One of many good presentations at this year’s MKS Umetrics users meeting in Chicago was given by Roland Bienert, Senior Scientist for process analytical technology (PAT ) and Sensors & Chemometrics at Sartorius Stedim Biotech. Bienert talked about various new applications of near infrared spectroscopy (NIR) for bioprocessing.
After the meeting, he discussed how he and his colleagues are using NIR and multivariate data analysis (MVDA) to enable improved bioprocess monitoring. Here is some of what he had to say.
PhM: What’s new in terms of applying NIR for bioprocess monitoring? Is anything being done now that wasn’t accepted practice a few years ago?
RB: In the early stage of the PAT initiative, people strove to measure single analytes online as accurately as they could using offline lab methods. That was, of course, hard to achieve, especially for low-concentration analytes such as glutamate or ammonia.
The new perspective is that NIR spectroscopy can do much more for process understanding. For instance, it is capable of measuring a metabolism sum-parameter, which does not rely on highly accurate single parameters but on the overall changes in spectra due to metabolite accumulation. And isn’t that what we are most interested in? This parameter has not been accessible via offline sampling.
In order to take the true Quality by Design (QbD) approach and learn about our processes, we do not necessarily need a number of isolated analyte concentrations. Instead, we need to combine the data we have acquired. This approach allows us to establish new tools like batch trajectories and endpoint determination for real-time release, which provide insights into the most important trends in an easy-to-interpret manner.
PhM: How about fermentation specifically—what improvements have been made in the application of NIR?
RB: From a technical point of view, it is a big advantage to be able to use a free beam spectrometer with a standard Ingold port adaptor, as we can today, instead of having to rely on fiber optics. The large free aperture, combined with a very high light yield and no moving parts within the spectrometer, results in excellent spectra quality throughout the whole fermentation. Even in extremely rough environments like large-scale bacterial fermentations, a free-beam spectrometer can automatically filter large fluctuations arising from air bubbles, for instance. This is possible since free beam spectrometers use diode array detectors instead of scanning techniques, which results in very fast measurements.
PhM: Do you see free-beam NIR as a substitute for fiber optic bioprocess monitoring, or do you see them each having strengths and weaknesses?
RB: With our new Ingold port adaptor, free beam spectrometers are surely a substitute for fiber optics NIR systems. Visiting the Achema 2012 fair recently, we observed a lot of progress in this field and a clear commitment to this kind of spectrometer. Comparing our system to fiber optic technologies such as MIR—as we have done recently in a project in cooperation with TCI Hannover—they are adding both their benefits to a complete package. The MIR system can be used to monitor small molecules, which can be used for feed control, whereas NIR in our solution is much more powerful in monitoring cell parameters and process trends.
PhM: Combined with MVDA, you can now use NIR to gain a much better sense of, for example, the activity of nutrients and metabolites during fermentation. What specific process parameters are more easily monitored today than even a few years ago?
RB: First of all, NIR spectra show a large number of variations which can be used to predict almost all the important parameters. The major challenge for vendors and manufacturers interested in NIR spectroscopy is to figure out which parameters can be directly measured by this technique and which are just based on correlations to some uninterpretable spectral changes. Therefore, Sartorius classifies between control parameters, which show directly changes within the spectra and are able to be discriminated from others.
Monitoring parameters, on the other hand, are in this context based on correlations and should be interpreted carefully. For parameter classification, Sartorius uses spiking experiments that break correlations between the analyte of interest and all the others.
With our approach, we can develop very strong predictions of important cell parameters, especially total cell count and viability. Those are both important during cell growth and production phase. Together with a reliable titer prediction, we have all important variables required to map the process during product formation.
PhM: What are some best practices for MVDA for fermentation processes? Or, conversely, what are some missed opportunities that you often see from manufacturers?
RB: Besides predicting the concentration of single analytes, MVDA can provide much more information from online data. Concerning NIR, for instance, we offer several tools for qualitative analysis of the process.
These tools are directly able to answer the most important questions, such as: Are the starting conditions consistent with earlier campaigns? Is the process alright? When is the perfect time for harvesting?
These are exactly the same questions that offline methods were supposed to answer several years ago, but they have not been suitable.
These questions can be perfectly addressed by qualitative process control tools. This starts with media classification just before inoculation in a statistically evaluated good-bad approach. The whole batch can be monitored as trajectories, which describe the process evolution and help to identify significant deviations in real time and thereby allow a guided sampling.
Finally, an endpoint determination is possible in real time, which results in a perfect product tailored for optimal downstream performance. Of course, the information from NIR-data can be supported by other sensors, which results in more robust models.
From my perspective, a misstep we are still facing is the sole focus on accurate analyte concentration as a universal remedy for process monitoring. As far as closed loop process control is concerned that is crucial, no doubt, but in terms of process optimization, a qualitative overview of the process is much more beneficial than an ammonia profile. The strong demand of monitoring analyte concentrations inveigled some vendors and users to over-interpret correlations between analyte concentrations and spectral information, which led to unstable calibration models.
Model robustness is another issue that is often concealed. A model based on one or two cultivation runs can not be used in a production environment, but shows nice error bars if an internal prediction set or cross validation is used. But to take batch-to-batch variations into account, a validation test set necessarily needs to be “unseen.” That means the prediction error must be calculated from one (or more) cultivation runs, which are not part of the calibration set. By doing so, all errors increase to a certain extent but reflect real life. However, this is the recommended way of method validation for any NIR application.
PhM: In your view, are manufacturers getting more sophisticated in, say, titer prediction or end-point determination of fermentation processes? What’s allowing them to do so?
RB: Both are highly requested and a lot of efforts have been made to set up robust models especially for titer. NIR spectroscopy is capable of high-titer prediction as well as highly suitable for end-point determination. For low-titer concentration and other hard-to-get parameters, the important step is to overcome the univariate sensor by sensor evaluation, but use a multivariate model with several sensors contributing valuable information regarding the desired parameter. The sensor information is usually available in real time. The crucial step here is data handling and automated combination for multivariate evaluation. Sartorius, in close cooperation with Umetrics, offers solutions for next level models using multiple sensors. It is not enough to offer either sensors or software. It is important to have the expertise from spectra acquisition to generation and interpretation of trajectory in order to offer a one-stop solution. This is the road Sartorius will take.
PhM: What is the focus of Sartorius Stedim’s PAT offerings?
RB: For years, Sartorius has offered a large variety of online sensors for both stainless steel and single-use bioreactors, including NIR but also microwave resonance, off-gas sensors, glucose-lactate analysers, refractometry and other technologies.
Despite this broad portfolio, for Sartorius, PAT does not just mean implementing a sensor. Our PAT unit is involved in numerous in-house projects that are part of an internal global PAT program initiated many years ago. We use PAT to optimize our own manufacturing processes and also help our customers implement suitable PAT solutions.
As a total solution provider, Sartorius offers the whole range of PAT tools. This includes PAT consulting starting with risk analysis and identifying critical process parameters. Thus, we implemented Umetric DoE tool MODDE in our own SCADA software, the multi-fermentation control software (MFCS), in order to run designed experiments in several bioreactors in parallel. We also support our customers if requested in terms of MVDA and process integration. Here we accompany or perform the generation of robust calibration models based on one or more sensors. In terms of MVDA, we cooperate closely with Umetrics leading in multivariate software tools.
PhM: Can you share a bit about your group’s work with TCI Hannover?
RB: We understand that not all players in the pharmaceutical industry want to evaluate a new sensor technology from scratch. Therefore, we are happy to have a strong partner in academia like the group of Prof. Thomas Scheper. This cooperation allows us to provide our customers an overview of the new NIR adaptor. The results of the Ingold port evaluation will be published in peer-reviewed journals. With this first step in sensor evaluation, we can now focus on process specific topics at the customer site.
