BACKED BY 1517 FUND · RAISED $100KLearn more →
EXPRESSIBILITY & VERIFICATION FOR ENZYME ENGINEERING

Confidence before
you commit to the bench.

Most de novo enzyme designs fail long before the plate: they don't express, they misfold, or they quietly break in the lab. Lumigenic wraps generative models, MD, and stability prediction in a verification layer that tells you which designs are worth making, and when to trust the simulation.

Request a pilot →See a sample report
Built for medical centers and experimental labs, with no in-house computational team required.
DESIGN · ENZ-4471REVIEW · 2 FLAGS
ExpressibilityPASS · 0.86
MD convergenceCAUTION
Protonation stateFAIL
Expression outlook is strong, but binding results are not yet trustworthy. Fix protonation and extend sampling first.
THE GAP BETWEEN IN SILICO AND THE PLATE

A design that looks good on screen isn't a design that works.

Generation is no longer the bottleneck. Trusting the output is. Here's where designs quietly fail, and exactly what we check before you spend wet-lab time on them.

Why designs fail

Won't express
The sequence scores beautifully in silico and yields nothing in the cell.
Misfolds & aggregates
A plausible predicted fold turns into insoluble protein in the tube.
Silent breakage
A mutation quietly kills activity, and nothing in the pipeline flagged it.
Simulations you can't trust
MD that never converged and mis-protonated residues read as real signal.
So teams fall back to directed evolution and brute-force screens.

What Lumigenic checks

Expressibility & developability
Predicts which designs are likely to express and be developable, not just which ones look good.
Fold & aggregation risk
Flags aggregation-prone regions and low-confidence structure before synthesis.
Function-preserving verification
Checks that the changes you made didn't quietly compromise the active site.
Simulation trust
Automatically detects non-convergence, mis-protonation, and common failure modes.
So you prioritize the designs most likely to work, and know what to trust.
STRUCTURED CONFIDENCE

One report. A verdict you can act on.

Every design comes back with a traffic-light verdict across expressibility, structure, function, and simulation integrity, plus a plain-language recommendation. No RMSD plots to interpret, no protonation logs to parse.

Pass: trust it, move forward.
Caution: a fixable issue, and here's what to do.
Fail: don't trust this result yet.
ENZ-4471 · lipase variant
Verification report · generated in 4m 12s
◆ REVIEW · 2 FLAGS
Expressibility
High predicted expression in E. coli, with no rare-codon or solubility penalties.
PASS · 0.86
Developability
No aggregation-prone regions detected; surface hydrophobicity within range.
PASS
MD convergence
RMSD not converged in 2 of 3 replicas. Extend sampling before drawing conclusions.
CAUTION
Protonation state
His41 mis-protonated at pH 7.4. Docking and binding scores are unreliable as run.
FAIL
Stability (ΔΔG)
Predicted marginally more stable than wild type.
PASS · +1.2
RECOMMENDATION

Re-run MD with corrected protonation before trusting binding results. Expression and stability outlook are strong, so it's safe to queue for synthesis in parallel.

A LAYER ON TOP OF THE TOOLS YOU TRUST

Generate, simulate, verify.

We don't replace your generative models or your simulations. We sit on top of them and tell you which outputs to believe.

01

Generate

Bring your own designs, or generate candidates with our models. Sequences, structures, and variants all welcome.

02

Simulate

Structure prediction, MD, docking, and stability run automatically, or point us at simulations you've already run.

LUMIGENIC
03

Verify

We score expressibility, audit every simulation for failure modes, and return one traffic-light report with a clear recommendation.

THE ENGINE UNDERNEATH · YAMI

A physics-aware model, doing an honest day's work.

Yami is our protein foundation model. It reasons over sequence, structure, and physical plausibility together, so its confidence estimates and failure-mode detection are grounded in chemistry, not just pattern-matching.

We'd rather tell you a design is uncertain than dress up a guess. That skepticism is the point.

Physics-awareChemical identity and physical regime built into the representation.
Structure-groundedSequence and structure reasoned about jointly, not as separate steps.
Calibrated confidenceScores tuned against real expression and assay outcomes.
EfficientRuns at a cost that makes routine verification practical.
FOR TEAMS WITHOUT A COMPUTATIONAL GROUP

Built for medical centers and experimental labs.

You shouldn't need a molecular-dynamics expert on staff to know whether a design is worth making. Lumigenic does the interpretation for you.

Traffic-light outputs

Pass, caution, or fail, each with a reason. No plots to read, no logs to parse.

↳

Nothing to set up

No pipelines to configure or tools to stitch together. Submit designs, get reports.

✎

Reports you can hand off

Plain-language summaries a PI, collaborator, or CRO can read without translation.

QUESTIONS

Fair questions, straight answers.

How is this different from tools like Tamarind or Schrödinger?
Those are excellent at generating and simulating, and we respect them and work alongside them. Lumigenic is a layer on top: we take generated designs and simulation outputs and tell you which ones to trust. We predict expressibility, and we automatically catch failure modes like non-convergence and mis-protonation that otherwise slip through. The difference is trustworthy, non-expert-readable outputs, not another modeling engine.
Do I need a computational biologist to use it?
No. That's the point. Outputs are traffic-light verdicts with plain-language recommendations, designed to be read by experimental teams and PIs without a dedicated computational group.
What exactly do you check in a simulation?
We detect non-convergence in MD, mis-protonation of catalytic and titratable residues, unstable trajectories, and other common artifacts that make downstream docking and binding results unreliable, then we surface them in the report instead of letting them pass silently.
Can I bring my own designs and simulations?
Yes. Submit your own sequences, structures, and existing simulation outputs, or generate candidates with our models. Either way you get the same verification report.
Do you run wet-lab validation?
We focus on computational design and verification, and partner with labs and CROs for wet-lab work. Our reports are built to hand directly to those partners.
Who's building Lumigenic?
A small team backed by 1517 Fund, with experience at Berkeley and across AI, biotech, and venture. We competed at the iGEM Startup Showcase in Paris, and we build with the skepticism that comes from having watched designs fail in the lab. Meet the team →

Run a pilot with us.

We're working with a small number of enzyme-engineering teams to verify their designs. Tell us what you're building and we'll show you what the reports look like on your own candidates.

✉founders@lumigenic.ai