes its positive results credible.
Read how we validate →
Two companies, cleanly separated. TriaX AI builds and owns the design engine and licenses it, program by program, to TriaX Therapeutics, which owns composition of matter on the resulting compounds and runs the biology. The engine stays reusable across future programs; each program carries its own composition IP.
The foundation model and the physics beneath it — geometry, energy and induced fit — together with the methods that turn a pair of protein modules into a ranked set of designed molecules. Retains the platform and its methods. Delivers designed, ranked leads with the provenance behind every judgement.
Takes licensed leads into synthesis and biology, and owns composition of matter on the conjugates. Runs discovery biology, DMPK and tox, and the work required to move an asset toward the clinic. Delivers the data that decides whether a designed molecule is a real one.
Licences are exclusive and field-limited, granted one program at a time. It is a deliberate structure: the platform company is not spending itself down on a single asset, and the pipeline company is not renting a capability it cannot control. Each side owns the thing it is actually good at.
See the portfolio →
The engine reasons about interaction geometry rather than a single target family, so the same design process carries across areas — and across both directions of proximity, whether the goal is to degrade a protein or to hold a complex together.
Tell us about your interaction of interest — the target class, the modality you're after, where the program is today — and we'll follow up with next steps for early access.