Skip to main content

Describe your process. Studio builds the model.

Tell Studio whether you run a batch, fed-batch or continuous process, describe it in silico and bring your data. Studio searches the model settings, ranks the candidate models on your runs and keeps the best, so you get a model that predicts process performance and product quality without building it by hand.

Book a demo

A process model without the modelling work

You bring the campaign data you already have and describe your upstream process in the terms you already use. The reactor model workflow in Novasign Studio turns that description into a process model, so your time goes into the process rather than into the model.

With the model trained, you can try conditions that have never been run in a lab and see how the process would have behaved. The model suggests which conditions are worth testing next and which setpoints move you towards your objective, and you decide which of them are worth a vessel.

You define the model once and retrain it on every new experiment, from early development across scale-up to manufacturing.

Your host, your product, one workflow

The workflow models any host and any product class. A CHO fed-batch, a viral vector production run and a Komagataella phaffii fermentation are described with the same workflow, and what changes is which performance metrics and quality attributes you are interested in. The way you describe the reactor stays the same.

Modalities also differ in how a run is staged. An induction, a carbon switch, a temperature shift or a transfection each opens a new phase in the bioprocess, and every one of them can be defined in the model. These are examples of what is supported.

Mammalian cell culture

Performance metrics
Viable cell density, product titer
Quality attributes
Charge variants, glycoforms, aggregation
Process events
Temperature shift, feed start

Cell and gene therapy

Performance metrics
Viable cell density, viral titer
Quality attributes
Full-to-empty ratio, dsDNA, host-cell protein
Process events
Transfection, infection at a chosen multiplicity

mRNA

Performance metrics
RNA yield
Quality attributes
Integrity, capping efficiency, poly(A) tail length, dsRNA
Process events
Template addition

Microbial fermentation

Performance metrics
Biomass, product titer
Quality attributes
Aggregation, host-cell protein
Process events
Feed start, induction

Precision fermentation

Performance metrics
Biomass, product titer
Quality attributes
Aggregation, host-cell protein, glycosylation in yeast hosts
Process events
Induction, carbon switch

Microalgae and cyanobacteria

Performance metrics
Biomass, chlorophyll a, polysaccharide titer
Quality attributes
Released share of total carbohydrate
Process events
Light intensity changes

Any attribute you measure can be used as a quality attribute. Contact Novasign if your modality is not mentioned.

You describe the process. Studio sets up the model.

Setup has two halves, your upstream process and the machine learning.

In the words you already use

Each half keeps its own vocabulary: process parameters and quality attributes for your process, loss functions and checkpoints for the machine learning.

The bioprocess side: parameters, components and feeds

Whichever controller software your runs were recorded with, the data can come in. Exports from Sartorius Ambr, Eppendorf DASGIP or Thermo Fisher systems come in as CSV or Parquet files, a Qubicon system connects directly, and custom-built infrastructure connects through the custom-code node. Historian exports come in as CSV or Parquet files.

Raw process data rarely arrives ready to model, which is why data processing is part of Novasign Studio in its own right: the dataframe nodes that clean and align a campaign run in the same workflow. What matters for the user is that the processing is reusable. You set it up once and it runs again on the next campaign, the same way the reactor model does.

The setup then walks you through your process.

Reactor Model
1

Initial configuration

Working volume, the time base, and the run IDs that separate one run from the next, so a whole campaign arrives as one table.

2

Process parameters

The parameters that are recorded and actively set during the run, and that the model should account for: temperature, pH, dissolved oxygen and stirrer speed, among others.

3

Components

Everything that is in the reactor or can enter it, whether present at the start, added or removed throughout the process or produced over time. Set cells, substrates, products and define material attributes, performance metrics and quality attributes.

4

Feeds

Every addition that was recorded, from a substrate feed or base to a one-off event like an induction or a transfection. A feed that differed between runs is a column rather than a number.

Every column is defined with its unit, °C, % or g/L, so the mass balances stay dimensionally consistent and nobody has to rescale a column by hand to make them add up.

The machine learning side: model training, tuning and selection

Trials save checkpoints as they train, and the model you carry forward is selected from them, as for every Studio model.

  1. 1

    Train

    Training fits the model to your data, with defaults chosen for reactor processes rather than for machine learning in general.

  2. 2

    Tune

    Tick tunable next to a setting and the tuner sweeps it in one pass instead of one candidate at a time.

  3. 3

    Select

    The selector chooses between the candidates, automatically in most cases or against criteria of your own.

Model Selector

Try the runs you have not done

With a trained model you can simulate runs that were not performed. The digital twin takes the trained reactor model, adds a simulation configuration, and returns trajectories for everything the model was trained on: cells, substrates, product, and the quality attributes you measured, whether that is the glycoform of a monoclonal antibody or the viral titer of a viral vector process.

The decisions the digital twin helps you make

Which decision you face depends on the phase of the lifecycle the process is in.

  • Screening

    Do these process parameters influence the process at all?

    Vary them across the space and watch what the performance metrics and the critical quality attributes do. A parameter that moves nothing is one fewer to characterize.

  • Optimization

    Which conditions give the best performance at acceptable quality?

    State the objective and the ranges to optimize within, and run the simulations inside that space. Every in-silico run can be read in a table or a chart. The visualizations and statistics show you which experiments are worth performing. Once the model has earned your trust, the workflow proposes the next set of experiments for you.

  • Characterization

    How robust is the process inside its design space?

    Restrict the acceptable ranges of your critical quality attributes and what remains is a design space. Vary the process inputs inside it, and how much the outputs move is your answer on robustness.

  • Scale-up

    Which conditions are suitable for the move to a larger vessel?

    Of the conditions that give acceptable results at lab scale, which ones are physically and biologically possible at scale? The hybrid model answers for the culture, and the CFD model of the target vessel answers for the vessel.

Configuring a simulation

A simulation adds the initial state of the reactor and the phases of the process to what the trained model already knows. Letting the configured phases drive the prediction is in development.

The initial state is the filling level and the media the reactor starts with. Changing it is how you simulate a different scale or a different media composition.

Your vessel enters the model through computational fluid dynamics (CFD). CFD models of your vessels, from Novasign’s technology partner SimVantage, give the kLa, shear and mixing of every run, and the Reactor feeds them into the hybrid model. The hybrid model learns what the culture does, and the CFD describes what the vessel does to it. See how Arkeon scaled up from 2 L to 150 L.

A phase is a section of the upstream process, and which ones you have depends on the modality: pre- and post-transfection in gene therapy, pre- and post-temperature-shift in cell culture, batch, growth and production in a microbial fermentation. Each phase is defined by three things.

1

Process parameters

Every phase has its own set of setpoints.

2

Feeds and outflows

Feeding strategies cover continuous feeds from constant through exponential, spiked feeds, and feeds that react to the current process state. Outflows are defined the same way, whether continuous like a perfusion culture or one-off like periodic sampling.

3

Stop criterion

A time, a volume, or a concentration target. The first to fire ends the phase and starts the next.

Reactor Process Simulator

Generating the in-silico runs

The digital twin lets you run the process in silico for a new set of inputs.

The example below is a simulation built on data from a viral vector process in which pH, temperature and multiplicity of infection were investigated systematically. Move the sliders and the twin simulates cells, product and the impurities, dsDNA and host-cell protein, for each condition.

Viral Vector Use Case

Simulation inputs

7.2
33 °C
0.003

pH 7.2; Temperature (°C) 33 °C; Multiplicity of infection 0.003; Cells on micro-carrier highest predicted 1.12 ×10⁶ cells/mL at 80 h; Virus titer highest predicted 6.923 log₁₀ TCID50/mL at 160 h

Highest predicted value

  • Cells on micro-carrier

    1.12 ×10⁶ cells/mL

    highest predicted · at 80 h

  • Virus titer

    6.923 log₁₀ TCID50/mL

    highest predicted · at 160 h

The data comes from a collaboration with Bespark*bio: Bespark*bio performed the experiments and Novasign created the model.

In collaboration with Repligen, the digital twin was used to predict the glycoform and glycation of a biosimilar when temperature and target glucose concentration change. The twin shows which conditions yield the required ranges of the critical quality attributes, and that defines the design space for the process while optimizing for product titer.

Design Space Finder

Quality criteria

G0F/G1F glycoform (%)10.3 – 29.2
Glycation (%)0.0 – 19.0
Predicted IgG titer (g/L)1.08 – 1.63

Highest predicted value

  • Predicted IgG titer

    1.63 g/L

    highest predicted titer inside the current windows

A grid of predicted IgG titer in grams per litre over 15 temperatures after the shift, from 30 to 37 degrees Celsius, and 9 glucose setpoints from 1 to 5 grams per litre, at 320 hours of process time, one simulation from each of the five runs per cell. A cell is kept where at least one of its runs has its predicted G0F/G1F glycoform and glycation inside their windows, it shows the mean titer of its five runs, and cells below the titer window stay empty. The design space is drawn as one connected region with no failing condition inside it, so a passing condition cut off from that region stays empty too. The windows open on example limits, G0F/G1F from 10.3 to 29.2 per cent, glycation up to 19 per cent and titer from 1.1 grams per litre, where the acceptable conditions form a wedge at low glucose and high temperature after the shift, and they can be moved. A readout gives the highest predicted titer inside the current windows. Highest predicted value: Predicted IgG titer 1.63 g/L, highest predicted titer inside the current windows.
Predicted titer and the current windowsPredicted IgG titer 1.08–1.63 g/L; Highest predicted value 1.63 g/L
Questions about reactor modelling

Start with one process question. See what your data can do.

Bring one process and the campaign data you already have. We work with you on a scoped proof of concept, so what comes out is a model of your process rather than a demonstration on somebody else’s. Not a standard unit? Scientific modelling models it with you.

Book a demoGet in touch

From the reactor

Online sensors, soft sensing and what they make possible in the vessel.

All posts