Bioprocess Modeling Regulators Can Follow: Webinar Recap

On 19 August, Novasign hosted the second session of its three-part webinar series, AI for Bioprocess Optimization That Regulators Can Trust: From Raw Data to Filing-Ready Outputs. CEO Mark Duerkop presented, with Johannes Pilz moderating questions from the audience.

The framing question was not whether process models work. It was whether a model can be defended: to a colleague, to an auditor, to an agency. In a regulated industry, a prediction that cannot be traced back to its raw data is difficult to act on, no matter how accurate it is.

It is the workflow, not the model

The central argument of the session was that modeling in bioprocessing has historically failed on integration rather than on mathematics. Experiments are planned in one tool, executed on hardware that speaks a different language, and the resulting online and offline data end up in separate places. Each handover loses knowledge and introduces the possibility of error.

Novasign Studio is built around closing those gaps: planning experiments, pushing them to the hardware, retrieving data automatically, and returning a decision, all within one environment. The practical consequence is that every step leaves a record. Studio tracks who imported which dataset, what preprocessing was applied to it, and which model structure was selected on what criteria. That history can be exported as a report, and the model structure itself remains open to inspection rather than sitting behind a vendor black box.

Mark contrasted this with the multi-linear regression models routinely used in filings today. Those describe a response surface at a single endpoint. A time-resolved hybrid model describes the whole trajectory, which makes the result less sensitive to a measurement error at the final timepoint.

Reducing experimental effort, and proving you can

Across projects, the iterative modeling workflow has reduced experimental effort by roughly 35 to 70 percent compared with classical DoE, with data quality being the main determinant. Mark was explicit about the caveat. If you do not run the experiments, you have to demonstrate that the model predicts that region correctly, and the one or two validation runs needed to show that eat into the saving.

In a viral vector characterization project with Bespark*bio, the workflow reduced runs from 22 to 8, cutting three months from the characterization timeline and around €40,000 in raw materials. The underlying approach to reducing experimental burden is described in Biotechnology Journal (2020) and Frontiers in Bioengineering and Biotechnology (2021).

Building quality into biosimilar development

The clearest illustration of the regulatory theme was a biosimilar case run on the NIST CHO cell line. That line was chosen deliberately, so that the data could be published.

The hypothesis was a familiar one: glucose concentration drives glycosylation. Experiments were run at 1, 3 and 5 g/L with enzymatic glucose control, and glycoforms were measured by capillary electrophoresis coupled to mass spectrometry. Those measurements fed a hybrid model, which in turn fed a digital twin.

What the twin produces is an inverse answer. Instead of asking what a given set of conditions yields, the process engineer defines the CQA corridor the biosimilar has to hit, meaning the acceptable range for each glycoform. The system then reduces the experimental space to the conditions capable of reaching it. In this case, high glucose combined with high temperature dropped out entirely, while low glucose with higher temperature was favorable. Related work on real-time monitoring of glucose depletion is published in Biotechnology Letters (2021).

Filtration, scale-up and continuous processing

Three further showcases rounded out the session.

For ultrafiltration and diafiltration, the workflow is embedded directly into Repligen hardware. Inline concentration measurement allows a model to be calibrated from a single scouting run, which is enough to compare membranes against a target concentration within a couple of days. The modeling basis for this is published for both crossflow ultrafiltration and single-pass TFF.

For scale-up, Novasign combines its cellular models with CFD simulations from SimVantage, so that the microenvironment a cell experiences in a large vessel is accounted for before the first large-scale run. Related work is published in Processes (2021) and Bioengineering (2024).

Finally, the ECOnti consortium demonstrated a fully integrated continuous protein A process running for 30 days at 3 g/L/day, in which the digital twin predicts chromatography column loading and triggers switching through the automation layer.

Next session

The third and final session of the series takes place on 16 September. Registration details are on our LinkedIn page and here on the website.

If you would like to work through one of these cases against your own data, get in touch. We can set up a Studio instance for a two-week trial, or scope a short feasibility study.