Antiverse

Generative design is what the company sells and what every partner is buying; Nxera licensed the output of the models, not access to a laboratory. The grade holds, but one thing about this vendor deserves recording rather than glossing. Antiverse leads its own homepage with a data advantage rather than with a model, and its stated edge is a stack: training data purpose built for this target class, machine learning generated libraries, and more than 240 proprietary stable cell lines.

That cell line estate is a durable asset that is not a model and would retain value if the models were replaced. The company is model native, and it is more honest than most about the fact that the models rest on something physical.

The structure is right and the evidence for it is missing. Antiverse describes a lab in the loop, with an integrated dry and wet laboratory, designs built and validated in house at what it calls therapeutic strength, and prioritised clones characterised in about a month because they hold their natural conformation. That is the same architecture that earned a higher grade at two neighbours in this lane, where every model output terminates in a physical measurement.

The difference is that those vendors published the numbers from inside the loop, showing how often designs fail as well as how often they land, and Antiverse publishes none. A single aggregate multiple against a naive library does not describe an oversight process. The loop is asserted rather than evidenced, and no statement exists on who selects which designs are built or what happens when the model and the assay disagree.

The components are named, which is more than a platform label, and none of them is documented. Antiverse identifies generative de novo design working from either sequence or structure, machine learning generated epitope specific libraries, and programmable cell line engineering over more than 240 stable lines, and it dates its training data to 2017 and scopes it to transmembrane targets. A reader can therefore describe the approach in outline.

Beyond that there is no architecture, no model class, no training procedure, no version, no benchmark and no publication, and the single quantified claim rests on the weakest available comparator. The gap between naming a stack and disclosing one is the whole distance between this grade and the top of the axis.

Data provenance is stated at a level most of this lane avoids. The company says its training data was purpose built for this target class rather than assembled from general protein sources, dates that effort to 2017, and quantifies the physical side of it at more than 240 proprietary stable cell lines, so a reader knows in outline where the inputs come from and that they are generated rather than licensed. That is the useful half.

The rest is absent: no compute or hosting provider, no foundation model provenance, no open source component inventory, no third party licence, and no account of whether the generative models are built from scratch or adapted from public protein models, which is the single question a technical buyer would ask first.

The thinnest evidence base of the antibody design vendors graded here, and the gap is publication. Two dedicated passes on 29 August 2026 across search and the full company site located no paper, no preprint and no conference abstract; the site carries no publications section at all.

What exists is partner activity, some of it meaningful: Nxera signed a licensing agreement rather than a research collaboration, work with GlobalBio produced candidates against PD-1 that entered preclinical development, and the Cystic Fibrosis Foundation entered a research agreement. What is claimed is not checkable.

The headline performance figure, a 2000 times greater hit rate for functional agonist antibodies against a Class A receptor, is measured against a naive library, which is the weakest baseline available and makes the multiple sound larger than the achievement; no comparison to an immunised repertoire, a display campaign or a competing computational method is offered. And the most quoted claim, therapeutic grade antibodies in under four months, comes from the managing partner of the fund that led the round rather than from the company or a partner.

No protected health information is in scope. The stewardship question that applies is about accumulation, and this vendor states the premise more plainly than most: its competitive advantage is training data built for this target class since 2017, meaning the value of the asset rises with everything that passes through it.

Five partners have run programmes against that backdrop and nothing published states what a partner's target work contributes to the shared data, how programmes are separated, or what persists after an engagement ends. Separately, and consistent with a finding recorded against another vendor in this lane, the first Nxera project is explicitly aimed at antibodies that activate a receptor rather than block it, and no statement of sequence screening, dual use policy or release control appears on any company surface. Recorded as a located absence rather than as an allegation.

Graded neutrally because the obligation does not arise on two grounds. The platform operates on receptor structures, antibody sequences and cell based assay measurements, with no protected health information in scope, and the company is a United Kingdom entity whose data handling falls under British data protection law and the Information Commissioner rather than under the United States framework this axis was written against. No business associate agreement is required and none is published.

Two dedicated passes on 29 August 2026 across a targeted domain search and a direct review of every navigation section and both legal documents found no security page, no trust centre, no certification and no attestation. The only security language located runs the other way: the terms of service state that the website is provided as is and as available and that the company does not guarantee it will be secure or free from harmful components.

That is standard drafting for a marketing site and it is the whole of what this vendor publishes about security, while it holds target information for five partners including a licensing counterparty and a disease foundation.

Nothing to clear and nothing adverse. All programmes are preclinical, with wholly owned candidates targeted only for later stage preclinical development by 2027, and as a United Kingdom company Antiverse would engage the Medicines and Healthcare products Regulatory Agency and European authorities before the FDA in any case. Two dedicated passes on 29 August 2026 located no enforcement action, designation or adverse standing. As throughout this lane, the grade records an absent obligation rather than an absent disclosure.

Nothing exists on any published surface as of 29 August 2026. The site runs to science, platform, pipeline, partnerships, about, team, careers, news and contact sections and none of them addresses model governance, evaluation practice, failure modes or responsible use.

There is no model card and no statement of the conditions under which the approach degrades, which matters here because the company's stated advantage is data specialised to one target class and the obvious failure mode of a specialised model is everything outside its specialisation. Neighbours in this lane that publish nothing on governance at least publish method papers that expose their evaluation to scrutiny. This vendor publishes neither, so nothing stands in the gap, which is the same basis on which another vendor in this lane holds this grade.

The strongest legal surface graded in this lane, which says as much about the lane as about the vendor. Antiverse publishes an actual terms of service, updated 18 August 2026, running to thirteen clauses with a limitation of liability, an intellectual property allocation, warranty disclaimers, third party linking terms, severability, governing law and exclusive jurisdiction in England and Wales, and a registered company number and office.

Every other vendor graded in this cluster offers either a privacy notice or nothing. Three qualifications keep it at this grade rather than higher. The document governs use of the website and not the design work, so it reaches none of the questions a pharmaceutical partner would have. It expressly disclaims that anything stated about the technology or pipeline is a representation, warranty or guarantee of outcome, which neutralises the performance claims elsewhere on the site.

And it is visibly unfinished: an unfilled template placeholder reading link to Privacy Policy sits in the privacy clause, and two section headings are misspelled, in a document dated eleven days before this record was written.

No electronic health record surface exists and none would be appropriate for a preclinical design company, so this is graded neutrally under the convention for this lane. Recorded in its place: the Nxera agreement is a genuine two way scientific integration rather than a service arrangement, pairing Antiverse's machine learning generated epitope specific libraries with Nxera's platform for receptor selection, validation and structural determination, so each party's technology depends on the other's output. No software product, application programming interface or connector exists, because nothing here is sold as software.

Nothing is published about where partner target information is held, under what tenancy, with what retention or with which subprocessors, and no software is deployed to anyone, so there is no customer environment to describe. Graded neutrally under the convention for this lane, now the sixth consecutive record in which that convention rather than the evidence is setting the grade.

Antiverse does carry more jurisdictional anchoring than its neighbours: a registered company number in England and Wales, a stated registered office, and terms fixing governing law and exclusive jurisdiction in the courts of England and Wales. That establishes where a dispute would be heard. It says nothing about where the data sits.

Partial and below the lane. Three partners are named with dates, and the Nxera structure is described in kind, an upfront payment plus research funding plus milestone eligibility with the licence sitting on Nxera's side, which tells a reader how the economics work. But no figure attaches to any agreement, so there is no upfront, no ceiling and no milestone value anywhere in the public record, and two of the five partners are identified only as a top 15 and a top 30 pharmaceutical company.

Funding is better disclosed, with a $9.3 million Series A in March 2026 named to its lead and syndicate against a stated total above $20 million. One conflict remains open: a trade report states an initial £50 million at launch from government funds and the regional authority, which cannot be reconciled with that total and is recorded in the vendor description as unresolved.

Narrow by choice and deep in one place. The target class is transmembrane proteins, principally G protein coupled receptors and ion channels, and the company does not claim to work outside it. Formats are broad within biologics, covering any construct with a heavy chain including full antibodies and single domain formats.

Indication reach comes through partners rather than through a wide internal pipeline: oncology through the PD-1 work, cystic fibrosis through the CFTR programme, and undisclosed targets under the pharmaceutical agreements. The company states an intention to bring its first wholly owned candidates into later stage preclinical development by 2027, so everything here sits well short of the clinic.