Xenonex

Protein structure · Binders · Developability · Variants

Predict the protein.
Sign the answer.

What we add is the part nobody ships: the evidence behind every answer, and a record somebody is willing to put their name on. The models underneath are the ones everyone runs. We did not set out to replace them.

Built on our own Super AI Engine  ·  Japan  ·  Hong Kong

  1. 00The subject

    One antigen, one binder, and a decision somebody has to sign.

    Every coordinate here is a published experimental structure from the Protein Data Bank: a cell-surface antigen with a Fab fragment bound to it, determined by X-ray crystallography at 3.00 Å resolution. Even the reference has gaps: 44 of its 716 residues were not modelled in the deposited coordinates. Someone else’s work, shown to demonstrate a method. Nothing on this screen is a Xenonex result.

  2. 01The models

    AlphaFold 3, ESMFold, Boltz-2, Chai-1. Remarkable, and now everyone has them.

    These are extraordinary pieces of work, and we run them rather than compete with them. Three of the four publish their weights under permissive licences; the best-known does not. A structure now comes out in an afternoon. That part of the problem is finished.

  3. 02What the metrics report

    The same structure, coloured by what it actually reports.

    The per-residue confidence score (pLDDT) describes local geometry, and describes it well. Interface metrics exist too: ipTM, and a predicted aligned error for every pair of residues. The field is actively refining them. What none of them sets out to report is whether the answer is right; they report how sure the model is. In the 2026 binder competition, a filter applied after the fact nearly doubled the hit rate those scores had produced. Reading confidence as biology is our mistake, not the model’s.

  4. 03Where the work is

    Pull back, and the answer is a speck in the search.

    The sequence databases searched for a single target run past 2 TB, and thousands of candidate structures are generated: 2,000 to 72,000 per target for a top-five entry in CASP16, the 2024 open assessment. Nearly all the computing, and all of the doubt, lives out here rather than inside the model.

  5. 04What we add

    Six ways to ask. One copy of the data.

    Connections, likeness, meaning, cause, exact wording, and reach into data we don’t hold. Six answers over one copy, ranked on the evidence behind each one. In that same assessment, teams picked five structures out of 8,040 and the best one they chose ranked 147th. The models had already produced the right answer. Nobody could find it.

  6. 05Where everyone is stuck

    CASP17 asks for exactly this: an immune complex.

    Its call for targets names immune complexes as a major failure area, alongside ligand complexes with unfamiliar chemistry, and nucleic acids. All three are cases with almost no comparable structures to align against. The problem is finding what is already known, not building a bigger model. It is where one shared store helps these models instead of replacing them.

  7. 06The signed answer

    Every residue, back to the evidence that justifies it.

    Which alignment, which template, which assay, which standard operating procedure. Each is pinned to the version it was resolved against. The models hand you a structure. This is what turns it into something a person is willing to sign.

00 / 46 s

This argument is animated in 3D where the browser allows it. Here it is written out instead, in the same seven steps.

01What we do

Four jobs, and the same bottleneck in each.

Generating candidates stopped being hard. Choosing between them, and defending the choice afterwards, did not.

  1. 01

    Structure and complexes

    Single chains, assemblies, and above all the interface where a binder meets its target. CASP17’s call for targets names that interface as a major failure area for current methods, and it is the case with the commercial weight behind it. We run the co-folding models on it and set the result beside every deposited complex with a comparable interface.

  2. 02

    Binder triage

    Ranking thousands of designed binders down to the dozen worth making. In the 2026 Bits to Binders competition, 12,000 designs against one target were screened and 707 came through, under 6%. Team hit rates ran from 0.6% to 38.4% on the same public models, so how you order candidates matters more than which model made them. We order on the evidence behind each one, not on the score alone.

  3. 03

    Developability

    Whether a promising binder can actually be made: aggregation propensity, immunogenicity, expression level, storage stability. Each is cheaper to predict than to discover in a batch record, and for each we can show the derivation.

  4. 04

    Variants

    What a single amino-acid substitution does to conformation, binding affinity and thermostability. This includes substitutions that exist only in a customer’s own records, answered in place, with no data transfer.

02Finding candidates

Find the one worth making. Sooner.

In CASP16, the 2024 open assessment, 8,040 candidate structures were produced for a single target and each team could submit five. The best structure anybody submitted ranked 147th. The right answer was already in the pile; finding it is what failed. Most systems search the way that produces this: several separate stores, each holding its own copy, queried one after another. Copies drift, every seam is a place an answer can be confidently wrong, and every question pays for the round trip. Our engine holds one copy and answers six ways over it at once.

Six ways to ask the same store, in one pass, with nothing duplicated between them.

  1. 01

    Connections

    Follow one thread as far as the question goes: this target → its protein family → its structural homologues → every deposited structure → the primary literature.

  2. 02

    Likeness

    Find the functional homologues, including remote homologues that keyword matching misses, and constrain by explicit annotation in the same query, because similarity lives where the annotations live.

  3. 03

    Meaning

    Declare once that a single-chain variable fragment is a binding domain, then ask for binding domains and get every one of them.

  4. 04

    Cause

    Four kinds of link hold the structure together: contains, leads to, is near, is a property of. The biology is named on top of those four, so the same relationship means the same thing in every dataset it came from.

  5. 05

    Exact wording

    Standard operating procedures, deviation reports and batch records, searchable verbatim. When an inspector asks where a number came from, “something similar” is not an answer.

  6. 06

    Data we don’t hold

    Query a partner’s sequencing archive or a hospital’s records while the question is still running, with no data transfer and no copy retained. Most of it cannot leave the site where it was generated.

Evidence this works · another team’s research

In a 2026 study in the Journal of Biomedical Informatics, grounding a general-purpose AI assistant in a controlled medical vocabulary cut its wrong answers on clinical questions from roughly half to under 2%.

03What comes back

A structure is not an answer. An answer has a record.

Every model returns coordinates and a score. The question a quality assurance lead actually asks is the one nothing returns: how do you know?

What a model returns

  • Coordinates.
  • One confidence value per residue.
  • A file, with no way to ask where any of it came from.

What comes back here

  • All of that, plus where each part of it came from: which alignment, which template, which assay, which standard operating procedure.
  • A verdict on what the answer is fit for, kept separate from the score.
  • A refusal, when the evidence doesn’t support an answer. An instrument that always answers is not measuring anything.
  • A run that can be rebuilt years later against the exact versions it used.

04Where it matters

It matters most where somebody has to sign.

A prediction used in a laboratory can be wrong and cost a week. The same prediction used where a product is released can be wrong and cost the batch. In a 2025 EBMT survey of European centres, 18.9% of CAR-T manufacturing attempts ended in an out-of-specification product or an outright failure. The most common causes were low cell dose and low viability, both driven by the patient’s own apheresis material. That product is then released for infusion case by case, on a benefit–risk assessment agreed with the regulator. Justifying that decision is not a modelling problem. It is an evidence problem.

On 1 May 2026 a State Council regulation took effect in China. It is binding law, not guidance, and it opens a second route: cell, gene and stem-cell treatments can now be used in around 1,700 tertiary Class A hospitals without going through drug registration. It was not written to loosen anything. It narrows who qualifies and raises the penalties.

A route under closer watch, at scale, needs auditable evidence for every batch and every patient. The US regulator’s draft framework of January 2025 for AI in regulatory decisions asks for the same thing, in seven steps. For a platform built to explain itself that is not a headwind. It is the reason one is needed. It is also why the regional company sits in Hong Kong.

05The team

Built by people who have done this before.

Policy, cell therapy, clinical practice and capital: the four things a Bio AI platform for regulated medicine has to survive contact with.

  1. Founder & CEO

    Allen Chan

    Leads the company, and holds the architecture the platform is built on.

  2. Fund Manager · Strategy Advisor

    Vincent Lee

    A fund manager by background, advising on strategy and capital across Greater China and Asia.

  3. CTO

    Jackson Cheung

    Builds the AI solutions the platform runs on, and the data science underneath them.

  4. Advisor · Policy

    Motohiro Asonuma

    Advised the Japanese government on medical system reform for seventeen years, and sat on the ministry committees that drafted Japan’s regenerative-medicine safety legislation.

  5. Advisor · Cell therapy

    Hiroshi Terunuma

    Developed immune-cell and stem-cell therapies and a stem-cell culture supernatant, with more than fifty published papers and years of clinical experience using them.

  6. Advisor · Clinical

    Takashi Kamigaki

    A gastrointestinal cancer surgeon, now a professor of cell and immunotherapy, who has run clinical research on immune-cell and stem-cell treatment.

The engine is built.
The biology is the work.

In build, in Japan and Hong Kong, with the first pilots ahead of us.