A complete discovery pipeline: generative design from a named target, structure prediction on every candidate, a fail-closed governance gate, a selection engine, and a decision record that survives an audit. Generation substrate is a commodity we deliberately license in rather than build.
Generation has become a commodity and we treat it as one. What we build, protect and improve is generative design tuned to your target, structure-informed ranking, and the record that proves each decision was made safely.
Open protein language, structure and design models that produce candidate sequences.
Fine-tuned models that take your named target and produce a candidate library — trained on curated data for the target class in question. The layer that turns a target into candidates.
Ranking and build-list construction across the governance-cleared pool, informed by Boltz-2 structure prediction and interface-confidence scoring on every candidate. Decides what is worth building.
Biosafety gate, freedom-to-operate attribution and a tamper-evident decision record.
Interpretation and delivery of results in a form your scientists can act on.
The internal methods of the selection engine are held as trade secret and are not described here or in diligence materials. What we publish is what it does, not how it does it.
A candidate is an amino acid sequence — a protein or a peptide. You order it as DNA, express it, and assay it. Every step after design costs real money, so the question we answer is which sequences are worth that spend.
Designs from open models, from a generator you already use, or from ours. We are generator-agnostic by design, so nothing about working with us requires you to change how you make candidates.
Nucleic acid enters on the screening side, because DNA is what actually gets synthesised.
Named as roadmap, not as capability. We are precise about the difference.
Our selection and governance layers do not know what a molecule is. They operate on candidate features, constraint signals and decisions — not on molecule class. That machinery carries to new modalities without redesign. What changes each time is the generation and feature layer beneath it, and we say plainly which ones are wired.
Structure prediction isn't a side demo — it runs on every candidate in the pipeline, between generative design and selection. These schematic demonstrations walk through the sequence of events a real campaign passes through, compressed from a timescale no instrument can watch directly. Pick a demonstration and press run.
A designed amino acid sequence collapses into a three dimensional shape. Contacts form between residues far apart in the chain but close in space.
These are schematic visualizations of the events a campaign passes through, rendered for clarity at a speed the underlying physics does not occur at. They are not real-time instrument output, and not structure predictions of specific candidates — for real Boltz-2 structure predictions, see Structures.
Most discovery platforms treat sequence screening as internal compliance. We built it as the product — it runs before any candidate is eligible for selection, and every batch ships with a sealed, auditable record.
Illustrative reconstruction using synthetic identifiers. No customer sequence or result is shown.
United States policy is moving steadily toward requiring that AI-designed biological sequences be screened before synthesis, and that the screening be documented and auditable.
Each candidate carries an attribution of what it derives from, and that attribution gates whether the candidate can advance at all. You find out that a candidate is encumbered before you build it, rather than during a licensing negotiation two years later.
Decisions are sealed into a tamper-evident chain. A third party, a regulator, a partner or your compliance group can verify independently that the record has not been altered since it was written. Verification does not require access to our systems.
A candidate is eligible only when every required signal returns an affirmative clearance. A missing signal, a timeout, an error or an incomplete record all produce a denial. There is no path through the gate that reaches approval by accident, and that property has been verified across each failure mode independently.
We use the Common Mechanism, the open-source synthesis screening standard, running on infrastructure we operate with the full reference database set. We deliberately did not build a private screening method. Our contribution is the enforcement and the evidence around it, which is the part that has been missing.
We work in discrete campaigns against a named target. Every engagement is structured so that the result is established by measurement rather than assertion, and so that your team can audit how each decision was reached.
We scope the target, functional bar and control arm with your scientists — agreed before any work starts.
Candidates come from open models, models tuned for your target class, or a generator you already use. We never lock you to ours.
Every candidate passes the fail-closed biosafety gate and carries freedom-to-operate attribution before it is eligible for selection. Nothing reaches your build list without a sealed decision record.
The selection engine ranks the cleared pool into one ordered, auditable build list sized to your lab capacity — regardless of which generator a candidate came from.
Results come back against your control arm, reported whether they favor us or not, and inform the next build list.
We scope the target, the functional bar and the control arm with your scientists. The comparison is agreed before any work starts, so the outcome is unambiguous to both sides.
Candidates are generated from open foundation models, from models tuned for the target class, or from a generator you prefer to use. We are indifferent to the source and never lock you to ours.
Every candidate passes the fail-closed biosafety gate and carries freedom-to-operate attribution before it is eligible for selection. Nothing reaches your build list without a sealed decision record.
The selection engine ranks the governance-cleared pool and produces one ordered, auditable build list sized to your lab capacity — regardless of which generator a candidate came from.
Results come back against the control arm you set at intake. That outcome informs the next build list, and the comparison is reported to you whether it favors us or not.
A ranked build list sized to your lab capacity. A sealed, third-party-verifiable screening record for every candidate. Freedom-to-operate attribution on every sequence. And a measured comparison against your baseline — reported whether it favors us or not.
| Rank | Candidate | Source | Biosafety | FTO | Provenance hash |
|---|---|---|---|---|---|
| 1 | SAMPLE-0417 | RFdiffusion | Clear | Clear | cc9100a525a833c6a39ea0d7b8ebfb4d4d5e0f496f2e726041f0999bdd8c720c |
| 2 | SAMPLE-0298 | LillixBio design | Clear | Clear | f2c2bf821f4b1ad813b63703712ea6959882f479a61f978dc57b7b7a7e9a8485 |
| 3 | SAMPLE-0165 | ProteinMPNN | Clear | Clear | 502322879714c4dcbc6083043977d32b0e803ce6b8e03f360c9015aadaee5459 |
| 4 | SAMPLE-0512 | Customer pipeline | Clear | Clear | d687dc831f05daa7b276042f6f1c1e9e0e419e494c6384553ef52c3a56765898 |
| 5 | SAMPLE-0033 | RFdiffusion | Clear | Clear | 5fe1c037560a2dcff6c6e15a664d87045e440d78d11b5312d81f878c08a9dbec |
Recomputed live from the values shown above, using SHA‑256 in your browser — nothing is sent to a server. Candidate, source, biosafety, and FTO cells are editable: try changing one and watch the status change. Full scheme below.
What this proves, and what it doesn't: a match means this record hasn't been altered since it was sealed. It does not prove who produced the list, and it can't stop someone who holds the full record from generating a different chain from scratch.
Illustrative example built to show the structure of the artifact. Candidate identifiers, hashes, source labels, and the sealer name are fabricated for demonstration and do not represent any real customer engagement, target, or molecule.
Enough to reproduce every hash above independently, using any standard SHA‑256 implementation. No LillixBio code or systems required.
V, take the UTF‑8 byte length of V as a decimal number, a colon, then the UTF‑8 bytes of V — e.g. Clear becomes 5:Clear. This length prefix is what makes it unambiguous: there is no separator character a field could contain that would let a byte shift from one field into the next and still produce the same hash input.target, generated, evaluated, cleared, and excluded header fields, in that order, and take SHA‑256. Render as lowercase hex.rank, candidate, source, biosafety, and fto fields, and take SHA‑256. Row 1's previous hash is the genesis hash; every later row's previous hash is the row hash directly above it — this is what chains the rows together, so editing any row changes every row hash from that point on.Field values are read exactly as rendered — exact casing and punctuation, leading/trailing whitespace trimmed only. The control above recomputes and checks all of this live, in your browser, from the values shown on this page.
Methodology, the fragmentation problem, and the third-party comparator data live on Evidence.