Galux

GaluxDesign is the product. The model proposes the protein or antibody, and the company frames the whole enterprise around designing proteins from first principles rather than screening for them. De novo antibody design is put forward as the demonstration that the model works. There is no separate asset that would remain if the model were taken away.

The engine takes a target and proposes a protein or antibody. What a partner cannot see is the control around it: where a scientist reviews or overrides a design before it advances, what the company does when a prediction fails in the lab, and which steps run without a person in the loop. The pipeline makes clear that designs are tested before they go forward, but that is inferred from practice rather than set out as a mechanism a collaborator could rely on.

The model is named and its idea is described, but the technology is not identified. GaluxDesign is said to combine AI with physical principles to design proteins from first principles, which conveys the approach without an architecture, a parameter count, or a characterization of the training data. A technical preprint is linked, so a reader can go deeper, but no version or update practice is published and the production system behind a given design is not pinned to a date.

Built in house, the platform names no outside model provider between a target and a design. Nothing further is established: no compute or hosting provider is named, no third party components or licenses are listed, there is no subprocessor list, and nothing describes what governs a partner's target data once it enters the system.

There is published research and there are named partners, with no independent validation yet. Galux posted a technical preprint on its de novo antibody design and points to pointed results from it: an antibody reported to tell apart a form of EGFR that differs from the normal protein by a single amino acid, and de novo antibodies against six targets produced in one design pass.

It works with two named pharmaceutical companies, AstraZeneca from June 2026 and Dong-A ST from October 2026 on antibodies for antibody drug conjugates. What is missing is a peer reviewed result and a clinical candidate: the preprint has not been through review, its full figures are not restated on the company's public pages, and Galux names no program of its own in a trial.

The inputs here are protein targets and sequences rather than patient records, so no protected health information is at stake. The question that matters to a pharmaceutical partner is a different one: once its target data enters the platform, is that data retained, used to train the model, or walled off from other work. Galux does not say. Its privacy policy describes the website, not the design service, so the handling of collaboration data is left undescribed.

Galux designs proteins from target and sequence data, not patient records, so the United States health privacy rule has nothing to attach to in what it processes. The company states no business associate position, and its only privacy document covers the website rather than the design service.

A pharmaceutical partner's security team, before putting target data on an outside platform, looks for an independent attestation it can open: a SOC 2 report, an ISO 27001 certificate, a HITRUST certification, or at least a trust portal or security page. Galux, which does hold that partner target data, publishes none of these, and no outside review of how its environment is secured is available to read.

No device claim is made, and the product is scoped to match. Galux designs candidate molecules at the discovery stage, so any regulatory position belongs to the molecules and their developers rather than to GaluxDesign. There is no clearance to hold and none is claimed.

A buyer weighing a design model wants to know how its behavior is checked: how reliably it performs, how failure modes are tracked, whether a responsible AI policy sets any rules. Galux addresses none of this on its pages. A design accuracy preprint on bioRxiv speaks to performance and may bear on the question, but its figures stay in the preprint and are not carried onto the company's public pages, so a reader reaches them only by going to the paper.

When a design does not hold up in the lab, a partner wants two things: a sense of how often that happens and a route to recourse. Galux states its accuracy in words rather than as a rate with a method and a denominator, so there is no error figure to plan against, and no guarantee, indemnity or correction route travels with a design. Where that risk lands is settled inside private collaboration agreements rather than on the public pages.

What the platform produces is designs, not software a buyer wires into a stack, so there is no clinical record surface to speak of, which suits a discovery platform. Even on the research side it stands alone: no electronic lab notebook, laboratory information management system or research data platform is named as connected, and no connector is offered to a partner who might want the output to flow into their own systems.

The platform runs for Galux and its partners rather than being deployed to anyone, so there is no tenancy or hosting model to evaluate. Location is implied by the company's base in Seoul rather than committed, and nothing published names where partner target and design data rests or how one collaboration's data is kept apart from another's.

Galux earns through partnerships rather than selling access, so there is no published price and none would be expected. The one public financial figure is a $29 million Series B, which is investment raised rather than a number a buyer could pay, so how the company funds itself is visible while the cost of a program is not.

Coverage is named at the level of what the platform designs, with specific examples behind part of it. Galux puts de novo antibody design forward as its demonstrated capability, and the examples are concrete: an antibody reported to bind a single amino acid variant of EGFR without binding the normal protein, a case that turns on fine specificity, and a claim of de novo antibodies against six targets in a single design pass.

A named collaboration with Dong-A ST aims the platform at antibodies for antibody drug conjugates, with broader protein design alongside. What is not established is therapeutic reach: Galux names no disease area, indication or clinical program of its own.