Real‐time online monitoring of insect cell proliferation and baculovirus infection using digital differential holographic microscopy and machine learning

Real‐time, detailed online information on cell cultures is essential for understanding modern biopharmaceutical production processes. The determination of key parameters, such as cell density and viability, is usually based on the offline sampling of bioreactors. Gathering offline samples is invasive, has a low time resolution, and risks altering or contaminating the production process. In contrast, measuring process parameters online provides more safety for the process, has a high time resolution, and thus can aid in timely process control actions. We used online double differential digital holographic microscopy (D3HM) and machine learning to perform non‐invasive online cell concentration and viability monitoring of insect cell cultures in bioreactors. The performance of D3HM and the machine learning model was tested for a selected variety of baculovirus constructs, products, and multiplicities of infection (MOI). The results show that with online holographic microscopy insect cell proliferation and baculovirus infection can be monitored effectively in real time with high resolution for a broad range of process parameters and baculovirus constructs. The high‐resolution data generated by D3HM showed the exact moment of peak cell densities and temporary events caused by feeding. Furthermore, D3HM allowed us to obtain information on the state of the cell culture at the individual cell level. Combining this detailed, real‐time information about cell cultures with methodical machine learning models can increase process understanding, aid in decision‐making, and allow for timely process control actions during bioreactor production of recombinant proteins.


| INTRODUCTION
The baculovirus-insect cell expression vector system (BEVS) is used for the production of recombinant proteins and is one of the most used eukaryotic expression platforms to produce virus-like particles (VLPs). [1][2][3][4] With this expression platform, VLP quantities comparable to those achieved in yeast can be produced. In addition, with BEVS it is possible to perform post-translational modifications (PTMs), such as glycosylation, comparable (but not identical) to that of mammalian cells. [5][6][7] The baculovirus vector constructs required for the expression of the desired proteins (e.g., VLPs, viral glycoproteins, gene therapy vectors, enzymes, and biologicals) can be generated in a matter of weeks. The ability to quickly express the desired protein and scale up the production makes the BEVS an excellent platform to respond to emerging virus threats, as witnessed recently during the global COVID-19 pandemic and earlier during outbreaks of the Zika virus. [8][9][10] When scaling up the BEVS in bioreactors, it is important to determine three key parameters that interact with each other; the cell den- On the other hand, performing the harvest too early leads to reduced yields as well. Harvesting at the right TOH is therefore critical for obtaining optimal yields of a high-quality product. Being able to precisely determine the values for CCI, MOI, and TOH during process development is essential for obtaining optimal volumetric productivity and product quality in the final production process.
Key process parameters describing the state of the culture, like viable cell density, viability, and infection state of the culture, are usually measured only once or twice per day. As a result, the time resolution (once every 12-24 h) is low and important information to determine for example cell growth or infection stage may be obtained too late or even completely missed. This leads to suboptimal control and potential failure of production runs. 11 In addition, insufficient information is obtained for a proper mechanistic understanding of the system. For example, when developing mathematical models based on scattered, offline sample data, the low time resolution of the data will impact the quality of these models. Finally, such manual sampling also introduces the risk of contaminating the culture and is prone to operator variance leading to less reliable datasets. Thus, online and realtime measurement of key parameters like viable cell density, viability, and infection state of the culture appears to be important for a proper understanding of the BEVS and timely control of the production process.
Currently, several physical probes are available to allow online measurements of biomass, such as dielectric spectroscopy and light scattering. Dielectric spectroscopy measures viable biomass based on biomass capacitance, whereas light scattering methods measure total biomass. [12][13][14] These methods are well suited for measuring cell biomass, however, they cannot directly measure the viability or infection   stage of a single cell. Image-based cell culture monitoring is a more   direct approach, where cells can be visualized individually to extract both the cell density and information on the state of the cells. An example of such an online imaging tool is online double digital differential holographic microscopy (D3HM). 15,16 With this technique, the cell density is measured as well as a large number of optical parameters among which cell diameter and circularity, as well as quantitative parameters associated with the light phase and light intensity of each cell. These parameters can be related to the physiological states of the cells using machine learning models and specific training data sets.
There is currently only one study that demonstrated the ability of online double digital differential holographic microscopy to monitor the BEVS in bioreactors. 17 In that study by Pais et al. (2020), two bioreactor batch runs of Sf9 cells in SF900 II medium were performed where the insect cell concentration was measured online and the viability and AAV production titer was predicted. One growth batch without baculovirus infection and a single AAV production run using a two-baculovirus infection strategy at an MOI of 0.05 TCID 50 /cell were monitored.
The current work aims to obtain more insight into the performance of online digital differential holographic microscopy to monitor baculovirus-infected cell cultures in bioreactors. Several baculovirus constructs were included, producing various recombinant proteins, and under varying process conditions, using the ExpiSf9 cell line as a model. 18 First, training data sets were generated to develop machine learning models. Then the performance of the D3HM tool and machine learning model was evaluated during online monitoring of bioreactor runs. Results showed that after training the machine algorithm with a training data set, the cell density, cell viability, cell diameter, and the fraction of infected cells could be accurately determined for a variety of bioreactor processes. Furthermore, the continuous online measurements allowed for the construction of high-resolution time-series profiles of these parameters. These high-resolution timeseries profiles gave more insight into the state of the cell culture inside the bioreactor. Infected cells could be detected earlier compared to offline methods and the effect of process interventions such as feeding became distinguishable. Improved training data sets can further increase the accuracy of the online prediction, allowing for more advanced process control strategies and increased process understanding of recombinant protein production processes.

| Analytical methods
Cells were counted by trypan blue exclusion using a TC20 Automated Cell Counter (Bio-Rad) or by manual counting using DHC-F01 cell counting chambers (INCYTO). Online measurements of viable and total cell density, viability, and infected cells were performed by differential digital holographic microscopy (D3HM) using iLine F holographic microscopes (OVIZIO). Data acquisition and quantitative data analysis were performed using OsOne software (OVIZIO). Infected cells were visualized by expression of GFP or mCherry detected by a C6 Plus Flow Cytometer (BD Accuri). SARS-CoV-2 S1 subunits were quantified using a SARS-COV-2 Spike S1 Protein ELISA kit (AssayGenie).

| Verification run of baculovirus infection at low MOI
After the machine learning model was trained using this training data set, the model was tested using new verification data sets. ExpiSf9 cells were infected with AcBac-2FL at an MOI of 0.01 at a CCI of 1.1 Â 10 6 cells/ml. The bioreactor culture was monitored by the iLine   (Figures 4b and 5b). With the iLine F, infected cells could be  (Figures 4b and 5b).

| Verification of baculovirus-infected insect cell culture with butyrate addition
The high resolution of the online measurements ideally would allow for spotting small process deviations that might be invisible with daily manual sampling. The sensitivity of the online microscope to spot temporary effects of butyrate addition was, therefore, investigated.
Butyrate was added to improve baculovirus gene expression. 18,33 ExpiSf9 cells were grown until they reached a density of 4 Â 10 6 cells/mL at 72 hpi, at which point butyrate was added at a final concentration of 2 mM. Immediately afterward, cells were infected with an MOI of 0.01 with a baculovirus construct containing GFP behind the polyhedrin promoter. Online measurements of viable cell density and total cell density, and online predictions of cell viability and infection state were compared to offline measurements. Offline measurements of cell densities and cell viability showed a similar trend to online measurements although offline-measured cell densities were lower (Figure 6a). The online measurements showed a clear peak in the viable cell density and a sharp switch from growth to death phase, whereas the manual counts showed a plateau in the viable cells followed by a decrease in viable cells. This demonstrates once more that due to the high resolution of online measurements, the exact moment when the cells stop growing can be determined. Together with the virus, butyrate was added to the cell culture at 72 hpi to enhance GFP production. A temporary slowdown of cell growth was visible between 72 and 85 h in the high-resolution data of the online viable cell and total cell density measurement (Figure 6a). This is in agreement with the fact that butyrate can cause a cell-cycle arrest. 34 Note that this arrest is not visible in the low-resolution offline data, showing the importance of having a high-resolution online measurement.
Online monitoring of the infected percentage of cells showed a small peak around 80 hpi after which the fraction of infected cells stays slightly elevated at about 5%, followed by a sharp increase around 120 hpi (Figure 6b). Such a small peak was not expected when infecting with an MOI of 0.01. It could be caused by the diameter increase of the cells directly after butyrate addition since the prediction algorithm mainly makes predictions based on the size of the cell. This indicates the potential limitation of the current model of having a strong dependency on certain parameters when using a machine learning modeling approach. Generating additional training sets specifically aimed at breaking this correlation, for instance by changing culture osmolality, could potentially improve the accuracy of the infection prediction algorithm. The small peak around 80 hpi was not clearly visible with the offline measurements GFP fluorescence at this point is unexpected as the polyhedrin promoter becomes active at 20 hpi. 26 3.5 | Online monitoring of SARS-CoV-2 spike protein production using a low MOI To follow the infection process for non-reporter-secreted proteins, the holographic microscope was used to monitor a SARS-CoV-2 spike protein production process (Figure 7). In contrast to reporter proteins such as GFP and mCherry, which accumulate inside the cells, 35 the S1 subunit of the SARS-CoV-2 spike protein is excreted into the medium. 9 To produce SARS-CoV-2 S1, insect cells were infected with an MOI of 0.01 at 70 hpi. S1 was detected for the first time in the culture fluid at 140 hpi and S1 concentration continued to increase until harvest at 190 hpi. Using the holographic microscope and the previously calibrated infected cell detection algorithm, the moment of 100% infected cells was detected at 155 hpi (Figure 7b). Peak viable cell density was reached a little bit earlier, at 150 hpi (Figure 7a). The The ideal TOH depends on product concentration and product quality. Since both parameters are relatively difficult to measure in real-time, the online measured viability, diameter, and percentage of infected cells can be used as indicators for determining the best moment of harvest.
Low viability can affect protein quality, as proteases are released from dead cells. 36,37 Harvesting too early, however, can negatively affect product yield. Sander et al. 38

| CONCLUSION
Online double differential digital holographic microscopy (D3HM) and machine learning modeling was used for real-time monitoring of insect cell proliferation and baculovirus infection in bioreactors. Training and verification data sets were generated using a purpose-built F I G U R E 7 Time-course profiles of a SARS-CoV-2 S1 production run. ExpiSf9 cells were infected with a baculovirus expressing SARS-CoV-2 S1 subunits at an MOI of 0.01. The gene for expression of SARS-CoV-2 S1 was located behind the very late polyhedrin promoter.