This page carries the dense material: the fragmentation problem in full, campaign economics, why structure is consequential, the structural difference from integrated discovery companies, and the honest state of the roadmap.
Generative design got cheap enough that the bottleneck moved. What changed is who has to do the choosing — and right now, everyone is doing it separately, from scratch.
A postdoc wires together a generator, a filter script and a spreadsheet. When it fails — a bad candidate ships, a target turns out encumbered — the failure dies in that one notebook. The lab down the hall repeats the same mistake, because nothing about the first one was recorded anywhere they could find it.
Selection runs on whoever is most senior in the room that week. It works, until it doesn't scale past that person — and there is rarely a written record of why one candidate was chosen over another, which is exactly the record a board, an acquirer or a regulator eventually asks for.
Larger organizations have the resources to build internal tooling, and do — program by program. The result is incomparable across teams and invisible above the team level: no one two floors up can see how one program's selection discipline compares to another's.
Strip away the scale and the budget, and an AMC core facility, a Series A biotech and a pharma discovery group are running the same broken pattern: generation outpaced the tooling to choose and govern what gets built. That's why we built one engine instead of three separate products — DECIDE, CLEAR and OPERATE are three doors into it.
Generation got good.
That made the choosing harder.
Not yet measured. On a named target, against a control arm agreed with the customer in advance — published here once the run is complete.
We're building this evaluation discipline as a standard the whole field can measure generators against, not just a number for this page.
Computational hit rates routinely overestimate what the bench confirms — that gap is the whole game. Enter your campaign numbers below; the rate is yours to set, not a figure we're asserting.
A generated candidate is a string of amino acids. Two candidates differing by a few residues can behave completely differently once made — which is why a defensible build list can never come from sequence identity alone.
Two candidates differing by a handful of residues can behave completely differently once expressed in the lab. That variability is why generating more candidates does not, by itself, improve a campaign.
Governance runs first. A candidate that does not clear biosafety and freedom-to-operate screening is not ranked, not scored and not shown, regardless of how attractive it looks.
Candidates from any generator enter the same governed ranking and produce a single ordered, auditable build list sized to your lab capacity.
AI-driven biologic discovery works — the funding and pharma deals already prove that. What's less visible is that most well-funded platforms are drug companies running an internal pipeline. Bring them a target, and you're bringing it to a future competitor.
Almost every well-funded company in this field also develops medicines internally. That is not a criticism of their science, which is often excellent — it is a description of their business model. When you bring them a target, you are bringing it to an organization with internal programs, internal therapeutic priorities, and reasons to be in your space that have nothing to do with yours.
LillixBio never puts a therapeutic asset on its balance sheet — discoveries are licensed or spun out, never carried by us. That's structural, not a promise, and it means we're never quietly building against you.
LillixBio is built to be modality-general. Phage and antigen work is a large, underserved candidate space — real problems that give the clearest demonstration of the selection and governance layers before carrying them to other modalities.
These are consequential problems with real demand that has been comparatively underserved by platforms concentrated on antibodies and small molecules, which makes them the clearest place to demonstrate the selection and governance layers before carrying them outward to other modalities.
Vaccine antigens — the same selection and governance layers, applied to immunogen candidate sets.
Engineered proteins, enzymes and binders — function-directed design where a control arm and a clear functional bar already exist.
Antibodies and nanobodies — a crowded field for generation, and one where governed selection is still uncommon.
Neoantigen and oncology applications — the long-horizon extension of the antigen engine, and the reason that engine stays in-house.
Antigen-receptor seating is a structural question, not just a sequence one. Schematic only.
A few residues at the binding interface can change affinity and specificity completely. Schematic only.
Function-directed design, where a clear bar and a control arm already exist. Schematic only.
Schematics of the biology the platform is pointed at. They illustrate the classes of problem the selection and governance layers address, and are not structure predictions of specific molecules or live platform output.