Earendil Labs

The company calls itself AI native, and the description holds up. A foundational model trained on protein sequence, structure and interaction data is what proposes each antibody, predicts whether it will bind and where, and estimates its developability before any lab work begins. The high throughput biology exists to test and refine what the model proposes. Take the model away and what is left is a conventional antibody lab, which is the opposite of how the company has built and funded itself.

The design loop is described in outline: the model proposes antibodies, the lab screens them for binding and function, and the results feed back to refine the next round. That is an oversight structure of a kind, a prediction checked against an experiment. What the company does not spell out is any mechanism inside it: when the model's output is trusted or overridden, what happens when an experiment contradicts a prediction, or which decisions a person makes rather than the system. The loop is a diagram rather than a rule.

Earendil describes the model by what it does: a foundational model trained on sequence, structure and interaction data that predicts whether an antibody binds a target and where, proposes sequence changes, and estimates developability properties such as stability, immunogenicity and half life before the bench. The functions are clear; the technology under them is not.

There is no architecture, model class, parameter count or training procedure, the training data is named only as those three data types, and no version practice ties the system behind a given program to a date.

The model is built and run in house, with no outside provider between a target and a design, so the chain is short. What a partner's review would still want sits around it: the compute or hosting provider, any third party components or licenses, a subprocessor list, and the terms that govern a partner's target data once it enters the platform. With the model itself unnamed, none of that is on the record.

The evidence is commercial rather than clinical, and it is substantial. Sanofi licensed two named bispecific programs, HXN-1002 and HXN-1003, for $125 million upfront in April 2025 and signed a broader discovery collaboration in January 2026, a partner committing real money to specific molecules the platform produced. A February 2026 agreement with WuXi XDC brings in payload and linker technology for antibody drug conjugates, extending the work beyond bare antibodies. The company reports more than 40 therapeutic programs and names its lead, HXN-1001, as ready for Phase 2.

What is missing is the clinical and the independent. No program is confirmed to have entered a trial, so every milestone is preclinical, and there is no peer reviewed paper describing how the platform works or how well. The validation a reader can check is the deals, not the data behind them.

The question a pharmaceutical partner would press is how its target and program data is handled: whether it is retained, used in training, or kept apart from other partners' work. Earendil answers none of it. No protected health information is at stake, since the inputs are protein targets and sequences rather than patient data, and no data handling statement reaches the platform.

No business associate status appears anywhere, and no privacy document on the site reaches the platform. The inputs are protein targets and program data rather than patient records, so United States health privacy rules do not touch what Earendil handles, and the company states nothing either way about them.

Earendil holds target and program data from pharmaceutical partners under major licensing agreements, and the first thing a partner's security review asks for is an independent attestation covering where that data sits. There is none: no SOC 2 report, ISO 27001 certificate or HITRUST certification, no trust portal and no security page.

The platform makes no device claim and is scoped to match. Earendil designs drug candidates, so the regulatory position that matters belongs to the molecules, not to the AI that designed them. The programs are preclinical, with the lead described as ready for Phase 2 rather than cleared or in a trial, and any investigational new drug filing that follows sits with Earendil or its partner as the drug's sponsor. There is no medical device here to clear.

The platform is said to estimate properties such as immunogenicity and half life, which is an account of what the model does, not of how reliably it does it. No benchmark, failure analysis or audit of model behavior appears, and no responsible AI policy sits alongside it.

A buyer cannot tell from the public record how often an Earendil design fails. The company reports its deals rather than its hit rates, so there is no success or failure figure with a method and a denominator, and nothing behind a design that does not work: no guarantee, indemnity or correction route. Risk sits inside private licensing agreements.

The platform is an internal engine that produces drug programs, not software a partner connects to, so there is nothing to integrate: no clinical record surface, and no electronic lab notebook, laboratory information management system, research data platform or connector on the research side either.

The platform runs for Earendil and its affiliate rather than being deployed to anyone, so there is no tenancy model or hosting option to evaluate. Location is implied rather than committed: the company is based in Delaware with research operations in China, and the model and the lab are described as one integrated engine it owns. Nothing published names where partner program data rests or how one collaboration's data is separated from another's on that infrastructure.

Earendil earns by licensing programs rather than selling access, so there is no price to quote. What is public is the scale of its deals: $125 million upfront from Sanofi in April 2025 against up to $1.8 billion, a January 2026 discovery collaboration worth up to $2.56 billion, and $787 million raised. Those are contract ceilings and investment figures, so a buyer can see how the company makes money without being able to size the cost of a program.

Coverage is named at the level of disease area and target, with programs behind a good part of it. The immune and inflammatory work spans inflammatory bowel disease, ulcerative colitis, Crohn's disease, asthma, COPD and skin inflammation, and the oncology work covers colorectal cancer, small cell lung cancer and other solid tumors. Named targets run through it: TL1A, α4β7 and IL23, carried by the HXN-1001, HXN-1002 and HXN-1003 programs. What is not established is clinical reach, since no program is confirmed to have treated a patient in a trial yet.