4.1 Discussion
The ability to determine, in real-time, cells’ health, therapeutic potential, or developmental state is paramount to developing effective feedback control for advanced therapy production. From a manufacturing standpoint, detecting cells in their proliferative, confluent, or differentiated states is critical to understanding the processes governing their production.30, 31 As a first step towards real-time monitoring of primary cell cultures, DSP was applied to MC3T3 cells for rapid ESI-MS, followed by PCA of the collected spectra to determine if features in the raw spectra corresponded to differences in the secretome, and therefore the cell state. This study was used to assess whether the known spatial heterogeneities in biomolecular content affect a PAT output. In this sense, the negative results from bulk samples (Fig. 3A and 4A) are just as important as the positive results from localized sampling (Fig. 3B and 4B). Depending on where a sample is taken in a bioreactor, the resulting PAT output is altered. An effective PAT should then capture these heterogeneities in the culture environment.
PCA of input spectra from bulk samples, in Figures 3A and 4A, shows no clustering, which indicates that the input spectra are not statistically distinguishable despite the fact that cells had undergone proliferation to confluence and then differentiated between time points 1 and 6: the cells’ secretome was expected to have changed. The data in Figs. 3A and 4A indicate that ESI-MS is unable to distinguish between cell types using sampling from the bulk away from the cells. Only DSP analysis of locally garnered samples produced clustering in ESI-MS output. This suggests that PAT output depends on spatial context, and that these spatial variations are not negligible and should be accounted for when designing sensor systems.
Unlike bulk analysis, PCA using spectra from localized sampling as an input revealed clusters which are associated with cells in their undifferentiated and differentiated states in figures 3B and 4B. An important feature within figures 3B and 4B is the great deal of variation amongst the same cell group along the horizontal axes (principal components 1 and 2). One explanation for this variation is that every sample taken from the cell culture was taken from a spatially disparate location. The variation observed along the horizontal axes in these cluster plots may reveal subtle changes in the secretory profile that are revealed when sampling at different locations in the well. Of course, there are a number of other factors that could contribute to this variation within the same group, including the fact that the cells in the undifferentiated and differentiated cultures experienced different culture conditions, and that the cells were expanding during these experiments.
In the PCA scatter plots in figure 3B and 4B, separation between cells in their undifferentiated and differentiated states is along the third, orthogonal, principal component 3 (vertical axis). The separation along a single axis suggests that even with variation within subgroups, the dimensionality can be reduced to a single component that captures a significant amount of variation in the input spectra which can be used to predict cells in their undifferentiated or differentiated state. The additives used to induce differentiation in the MC3T3 cells are expected to have been removed due to their small size (MW <300 Da). Therefore, it is unlikely that the presence of the additives contributed to the separation observed between the different cultures shown in figure 3B.19, 20 In order to remove the potential variation due to different culture conditions, the differentiated cell group was analyzed in an early time point vs a late time point to again compare cells in their undifferentiated and differentiated states. This comparison mitigates the effect that different media compositions could have on the analysis. In fact, each sample was taken from the same well, so this experimental design removes any changes in cell culture conditions that could contribute to variation in the data. The resulting PCA cluster plot with input spectra from early and late time points from the differentiated cell group (Fig. 4B) separates well along the vertical axis (principal component 3), corroborating the assumption that the clustering observed in figure 3B was most likely due to cell differentiation and not culture conditions. Furthermore, this analysis emphasizes a central finding of this work which is that cell bioreactors are highly heterogeneous in nature, and PAT measurements are dependent on where a sample is taken.
Along principal component 1 in figure 4B the undifferentiated cell group (red dots) separates into two subgroups, corresponding to time points 1 and 2, which are demarcated with ovals highlighting different samples. The cells were seeded such that they were proliferative early in the culture process and later reached confluence. Since the subgroup separation in figure 4B correlated with proliferation, it was expected that similar variation could be observed in the undifferentiated culture. Therefore, PCA was applied to time points 1 and 2 versus time points 5 and 6 for the undifferentiated cell group (Fig. S3). In this case, bulk sampling once again resulted in no separation, and for local sampling the separation between the two groups is along principal component 1. This suggests that the variation along principal component 1 may be due to a change in the cell secretome during expansion, which has been shown to correlate with preosteoblasts in a state of proliferation (early) or confluence (late).32 In other words, while principal component 3 correlates well with cell differentiation, principal component 1 correlates with cell proliferation. Future studies will be designed to create spatial maps/scans of the cell culture which will help to investigate cell-to-cell heterogeneity throughout the culture as well as suspected sources of heterogeneities, such as edge effects in 6-well plates.
In addition to the experiments to probe spatial effects in more detail, it would be of fundamental interest to identify target biomolecules which are known to correlate with cells in specific states. One drawback of the experimental design used here is the MS used for real-time analysis was not capable of MS/MS (tandem mass spectrometry) for potential chemical IDs. Coupling DSP with an MS system capable of data-dependent-analysis (DDA) for feature identification will significantly increase the potential for CQA identification. However, the PCA output did enable some exploratory work towards chemical identification. Offline HPLC-MS was performed on aliquots of conditioned media which were gathered during media change and frozen at -40 ͦC until the end of the study. HPLC was carried out on media from both cell types at time points 1, 2, 5, and 6 to identify candidate differentiation biomarkers and to quantify differences in biomolecules between timepoints. The candidate biomolecules identified with HPLC were then manually compared to m/z values with the highest contribution to variance in the PCA data that are hypothesized to correlate with cell state (i.e., differentiated vs undifferentiated). Since the mass spectrometer used for DSP analysis was not equipped with tandem mass spectrometry, fragmentation patterns could not be matched to the HPLC-MS data. Therefore, the m/z values from the PCA loading data were matched to potential chemicals from HPLC-MS based on accurate mass alone, resulting in tentative IDs only. Multiple candidate molecules were identified using this approach, which suggests that DSP may not only generate cell culture “fingerprint” spectra, but also help to identify which detected biomolecules correlate with cell state (supplementary information, table S1). For more complete identification of biomolecules, DSP can be used with inline MSn for fragmentation data to generate candidate IDs before HPLC is carried out. Further studies will elucidate how PCA loading data correlates with quantitative HPLC data for independent identification of CQA biomarkers using DSP based analysis.