In Situ Microscopy with Real-Time Image Analysis Enables Online Monitoring of Technical Protein Crystallization Kinetics in Stirred Crystallizers

Abstract

Controlling protein crystallization processes is essential for improving downstream processing in biotechnology. This study investigates the combination of machine learning-based image analysis and in situ microscopy for real-time monitoring of protein crystallization kinetics. The experimental research is focused on the batch crystallization of an alcohol dehydrogenase from Lactobacillus brevis (LbADH) and two selected rational crystal contact mutants. Technical protein crystallization experiments were performed in a 1 L stirred crystallizer by adding polyethyleneglycol 550 monomethyl ether (PEG 550 MME). The estimated crystal volumes from online microscopy correlated well with the offline measured protein concentrations in solution. In addition, in situ microscopy was superior to offline data if amorphous protein precipitation occurred. Real-time image analysis provides the data basis for online estimation of important batch crystallization performance indicators like yield, crystallization kinetics, crystal size distributions, and number of protein crystals. Surprisingly, one of the LbADH mutants, which should theoretically crystallize more slowly than the wild type based on molecular dynamics (MD) simulations, showed better crystallization performance except for the yield. Thus, online monitoring of scalable protein crystallization processes with in situ microscopy and real-time image analysis improves the precision of crystallization studies for industrial settings by providing comprehensive data, reducing the limitations of traditional analytical techniques, and enabling new insights into protein crystallization process dynamics.

1. Introduction

Protein crystallization is fundamental in structural biology, biotechnology, and drug development, as it plays a crucial role in obtaining high-resolution molecular protein structures via X-ray and producing purified proteins on a large scale [1,2,3,4]. In an industrial context, technical protein crystallization applying stirred crystallizers has become essential for large-scale purification and process optimization, as it provides a scalable method for generating high-purity protein crystals efficiently. Technical protein crystallization becomes more and more important in biomanufacturing, where consistent protein crystal quality and reliable yields are essential [2,5,6]. The success of a crystallization process depends on several variables, such as precipitant type and concentration, temperature, sample concentration, and protein solubility [2,7,8,9]. In addition, reaching a supersaturation state without protein precipitation is a key challenge for crystallization processes since the process is generally mass transport and diffusion-limited, directly affecting nucleation and crystal growth [10]. To address this challenge, it is essential to balance hydrodynamic and mass transfer phenomena by optimizing agitation conditions, examining a broad range of mass transfer regimes and maintaining a uniform energy input in stirred crystallizers. This is vital for ensuring consistency in terms of overall yield and purity on large-scale production environments due to the hydrodynamic stress on the protein crystals [11,12] At the same time, an insufficient energy input may lead to an inhomogeneous mixing [11]. To address these challenges in an industrially relevant setting, a comprehensive analysis of various crystallization conditions and the implementation of process analytical technologies (PATs) are required [10,13].

Nonetheless, even with these advanced strategies, the fundamental understanding of nucleation remains limited [10,13,14,15,16]. Thus, a deeper understanding of the protein structure at the crystal contacts and protein–protein interfaces seems crucial for identifying the mechanisms of protein crystallization, investigating suitable conditions, and developing new applications [13,15,17].

Recent studies demonstrated that rational crystal contact engineering is a promising approach in this field. This involves the strategic exchange of single amino acids at crystal contact sites, particularly for surface entropy reduction (SER) [5,6,18,19]. By reducing disorder on the protein surface, this method enhances crystallizability, as defined by an increase in the number of crystals (higher nucleation rate), a reduction in induction time, and a decrease in the time required to reach equilibrium concentration [5,6,20].

In addition, recent work has focused on optimizing electrostatic interactions and hydrophobic regions between protein–protein interfaces and crystal contacts. Here, the exchange of the amino acids threonine by glutamic acid of the alcohol dehydrogenase from Lactobacillus brevis (LbADH), resulting in mutant T102E, should be mentioned as an example. This single amino acid exchange has led to more stable protein crystals and higher crystallization efficiency, underlining the potential of rational crystal contact engineering to improve protein separation and purification processes in biotechnology [6,20,21,22].

Several analytical in situ techniques are known to investigate and monitor crystallization processes in detail. In particular, spectroscopy in the near-infrared (NIR) and ultraviolet-visible (UV-vis) range is often used for real-time monitoring and control of protein concentration [23,24]. Interferometry also enables the precise measurement of changes in protein concentration and solubility during crystallization [25,26,27]. Furthermore, integrated ultrasonic measurement techniques and focused beam reflectance measurement (FBRM) probes provide valuable data for controlling crystallization conditions [28,29,30,31]. These methods are widely known for their exceptional sensitivity, leading to enhanced product quality while also being non-invasive [23,28]; whereas in cases of crystallization in turbid or opaque solutions, the effectiveness of these methods may be compromised [28]. However, these advancements in monitoring technologies not only improve our understanding of crystallization kinetics but also support the reproducibility and scalability required in industrial applications.

In recent years, the field of crystallization process control has also been extended by integrating machine learning algorithms. A notable example is the deep learning algorithm MARCO (Machine Recognition of Crystallization Outcomes), developed to recognize crystal growth patterns in bright field images [32]. In addition, machine learning-based methods for the recognition of protein crystals in microscopic images have been researched [33]. Moreover, the application of machine learning for analyzing in situ microscopic images using convolutional neural networks (CNNs) has enabled the detection of unwanted crystals in slurry samples, facilitating the implementation of corrective actions during the crystallization process [34].

However, this example shows that the aforementioned monitoring techniques require complex solutions to deal with atypical protein behavior, including the precipitation of proteins. It is important to note that these amorphous precipitations, influenced by the protein and solution properties, can complicate the accuracy of the analytical method [4,35,36].

This study focuses on coupling a machine learning-based crystal detection software with an in situ microscopy probe. The aim is to identify exemplarily the batch crystallization kinetics of an alcohol dehydrogenase from Lactobacillus brevis (LbADH) and its rationally designed mutants Q207D and T102E [37]. Using a scalable 1 L stirred crystallizer, this setup enables automated real-time image analysis to correlate the online estimated protein crystal volume with the offline measured protein concentration under standardized crystallization conditions. Such a scalable approach aligns with industrial needs, supporting real-time process control to ensure consistent product quality and yield. Therefore, advantages over conventional analysis methods will be discussed, particularly in avoiding incorrect measurements due to amorphous protein precipitation.

2. Materials and Methods

2.1. Protein Mutagenesis, Production and Purification

In this work, the LbADH (PDB ID: 6H07, [38]) wild type (WT), and the LbADH mutants T102E (PDB ID: 6Y0S, [38]) and Q207D (not published) were investigated. The crystallization properties and structure of the LbADH and its mutants were previously characterized through X-ray diffraction and electrophoretic analysis, confirming their crystallizability, structural integrity, and sufficient resolution. These validations support the reproducibility and comparability of crystallization kinetics and yields obtained in this study. Specifically, the T102E mutation enhances crystal packing through stabilized salt bridge interactions, without compromising resolution, while Q207D preserves overall crystal integrity [5,39].

For this purpose, the cloned plasmid pET28a_LbADH_GSG_His₆ (LbADH WT) constructed by Nowotny et al. [5] was used, as well as the plasmids of the mutants LbADH T102E and LbADH Q207D produced by site-directed mutagenesis according to a QuickChange PCR protocol [37].

Utilizing chemically competent E. coli BL21(DE3) cells, heterogeneous protein production was initiated by transforming the cells with plasmids of the WT or the mutants via heat shock. The cells were then transferred to the first preculture stage in 13 mL tubes, which contained 5 mL lysogeny broth (LB) medium with 35 µg mL⁻¹ kanamycin, and incubated for 24 h (180 rpm, 30 °C). Subsequently, 1 mL cell broth was aliquoted into eight 2 L shake flasks, without baffles, containing 250 mL LB medium with kanamycin (35 µg mL⁻¹) to inoculate the second preculture (incubated for 16 h at 250 rpm and 37 °C).

A high-cell density cultivation on a 50 L scale in a stirred tank reactor (LP75; Bioengineering AG, Wald, Switzerland) was then inoculated with the second preculture to achieve an initial OD₆₀₀ of 0.5 in 34 L of defined Riesenberg mineral medium [40]. The protein production process was split into a batch phase (5 g L⁻¹ glucose, 4 h, 37 °C), an exponential fed-batch phase with 500 g L⁻¹ glucose and 12.5 g L⁻¹ MgSO₄ in the feed (µ_set = 0.15 h⁻¹, 22 h, 37 °C), and the production phase with the addition of 500 µM isopropyl ß-D-1-thiogalactopyranoside and constant glucose feeding (2.7 g L⁻¹ h⁻¹ glucose, 24 h, 30 °C), according to Schmideder et al. [41]. The pH was controlled to pH 6.8 by adding NH₄OH (25% (v/v)). A cascade control (stirrer speed, aeration, pressure) set the DO to 30% air saturation, using 400 rpm, 25 L h⁻¹, and 0.2 bar as initial conditions. The resulting cell dry weight (CDW) reached a maximum concentration of 120.66 ± 1.31 g L⁻¹ after a process time of 50 h.

The cell suspension was directly transferred to a high-pressure homogenizer (Variete NS3015H; GEA Niro Soavi, Parma, Italy) operated at 900 bar with a flow rate of 100 L h⁻¹. The disrupted cell broth was then adjusted to pH 7.5 before adding 500 mM NaCl, 20 mM imidazole, 10 mM NaH₂PO₄, and 10 mM Na₂HPO₄. In the following, microfiltration and diafiltration of the cell lysate with 500 mM NaCl, 20 mM imidazole, 10 mM NaH₂PO₄, 10 mM Na₂HPO₄ (pH 7.5) was performed using a gravimetric tangential flow pump (SpectrumLabs KMPi TFF System, Waltham, MA, USA) and a 750 kDa MWCO hollow fiber membrane module (UFP-750-E-4X2MA, Cytiva, Marlborough, MA, USA) with a concentration factor (C_f) of 2, a diafiltration factor (D) of 1.5 and a constant transmembrane pressure (TMP) of 0.3 bar.

For purification, the protein was then loaded onto a 1 L-His-trap nickel affinity column (PureCube 100 Ni-NTA Agarose; Cube Biotech, Monheim, Germany), which was pre-equilibrated in a binding buffer (500 mM NaCl, 20 mM imidazole, 10 mM NaH₂PO₄, 10 mM Na₂HPO₄, pH 7.5). Subsequently, a washing step was performed with the binding buffer, followed by eluting the bound protein with an elution buffer (500 mM NaCl, 500 mM imidazole, 10 mM NaH₂PO₄, and 10 mM Na₂HPO₄, pH 7.5). Due to the following concentration step, no specific fractionation was carried out, and the resulting approximately 2 L of purified protein was stored at −20 °C with 5% (v/v) glycerol for further use. For the individual crystallization experiments, the purified protein solutions were dialyzed after thawing on ice using a tangential flow unit (Sartoflow Beta plus, Sartorius Stedim, Goettingen, Germany) with a 10 kDa MWCO membrane cassette (Sartorius Sartocon Hydrosart, Sartorius Stedim, Goettingen, Germany), according to the protocol of Hebel et al. [11], against the protein buffer (20 mM HEPES-NaOH, 1 mM MgCl₂, pH 7.0). Finally, a particle filtration was performed with a 0.22 µm polyethersulfone filter (Steritop MILLIPORE Express, Merck Chemicals GmbH, Darmstadt, Germany). The protein purification yields are shown in Supplemental Figure S1.

2.2. Protein Crystallization and Standardization of Conditions

All crystallization experiments were performed in a 1 L stirred tank reactor made of glass (inner tank diameter D = 0.12 m; tank filling height H = 0.12 m) with a rounded bottom and a three-bladed segment impeller (agitator tip diameter d = 0.06 m; impeller blade height h = 0.04 m), which provided an upward flow near the stirrer axis [11]. The in situ microscopy probe (EasyViewer 400, Mettler-Toledo, Giessen, Germany) was installed in the 1 L reactor at an angle γ of 29° to the stirring axis, ensuring the position of the probe cleft (d_c = 0.005 m) was above the impeller (Figure 1).

The stirrer was controlled by a custom-built control unit [18] based on a micro-stepping motor driver module (TMCM6110, Trinamic, Hamburg, Germany), which ensured a smooth rotational speed.

For the experiments involving an in situ microscopy probe, the stirrer speed was set to 50 rpm. The temperature was held constant at 20 °C in all experiments with a double jacket and a temperature-controlled water bath (1157P, VWR, Darmstadt, Germany).

An initial protein concentration of 5 g L⁻¹ was used. The concentration was adjusted by dilution with the protein buffer and was verified spectrophotometrically at 280 nm.

The crystallization buffer composition consisted of 0.1 M Tris-HCl at pH 7.0, 50 mM MgCl₂, and 100 g L⁻¹ PEG polyethyleneglycol 550 monomethyl ether (PEG 550 MME). 100 g L⁻¹ PEG 550 MME ensured the crystallization of 5 g L⁻¹ of all LbADH variants within 24 h, allowing the system to reach thermodynamic equilibrium at 20 °C, as indicated by the stabilization of the soluble protein concentration.

0.5 L of purified protein solution and 0.5 L of crystallization buffer were filtrated using a 0.22 µm polyethersulfone filter at 20 °C. The filtered protein solution was first added to the crystallizer. Then, the batch crystallization process was initiated by adding the crystallization buffer into the stirred system using a funnel at a rate of ~4 mL s⁻¹. An immersion tube was used for manual sampling. Before further analytical processing, the samples (1 mL) were centrifuged for 10 min at 15,000× g (5415R, Eppendorf AG, Hamburg, Germany).

To calculate the mean power input $\overline{\epsilon }$ [W kg⁻¹] of the stirred 1 L crystallizer, the data set of Smejkal et al. [42] was utilized, since the identical reactor was used. This data set was generated by measuring the torque M [kg m² s⁻²] between the motor and stirrer axis at varying rotational stirrer speeds n_s [s⁻¹] at constant filling volume $V$ [m³] and the density of the liquid phase $\rho$ [kg m⁻³]. Smejkal et al. [42] also provided an estimation of the maximum local energy dissipations ${\epsilon }_{\max }$ [W kg⁻¹] by determining the characteristic size distribution of silicone oil droplets (Equation (1)) [43]:

$${\displaystyle \frac{{\epsilon }_{\max }}{\overline{\epsilon }}}={\displaystyle \frac{\beta }{ {\left(A·{n_s}^B\right)}^{\alpha }}}·{\displaystyle \frac{ V·\rho }{M·2·\pi ·n_s}}$$

(1)

with $\alpha$ = 1.415, $\beta$ = 12.169 [m² s⁻³], A = 3.306 × 10⁵ [s⁻¹], and B = −1.249.

2.3. Protein Analytics

The solved LbADH protein concentration was determined spectrophotometrically at 280 nm (N120 NanoPhotometer, Implen GmbH, Munich, Germany), based on a molar extinction coefficient of 19940 M⁻¹ cm⁻¹ (calculated with Protparam; [44]). The protein yield and concentration during the purification steps were determined densitometrically using a calibrated sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) and analyzed using the software ImageJ (Version 1.54g, National Institutes of Health (NIH), Bethesda, MD, USA) (results shown in Supplemental Figure S1).

The non-soluble protein concentrations c_X(t) are obtained by a mass balance of the batch crystallization process (Equation (2)).

c_X(t) = c₀ − c(t)

(2)

In this context, c₀ represents the initial soluble protein concentration of the supernatant, and function c(t) represents the soluble protein concentration at a certain point in time t.

2.4. Protein Crystal Detection by Real-Time Image Analysis

To observe crystal growth in real-time, photomicrographs were recorded at 60 s intervals (sampling rate f_s of 0.16 Hz) using the imaging tool of the in situ microscopy probe and the corresponding iC Vision software (iC Vision Version 8.1, Mettler-Toledo, Giessen, Germany). The probe was operated in backscattering mode with eight laser beams.

The resulting photomicrographs were automated, exported online in real-time from the iC Vision software, and uploaded to an image crystal detection tool. A deep learning method based on the S²A-Net-oriented object detection (OOD) model established by Han et al. [45] enabled automated image analysis. This model was specifically adapted by Bischoff et al. [33] using a large synthetic dataset of photomicrographs with protein crystals in suspension. Additionally, a series of stochastic transformations was applied to the image files to enhance the robustness of the final image recognition model against various confounding factors [33]. Validation of the model was conducted both through a validation dataset and experimentally by time-resolved analysis of microscope images of protein crystals in suspension, which were compared with simultaneous measurements of the supernatant [33].

The performance of this state-of-the-art crystal detection was additionally improved by online augmentation of the input images using two additional image versions (90° rotation and horizontal flip). The probability of accurately detecting all crystals should be increased by evaluating the individual photomicrographs several times in the same image analysis step. The resulting images were analyzed using the detection software regarding single crystal size and overall crystal number.

The total observed crystal volume V_obs(t_n) of N_c(t_n) crystals in the n-th image was then approximated from the length l_i and the width w_i of each detected crystal according to Wu et al. [46] (Equation (3)). In this context, microscopic observations were used to approximate the crystal morphology as a rectangular shape, without specifically considering the solvent content.

$$V_{\mathrm{obs}}(t_{\mathrm{n}}) = {{\displaystyle \sum }}_{i=1}^{{\mathit{N}}_c\left({\mathit{t}}_{\mathrm{n}}\right)} {\mathit{l}}_{\mathit{i}}\left({\mathit{t}}_{\mathrm{n}}\right)· {{\mathit{w}}_{\mathit{i}}}^2\left({\mathit{t}}_{\mathrm{n}}\right)$$

(3)

Following the automated image analysis and evaluation, the crystal volume could be displayed in real-time (delay of 30 s) via a custom-developed graphical user interface. The different threshold values of the detection software used, and their measurement range limits are listed in the Supplemental Table S1.

2.5. Empirical Modeling of Batch Crystallization Process Kinetics

Logistic functions were used to estimate either the non-soluble protein concentration c_X(t) (Equation (4)) or the total crystal volume V_obs(t_n) (Equation (5)) over process time. This is a common empirical description [46,47], as logistic functions are able to cover an initial phase of a batch crystallization process with increasing crystal formation, followed by exponential crystal growth, and finally, reduced crystal growth approaching the thermodynamic equilibrium.

$$c_{\mathrm{X}}\left(\mathit{t}\right) = {\mathit{c}}_{\mathrm{X},\max } \{1+\exp [-k_{\mathrm{X}}·(\mathit{t}-{\mathit{t}}_{1/2,\mathrm{X}})]{\}}^{-1}$$

(4)

$$V_{\mathrm{obs}}\left({\mathit{t}}_{\mathrm{n}}\right) = {\mathit{V}}_{\max }\{1+\exp [-k_{\mathrm{V}}·(\mathit{t}-{\mathit{t}}_{1/2,\mathrm{V}})]{\}}^{-1}$$

(5)

The estimation of the crystal growth factors k_i, the maximum protein concentration c_X,max, and the maximum observed crystal volume V_max, and the inflection points t_1/2 were achieved by non-linear parameter identification (MATLAB Version: 23.2.0 (R2023b), MathWorks Inc., Natick, MA, USA).

To correlate the non-soluble protein concentration with the total crystal volume in a batch crystallization process, the functions V_obs(t_i) and c_X(t_i) were normalized to the range [0, 1], and the root mean square deviation (RMSD) over the n-th image was calculated to quantify the deviation between the two functions over time (Equation (6)). The RMSD is given by:

$$\mathit{RMSD}=\{{\mathrm{n}}^{-1}·{{\displaystyle \sum }}_{i=1}^{\mathrm{n}}·[c_{\mathrm{X}}({\mathit{t}}_{\mathit{i}})-V_{\mathrm{obs}}({\mathit{t}}_{\mathit{i}}){]}^2{\}}^{1/2}$$

(6)

with the normalized values of the non-soluble protein concentration c_X,norm(t_i), and the observed crystal volume V_obs,norm(t_i).

3. Results and Discussion

3.1. Optimum Stirrer Speed and Maximum Energy Dissipation Rate

First, the optimum rotational stirrer speed of the 1 L crystallizer was determined at which further increases in power input did not accelerate the onset of LbADH WT crystallization. In addition, the ratio of maximum local energy dissipation ${\epsilon }_{\max }$ to mean power input $\overline{\epsilon }$ was calculated. For this purpose, three batch crystallization experiments with initial protein concentrations of 5 g L⁻¹ and crystallization agent concentrations of 100 g L⁻¹ PEG MME 550 were carried out with the LbADH WT, whereby only the rotational stirrer speed was varied between 50 rpm and 100 rpm (Figure 2a). Lower rotation rates were not investigated, as an accumulation of crystals at the bottom of the crystallizer was observed, and homogeneous mixing at the start of the process would not have been possible in less than approx. five minutes, according to Smejkal [48]. The applied crystallization conditions were determined based on a previously conducted parameter study (results shown in Supplemental Table S2).

The final crystallization yields were identical within the estimation error (82–86%), indicating that the thermodynamic equilibrium was achieved within an estimated process time of around 16 h independent of the stirrer speed tested.

By applying the optimum rotational stirrer speed of 50 rpm to Equation (1), the maximum local energy dissipation was calculated to be 0.19 W kg⁻¹, with a corresponding mean power input of 2.37 ± 0.84 mW kg⁻¹ [48], resulting in an estimated ratio of ${\epsilon }_{\max }/\overline{\epsilon }$ = 80.15 ± 28.29 with the LbADH WT protein. The high standard deviation is caused by the high inaccuracies of the mean power input measured by Smejkal at 50 rpm [48]. Comparing this ratio with batch crystallization data of lysozyme with the same stirred crystallizer (optimum stirrer speed of 200 rpm) resulted in no significant differences (${\epsilon }_{\max }/\overline{\epsilon }$ = 73.10 ± 2.60).

When comparing the optimum stirring speeds of lysozyme and LbADH, it is important to consider the impact of the different crystallization agents used, as these play a critical role in influencing crystallization processes [7]. For the LbADH, the viscosity-enhancing polymer PEG was employed [36], whereas lysozyme was crystallized using an aqueous sodium chloride solution [42]. PEG is known to decrease protein solubility and enhance protein–protein interactions, promoting nucleation events and crystal growth [7,49,50]. This was also observed in lysozyme crystallization with PEG [51]. This balance between nucleation promotion and increased viscosity likely influences the crystallization dynamics. Additionally, PEG can alter solute diffusion, leading to faster diffusion rates than predicted by the Stokes–Einstein equation [10,52]. This increase in nucleation events, along with the anomalous diffusion behavior [10], likely contributes to the lower optimum stirrer speed observed for the LbADH crystallization.

3.2. Monitoring of LbADH WT Crystallization Processes with In-Situ Microscopy

With the use of the S²A-Net OOD model that was trained from random initialization using a specialized synthetic dataset for protein crystallization processes by Bischoff et al. [33], which was further complemented in this work, individual protein crystals can be identified in photomicrographs provided by the in situ microscope probe. The model records the number of crystals, as well as their lengths and widths. Batch crystallization experiments were carried out in duplicates in the stirred 1 L crystallizer with the LbADH WT, as the standardized and well-controlled conditions ensured that duplicates were sufficient to achieve reliable and reproducible results.

The offline measured protein concentration dynamics of the LbADH WT are shown in Figure 3a (blue dots with min–max values). In addition, the dynamics of the average crystal volume V_obs(t_n), online, calculated based on the results of the real-time image analysis of the photomicrographs of the in situ microscopy probe, are plotted in the same graph. The average crystal volume data of the two individual batch crystallization processes (dark and light gray) and the mean crystal volume (orange) indicate the increase in protein crystal volume with process time, which corresponds to the decrease of the protein concentration in the supernatant.

Figure 3b shows the dynamics of the corresponding non-soluble protein concentrations (blue dots). The offline data were interpolated with a logistical function (blue line). The same logistic interpolation was applied to the average crystal volume dynamics (yellow line). Both interpolation curves are almost identical, as indicated by the identified crystal growth factors of k_X = 0.73 h⁻¹ and k_V = 0.71 h⁻¹, respectively, as well as a root mean square deviation of almost zero (0.01).

The online estimated average crystal volumes showed unexpected behavior at the beginning of the batch process (decrease between 4–6 h), but these deviations were within the estimation error of the non-soluble protein concentrations derived from the offline measured protein concentrations in the supernatant (Figure 3b). Increased stochastic deviations in the online estimated average crystal volumes occurred as far as the thermodynamic equilibrium was approached after a process time of ~12 h. An exemplary photomicrograph recorded by the in situ microscope probe at the end of the batch crystallization shows agglomerates of the individual protein crystals (Figure 3c, right). According to the photomicrographs, these agglomerates and cross-linked crystals are not homogeneously distributed in the solution. This results in noise in the crystal volume signals, whereby the image analysis software can no longer distinguish all the individual crystals in contrast to a photomicrograph after 8 h (Figure 3c, left). This limitation is due to the evaluation parameters of the software being optimized specifically for individual crystals and cannot be dynamically adjusted during the process. The detected crystals are highlighted here in color to disambiguate overlapping crystals (green, blue, purple) by the software.

Furthermore, it was observed in an additional experiment that local inhomogeneities led to amorphous protein precipitation. This phenomenon occurred when the crystallization buffer was added to the stirred system (50 rpm) with factor 2 increased feed rate (see Section 2.2). The rapid addition resulted in a heterogeneous distribution within the solution, creating local concentration differences in the stirred crystallizer.

Figure 4a shows the average observed crystal volume (orange) and the non-soluble protein concentration (blue dots) of an LbADH WT crystallization with initial amorphous precipitation, as well as the corresponding logistic interpolations (orange and blue). Exemplary photomicrographs of the precipitation at the beginning (left) and end of the process (right) are depicted in Figure 4b.

Due to the initial amorphous protein precipitation, the non-soluble protein concentration was higher at the beginning of the batch process (Figure 4a). However, no protein crystals were initially detected in the microscopic images (Figure 4b, left). The onset of crystal growth was observed by the real-time image analysis of the photomicrographs after ~4 h. The interpolation curves are different (k_X = 0.25 h⁻¹, and k_V = 0.57 h⁻¹) with similar inflection points with t_1/2,X of 7.84 h, and t_1/2,V of 8.17 h, respectively. However, the root mean square deviation was high (0.12) due to the initial amorphous protein precipitation. After reaching thermodynamic equilibrium within ~12 h of process time, increasing deviations in the online estimated average crystal volume occurred. An exemplary photomicrograph from the end of the batch crystallization also shows the agglomeration of protein crystals that accumulated along the amorphous precipitates (Figure 4b, right). As a result, the image analysis software could no longer detect all crystals (marked in green and blue) in the microscopic images, causing noise in the crystal volume signal.

Compared to the stirred 1 L batch crystallization of the LbADH WT without initial amorphous precipitation (Figure 3), the crystallization yield Y decreased by 7.6% from 82.9% to 75.3% after 24 h. At the beginning of the process, 0.78 g L⁻¹ of the LbADH WT protein precipitates. However, by the end of the process, only 0.38 g L⁻¹ less protein was found in crystalline form, suggesting that a certain proportion of the amorphous protein precipitation can still be transformed into crystals during crystallization. The transformation of amorphous precipitate into protein crystals can occur either through a phase transition, where the amorphous precipitate reorganizes itself into a crystalline structure. Alternatively, it can happen by Ostwald ripening, where smaller particles dissolve, and their material is redeposited onto larger crystals due to differences in solubility between small and large particles [53].

3.3. Monitoring of Crystallization Processes with LbADH Mutants

The real-time image analysis was also used to investigate stirred batch crystallization processes on a 1 L scale with the LbADH mutants T102E and Q207D.

The protein concentration dynamics of the supernatant and the average observed crystal volumes (dark and light gray) of the LbADH mutants Q207D (green) and T102E (red) are shown in Figure 5 (left). The dynamics of the mean crystal volume (orange) and the non-soluble protein concentration are depicted in Figure 5 (right), along with their corresponding fits. For both mutants, the real-time image analysis detected an increase in crystal volume over time, corresponding to a decrease in protein concentration in the supernatant and an increase in the non-soluble protein concentration.

The estimated crystal growth factor k_X increased from 0.73 h⁻¹ (LbADH WT) to 1.18 h⁻¹ (Q207D) and 23.44 h⁻¹ (T102E). A similar trend was observed for the crystal volume, with crystal growth factors k_V of 0.71 h⁻¹ (LbADH WT), 1.00 h⁻¹ (Q207D), and 8.84 h⁻¹ (T102E). In both cases, a simultaneously decreasing inflection point was noted. These data clearly show faster crystallization kinetics of the LbADH mutants.

The measures of determination (R²) of the modeled non-soluble protein concentration support the use of the logistical function (0.99 (LbADH WT), 0.98 (Q207D), and 0.99 (T102E)). In contrast, lower coefficients of determination (R²) were noted for the observed crystal volume (0.98 (LbADH WT), 0.84 (Q207D), and 0.55 (T102E)), which are attributed to increased stochastic deviations over longer observation periods due to the formation of protein crystal aggregates, as shown before.

When comparing the empirically modeled non-soluble protein concentration and the observed crystal volume using the root mean square deviation (RMSD), the proportionality becomes obvious in all cases (0.01 (LbADH WT), 0.02 (Q207D), and 0.03 (T102E)). The slightly enhanced RMSD value of the functions with the LbADH mutant T102E occurs due to the limited number of data points in the inflection point region. However, these errors clarify that the real-time in situ image analysis used here can accurately display different crystallization kinetics, particularly until thermodynamic equilibrium is reached, which is especially critical for industrial applications.

Table 1. Comparison of the crystallization variables and results with the starting concentration c₀, the equilibrium concentration c_eq after 24 h, and the yields Y of the crystallizations carried out in a stirred 1 L scale of the LbADH WT, the mutants T102E, and Q207D. The batch crystallization data of the LbADH WT with initial amorphous precipitation (AP) added for comparison. The estimated parameters of the logistic fit functions are also shown (maximum concentration c_max,X, maximum observed volume V_max, crystal growth factors k_i, and inflection points t_1/2,i, respectively). Furthermore, the coefficient of determination R² and the calculated root mean square deviation (RMSD) according to Equation (6) are given to illustrate the deviations.

Lactobacillus brevis ADH			WT	Q207D	T102E	WT (AP)
Crystallization parameters	c0	[g L−1]	4.98 $\pm$ 0.22	4.91 $\pm$ 0.28	5.05 $\pm$ 0.07	5.00 $\pm$ 0.00
	ceq	[g L−1]	0.85 $\pm$ 0.04	0.85 $\pm$ 0.34	0.18 $\pm$ 0.00	1.23 $\pm$ 0.00
	Y	[%]	82.86	82.65	96.43	75.33
Logistic parameters cX(t)	cmax,X	[g L−1]	4.03	3.75	4.68	3.87
	kX	[h−1]	0.73	1.18	23.44	0.25
	t1/2,X	[h]	7.70	2.48	0.23	7.84
	R2	[-]	0.99	0.98	0.99	0.97
Logistic parameters VObs(t)	Vmax	[µm3]	59,337	63,009	113,700	58,061
	kV	[h−1]	0.71	1.00	8.84	0.57
	t1/2,V	[h]	7.85	2.3074	0.1941	8.17
	R2	[-]	0.98	0.84	0.55	0.98
RMSD		[-]	0.01	0.02	0.03	0.12

summarizes all details of the crystallization variables and results, as well as the estimated functional parameters of the logistic interpolation curves of the protein crystallizations carried out.

3.4. Final Crystal Size Distributions

Figure 6 shows the online measured length and width of the final protein crystals, along with the interpolated cumulative distributions (intervals of 2.5 µm). The crystal size distributions of the LbADH WT (blue) and its mutants Q207D (green) and T102E (red) are compared, whereby the interpolated cumulative distribution function for the crystal lengths and widths was obtained by determining the cumulative frequency (sum of the counts for each increment), and then normalizing each cumulative sum by dividing it by the total count.

The final LbADH WT crystals showed a broad crystal size distribution with a length between 10–50 µm and a width between 3–15 µm, whereas the final crystals of the LbADH mutant T102E are characterized by narrow size distributions (length between 18–28 µm, and width between 3–10 µm, respectively). The final crystal sizes of the LbADH mutant Q207D exhibited a broader distribution of crystal lengths and widths, showing greater variability and a tendency towards larger crystals compared to the mutant T102E but smaller crystals with respect to the LbADH WT.

The absolute count of the final number of protein crystals per liter after 24 h, with 13,683 crystals of the LbADH WT, 64,917 crystals of the mutant Q207D, and 119,211 crystals of the mutant T102E, respectively, indicates an increase in nucleation events.

Thus, the LbADH mutant T102E outperforms the other proteins concerning all batch crystallization performance indicators as follows: highest yield, fastest crystallization (highest crystal growth factors with smallest inflection points), narrowest final crystal size distributions, and highest final number of protein crystals.

In contrast, the differences between the crystallization performance indicators of the LbADH WT and the mutant Q207D are not as consistent: the mutant crystallizes faster (higher crystal growth factors with smaller inflection points), the final crystal size distributions are smaller, and the final number of protein crystals is higher, but there is no difference in the yields (Table 1).

The protein and its mutants were selected based on their varying changes in Gibbs free energy ∆∆G between mutant crystallization and wild type crystallization, which were obtained by molecular dynamics (MD) free energy simulations [37,39]. The calculated ∆∆G of the mutant T102E is approximately −14.4 kJ mol⁻¹, whereas the ∆∆G of the mutant Q207D is +2.4 kJ mol⁻¹. A negative ∆∆G should theoretically result in improved crystallizability compared to the LbADH WT, whereas a positive ∆∆G will theoretically not improve protein crystallization [39].

The predicted improved crystallizability of the LbADH mutant T102E compared to the LbADH WT was clearly demonstrated with respect to all crystallization performance indicators, but the experimental outcome with respect to the other LbADH mutant Q207D was not as expected from the MD free energy simulations. The similar crystallization yields of the LbADH WT and the mutant Q207D may be attributed to the relatively low numerical value of ∆∆G ≈ +2.4 kJ mol⁻¹.

This clearly demonstrates that experimental validation of theoretical simulation results is imperative. In situ microscopy with real-time image analysis has been shown to be a valuable tool for the online monitoring of the relevant protein crystallization performance indicators.

4. Conclusions

The in situ microscopy probe coupled with real-time machine learning-based image analysis for the online detection of protein crystals offers significant advantages over traditional methods in protein crystallization monitoring. This approach enables continuous, real-time observation of the crystallization process, allowing for the direct visualization of crystal formation and morphology without separating solid and liquid phases. This enhances practicality and reduces procedural steps [31,43].

In addition, the method provides detailed information on crystal size, shape, and morphology, which is crucial for understanding crystal mass distribution and improving product quality [22,43], as well as therapeutic efficiency [54].

Compared to conventional techniques, like UV/Vis measurements with centrifugation, the in situ method minimizes the impact of environmental factors such as temperature and viscosity fluctuations. This leads to more consistent and accurate measurements, even in the presence of amorphous precipitates [31,36]. Furthermore, it avoids the limitations of methods like FBRM, which primarily track size distributions and can be negatively affected by amorphous precipitates due to shifts in absorption wavelengths and changes in the protein environment [30,36].

Ultimately, the in situ microscopy method improves the precision of crystallization studies by providing real-time, comprehensive data and reducing the limitations of traditional analytical techniques [31,43]. In addition, the scalable 1 L crystallizer setup demonstrated here offers significant potential for industrial applications, where real-time control of crystallization kinetics is essential for maintaining consistent product quality and yield.