Model your TFF process

The tangential flow filtration (TFF) model workflow in Novasign Studio is the framework you build your TFF model in, covering batch, fed-batch and continuous concentration, batch and constant-volume diafiltration, and recovery. You bring the run data you already have, the workflow walks you through describing the membrane and the dissolved components, and the model applies to a product from any modality.

Once the model is trained you can run scenarios in silico, conditions that have never been run on a membrane, and see how the process would have behaved. From there the model suggests which conditions are worth testing next and which setpoints, buffer volumes and excipient concentrations move you towards the objective you are optimizing for, and decide which of them are worth performing.

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

The workflow describes every operation a TFF step is built from with the same handful of concepts: the membrane, the molecules in the feed and the buffers that are added. What changes between a concentration and a buffer exchange is which buffers are added, and the way you describe the membrane stays the same.

Concentration, diafiltration and recovery phases define the TFF process, and every one of them can be defined in the model, optimized and used for what-if scenarios to improve the process.

Every component is defined with its own rejection coefficient, from fully retained to freely passing, so the same balance that concentrates a product also lets a salt through. Contact Novasign if your operation is not mentioned.

In TFF the experiment is part of the modelling method, which saves a significant amount of time. A TFF model trains on concentration-diafiltration-concentration (C-D-C) runs, the ones where the process parameters are held at their setpoints while the retentate concentrates, and if you already have those it will use them. Each experiment allows the model to learn the influence of one set of conditions. This can be improved so that one experiment provides information about multiple conditions.

The recurrent nature of TFF opens a shorter route. The retentate and the permeate can be returned to the vessel, so the retentate concentration stays constant. That makes the material attributes and the process parameters controllable, and the data can be screened exactly where it is needed: transmembrane pressure and feed flow move across the full range the device, the membrane and the product quality allow, inside a single run. Because the parameters move together rather than one at a time, what comes out is a design space that takes interactions into account.

That is an intensified design of experiments, and it is the default way Novasign Studio operates in TFF modelling. Training data is collected in hours instead of over several days, and the resulting model has already seen the conditions a production run can reach.

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

In the words you already use

The bioprocess side: parameters, components and buffers

Every column is defined with its unit, bar, mL/min or min, 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

With a trained model you can simulate runs that were not performed. The digital twin takes the trained TFF model, adds a simulation configuration, and returns time-resolved simulations of everything the model was trained on: the process parameters, the product and excipient concentrations, and the quality attributes, whether that is the time-resolved prediction of relevant parameters in a C-D-C process or the time the entire process takes, so you can reduce time in the lab.

What you ask the digital twin for

Which question you put to the twin depends on the phase of the lifecycle the process is in.

Configuring a simulation

The common process setups are preconfigured, so a standard concentration and buffer exchange is a matter of picking one, while a sequence that was not executed in the lab can be built up from scratch, phase by phase.

Generating the in-silico runs

The same simulation can be read at the end of the process. The process optimizer shows the time to the final concentration across transmembrane pressure and feed flow, one diafiltration concentration at a time, marks the shortest time inside that slice and plots how that shortest time changes across every diafiltration concentration. Once the protein is concentrated far enough for a gel layer to form on the membrane, the flux drops, so the shortest time, 161 minutes, sits at a diafiltration concentration of 80 g/L and not at the top of the range. Which condition you run is yours.