TrellisbyLemnisca

The shortest path for your bioprocess

from plate to pilot to production

Every bioprocess is a large, partially observed system. Trellis builds one evolving model of yours, so each experiment adds the evidence your next decision actually needs.

Book a callSee how it works
Plate to pilot to production: a route climbing a response surface
The problem

You are optimising a process you can only partially observe.

From a limited number of experiments you still have to work out which parameters matter, how they interact, what drives your CQAs, and where the process is robust.

Most of the design space will never be tested. So a condition that looks good may simply be the best one you happened to sample.

Limited experiments

The design space is far larger than the laboratory can physically sample.

Interacting variables

Understanding one variable in isolation rarely explains the process.

Fragmented learning

Design, evidence, assumptions, models and decisions end up in different systems — or different people.

How it works

A learning loop the scientist drives.

Trellis connects experiment design, data, modelling and next-experiment selection into one continuous workflow — then recommends the run most likely to produce evidence that matters for the next decision.

You define the objective, factors and constraints, review what the model has learned, and decide whether it needs refining before the next experiment runs.

  1. 01Frame
  2. 02Design
  3. 03Integrate
  4. 04Model
  5. 05Optimize
  6. 06Learn
The evidence

Same process, same objective, different learning efficiency.

OFAT, DoE and Trellis were benchmarked on the same virtual 12-factor CHO fed-batch process. Trellis reached effectively the full known optimum; DoE plateaued near 77%, OFAT near 50%.

The difference is where the experiments go. OFAT searches locally, DoE spreads broadly and wastes runs far from the optimum, Trellis explores uncertain regions and concentrates where better performance is most likely.

OFAT
50%
DoE
77%
Trellis
99%
KNOWN OPTIMUM050100EXPERIMENTS RUN% OF KNOWN OPTIMUM TITREOFAT50%DOE77%TRELLIS99%
Plate to pilot to production: a route climbing a response surface
Start here

Bring us the development decision you're still uncertain about.

Start with one real decision, not a platform migration. We configure Trellis around it and you keep every assumption in view.

  1. 01
    Bring one processA live process with an important, unresolved question.
  2. 02
    Define the objectiveAgree the objective, available evidence, practical experiments and success measures.
  3. 03
    Run a focused pilotConfigure Trellis, generate new evidence, review what the model learns, decide next steps.
Start with one decision

Questions we get asked

That is the case Trellis is built for. It combines whatever experimental evidence you already have with published process knowledge, so the first model is grounded even when the run count is small — and every run after that sharpens it.

DoE fixes the whole experimental programme up front and spreads runs evenly across the design space. Trellis chooses each round from what the current model knows, balancing exploration of uncertain regions against exploitation of promising ones. On the same benchmark process that difference was roughly 77% versus effectively the full optimum.

No. It recommends. The scientist defines the objective, factors and constraints, reviews what the model has learned, and decides whether the model needs refining before the next experiment runs. Assumptions, choices, uncertainty and recommendations all stay visible and reviewable.

Nothing resets. The same model carries forward, now also accounting for scale-dependent behaviour. That is the whole point of the approach: the stage where experiments get most expensive is the stage that inherits the most understanding.

The published benchmark is a 12-factor CHO fed-batch process, but the method is not specific to CHO. If your process has parameters that interact, responses you care about, and fewer experiments than questions, it applies.

One live process with an unresolved question, an agreed objective and success measure, and a short round of experiments. You review what the model learned and decide the next step from there.