Advancing 2D fluorescence online monitoring in microtiter plates by separating scattered light and fluorescence measurement, using a tunable emission monochromator

Online fluorescence monitoring has become a key technology in modern bioprocess development, as it provides in‐depth process knowledge at comparably low costs. In particular, the technology is widely established for high‐throughput microbioreactor cultivation systems, due to its noninvasive character. For microtiter plates, previously also multi‐wavelength 2D fluorescence monitoring was developed. To overcome an observed limitation of fluorescence sensitivity, this study presents a modified spectroscopic setup, including a tunable emission monochromator. The new optical component enables the separation of the scattered and fluorescent light measurements, which allows for the adjustment of integration times of the charge‐coupled device detector. The resulting increased fluorescence sensitivity positively affected the performance of principal component analysis for spectral data of Escherichia coli batch cultivation experiments with varying sorbitol concentration supplementation. In direct comparison with spectral data recorded at short integration times, more biologically consistent signal dynamics were calculated. Furthermore, during partial least square regression for E. coli cultivation experiments with varying glucose concentrations, improved modeling performance was observed. Especially, for the growth‐uncoupled acetate concentration, a considerable improvement of the root‐mean‐square error from 0.25 to 0.17 g/L was achieved. In conclusion, the modified setup represents another important step in advancing 2D fluorescence monitoring in microtiter plates.


| INTRODUCTION
More than 50 years ago, the first use of fluorescence measurements for quantitative spectroscopic determination of intracellular fluorophores (Chance, 1952) and the detection of metabolic changes in microorganisms (Harrison & Chance, 1970) was reported. Ever since then, fluorescence online monitoring has widely been established as an effective tool for bioprocess development. Due to the growing interest in generating process insights and the call to implement process analytical technologies (FDA, 2004), the application of appropriate machine learning evaluation algorithms for data exploitation is regularly reviewed (Claßen et al., 2017;Faassen & Hitzmann, 2015;Helleckes et al., 2023;Lourenço et al., 2012;Mowbray et al., 2021;Rathore et al., 2011;Reyes et al., 2022).
The basic components of spectrofluorometers include a light source, two elements for wavelength selection of the excitation and the emission wavelength, and a photosensitive element for measurement of the light intensity (Lakowicz, 2006). Various designs for online 2D fluorescence monitoring have been presented and underlined the strong impact of the composition on its usability in bioprocess development. The most well-known setup for stirred tank reactors, the BioView (Delta Optics), consists of a xenon flash lamp and two filter wheels for excitation and emission, with each containing 15 single-wavelength filters. A photodiode is used to record the intensities for the 120 filter combinations, resulting in 2D spectra with an excitation wavelength (λ ex ) range between 290 and 550 nm and an emission wavelength (λ em ) range between 310 and 590 nm (Marose et al., 1998). In combination with methods of multivariate data analysis (MVDA), the spectroscopic setup has successfully been applied for the monitoring and control of various microbial and eukaryotic cultivation processes for nearly 25 years (Bayer et al., 2020;Bonk et al., 2011;Haack et al., 2007;Johansson & Liden, 2006;Ödman et al., 2010;Skibsted et al., 2001). In a more recent 2D fluorescence monitoring system for stirred tank reactors, the continuous xenon lamp was replaced by three separate highpower LEDs at λ ex of 280, 365, and 455 nm. The emitted light is dissected by an optical grating before continuous spectra are measured for λ em between 190 and 980 nm by a back-thinned charge-coupled device (CCD) camera. The large amount of spectral data recorded for various batch and fed-batch cultivation processes of Chinese hamster ovary cells was again evaluated by methods of MVDA. While the principal component analysis (PCA) was used for identifying deviations from "golden batch" trajectories, partial least square (PLS) regression models were generated for predicting process variables, such as total and viable cell count Graf et al., 2019).
Also in microbioreactors, fluorescence monitoring is widely applied, as its noninvasive character allows the determination of critical process parameters without interrupting the experiments for assays, that potentially consume the culture broth (Schäpper et al., 2009). Starting from repetitive measurements in microtiter plate (MTP) readers (Zimmermann et al., 2004), the BioLector system has been established for monitoring of up to 96 parallel cultivations (Samorski et al., 2005). Different spectroscopic setups including both optical filters (Huber et al., 2009;Kensy et al., 2009) and gratingbased monochromators (Ladner, Held, et al., 2016;Wandrey et al., 2016) have been presented for measuring the fluorescence through the transparent bottom of a MTP. However, for all the presented systems, photodiodes, or photomultipliers were implemented, thus, limiting the number of measured wavelength combinations due to time constraints. Only by combining a grating-based emission monochromator with a CCD camera, comprehensive online 2D fluorescence monitoring of 48 parallel cultures with up to 40,000 single intensities per spectrum was enabled . This groundbreaking advancement of the measurement concept was applied in multiple studies to investigate the robustness of PLS modeling for screening experiments  and the possibility to reduce the required sampling for model calibration (Babor et al., 2023;Berg et al., 2022;Paquet-Durand et al., 2017).
Due to its physical nature, the scattered light is usually considerably stronger than the emitted fluorescent light (Samorski et al., 2005). While for single wavelength measurement the settings can be adjusted accordingly, for the implementation of a CCD array, the two signals of scattered and fluorescent light are measured simultaneously. To mitigate the risk of oversaturation of the CCD array, the setup of  used short integration times (t int ). This provided a good resolution for the scattered light, but limited the signal intensities for the fluorescence area. To overcome this limitation, this study presents a novel, more flexible spectroscopic setup. The introduced tunable emission monochromator allows for a precise manipulation of the emission spectra range. Thereby, the inherently stronger scattered light signal can be excluded for the fluorescence measurement and considerably longer t int can be realized, without saturating the CCD detector. After verifying the envisioned functionalities, the capabilities of the new monitoring device are demonstrated during two Escherichia coli cultivation experiments with varying glucose or sorbitol supplementation. The recorded spectral data is evaluated by applying MVDA methods of PCA and PLS. A direct comparison of the models generated for the data based on long t int with models based on short t int is performed to evaluate the qualitative and quantitative improvements due to the increased fluorescence sensitivity. In a final discussion, the setup is critically discussed with respect to further increase the robustness of the method.

| Microorganisms and media
All experiments were conducted with an E. coli BL21(DE3) strain containing a pET22b(+) expression vector encoding for a Bacillus subtilis lipase A (BSLA) (van Pouderoyen et al., 2001) and an ampicillin resistance. Throughout this work, the strain will be abbreviated as E. coli BL21 BSLA. The strain was stored as cryo stocks at −80°C.
All cultivations were conducted in a modified Wilms-MOPS medium (Scheidle et al., 2007;Wilms et al., 2001). The base solution of the medium consisted of 7 g/L (NH 4 ) 2 SO 4 , 3 g/L K 2 HPO 4 , 2 g/L Na 2 SO 4 , 41.85 g/L (N-Morpholino)-propane sulfonic acid (200 mM MOPS), 0.5 g/L MgSO 4 ·7H 2 O, 0.01 g/L thiamine hydrochloride, 0.1 g/L ampicillin. The pH value was adjusted to 7.5 using 2 M NaOH. Sterile trace element solution was added with 1 mL/L. The trace element solution consisted of 1.98 g/L CaCl 2 ·2H 2 O, 0.48 g/L CuSO 4 ·5H 2 O, 0.54 g/L CoCl 2 ·6H 2 O, 41.76 g/L FeCl 3 ·6H 2 O, 0.3 g/L MnSO 4 ·H 2 O, 33.4 g/L Na 2 EDTA (Titriplex III), 0.54 g/L ZnSO 4 ·7H 2 O. Sterile carbon sources were added as described for each experiment. All medium components were sterilized separately by autoclaving or filtration. GmbH) (V L = 0.8 mL, d 0 = 3 mm, n = 1000 rpm). The MTP was placed on an orbital shaker (LS-X; Kuhner shaker GmbH), which was located in a temperature regulated incubation chamber (T = 37°C). For the sampled cultures, a second, non-monitored MTP was placed on the same shaker tray. Both MTPs were sealed with gas-permeable membranes (AeraSeal film; A9224-50EA; Sigma-Aldrich Chemie GmbH) and covered by separate aeration hoods made of acrylic glass. The aeration hoods were supplied with humidified air at an aeration rate of 1 vvm. For the experiments with additional monitoring of the respiration activity, a separate, identically filled MTP was incubated in a µRAMOS device . Here, a polyolefin-based, gas-permeable membrane (900371-T; HJ-Bioanalytik GmbH) was used as sterile barrier, which provided a comparable maximum oxygen transfer capacity (Sieben et al., 2016). Aeration with humidified air was provided at a rate of 1 vvm.

| Spectroscopic setup
The spectroscopic monitoring system presented in this study is comparable to the setup described by  and is shown in Figure 1a The spectroscopic online signals of the cultivations were recorded using a LabVIEW-based software (National Instruments Germany GmbH). The framework of the software was initially developed by ZumoLabs GmbH and was first used by Wandrey et al. (2016). To enable the measurement of the 2D fluorescence spectra, the available software was adapted in-house.

| Measurement concept and implementation
In an approach to overcome the problem of low fluorescence sensitivity of the previously presented prototype , the newly developed monitoring device of this study employs a tunable emission monochromator (Figure 1a, dashed box), which allows a precise manipulation of the grating angle (α I , α II ) and a variable wavelength allocation onto the CCD camera ( Figure 1b

| Offline sampling and analytics
For the experiments including offline sampling, samples were taken from a MTP incubated in parallel. Therefore, the respective aeration hood provided a common headspace of around 160 mL above the MTP and sampling ports above each well. The ports were covered with airtight sealing tape during cultivation, which was temporarily removed during sampling. A visualization of the setup is shown in Supporting Information: Figure 2. A micropipette was used to pierce the sterile membrane of the to-be-sampled well and to withdraw the F I G U R E 1 (a) Schematic spectroscopic setup and (b) measurement principle of the 2D fluorescence monitoring system, implementing a tunable emission monochromator. The schematic in (a) is adapted from . The light of a specific wavelength is selected by an excitation monochromator and sent to the transparent bottom of a continuously shaken MTP via a Y-shaped optical fiber. The backscattered and emitted fluorescence light is collected by the Y-shaped optical fiber and transmitted to an emission monochromator (a, dashed box). The incident light is dissected by the second monochromator and the emission spectrum is recorded by a CCD camera. (b) illustrates the novel functionality of the tunable emission monochromator. By changing the grating angle α, the wavelength allocation to the light-sensitive elements of the CCD array (area between the white triangles) is manipulated. Thereby, measurements including (b, left) and excluding (b, right) the scattered light (blue light fraction) are enabled. For the measurement excluding the scattered light, the fluorescent light sensitivity can be increased by increasing the t int of the CCD camera, without reaching the detection limit of the CCD array. CCD, charge-coupled device; MTP, microtiter plate. sample volume. Initial samples (t = 0 h) were prepared from inoculated medium, that was distributed across the MTPs. The sampling broth was analyzed for pH value (HI21; HANNA instruments Inc., electrode: InLab Solids; Mettler Toledo GmbH), OD 600 , and carbon source concentrations. Except for a temperature of 60°C used during HPLC measurement, the procedure for the offline measurements was identical to the previous studies (Berg et al., 2022. The linear correlation between OD 600 and CDW is given in Supporting Information: Figure 3.

| Spectral data merging and processing
With changing the position of the emission grating, the wavelength allocation on the sensor elements of the CCD camera changes ( Figure 1b). Therefore, right after the measurement, linear interpolation was used to transform the intensity data of each recorded emission spectra from a uniform λ em scale, so that the spectral data can be saved in a common data matrix. In this process, the resolution was reduced from 0.26 to 0.5 nm. Additionally, all intensities with an λ em above 700 nm were discarded, as no biogenic fluorescence was expected in this spectral range (Faassen & Hitzmann, 2015).
The subsequent data processing, and MVDA were performed using MATLAB 9.10.0.1710957 (R2021a) Update 4 and the opensource toolbox mdatools (Kucheryavskiy, 2019). The post-experimental spectral processing steps were comparable to those described in Berg et al. (2022Berg et al. ( , 2023. The detected spectral range was reduced to include only intensities of λ em between −20 and +270 nm relative to the λ ex . For each λ ex , the fluorescence fraction of the emission spectra (+15 to +270 nm) was smoothed using a moving average filter with a size of 25 intensities (equivalent to 12.5 nm). The scattered light was not smoothed to avoid distorting effects. To further allow fast data handling, the data set was further reduced to include only every fifth emission intensity (equivalent to 2.5 nm). Finally, to account for intensity offsets between the wells, the intensity (I) of each measurement cycle was referenced to the intensity of the first measurement cycle (I 0 ) of the respective well by subtraction (I−I 0 ).
Merging of the data measured at different t int was conducted by extracting the intensities of the respective measurements to a new data set. For the scattered light, the λ em range between −20 to +12.5 nm relative to the λ ex was used. For the fluorescence, the λ em range between +15 to +270 nm relative to the λ ex was used. In case no data was recorded at a specific λ em, the intensity was set to zero.
The spectral intensities from +10 to +15 nm relative to the λ ex were replaced by zero (Supporting Information: Figure 4). The subsequent data processing was generally identical to the procedure described above. However, to avoid a distorting effect, smoothing was conducted starting from +20 nm respective to λ ex .

| MVDA
The MVDA was conducted identically to the descriptions in Berg et al. (2022Berg et al. ( , 2023. PCA was performed using singular value decomposition (Wall et al., 2005). The data sets used for PCA and PLS consisted of spectral data from all monitored cultures in triplicates. The offline values used for PLS modeling were extracted from a linear interpolation of the sampling values. The exact composition of the calibration and prediction data sets is provided in the results section. The models were generated using the SIMPLS algorithm (de Jong, 1993). In contrast to the approach presented in Berg et al. (2022Berg et al. ( , 2023, the number of latent variables (LVs) used to generate the PLS models of this study was arbitrarily chosen to five. This facilitates the comparison and interpretation of the models, generated for different measurement F I G U R E 2 Measurement procedure for generating 2D spectra by alternating spectral measurements including and excluding scattered light. The schematic, out-of-scale 2D spectrum in (a) shows the scattered light area in orange and the fluorescent light area in dotted white. The gray area represents the spectral area not measured by the new setup. The ① blue and ③ green solid horizontal lines represent two measurements, including scattered light, at two increasing excitation wavelengths (λ ex1, λ ex2 ). The ② red and ④ yellow dashed horizontal lines represent the wavelengths covered for the measurements, excluding the scattered light. For the generation of 2D spectra, alternating measurements including and excluding scattered light are conducted, resulting in repeated movement of the emission monochromator indicated by the purple arrows. The temporal measurement sequence is shown in (b). The movement of the excitation and the emission monochromator is represented by the black and purple sections, respectively. The time for measurements including the scattered light (① blue, ③ green) is shown in orange. The time for measurement excluding the scattered light (② red, ④ yellow) are shown in dotted white. The widths of the sections are not to exact scale. settings. Model evaluation was conducted based on the calculation of the root-mean-square error (RMSE).
For the sorbitol and glucose concentration, before PLS model generation, the residual carbon source concentrations were converted to the total consumed carbon source concentrations. Thereby, the initially referenced spectra, for which all intensities were zero, were correlated to an identical concentration (i.e., 0 g/L). This considerably improved the modeling performance (data not shown).
After applying the generated PLS model to the spectral data, the respective carbon source concentrations were recalculated based on the initially supplemented concentration.

| RESULTS
3.1 | Functional commissioning of the spectroscopic monitoring system

| Assessment of functionality and reproducibility
In a first commissioning verification of the described spectroscopic setup, two wells were filled with Wilms-MOPS medium and incubated for 8 h. Since one of the wells was additionally inoculated with E. coli BL21 BSLA, the signal response to biologic and nonbiologic changes could be assessed. Therefore, in a first step only the intensities of the scattered light were considered, to assess the reproducibility of the measurements. In a subsequent step, the functionality of individual measurements including and excluding scattered light was analyzed.
The scattered light spectra recorded of the non-inoculated medium and the E. coli BL21 BSLA are shown in Figure 3a The resulting scattered light spectra of the non-inoculated medium were a result of the wavelength-dependent optical properties, such as the spectral output of the light source, the grating efficiency, and detector sensitivity (Aikens et al., 1988;Lakowicz, 2006;Myers et al., 2013) or the polystyrene-made bottom of the MTP (Beckman Coulter GmbH, 2021; Torkelson et al., 1983).
Generally, low intensities were recorded for the UV light range below to 320 nm, while the highest values were measured at 420 and 520 nm. With an average wavelength-specific deviation of 1.26%, a very reproducible measurement was determined. The referenced intensities in Figure 3b suggest a slight systematic increase over time.
At a scattered light wavelength (λ Sl ) of 400 nm, the maximum increase accounted for up to 900 a.u., which represents an increase of 5% compared to the respective first measurement. Although the initial shape of the scattered light spectra of the E. coli BL21 BSLA culture in Figure 3c was comparable to the non-inoculated medium, the intensities show a wavelength-dependent increase over time. For example, while the intensities between 300 and 480 nm started to decrease after 4 h, for the spectral range above 600 nm, the intensities stagnated. The maximum measured values of 33,900 a.u.
were recorded at a scattered light wavelength of 520 nm. This represents a maximum difference of around 14,550 a.u. and, thus, a F I G U R E 3 Scattered light measurements of (a, b) non-inoculated Wilms-MOPS medium and (c, d) Escherichia coli BL21 BSLA culture over the cultivation time. Figure

| Comparative analysis of integration times and scattered light exclusion
In the previous experiments, the measurements including scattered light were conducted at a t int,Sl of 60 ms. Additionally, also measurements excluding the scattered light with a t int,Fl of 1500 ms were conducted. According to previous dilution series experiments (Supporting Information: Figure 5), this t int,Fl provides a good balance between improved fluorescence sensitivity and a prolonged measurement time, which was calculated to 2.8 min per spectrum in one well (Supporting Information: Table 1).
Exemplary emission spectra, recorded after 3.5 h, are shown in Besides strong scattered light intensities, for the spectra in This clearly underlines the advantage of the extended integration time.

| Application studies of intensified 2D fluorescence spectroscopy in MTP cultivations
With the improved fluorescence measurement sensitivity, the presented two tunable monochromator-based system holds significant advantages over the previously described CCD-based systems . To demonstrate the capabilities of the intensified 2D fluorescence monitoring system during the application of MVDA, two additional E. coli BL21 BSLA cultivation experiments were conducted. In the first experiment, PCA was applied to the spectral data, recorded for a sorbitol concentration variation study. Whereas, in the second experiment, spectral-based PLS regression models were generated for offline sampling data of a glucose concentration variation study. As for both experiments, the spectral data was recorded at a t int of 60 and 1500 ms, a qualitative and quantitative assessment of the improvement was possible.

| Application of PCA for a sorbitol concentration variation experiment
The sorbitol concentration variation experiment comprised the monitoring of 14 wells. Besides a general supplementation with 15 g/L glucose, four different sorbitol concentrations ranging between 0 and 10.1 g/L were supplemented in triplicate cultures.
To maintain a feasible measurement cycle time at this elevated throughput, the λ ex range for the fluorescence measurement was limited between 280 and 500 nm. Further, the scattered light was reduced to measurements at λ ex of 600 and 620 nm. This adjustment resulted in a measurement cycle time of 12.7 min per 14 wells (Supporting Information: Table 1).
The results of the cultivation experiment are presented in Figure 5.
In Figure 5a To compare the spectral data more holistically, the dimensionalityreducing PCA was applied to the two available data sets measured at a t int,Sl 60 ms and a t int,Fl of either 60 (Figure 5a With an explained spectral variance of 99.93%, the scores of the first PC1 (Figure 6a) of the data set recorded at a t int,Fl of 60 ms, were qualitatively comparable to the scattered light signals (Figure 5c).
Consequently, the scores of higher PCs accounted for a spectral variance below 0.05% and included an increasing degree of noise.
Only for PC2 (Figure 6b), the signal dynamics described a consistent progression, from which the final depletion of the supplemented sorbitol can be anticipated. For a t int,Fl of 1500 ms, the scores of PC1 (98.53%, Figure 6f) were not comparable to a singular signal shown in (1.31%, Figure 6g) described the decelerating trajectory observed for the fluorescence signals in Figure 5c,e. The PC3 scores (0.07%, Figure 6h) of the sorbitol-supplemented cultures described a small initial decrease, before steeply increasing during sorbitol consumption. In parallel to the signal increase in Figure 5c,d, also the PC3 score started to decrease upon sorbitol depletion. While the PC4 score (0.04%, Figure 6i) did not provide additional information, PC5 (0.01%, Figure 6j) showed an inverse progression of PC3. During sorbitol consumption, the scores of PC5 increased, whereas after sorbitol consumption the scores decreased again. Interestingly, this decrease was also observed for the culture not supplemented with sorbitol ( Figure 6j, blue line).
In addition to the PCA, PLS regression models were generated from the available spectral and offline data of the cultures supplemented with 0.0 g/L sorbitol and 10.1 g/L. The prediction data set consisted of the spectral data of the cultures supplemented with 3.3 and 6.8 g/L sorbitol.
The results of the PLS models, generated with five LVs, are presented in Supporting Information: Figure 7. In agreement with the scores of the PCA, the increased t int,Fl resulted in a considerably more consistent calculated trajectories. The most qualitative improvement was obtained for the sorbitol (Supporting Information: Figure 7d,j). Here, the increased t int,Fl reduced the RMSE of the calibration data set (RMSE Cal ) from 1.37 to 0.33 g/L, which corresponds to a relative improvement of more than 75% (Supporting Information: Table 2). For the other offline parameters, increasing the t int,Fl reduced the RMSE Cal by between 50% (glucose) and 17% (ethanol). For the prediction data set, only the initial of data was available for model validation. Therefore, the RMSE Pred was not calculated.

| Data-driven PLS modeling for glucose concentration variation experiments
In the second experiment, PLS regression modeling was conducted for an experiment with E. coli BL21 BSLA cultures supplemented with four different glucose concentrations ranging between 4.2 to 14.9 g/L. As especially for the low glucose supplementation, the growth phases were expected to be short, the λ ex were further reduced to ranges between 300−380 and 440−480 nm. Thereby the measurement cycle time for the 14 monitored wells was reduced to 9.8 min (Supporting Information: Table 1). The online and offline data measured for the glucose concentration variation experiment are shown in Supporting Information: Figure 8. The OTR trajectories of F I G U R E 5 Online signals and offline sampling values for Escherichia coli BL21 BSLA cultures, supplemented with constant glucose and different sorbitol concentrations. The 2D spectra in (a) and (g) were recorded for the culture supplemented with 10.1 g/L sorbitol after 9 h of cultivation (vertical dashed line). The fluorescence intensities were recorded at a t int,Fl of (a) 60 or (g) 1500 ms and were not referenced to the first measurement cycle. Gray color indicates non-measured spectral area. Red crosses indicate the location of the respective scattered light. Symbols in (g) indicate wavelength combinations visualized in (c−f) for a t int,Sl of 60 ms and a t int,Fl of 1500 ms. The symbols in (b) represent every fourth measured OTR value. Shaded areas in (b−f) show standard deviations of triplicate cultures, whereas the non-inoculated medium was measured in only one well. Barely visible standard deviations indicate very good reproducibility. Symbols in (h−m) show offline data determined by sampling in singlets. Scattered light measurement: λ Sl = 600-620 nm, step size = 20 nm, t int,Sl = 60 ms, multiacquisition Sl = 5. Fluorescence measurement: λ ex = 280-500 nm, step size = 20 nm, t int,Fl = (a) 60, (d−g) 1500 ms, multiacquisition Fl = 1. Cultivation conditions: 48-well microtiter plate with round geometry, medium = Wilms-MOPS (200 mM), carbon source = 10 g/L glucose, V L = 800 µL, d 0 = 3 mm, n = 1000 rpm, T = 37°C. BSLA,Bacillus subtilis lipase A; OTR, oxygen transfer rate.
the cultures with at least 6.8 g/L glucose (Supporting Information: - For the culture with 4.2 g/L glucose, the oxygen limitation was barely detectable both in the OTR or acetate concentrations. Consequently, no secondary growth phase was observed. F I G U R E 6 Scores of the first to fifth principal component (PC1−PC5) calculated for 2D spectra of Escherichia coli BL21 BSLA cultures, supplemented with constant glucose and variable sorbitol concentrations. The scores are based on the 2D spectra recorded at t int,Fl of 60 (a−e) and 1500 ms (f−j), as shown in Figure 5a,g, respectively. Shaded areas describe the standard deviation of three replicates. The non-inoculated medium was measured in only one well. Percentage values in brackets describe the explained variance for the respective PC. Asterisks indicate axes inverted for better clarity. Scattered light measurement: λ ex = 600−620 nm, step size = 20 nm, t int,Sl = 60 ms, multiacquisition Sl = 5. Fluorescence measurement: λ ex = 280−500 nm, step size = 20 nm, t int,Fl = (a−e) 60 ms, (f−j) 1500 ms, multiacquisition Fl = 1. Cultivation conditions: 48-well microtiter plate with round geometry, medium = Wilms-MOPS (200 mM), carbon source = 15 g/L glucose, OD 600,t0 = 0.5, V L = 800 µL, d 0 = 3 mm, n = 1000 rpm, T = 37°C. BSLA,Bacillus subtilis lipase A.
The results of the PCA are presented in Supporting Information: - Figure 9. Comparable to the previous experiment, the t int,Fl of 1500 ms resulted in increased explained variances and better signal consistency for the higher PCs. Notably, the scores of PC2 of the longer t int,Fl (Supporting Information: Figure 9G) were comparable to PC4 scores of the short t int,Fl (Supporting Information: Figure 9d). However, the explained variances of 0.61% and 0.003% suggested a considerably higher spectral dominance in the data set with a t int,Fl of 1500 ms.
In an approach to quantitatively compare the spectral quality at the two t int,Fl , PLS regression modeling was conducted. Individual models were generated for all offline parameters and included either the spectral data sets recorded with a t int,Sl of 60 ms and a t int, Fl of either 60 (Supporting Information: Figure 8A) or 1500 ms (Supporting Information: Figure 8F). The calibration data set consisted of the spectral data for the cultures supplemented with 4.2 (blue filled squares) and 14.9 g/L glucose (black filled downward triangles). For model validation, the models were then applied to the prediction data set consisting of the spectral data of the cultures supplemented with 6.8 (green open circles) and 11.9 g/L glucose (red open upward triangles). A visualization of the PLS modeling results is shown in calibration (4.2 and 14.9 g/L glucose) and the prediction data set (6.8 and 11.9 g/L glucose). The highest deviations of the calibration data set were observed during the acetate consumption phase, whereas for the prediction data set, a decreased accuracy was observed towards the end of the cultivation. For the models based on the increased t int,Fl , the calculated trajectories reflected the measured values more precisely. In total, the RMSE Cal of the CDW was reduced from 0.29 g/L, which is equivalent to 3.9% of the parameter range, to 0.17 g/L (2.3%) (Supporting Information: Table 3). With a RMSE Pred of 0.37 g/L (4.7%) and 0.33 g/L (4.5%), the model performances for the prediction data set were slightly increased. Similarly, for the pH value the t int,Fl of 1500 ms led to the RMSE Cal being reduced from 0.044 (5.7%) to 0.032 (4.2%), while the RMSE Pred was decreased from 0.052 (6.8%) to 0.040 (5.2%). For the glucose concentration, increasing the t int,Fl reduced the RMSE Cal from 0.48 g/L (3.2%) to 0.34 g/L (2.8%), while the RMSE Pred was only slightly reduced from 0.85 g/L (5.7%) to 0.81 g/L (5.4%). The relatively low model performance of the prediction data set is a result of the inaccurately calculated progression for the culture supplemented with 6.8 g/L glucose. The systematic underestimation observed after 1.5 h presumably results from recalculating the total consumed glucose from the offline value before modeling (see Section 2). This procedure introduced a high dependency on the first measurement, and likely led to the observable offset. The PLS model of the acetate represented the biggest challenge, as over the cultivation time, the carbon source first accumulated before being consumed again. With a RMSE Cal of 0.20 g/L (14.1%) and a RMSE Pred of 0.25 g/L (17.8%), the PLS modeling performance was comparably low for the spectral data set recorded at a t int,Fl of 60 ms. While the model described the acetate accumulation appropriately, the accuracy during acetate consumption decreased. For the longer t int,Fl , a RMSE Cal of 0.14 g/L (10.1%) and a RMSE Pred of 0.17 g/L (11.8%) resulted. Thus, the values are still comparably high, which well-reflect the inaccurate description of the offline values during acetate consumption. Nevertheless, in contrast to the model for the short t int , the new model calculated a delayed, yet full depletion of the accumulated acetate.

| DISCUSSION
Given the increased fluorescence intensity signal and the improved results during PCA and PLS, the spectroscopical setup presented in this study confirmed the envisioned advantages. Nevertheless, during the experiments, several potential aspects for further investigation and improvement were identified.
The new monitoring setup offers a considerable experimental flexibility, by providing the possibility to adjust various measurement parameters, such as the covered spectral range, the number of λ ex or the chosen t int,Fl . However, the different settings have a significant impact on the measurement time per well, which results in limited monitoring throughput and, thus, represents the main limitation of the presented setup. For example, for the reduced spectral measurement settings of the last experiment, the measurement cycle time was calculated to 33.6 min for 48 wells (Supporting Information: Table 1). In contrast, for the setup by , a measurement cycle time of only 15 min was calculated for comprehensive 2D spectra. This even allows the monitoring of 48 individual E. coli cultures with a doubling rate as low as 20 min (Tuttle et al., 2021). However, as described before, the measurements were conducted at a short t int of 30 ms, which reduced the fluorescence sensitivity. In future studies, the flexibility of the new spectroscopic setup should, therefore, be utilized to elaborate on the optimal balance between the quality and the quantity of the recorded spectral data by varying the number of λ ex and the t int .
Besides reducing the amount of recorded spectral data, another possibility to decrease the measurement cycle time is by increasing the optical throughput of the setup. Noteworthy, although the new setup significantly increased the fluorescence sensitivity, not all expected fluorophores were detected. For example, no distinct fluorescence intensities were observed for fluorophores of lower quantum yields, such as amino acids or pyridoxine (Callis & Liu, 2004;Chen, 1965Chen, , 1967. Although this is partially derived from the optical components absorbing the UV-light, preliminary experiments with a 150 W xenon lamp also suggested a limited light flux of the 75 W xenon lamp. As the increase to a high-power light source demanded for additional cooling, the 75 W setup was chosen in this prototype.
However, in an industry-oriented setup, a switch to high-power LEDs could be advantageous. Besides being economically more attractive, the use of LEDs also enables faster changes of λ ex , as it eliminates the need for movement of the excitation monochromator. Nevertheless, this requires the availability of LEDs at the desired λ ex .
Another aspect posing a challenge for broad application of the current setup is the mechanical movement of the monochromators, which is needed to separate the scattered light from the fluorescence measurement. As an alternative, less maintenance-requiring setups, including a combination of transmission gratings and linear transmission bandpass filters or a fixed grating in combination with purposefully oversaturating the CCD camera, as shown by Graf et al. (2019), should be considered. However, especially the latter may decrease the lifetime of the light-sensitive CCD array. Thus, the implementation of a simpler detector, such as a photodiode line sensor array, could be considered.

| SUMMARY AND CONCLUSION
The 2D fluorescence monitoring system presented in this study represents a significant step in advancing this versatile spectroscopic technology. The introduced flexibility of the tunable emission monochromator resulted in a higher sensitivity for fluorescence measurements. The improved quality of the online data provided more diverse spectral dynamics, which were successfully utilized during PCA and PLS modeling. In conclusion, 2D fluorescence spectroscopy was shown to hold the potential to fundamentally change the way fluorescence monitoring is used in MTP and to align with current trends in advanced data analysis workflows. In future BERG ET AL. | 2937 studies, technical aspects such as the impact of a reduced number of λ ex on modeling results should be evaluated. Thereby, current issues of measurement cycle times and maintenance issues may be solved, resulting in a more robust, industry-oriented setup.

AUTHOR CONTRIBUTIONS
Christoph Berg designed this study, drafted the manuscript, devel-