LumiPredict

The problem

Every new fluorophore is still discovered the slow way.

Step 1

Synthesise

Route design, synthesis and purification for each candidate — before anything can be measured.

Step 2

Measure

Absorption & emission spectroscopy and quantum-yield determination on specialist instruments, solvent by solvent.

Step 3

Repeat

Most candidates miss the target photophysics — and the loop restarts from synthesis.

Dyes, probes, sensors and display materials hinge on four numbers — absorption wavelength (λabs), emission wavelength (λem), quantum yield (Φf) and extinction coefficient (ε) — that are only obtained after making the molecule, with biodegradability usually arriving last, when redesign is most expensive.

LumiPredict lets you pre-screen candidates first, so bench time goes only to the molecules worth making — from make everything, then measure to pre-screen, then commit to the bench.

The solution

A production web application that turns a structure into its photophysics.

  • Five properties — λabs, λem, Φf, ε and biodegradability.
  • Any identifier in — trade name, IUPAC, CAS, InChI, InChIKey or SMILES; solvent given the same way.
  • Solvent-aware — the medium is part of the prediction, not an afterthought.
  • Prioritise before the bench — rank candidates in the browser, take only the best forward.
  • Instant feedback — live structure rendering, results on screen, one-click CSV export.
lumipredict.cdi-sg.com
LumiPredict single-molecule prediction page: property selection and model settings

Live demonstration

Type the names you know — molecule and solvent.

Get a structure and five properties.

rhodamine b+ethanol →molecule & solvent names both resolved →rendered structure →five properties
Molecular Input panel: the molecule field resolved from the name rhodamine b and the solvent field resolved from the name ethanol, each with its rendered structure
Both names → the right structures. Neither was drawn or pasted.
Prediction Complete card: emission 584.08 nm, quantum yield 0.71, absorption 526.99 nm, extinction coefficient 4.66 log10, biodegradability 0.0000
All five properties, unedited from lumipredict.cdi-sg.com (v1.2.0).
Absorption 527 nm Emission 584 nm
Both inputs were typed as names — rhodamine b and ethanol — and resolved to SMILES automatically. Biodegradability 0.00 correctly flags a rhodamine dye as not readily biodegradable.

Batch prediction

Screen a whole library in one pass.

2,800 × 2 props ≈ 37 s
Throughput
measured in production
1 column
is all it takes
first column = molecules, any identifier
0
batches lost to one bad row
row-level fault isolation

Upload a CSV of names, CAS numbers and SMILES — mixed. Prioritise the whole set, then commit only the top candidates to synthesis.

Unresolvable rows come back flagged, never silently dropped.

Batch results: summary statistics and a results table of molecules, solvents and predicted properties

Prediction risk

How close is your molecule to what the models have seen?

Every prediction can be graded against the training set. The similarity view finds the nearest known molecules — their structures, how similar they are, and what they actually emit — and a complementary fragment view names any substructure the model has seen little of.

LOWMEDIUM HIGHUNKNOWN
Emission-wavelength distribution of the 50 nearest dyes, bars shaded by similarity
Emission of the 50 nearest dyes, shaded by similarity — clustered at a 582.86 nm mean.
Table of the nearest training dyes: structure, similarity 85 to 91 per cent, and measured emission
The nearest known dyes — 85–91% similar — emit at 570–640 nm, bracketing our predicted 584 nm for rhodamine B.

Security & privacy

Your structures stay yours.

Nothing is stored

Every prediction runs in your session and is returned to you. Your molecule is never written to a database or kept after the request.

Your data isn’t used to train

We never train models on your inputs, and we never share them. Contributing experimental data to improve the models is entirely opt-in.

Independently pen-tested

LumiPredict was externally security-reviewed and penetration-tested in 2026 — findings fixed and regression-tested. Access is by secure single sign-on.

Live in production today (v1.2.0), behind your organisation’s single sign-on.

Improves with every prediction

Your feedback trains the next model.

When you have measured data, close the loop right in the app — rate any prediction and attach the experimental value. Curated feedback feeds the next model release, so accuracy compounds for the properties and chemistries you actually work on.

  • Rate it. Mark each property Correct, Close or Incorrect.
  • Attach data. Add the measured value and its source — optional.
  • Models improve. Contributions are curated into the next training round.
In-app feedback panel: choose correction, confirmation or flag issue, then rate each predicted property and attach an experimental value
Strictly opt-in. Nothing is stored unless you choose to submit feedback — consistent with the session-based, no-storage default.

Beyond fluorescence

Need a property we don’t predict yet? Start a project with us.

01 Data

We source, curate and quality-control the data for your property — from literature, licensed sets and your own measurements.

02 Models

We train and validate custom property models on that data, with the same training-domain transparency you see here.

03 Application

We deliver a production web application — like LumiPredict — deployed securely on AWS, branded and integrated for your team.

A proven pipeline, not a one-off. LumiPredict itself was built exactly this way, in close collaboration with industry partners — data curated, models trained and validated, and the finished application deployed securely on AWS. We bring the same pipeline to your property.

If you’re interested in booking a demo or would like to learn more about pricing, please contact us at info@cdi-sg.com