Lumina Bio — Contract computational chemistry for hard targets

Why we exist

When a target fights back, we re-route.

Drug discovery is slow and expensive, and it stalls most often on the targets that matter most — the hard ones, the ones called undruggable. Many teams have a good target but no clear path to a molecule.

We route around that bottleneck with computational design and multi-layer validation. When one approach is blocked, we find another path — inhibition gives way to degradation, one chemistry to another — where conventional discovery stops.

We do not screen known libraries by brute force. Molecules are generated in the binding site and then cut down hard — the breadth is in the generation, the discipline is in what survives. Precision is the strategy.

The platform

One pipeline, from a target to a validated candidate.

Our proprietary platform runs a single, sequential pipeline: generate novel molecules, filter them through independent computational gates, and accumulate what we learn. Each stage is built to remove the false positives the previous stage could not see.

01 — Generate

Create

Novel molecules designed directly in the target's binding site — not screened from known libraries.

02 — Cross-validate

Confirm

Independent, orthogonal methods at every stage, so one signal never decides.

03 — Accumulate

Compound

Every result — including negative ones — feeds the platform and sharpens the next design.

01Generative design Create novel candidate molecules for the target.

02Co-folding structure prediction Rapidly triage promising binders from the rest.

03Structure-based docking Predict how a molecule sits in the binding site.

04Molecular dynamics Test whether the binding holds over time.

05Free-energy methods Quantify binding strength on a physical basis.

06ADMET & drug-likeness Evaluate absorption, metabolism, safety, and more.

07Synthesizability Confirm a real, viable route to make it.

08Novelty screening Check it is distinct enough to protect.

→ Validated candidate A molecule that has passed every independent gate.

Validation discipline

We trust no single signal.

A good docking score is not binding. Binding is not potency. Each layer must be independently confirmed — and this discipline is what separates a real candidate from a computational artifact.

Principle 01

Distrust single metrics

No one score decides. Binding prediction, dynamic stability, and free energy each say something different — and any of them can reverse the verdict.

Principle 02

Independent cross-validation

We confirm with methods built on different principles. Synthesizability is checked several ways; free energy is computed by two mathematical routes at once.

Principle 03

Prediction meets measurement

We compare predictions against real measurements and correct the next prediction with the gap. The platform grows more accurate the more it is used.

The real asset is not any single molecule. A candidate list ages; the ability to design better molecules — and the discipline to validate them — does not. Every project deepens the platform.

What the gates actually do

Three million in. Nineteen through the gates.

A recent multi-target campaign, stage by stage.

Structure-based generation 3,000,000

Passed novelty filter 424,387

Passed property filter 34,847

Docked against all three targets 34,755

Strong binding at every target 1,645

Completed co-folding prediction 1,644

Top tier, carried forward to extended dynamics 19

Every stage is an independent gate, and each one uses a different principle from the last. The overall pass rate is 0.0006%. We treat that number as a measure of rigor, not a weakness — a pipeline that passes most of what enters it is not filtering anything.

Scored against published experiments

We do not run experiments. We are graded by them.

Every case below had its answer already sitting in the literature, and we went at it with that answer closed. Our own programs stay undisclosed; the scoring runs on public benchmarks, so anyone can repeat it.

01

Imatinib resistance mutations, predicted blindFor six known resistance mutations in Abl kinase, we computed how far each one shifts drug binding. The measured answers were taken from ChEMBL and hashed shut before any calculation began, then opened once the predictions were fixed. Classification was correct in 4 of 5; mean error against experiment was 0.32 kcal/mol across 4 mutations — the same spread two laboratories get measuring the same mutation.

02

A known binder, found againA compound with experimentally confirmed binding went through the procedure as a control. Nine runs, nine calls, all correct. Nothing unknown enters the pipeline until the known answer comes back out of it.

03

A crystal structure held with no restraintsFrom crystallographic coordinates, 100 ns of dynamics with nothing restrained. Every interaction the structure paper identifies survived. Under restraints it would have held by construction; without them, it is evidence the system was built correctly.

04

Bilayer thickness against scattering dataWithin 0.2% of the measured value. This is the first quantity to drift when a system is assembled wrongly, which is why it is the first thing we score. We re-check it on every build and do not run one that misses.

05

The point where our method stops seeingTwo compounds whose measured cooperativity differs 34-fold are not separated by our membrane metric; the gap sits inside ±6.6 percentage points of measurement uncertainty. We publish the limit, not only the reach.

Pre-registration record

Analysis plan fixed before the data was opened

2026-11-03 17:11 KST
md5 14985ad4bc5e8f11e7bd98cb84dad7a8

For a multi-component lipid bilayer campaign, the observables, the averaging window, the convergence thresholds and the rule for extending a run were written down and hashed before a single production trajectory was analyzed. The analysis tool will not read the production directories unless it is handed that hash. The two pilot systems were analyzed before the plan was fixed, and they are labeled as such. The document is available in full under NDA and in diligence.

Every experimental value scored here was published and settled by other groups before we graded against it, so the answer was fixed in advance and there was nothing to tune toward. That is the harder test, and the reason the record is built this way.

Targeted protein degradation

It is a geometry problem, not a binding problem.

Inhibitors block a protein's function. Degraders remove the protein. That difference bypasses the feedback activation which limits inhibitors, and it is less vulnerable to resistance driven by target overexpression — and because degraders act catalytically, activity does not require sustained high occupancy.

The trade-off is difficulty. An inhibitor needs to bind one site. A degrader must hold the target and an E3 ligase at the same time, and the resulting ternary complex has to adopt a geometry that actually permits ubiquitin transfer. Lumina Bio has concentrated on treating that geometry problem computationally. The linker is where that geometry is decided — and on a membrane-anchored target, the geometry available to it is not the one you get in solution.

PROTAC — current focus Molecular glue — fourth research axis

We do not narrow a program by indication, modality or molecular weight. A degrader that breaks the usual property rules may still be the right molecule somewhere, and a large design is not disqualified on size alone. Reaching the cytoplasm is a separate problem from forming the complex once there, and we have started separate research on it.

Ternary complex predictionDeep-learning co-folding of the bound complex.

Stability under dynamicsExtended simulation of the predicted structure.

Benchmarked against published structuresMeasured against active-state complexes in the literature.

Exhaustive lysine scanUbiquitination sites mapped, not assumed.

Independent cross-checkConfirmed by a method built on other principles.

An honest boundary

We report these as geometric feasibility assessments, not verification. Computation evaluates whether a productive arrangement is possible; actual degradation efficiency is established experimentally, and we do not describe one as the other.

Membrane context

A degrader can look perfect in solution and be impossible on the membrane.

We find that out before you commit to synthesis. Most degrader work models the ternary complex in water, but many of the proteins that matter never see that state — they sit on a membrane in a defined orientation, and that orientation decides where an E3 ligase can and cannot reach. We build those systems, run them, and hand back a result you can check.

On a membrane, the whole assembly moves. A solution calculation never sees this state. Nanoseconds, compressed

Deliverable 01

Ternary dynamics in a bilayer

The whole system — target, degrader, E3 ligase and its adaptors — in an explicit lipid bilayer, with the composition stated in the report rather than left at an unreported default. We report whether the interactions that define the complex survive the run, residue by residue, with occupancy measured across the trajectory rather than read off one frame.

These systems are not small. A bilayer wide enough that the complex never meets its own periodic image, with explicit solvent and counter-ions throughout, runs to roughly six hundred thousand atoms.

Deliverable 02

How much of the trajectory admits the geometry

An E3 ligase arriving at an anchored target has to fit somewhere, and part of the space around that target is taken up by the membrane itself. We measure how much of a trajectory admits a given ternary geometry, and report where candidates stand relative to one another — which of them the membrane leaves room for, and which it does not.

This is the part solution-phase prediction cannot see.

Deliverable 03

System build, with the checks shown

If you run your own simulations, we will build and verify the system. Our protocol covers six areas: structural integrity, parameter assignment, periodic boundaries, equilibration, sampling, and the basis on which a verdict is reached. The last of those is where most of the damage hides — a ligand routinely passes every count-based check while being structurally incorrect.

The report is the deliverable, not an appendix to it.

Two orientations, not the convenient one

A membrane-anchored target does not hold a single pose. It samples several, and the pose you choose decides the answer you get. Building only the orientation that favors a complex is answering a question you have already decided.

So we build both of the states our sampling identifies, and we build them the same way: same composition, same construction procedure, same simulation engine, same equilibration protocol. Where the comparison is the deliverable, any difference in how the two sides were made is a difference you cannot subtract out afterwards.

If the two orientations agree, orientation is not the gate we suspected, and that is a result. If they diverge, it is a term solution-phase prediction cannot see at all.

The run finished. Does that mean it was right?

Every project we take on is put through a fixed set of checks before anything is trusted. These seven catch the kind of problem that never announces itself.

What we checkIf it goes unchecked

Is the analysis reading the current trajectoryA stale derived file is reused, values match to the last decimal, and the result reads as "no difference".

Are the replicas genuinely independentThree runs on the same seed are identical trajectories, reported as a reproducibility check.

Do repeats seeded from an earlier run still remember itFresh velocities on the same end structure pass a structural-distance check while the repeats still track where they started, and error bars that count them as independent come out too narrow.

Does the distance formula match the box shapeA rectangular formula on a non-rectangular box returns 106 Å inside a 140 Å box.

Does the pass criterion account for the equilibrated stateA judgment made on absolute coordinates discards replicas that in fact ran to completion.

Have periodic boundaries been resolvedRaw coordinates read as real displacement make a protein look embedded in the membrane when none of it is real.

Is every distance measured from the same reference planeTwo quantities carrying the same name, measured against different planes, get compared as though they were one.

None of these raises an error. The job exits cleanly and the numbers look ordinary. That is precisely why they survive into conclusions.

And the checks are checked

A test that cannot hold up does not reach your report

As a rule, a statistical test and its pass line are fixed and hashed before the data are opened, and the test is first run on synthetic data with a planted answer — so we know it raises false alarms only at its stated rate and catches what it is meant to. A test that does not clear that bar is demoted to a diagnostic and never carries a judgment.

What that looks like in practice: a drift test — flag a run if the second half of its averaging window differs from the first by more than two standard errors. Run against synthetic data, a flat cut-off gave false-alarm rates from 3% to 10% depending on the series, so a single threshold was strict for one observable and lax for another. Calibrating on the measured autocorrelation is circular, and an assumption-free bound is valid but separates almost nothing.

The limit is in the data, not the arithmetic: over a window this long the autocorrelation time cannot be pinned down from the same series well enough to calibrate on. So the test stands as a diagnostic, and the judgment rests on a comparison across independent starting configurations, which needs no autocorrelation model. Every version of the plan stays in the pre-registration, so you can see the analysis was never reshaped around the answer.

Checked against experiment — not against ourselves.

A method validated on its own results proves nothing. Ours is measured against published crystal structures and published experimental values, so anyone can pull the same references and check the arithmetic.

Independent benchmark

Does this just repeat what SPR already tells you?

It does not, and we checked. Taking three ternary complex structures published by an outside group, we computed our membrane compatibility metric for compounds whose cooperativity that same group had measured by SPR. Two of them differ 34-fold in cooperativity — α = 143 against α = 4.2 — and our metric could not tell them apart. The gap between the two values falls well inside the measurement uncertainty we later quantified, ±6.6 percentage points.

So the quantity we measure is not one the binding experiment already reports. It is orthogonal information, not a slower route to a number you can get from SPR.

It is also not a potency predictor — in that series, cooperativity dominated the outcome. Our contribution sits upstream of it: ruling out the geometries a membrane will not permit, before anyone commits to synthesis.

Ligand parameters with no hand correctionAcross 298–1,486 Da, including degraders with more than twenty rotatable bonds. No missing parameters, no substituted analogies.

Orientation behavior against the literatureReproduced. Published studies reached those states with microseconds of sampling; ours arrived within hundreds of nanoseconds. One of the states appears in every replica we have run; the other does not, and we say so rather than average them together.

Two orientations carried in parallelBoth membrane states are constructed from the same procedure and equilibrated under the same protocol, so the comparison between them rests on the orientation and not on how each was assembled.

Controls before rankingsWhere a target has published measurements, we run a control set alongside the candidates and measure the resolution of the method before we rank anything. Where no such set exists, we do not produce a ranking. We have discarded rankings made without one.

What survives more samplingExtending our trajectories by a factor of three cut the absolute compatibility values by as much as half. The ordering of compounds did not change, and neither did which pairs the method cannot separate. We report rankings and separability rather than absolute values, because those are what held.

Experimental follow-up

Where a project needs wet-lab confirmation, we arrange it through contract laboratories — yours if you have one you trust, ours if you do not — and fold the results back into the model.

What we measure is geometric feasibility, not degradation efficiency: the two are established by different means and we do not describe one as the other. The full verification checklist, and a plain statement of what this method can and cannot settle, are available on request or at a scoping call.

Services

Where our partners come in.

We develop our own pipeline — and we put the platform to work for other teams. Partners enter at different points; we meet them there.

Start here

Computational reliability audit

For teams already running their own simulations. You send the system, not the science: we put it through the same checks our own work goes through and return a reliability report — not a verdict on whether you were right, but a statement of how far each result can be carried into a decision.

What we take Run inputs and a portion of the trajectory. We do not take your compound structures. There is no molecule of yours on our side to protect, which makes the first engagement a short conversation rather than a long negotiation.

What you get back A reliability report: each check with its finding, the command to reproduce any of them yourself, and what each one bounds downstream. You are not asked to take our word for any of it.

What we do not give Any judgment on your compounds. That is a separate engagement and a separate conversation, and it does not start here.

One check needs even less than that. If your repeats were seeded from earlier runs, we test whether each still tracks its parent on dH/dλ, the energy the free-energy estimate is built from, report how many independent replicates the data is actually worth, and recompute the error bars by lineage. For that we need only the per-replica dH/dλ series and which run started from which: no structures, no coordinates.

And whether a named column in a tool measures what its name says. A column labeled crypticity, druggability or confidence can, under a given input configuration, collapse to a pure function of another column — so a ranking carries a name that is no longer measuring that quantity. We check it from the tool and its settings alone; your structures are never required.

We run one of these at a time. That is a limit, and it is also why the date we give you holds. Scope and timing are set at the start; write to us and we will tell you what it involves before anything is committed.

Once a set exists

Lead triage by dynamics

You have five candidates and budget to synthesize two. The question is not which looks best on a score, but which of them a method can actually tell apart. We answer the second question, because the first one is only worth asking after it.

What we take Your candidate set and the target structure, under NDA. Where the target has published measurements, we ask for nothing else — the control set comes from the literature.

What you get back A ranking from free-energy differences computed across replicas, with the resolution of the method stated beside it and the pairs it cannot separate named as such. Stability measures are reported, but only ever to rule a candidate out.

What we do not give Absolute affinities, and any ranking of candidates that fall inside the resolution we just measured. Where a target has no published control set, we do not produce a ranking at all.

We run the controls first and the candidates second, in that order, because a ranking is only meaningful once you know how finely the method can discriminate. It costs extra compute and it is the reason our rankings are shorter than they could be.

Entry 01

Strong chemistry, no in-silico

You have an excellent medicinal chemistry team but no computational design capability. We design candidates for your target; you synthesize.

Entry 02

You ran it. Can you trust it?

The simulation finished and the answer looks clean, but nobody has checked whether the system underneath it was built correctly. That is where most engagements with us begin.

Entry 03

Biology, meet computation

We connect target biology and molecular simulation in one integrated pipeline.

Start here

Computational Reliability AuditYour simulation, put through our checks. No compound structures required.

Core work

Lead Triage by DynamicsFree energy sets the ranking; stability measures only rule candidates out. Absolute values are not reported.

In-silico DesignNovel candidates designed for your target.

Hard-Target StrategyWorkaround design for targets that resist.

Synthesis CoordinationSynthetic routes and external CRO synthesis, coordinated.

Advisory RetainerOngoing in-silico support, month to month.

Milestone PartnershipMilestone-based, with joint-IP options.

Flagship capability

Membrane-Context SimulationOur hardest problem, and the one we have built the most around. Commissioned case by case rather than run off a price list.

Cryptic & Allosteric Site DiscoverySites a single static structure cannot show, found across hundreds of conformations from microsecond-scale dynamics and counted by how often they open — not asserted from one frame. We report nothing until the calibration stands.

Your data

Nothing has to leave your side

We run our own on-premise GPU infrastructure, so your project is never bottlenecked by rented compute and research data does not transit external cloud services. Structures and compounds can be processed entirely on premises; we do not require upload to a hosted service.

Your IP

You get the system, not only the answer

Coordinates, topologies, trajectories, analysis scripts and the verification report — enough to repeat the calculation yourself or hand it to another group. Rights to what a commissioned campaign discovers belong to you; the platform and its methodology remain ours. We set this out in contract, work under NDA, and never repurpose your target for anyone else.

Your name

Your name stays yours

We do not publish client names, logos or case studies, and we do not reference an engagement to anyone else. Nothing about your program appears on this site. What we put in public is method, scored against published experiments — so the work can be judged on its own terms rather than by whose name we are allowed to print.

Getting started

Start with a conversation.

If you have a target or a compound you would like to discuss, write to us. Scope and terms are set to the problem rather than to a template, and a short technical call usually settles whether computation will help before either side commits anything. Detailed data is exchanged under NDA.

Request a scoping call

Pipeline

What the platform has produced.

These are in-silico validated assets — evidence that the platform carries a candidate all the way through. They are not clinical-stage programs; wet-lab validation lies ahead.

Program	Target	Stage
Program A	Undisclosed	Provisional filed (KR, Aug 2026)
Program A — second indication	Same target, second indication	Covered by the same filing
Program B	Undisclosed	Deprioritized after free-energy triage
Program C	Undisclosed	Closed after free-energy triage

Stages: Design → Multi-stage validation → Free-energy → Patent filing → Synthesis → Binding assay → in-vivo PoC → Partnering

These are in-silico assets, not approved or clinical-stage medicines: nothing here has yet been tested in the laboratory. Disease areas, targets, structures and figures stay withheld while filings progress. A non-confidential summary — stage, data in hand, and what the next step requires, with no structures — is available on request.

Roadmap

What we are working toward.

What follows is what is running now and where the method extends next. These are objectives rather than commitments, and we state them at the resolution we can actually defend. Computational work has a way of returning answers nobody planned for, and when that happens we would rather change the plan than the account of it.

Underway now

Extend the paired-orientation comparisonTwo membrane-anchored systems, built identically and differing only in the orientation of the anchored target. The first round is behind us; what is running now adds independent lineages to it. The question is whether orientation governs whether a productive ternary geometry can form at all. Both answers are informative.

Turn that comparison into a reference caseA worked example with the build checks shown, so what the deliverable contains can be read in full before any scope is agreed.

Check whether the starting arrangement persistsA multi-component bilayer is being run from two starting arrangements, one mixed and one separated, each from independent seeds and each built and run the same way. If the two converge to the same membrane, the starting arrangement is not a gate, and that is the result. If they do not converge within the run length, we report that rather than averaging them together.

Where it extends

A published protocol behind the serviceThe membrane-context assessment is running now. What we are adding is a protocol published in full and a fixed verification report, so a partner can read exactly what was done rather than take our word for it.

Experimental cross-check of the geometric metricThe metric is benchmarked against published structures and published measurements. Where a laboratory measures a system we assessed in advance, we publish the comparison whichever way it falls.

Resistance-mutation prediction, under retrospective testFree-energy differences between wild type and mutant, scored against resistance mutations already reported in the clinic. We are assembling the retrospective record first. Until that record stands on more than a handful of cases, this is something we are checking, not something we sell.

Extension from PROTACs to molecular gluesThe geometry problem is the same one. The tolerances are tighter, and the membrane constraint matters more, not less.

Advance our own degrader program toward experimental confirmationStructures and target remain undisclosed until filings permit otherwise.

We have revised this roadmap before, and we expect to again. A plan that never changes in this field is a plan nobody is checking against results.

Company

Built from the ground up, for precision.

We believe the next generation of medicines can be designed — computationally, rigorously, and for the targets the world has given up on. We build that capability, and prove it one candidate at a time, through every gate that stands between a calculation and a claim.

Lumina Bio develops and operates the Andromeda platform in house. No single score decides an outcome here, and nothing advances on computation alone — that discipline, rather than any one molecule, is what we take to be the asset.

CompanyLumina Bio Co., Ltd.

CEOChoi, Jae Yong

LocationRoom 409, 204 Munhwa-ro, Naju-si, Jeollanam-do, Republic of Korea

FocusComputational chemistry and biochemistry · targeted protein degradation and molecular glues · membrane-context simulation

Business registration223-88-03724

Contact

Have a hard target?

For partnership, licensing, and scientific collaboration. Detailed compound data is exchanged only under NDA, and we reply within business days.

Prefer email? Reach us directly:

Inquiries [email protected]

Administration [email protected]

Your hardest target — designed through to the ligand.

Hand over one stage, or all of them

Multi-layer discipline

A way around the wall