The design-of-experiments procedure presented here allows the evaluation of different flocculants in terms of their ability to aggregate dispersed particles in plant extracts, thus reducing turbidity and the costs of downstream processing.
Plants are important to humans not only because they provide commodities such as food, feed and raw materials, but increasingly because they can be used as manufacturing platforms for added-value products such as biopharmaceuticals. In both cases, liquid plant extracts may need to be clarified to remove particulates. Optimal clarification reduces the costs of filtration and centrifugation by increasing capacity and longevity. This can be achieved by introducing charged polymers known as flocculants, which cross-link dispersed particles to facilitate solid-liquid separation. There are no mechanistic flocculation models for complex mixtures such as plant extracts so empirical models are used instead. Here a design-of-experiments procedure is described that allows the rapid screening of different flocculants, optimizing the clarification of plant extracts and significantly reducing turbidity. The resulting predictive models allow the identification of robust process conditions and sets of polymers with complementary properties, e.g. effective flocculation in extracts with specific conductivities. The results presented for tobacco leaf extracts can easily be adapted to other plant species or tissues and will thus facilitate the development of more cost-effective downstream processes for commodities and plant-derived pharmaceuticals.
Plants are widely used to produce food commodities such as fruit juices, but they can also be developed as platforms for the manufacture of higher-value biopharmaceutical products 1-3. In both cases, downstream processing (DSP) often begins with the extraction of liquids from tissues such as leaves or fruits, followed by the clarification of particle-laden extracts 4,5. For the manufacture of biopharmaceuticals, the costs of DSP can account for up to 80% of the overall production costs 6,7 and this in part reflects the high particle burden present in extracts prepared by disruptive methods such as blade-based homogenization 8,9. Although the rational selection of filter layers to match the particle size distribution in the extract can increase filter capacity and reduce costs 10,11, the improvement can never exceed the ceiling of absolute capacity defined by the number of particles that must be retained per unit of filter area to achieve clarification.
The ceiling can be lifted if fewer particles reach the surface of the finest filters in the filtration train, and this can be achieved if dispersed particles are mixed with polymers known as flocculants that promote aggregation to form large flocs 12. Such flocs can be retained further upstream by coarser and less expensive bag filters, reducing the particle burden reaching the finer and more expensive depth filters. The polymers must have safety profiles suitable for their applications, e.g. for biopharmaceuticals they must be compliant with good manufacturing practice (GMP), and typically they must have a molar mass >100 kDa and can either be neutral or charged 13. Whereas neutral flocculants generally act by cross-linking dispersed particles causing their aggregation and the formation of flocs with diameters >1 mm 11, charged polymers neutralize the charge of dispersed particles, reducing their solubility and thus causing precipitation 14.
Flocculation can be improved by adjusting parameters such as buffer pH or conductivity, and the polymer type or concentration, to match the properties of the extract 15,16. For tobacco extracts pretreated with 0.5-5.0 g L-1 polyethylenimine (PEI), a greater than 2-fold increase in depth filter capacity was reported in a 100-L pilot-scale process. The cost of this polymer is less than €10 kg-1 so its introduction into the process resulted in cost savings of about €6,000 for filters and consumables per batch 16 or even more when combined with cellulose-based filter aids 17. Even so, predictive models are required to evaluate the a priori economic benefits of flocculants because their inclusion can require hold steps of 15-30 min 16,18, resulting in further investment costs for storage tanks. However, there are currently no mechanistic models available that can predict the outcome of such experiments due to the complex nature of flocculation. Therefore, a more appropriate design-of-experiments (DoE) approach 19 was developed as described in this article. A protocol for the general DoE procedure has recently been published 20.
Small-scale devices are now available for the high-throughput screening of flocculation conditions 21. However, these devices may not realistically simulate conditions during the flocculation of plant extracts because the dimensions of the reaction vessel (~7 mm for wells on a 96-well plate) and the particles or flocs can be less than an order of magnitude apart. This can affect mixing patterns and thus the predictive power of the model. Furthermore, it can be difficult to scale down processes involving precipitation due to non-linear changes in the mixing behavior and precipitate stability 22. Therefore, this article outlines a bench-top-scale screening system with a throughput of 50-75 samples per day, yielding results that are scalable from the initial 20 ml reaction volume to a 100 L pilot-scale process 16. When combined with a DoE approach, this allows the predictive models to be used for process optimization and documentation as part of a quality-by-design concept.
The method described below may also be adapted to biopharmaceuticals produced in cell culture-based processes, where flocculants are also being considered as a cost-saving tool 23. It can also be used to model the precipitation of target proteins from a crude extract as part of a purification strategy, as demonstrated for β-glucuronidase produced in canola, maize and soybean 24,25. A detailed description of flocculant properties can be found elsewhere 16,26 and it is important to ensure that the polymer concentrations are either non-toxic or below harmful levels in the final product 11.
1. Develop an Adequate Experimental Strategy
2. Prepare the Flocculation Experiments
Figure 1: Plant extract flocculation workflow: process scale (left) and benchtop scale (right). Following protein extraction with aqueous buffers, dispersed particles of cell debris are aggregated by the addition of flocculants. The aggregates are then removed by a cascade of bag and depth filtration and the capacity of these filters along with the filtrate turbidities can be used directly to measure the efficiency of flocculation.
3. Flocculate the Plant Extracts with Different Polymers
4. Evaluate the DoE
5. Improve the Model and Verify the Predictive Power
Flocculation of tobacco extract with different polymers
The method described above was successfully used to develop a process for the flocculation of tobacco extracts during the manufacture of a monoclonal antibody (the HIV-neutralizing antibody 2G12) and a fluorescent protein (DsRed) (Figure 1) 16, and has since been transferred to other proteins including lectins, malaria vaccine candidates and fusion proteins (unpublished data). Typically, the application of flocculants reduced the turbidity of bag-filtered plant extract from ~6,000 NTU (10,000 NTU after extraction) to ~1,000 NTU. In an initial screening experiment, a 91-run IV-optimal design was used to test 18 different polymers in three different concentrations (because this factor affects flocculation efficiency 13,27) and observed flocculation over a ~12 hr incubation period (Figure 2A and B). The long incubation period can be important to identify meaningful time frames for the flocculation process. Also pH values of 4-8 were tested because these may be relevant in future processes due to the properties of specific target proteins 13,25,27,37. Among the 18 tested polymers, six were found to reduce extract turbidity after bag filtration in typical extracts with a conductivity of 25 mS cm-1.
The model was refined by excluding all ineffective polymers in two iterations and then including additional process parameters, such as conductivity in the 15-45 mS cm-1 range, an incubation time of 5-75 min and temperatures of 4-30 °C, to generate models suitable for a wider range of process conditions. The predictive power of the model increased after each iteration, resulting in a highly reliable model (Figure 3A).
After four iterations, the highly-charged cationic and branched polymer PEI was found to be the most efficient for the aggregation of dispersed particles in tobacco extracts. However, the efficiency of this polymer declined with increasing extract conductivity. The properties molecule size, charge, structure (branched or linear), charge density and degree of amine substitution (primary, secondary, tertiary or quaternary) were tested as factors in a DoE and the last two parameters had the largest effect. The details have been reported elsewhere 16. Based on this knowledge of polymer properties from the DoE results, five other polymers were selected with molecular characteristics similar to PEI (charge density >meq g-1 and quaternary amine). One of these five polymers demonstrated greater flocculation efficiency at higher conductivities (Figure 3B) 11.
As part of the DoE approach, it was confirmed that none of the PEIs affected product recovery under any of the tested conditions. Indeed the capacity of depth filters used subsequently to remove remaining dispersed solids increased by a factor of 3.2-5.7, reaching ~110 L m-2 depending on the filter type. These results were also confirmed in a 100 L pilot-scale process, in for which the application of flocculants reduced the clarification-related production costs by >50% and the total production costs by ~20%.
Figure 2: Efficiency of different flocculants under diverse process conditions. (A) Extract samples directly after flocculation and bag filtration can still appear turbid. (B) After settling for several hours, the turbidity of the same samples can be reduced significantly. However, turbidity values obtained immediately after filtration are often preferable because extended hold times may not be possible in large-scale manufacturing processes. (C) Flocculation is also effective when applied to plant extracts generated with a screw press instead of a blender as indicated by the clear red liquid at the bottom of the 50 ml tubes (the red color is due to the presence of the fluorescent protein DsRed). (D) Mixtures of different flocculants can also induce flocculation.
Figure 3: Modeling flocculation using a DoE approach. (A) The accuracy of the model predictions increased as the number of polymers in the model was reduced from initial screening to refinement even though the number of process parameters increased from two to five. (B) Switching the polymer type (here from one PEI to another) as a consequence of a change in process parameters (here conductivity) maintains efficient particle flocculation and corresponding low filtrate turbidity compared to non-treated control extract (solid red line). Error bars in A and B indicate standard deviations of model predictions. Dashed red lines indicate standard deviations of the non-treated extract (n = 10). Please click here to view a larger version of this figure.
Flocculation of tobacco extracts prepared with a screw-press
The flocculation results were also transferred from tobacco extracts prepared with a homogenizer to those prepared with a screw-press, which generated fewer dispersed particles in the mm size range but more particles in the µm size range. In a 29-run IV-optimal design, it was shown that PEI is also effective for this type of extract in a similar concentration range and that the recovery of target proteins is not affected (Figure 2C). This shows (i) that flocculation conditions identified for one type of feed stock can be to some extent transferred to other feed stocks, saving time during process development, and (ii) that the DoE strategy can be used to confirm this transferability not only for individual process conditions but over the entire design space.
Flocculation experiments with flocculant mixtures
Combinations of flocculants can be more effective than single polymers, e.g. due to more enhanced bridging between particles 12. Therefore, the method described above was adapted to accommodate the addition of two polymers (3.2) 26. Three non-synthetic polymers were tested alone, in combination with each other or combined with PEI. The most efficient flocculation of tobacco extracts was achieved with PEI alone, but a combination of PEI and chitosan or polyphosphates can reduce the concentration of PEI required. Furthermore, the DoE approach allowed us to identify the most effective polymer combinations when omitting PEI (with or without chitosan and polyphosphates), thus helping to define optimal flocculation conditions in processes where PEI is incompatible with the target protein, e.g. due to precipitation, as reported for βglucuronidase 24,25. Furthermore, the DoE was able to characterize a complex design space for which no mechanistic model was available (Figure 2D). Using the ANOVA tools of the DoE software it was possible to distinguish between reliable predictive models and poorly-evaluated counterparts (Figure 4).
The most important aspect to consider when setting up a DoE to characterize particle flocculation is that the design must in principle be able to detect and describe the anticipated or possible effects 36,38, e.g. the influence of pH, polymer type and polymer concentration 16. Therefore, it is important to evaluate the fraction of design space (FDS) before starting the actual experiments. The FDS is the fraction of the multidimensional experimental space (covered by the design factors, e.g. pH) within which it is possible to detect pre-defined differences between two experimental outcomes given a system of known variability, e.g. detecting a difference in turbidity of 250 NTU given a variability of 125 NTU. The FDS can be increased by augmenting the design with additional runs and should be ≥0.95 for designs intended to guide process control 36. Furthermore, if the number of runs does not permit the entire experiment to be carried out in a single day, blocks should be pre-defined in the DoE to account for batch-to-batch and day-to-day variability. When working with plant material, the inclusion of reference runs in each block (e.g. non-treated controls) helps to compensate for variability, allowing the comparison of data from several runs each normalized to their corresponding reference run. In this context, increasing the number of replicate runs in the DoE is also useful.
When large numbers of polymers are screened, it is advisable to use the individual properties of the flocculants, e.g. charge density and molecular mass, as discrete numeric factors rather than the polymers themselves as categorical factors. This reduces the number of experiments because experimental designs often need to be replicated for categorical factors, whereas additional levels of numeric factors only need a small number of extra runs. The information content of the experiment also increases and allows the identification of polymer properties that improve flocculation, e.g. a high charge density as found in the experiments described here. CCD and RSM experimental designs are useful to establish models with high predictive power, allowing the identification of robust processing conditions (e.g. to guide process control) and are typically used to follow up screening designs. If the number of factors and factors levels under investigation results in DoE with more than 400 individual experiments, it may be advisable to reduce the number of factor levels or switch to other design types because the number of sample that can be easily handled with the technique presented here is limited to ~100 per day.
From an experimental point of view, polymers must remain stable under the selected experimental conditions, e.g. they must not depolymerize at low pH. Careful preparation of the flocculant stocks in terms of concentration is also necessary to obtain reproducible results and high-quality models. In this context, the flocculant may need to be pretreated, e.g. swelling times or pH adjustment for chitin, to ensure complete solubilization, and thus to obtain a homogeneous solution. Highly viscous stocks should be avoided because this can cause pipetting errors when transferring the polymer to the extract. Many polymers can have a strong buffering effect and the stocks have extreme pH values, e.g. pH ~9.5 for 8% [w/v] PEI. This can affect the pH of the extract if the stocks are not pre-adjusted and will distort the experimental results. For example, if flocculation is more effective at high pH and a non-pH adjusted PEI stock is used then a DoE might suggest that high PEI concentrations are more effective. However, this effect will be caused by the higher pH caused by the larger volume of stock that was added, not by the increased polymer concentration per se. The stock concentrations used should also resemble those used in large-scale applications to avoid differing dilution effects between the scales that can affect the particle concentration and thus flocculation. Some clay-based flocculants such as kaolin contain a large number of fine particles themselves which can mask the flocculation effect, e.g. turbidity reduction after initial filtration, and other responses should be selected to evaluate the efficiency of these substances, e.g. downstream filter capacity.
For data analysis it is important to evaluate the collected results in terms of extreme values, misalignments and general consistency, e.g. extreme values can indicate a copy-paste error, a shift in the decimal place or a malfunction of equipment/analytical devices. A thorough analysis will ensure that only high-quality data are used for model building. During model building it is important to constantly assess the broad set of quality indicators provided by the DoE software. The most basic criteria are the R2, adjusted R2 and predicted R2 values, but normal residuals, residuals-vs-run and actual-vs-predicted plots (Figure 4) are even more important because they provide information about each run in an experiment rather than a sum parameter. Furthermore, the coherence of the final model and its predictions with the known mechanisms of flocculation should always be investigated. Major discrepancies between predictions and scientific expectations may occur because DoE models are only descriptive rather than mechanistic, e.g. models may predict extreme values at the edges of a design space reflecting the use of polynomial fitting algorithms.
Figure 4: Quality indicators of DoE models. The normal plot of studentized residuals should resemble a straight line as closely as possible (A) with only minor deviations (green arrows) acceptable for high-quality models. A curved appearance (C) with strong deviations (red arrows) from the ideal line (red) indicates a poor model, e.g. due to missing significant factors. Ultimately, the predicted and experimental (actual) values should match (B) and again follow a straight line. Deviations from the ideal line (red circle and dashed line) indicate poor model predictions (D). Please click here to view a larger version of this figure.
The DoE approach can help to characterize flocculation in complex feed stocks such as plant extracts, even if there are no existing data. The flocculation of tobacco extracts was optimized with a work load of 2 weeks and consumables costs of ~ €500. This reduced the number of depth filters required for a single pilot-scale batch involving ~800 L of plant extract by 60%, which achieved a corresponding reduction in consumables costs.
The flocculants were also applied to different plant extracts and to cell culture homogenates. Although the same flocculant was effective for all of these feed stocks, the polymer concentration had to be adjusted in order to accommodate the different concentrations of dispersed particles. Additionally, once an effective polymer has been identified, the filtration and/or centrifugation steps may need to be adjusted to match the different particle size distribution 11.
The method described here can easily be adapted to other feed stocks and is therefore also relevant for scientists and engineers developing clarification strategies for mammalian cell cultures and food/feed production processes. Especially plant-based processes will benefit from the intermediate sample volumes suggested here because plant extracts can contain particles up to 1 mm in diameter which are incompatible with microplate formats 21, e.g. because the mixing dynamics differ due to a particle diameter to vessel diameter ratio that is not representative of the process scale.
The authors have nothing to disclose.
I would like to acknowledge Dr. Thomas Rademacher for providing the transgenic tobacco seeds and Ibrahim Al Amedi for cultivating the tobacco plants. I wish to thank Dr. Richard M Twyman for editorial assistance and Prof. Dr. Rainer Fischer for fruitful discussions. This work was funded in part by the European Research Council Advanced Grant ”Future-Pharma”, proposal number 269110, the Fraunhofer-Zukunftsstiftung (Fraunhofer Future Foundation) and the Fraunhofer-Gesellschaft Internal Programs under Grant No. Attract 125-600164.
2100P Portable Turbidimeter | Hach | 4650000 | Turbidimeter |
2G12 antibody | Polymun | AB002 | Reference antibody |
Biacore T200 | GE Healthcare | 28-9750-01 | SPR device |
BP-410 | Furh | 2632410001 | Bag filter |
Catiofast VSH | BASF | 79002360 | Flocculating agent |
Centrifuge 5415D | Eppendorf | 5424 000.410 | Centrifuge |
Centrifuge tube 15 mL | Labomedic | 2017106 | Reaction tube |
Centrifuge tube 50 mL self-standing | Labomedic | 1110504 | Reaction tube |
Chitosan | Carl Roth GmbH | 5375.1 | Flocculating agent |
Design-Expert(R) 8 | Stat-Ease, Inc. | n.a. | DoE software |
Disodium phosphate | Carl Roth GmbH | 4984.3 | Media component |
Ferty 2 Mega | Kammlott | 5.220072 | Fertilizer |
Forma -86C ULT freezer | ThermoFisher | 88400 | Freezer |
Greenhouse | n.a. | n.a. | For plant cultivation |
Grodan Rockwool Cubes 10x10cm | Grodan | 102446 | Rockwool block |
HEPES | Carl Roth GmbH | 9105.3 | Media component |
K700P 60D | Pall | 5302305 | Depth filter layer |
KS50P 60D | Pall | B12486 | Depth filter layer |
Miracloth | Labomedic | 475855-1R | Filter cloth |
MultiLine Multi 3410 IDS | WTW | WTW_2020 | pH meter / conductivity meter |
Osram cool white 36 W | Osram | 4930440 | Light source |
Phytotron | Ilka Zell | n.a. | For plant cultivation |
Polymin P | BASF | 79002360 | Flocculating agent |
POLYTRON PT 6100 D | Kinematica | 11010110 | Homogenization device with custom blade tool |
Protein A | Life technologies | 10-1006 | Antibody binding protein |
Sodium chloride | Carl Roth GmbH | P029.2 | Media component |
Synergy HT | BioTek | SIAFRT | Fluorescence plate reader |
TRIS | Carl Roth GmbH | 4855.3 | Media component |
Tween-20 | Carl Roth GmbH | 9127.3 | Media component |
VelaPad 60 | Pall | VP60G03KNH4 | Filter housing |
Zetasizer Nano ZS | Malvern | ZEN3600 | DLS particle size distribution measurement |