Design of transfections: Implementation of design of experiments for cell transfection fine tuning

Abstract Transfection is the process by which nucleic acids are introduced into eukaryotic cells. This is fundamental in basic research for studying gene function and modulation of gene expression as well as for many bioprocesses in the manufacturing of clinical‐grade recombinant biologics from cells. Transfection efficiency is a critical parameter to increase biologics' productivity; the right protocol has to be identified to ensure high transfection efficiency and therefore high product yield. Design of experiments (DoE) is a mathematical method that has become a key tool in bioprocess development. Based on the DoE method, we developed an operational flow that we called “Design of Transfections” (DoT) for specific transfection modeling and identification of the optimal transfection conditions. As a proof of principle, we applied the DoT workflow to optimize a cell transfection chemical protocol for neural progenitors, using polyethyleneimine (PEI). We simultaneously varied key influencing factors, namely concentration and type of PEI, DNA concentration, and cell density. The transfection efficiency was measured by fluorescence imaging followed by automatic counting of the green fluorescent transfected cells. Taking advantage of the DoT workflow, we developed a new simple, efficient, and economically advantageous PEI transfection protocol through which we were able to obtain a transfection efficiency of 34%.


| INTRODUCTION
The transfection process is extremely useful in basic research to study the role of a specific gene through experiments of both gain of function (DNA transfection) or loss of function (small interfering RNA transfection), as well as for modulation of gene expression, mutational analysis and recombinant protein production (Kim & Eberwine, 2010). Moreover, transfection is fundamental in several therapeutic approaches based on gene delivery strategies (Neshat et al., 2020;Pfeifer & Verma, 2001), for the genetic modification of cells of human origin, for example, for the generation and engineering of the induced pluripotent stem (iPS) cells to be used as models of disease and/or in patient-specific cell therapy (Nishikawa et al., 2008;Shankar et al., 2020;Takahashi & Yamanaka, 2006) as well as in many bioprocesses for the production of clinical-grade recombinant biologics (Wurm, 2004).

Cultivated mammalian cells have become the dominant system
for the production of recombinant proteins for clinical applications because of their capacity for proper protein folding, assembly, and posttranslational modifications. Biotherapeutics are generally secreted by genetically engineered mammalian cell lines into their culture medium, they can then be purified to homogeneity and sterilized under well-defined and regulated conditions. In 1986, human tissue plasminogen activator (tPA, Activase; Genentech) became the first therapeutic protein from recombinant mammalian cells to obtain market approval. Since then, monoclonal antibodies, fusion proteins, and enzymes have been produced to treat various pathologies such as cancer, and metabolic and autoimmune diseases (Wurm, 2004). Their successes have driven substantial increases in clinical trials over the years, resulting in a steady growth of the number of recombinant protein therapeutics on the market, with over 100 products approved by the US FDA in 2015, most of which were produced by genetically engineered mammalian cells (Kinch, 2016). Recently, the productivity of mammalian cells cultivated in bioreactors has been greatly enhanced through improvements in media composition and process control. A great deal of effort has been applied to the establishment of methods for efficient product scale-up and cost reduction of manufacturing processes. The development of a manufacturing process for a recombinant protein in mammalian cells follows a well-established scheme starting with the transfer of the recombinant gene with the necessary transcriptional regulatory elements to the cells by transfection (Wurm, 2004).
Chemical transfection methods were the first to be used to introduce foreign genes into mammalian cells and are still the most widely used methods. The underlying principle of chemical transfection relies on the formation of a complex between positively charged chemicals (calcium phosphate, polymers, and lipids) and negatively charged nucleic acids. Such complexes are subsequently attracted to the negatively charged cell membrane and pass through it, probably through mechanisms involving endocytosis and phagocytosis. Compared with the nonchemical ones, chemical methods have the merits of relatively low cytotoxicity, ease of use, costeffectiveness, with no mutagenesis, no viral vector involvement, no size limitation on the packaged nucleic acid, and no safety problems (Kim & Eberwine, 2010). The Achilles' heel of chemical transfection is the transfection efficiency, it being lower than in biological methods based on viral vectors and also highly variable in response to different factors. Depending on the cell type and the molecule to transfect, the correct protocol has to be identified to maximize transfection efficiency and product yield.
Traditional experimentation in cell biology, as elsewhere, has typically been conducted using a one-factor-at-a-time (OFAT) approach, in which every factor (variable) is kept constant except for the factor under investigation that is varied with the resulting output being measured. However, the complexity of cell processes requires the simultaneous examination of several input variables that must be controlled. Design of experiments (DoE) is a mathematical method for planning, conducting, analyzing, and interpreting controlled tests of multivariable processes. DoE allows to identify the factors that significantly influence the desired output, their interactions, and ultimately the best combinations of factors that maximize the output and optimize the process (Montgomery, 2012;Myers et al., 2016).
DoE has been widely used to maximize yields and improve processes at an early stage, leading to major benefits in both product performance as well as management of resources (Grangeia et al., 2020;Mandenius & Brundin, 2008;Politis et al., 2017;Weissman & Anderson, 2015). In recent years, the use of the DoE approach as an alternative to the traditional OFAT method is definitely increasing in many fields of scientific research, including cell biology, biochemistry, and nanotechnologies, to identify the main factors controlling the scientific process and for modeling their effects (Bollin et al., 2011;Durakovic, 2017;Esteban et al., 2021;Lanati, 2018;Mancinelli et al., 2015;Narenderan et al., 2019;Papaneophytou et al., 2021;Papaneophytou, 2019;Tavares Luiz et al., 2021;Toms et al., 2017;Xu et al., 2020). In this scenario, we have been committed for many years in developing, applying, and validating models that can be useful for the management or research activities, helping in standardizing, and optimizing processes, thus improving data reliability and reproducibility Digilio et al., 2016;Mancinelli et al., 2015;Mascia et al., 2020).
In the present study, we focused on the application of DoE to the analysis, modeling, and optimization of a cell transfection protocol.
Previous DoE applications to cell transfection mainly focused on CHO or HEK cell lines, although they differed for the factors analyzed, the methodological approach proposed, the transfection output selected, and the relative measurements, ranging from the titer and quality of the produced antibody to fluo-cytometer or fluorimetric analysis of GFP expression (Agirre et al., 2015;Bollin et al., 2011;Cervera et al., 2015;Elshereef et al., 2019;Thompson et al., 2012). Ultimately, the outcome of these studies pointed to the suitability of cell transfection for DoE application in different contexts. As far as we could ascertain, our study applies for the first time the DoE method to noncommercial cell lines, specifically neural progenitor cells that are widely recognized as difficult to transfect (Alabdullah et al., 2019;Karra & Dahm, 2010). We choose to test polyethyleneimine (PEI) as a transfection reagent for its ease of use and its relatively low costs (Neuberg & Kichler, 2014) and to analyze the effect of PEI type, PEI concentration, DNA concentration, and cell density on cell transfection efficiency as measured by fluorescence imaging. We clearly defined a sequential approach, using first a two-level full factorial design to study the effect of each factor on transfection and all the possible factor interactions. Second, a response surface methodology was applied to identify the best factor combinations that optimize transfection as well as to develop a predicting model describing the relation between transfection efficiency and its most influential factors. This flexible operational flow, aimed at obtaining an effective, standardized, and reproducible cell transfection procedure, suitable for different cell types and transfection reagents, we called "Design of Transfections" (DoT).

| Cell culture and transfection
Mes-c-myc A1 (A1) cells are noncommercial immortalized progenitors derived from mesencephalon of mouse embryos at 11 days of development and infected with a replication-defective retrovirus bearing c-myc, expressing both markers of neural precursors as well as neuronal markers (Colucci-D'amato et al., 1999). A1 cells ( Figure 1a) were cultured in Minimum Essential Medium (Gibco™) and F12 medium (Gibco™) 1:1 (vol/vol) supplemented with 10% fetal bovine serum (Gibco™). In these conditions, the cells proliferate and maintain all the characteristics of neural progenitors (Colucci-D'amato et al., 1999). The plasmid chosen for transfection is pEGFP-N1 (Clontech) carrying the coding sequence for green fluorescence protein (GFP) downstream to the promoter of cytomegalovirus ( Figure 1a). To set the transfection protocol, we used as reference the protocol developed by Ming Hsu and UludaĞ (2012) to efficiently transfect primary tissue-derived cells (fibroblasts and bone marrow stromal cells) using PEI (Ming Hsu & UludaĞ, 2012), and adapted it to our cellular system (Figure 1b). We tested both 25 kDa branched (BPEI25) and 22 kDa linear PEI (LPEI22). Twenty-four hours before transfection, A1 cells were seeded in 24-well plates with 0.5 ml of medium/well at the densities fixed in the experimental design. Complexation between DNA and PEI was performed through a two-part mixing in culture medium. Moreover, for each factor analyzed, we tested different levels according to the DoE approach.
According to the experimental design, different volumes of PEI, linear or branched, (1 μg/ml, pH 7.0) (SIGMA), were added in a 1.5-ml tube to antibiotic-and serum-free culture medium to a final volume of 50 μl. In a second tube, the DNA volumes indicated by the experimental design were added to antibiotic-and serum-free culture medium to a final volume of 50 μl as well. The two solutions were vortexed and incubated for 10 min at room temperature (RT) to allow DNA and PEI to dissolve properly in the medium, then were mixed together. The resulting transfection solution (final volume of 100 μl) was vortexed and incubated 15 min at RT to allow the formation of PEI-DNA complexes and then was added drop by drop to the A1 cell culture, and left in the cell incubator at 37°C in 5% CO 2 and 20% O 2 .
The transfection medium was replaced after 16 h of incubation with normal culture medium.

| CellProfiler pipeline for transfection efficiency computation
Twenty-four hours after transfection, cells were fixed in 4% paraformaldehyde for 15 min at RT, and Hoechst (Thermo Fisher Scientific) counterstained following the manufacturer's instructions. Cells transfected with pEGFP-N1 plasmid acquired the ability to express GFP, and could then be visualized by means of fluorescence microscopy. Ten randomly chosen areas for each design run were captured as images and further processed using a specifically generated pipeline and the Cell-Profiler image analysis software. Single-channel 8-bit images were uploaded in the proper section of the CellProfiler software and a unique text pattern was chosen to identify each channel (e.g., green channel image = ch01; blue channel image = ch02). The identification of both Hoechst counterstained nuclei (in blue) and GFP positive cells (in green) was obtained through an unbiased segmentation algorithm and generation of a biunivocal relation between nuclei (parent objects) and GFP positive cells (child objects) to filter transfected cells. The percentage of transfected cells was calculated as the ratio between the number of child objects over the number of parent objects.

| Experimental design
First of all, factors to be tested for their effect on transfection efficiency were identified together with a suitable value range. The designs of the experiments were then generated through Minitab Statistical Software version 16 and version 19 (www.minitab.com; Minitab Inc.), following suggestions by the Quality Companion 3 by Minitab. The first experimental design chosen for this study was a two-level full factorial design, including 2 k different combinations (where k represents the number of factors analyzed) and two possible values or levels, high (+1) and low (−1) for each factor (Montgomery, 2012;Myers et al., 2016). With four factors, the design included 2 4 = 16 different combinations, that became 32, since we chose to perform the tests in duplicate. We then performed a successive protocol optimization by using the following response surface designs: (i) the Box-Behnken design (BBD), which includes a number of combinations to tests N = 2 k(k − 1) + C 0 , where k is the number of factors, and C 0 is the number of central points (Ferreira et al., 2007) and (

| Modeling and validation
The data obtained from the full factorial and both the response surface designs, keeping constant the optimized levels of PEI type and cell density, were combined and analyzed together by using MiniTab statistical software to generate a response surface plot and a contour plot. The plots describe how the combination of the influential factors affects transfection efficiency (the response output) and allow us to identify the setting of factors that maximizes the output. To validate the model, we tested, in triplicates, two different factor settings, corresponding to the optimal conditions found (6.5 μg/ml LPEI22, 1 μg/ml DNA, 25,000 cells/cm 2 ) and a suboptimal one (5 μg/ml LPEI22, 2 μg/ml DNA, 25000 cells/cm 2 ). The protocol used was the one described above. Two other contour plots reporting the transfection efficiency with respect to PEI/DNA ratio and DNA or PEI respectively were generated to identify the best ratio range.

| Experimental setting
Chemical transfection efficiency varies depending on cell type, genetic material to be introduced, and chemical method adopted, and is largely affected by several parameters, both quantitative and qualitative (Kim & Eberwine, 2010). Here, we present a DoT operational flow to identify the parameters that significantly affect transfection, and, then, to identify the best setting which maximizes transfection efficiency.
As proof of concept, we chose to transfect cells of neural origin that are widely recognized as extremely difficult to transfect and with a low response to traditional lipidic transfection methods (

| Factor 4 (quantitative): cell density
The density of cells during transfection is closely linked to the polymer and DNA concentrations. If cell density is low, the concentration of polymer would be relatively high compared with transfection in high-density cell culture and might affect cell viability (Mancinelli et al., 2015;Ming Hsu & UludaĞ, 2012). As in many transfection protocols, we chose to refer to the seeding density, which can be measured more precisely, even though it might not necessarily resemble the attached cell density, depending on culturing conditions, handling processes, and age of cell culture. The cell density levels chosen were the ones for which A1 cells were shown to be in a logarithmic phase of proliferation, fundamental for a good DNA uptake efficiency (Mancinelli et al., 2015). at 25,000 cells/cm 2 ; 11.27% MTE for LPEI22 against 8.38% for BPEI25 at 50,000 cells/cm 2 ). Finally, the three-way interaction among the type of PEI, cell density, and DNA concentration was also found significant (p = 0.014; data not shown), confirming that the optimal ratio between PEI positive charges and DNA negative charges for cell transfection changes depending also on cell seeding density.

| Factorial design and identification of significant factors and interactions
The interaction analysis clearly showed how the traditional OFAT approach, which does not take into account all the possible interactions among factors, might severely limit the optimization of a multivariable assay, possibly leading to the identification of a relative maximum output thus hiding the potential best achievable one.
Overall, the main effect and interaction analysis ultimately determined LPEI22 as the PEI conformation to use with A1 cells.

| Optimization design
Once fixed LPEI22 as the PEI conformation to use, PEI concentration, DNA concentration, and cell density, all significant with respect to the output, were analyzed in the subsequent optimization design to obtain a model describing the transfection efficiency variability as a combination of interacting significant factors. For the optimization, we took advantage of the response surface designs, mainly the BBD and the CCD, used to identify the points of absolute maximum and to highlight possible nonlinearities, by adding points to the factor space analyzed (Lanati, 2018;Myers et al., 2016). The addition of more points to the factor space strengthened the analysis and let us model more precisely the relationship among the output variable and the input factors.
We first ran a BBD, which for three factors requires fewer combinations than the CCD. The BBD takes into account the midpoints of edges of the process space as well as the center levels ( Figure 3a), and, for a three factors' analysis, includes 15 runs (instead of the 20 required for CCD), becoming 30 with replicates, as summarized in Figure 3b. All the combinations analyzed and the relative transfection efficiencies are reported in the bar plot in Figure 3c and detailed in Figure S1a. The statistical analysis showed normal distribution and variance of the residuals ( Figure S1B). Noteworthy, following the analysis of the response surface regression (Figure 3d), the square term of the DNA concentration showed significance (p < 0.00001), indicating the presence of a statistically significant curvature in the main effect of this factor on the transfection output.
Conversely, the square term related to PEI concentration was not significant (p = 0.502), indicating that the relation between PEI concentration and transfection efficiency is linear within this interval ( Figure 5d). These data were clearly shown in the main effect plot in Figure 3e, highlighting for DNA concentration (but not PEI concentration) a significant deviation of the effect at the center point (MTE = 10.12%), with respect to the average mean response of the factors at their low and high levels, that is 6.71% and 2.33%, respectively ( Figure 3e). Moreover, with the addition of the midpoints, cell density, whose standardized effect was the least significant in the factorial experiment (Figure 2c,d), no longer showed a significant effect on the experimental output (p = 0.421, Figure 3d). Therefore, this factor did not require any additional optimization and was set constant at the lower level of 25,000 cells/cm 2 to save time and experimental material. The variation of the transfection output depending on the simultaneous variation of both PEI and DNA concentration (keeping cell density at 25,000 cells/cm 2 ) was reported in both a three-dimensional response surface plot ( Figure 3f) and a twodimensional contour plot (Figure 3g). The contour plot highlighted a response surface "rising ridge," with the best output lying at the edge F I G U R E 3 (See caption on next page) of the plot at the high level of PEI concentration and centered mainly in the lower half of the analyzed DNA concentration range.
As a further step, we refined the analysis by focusing on lower levels of DNA concentration and slightly higher levels of PEI concentration (up to 7.75 μg/ml), although corresponding to a decreased cell viability moving close to 50% (Mancinelli et al., 2015). The other two factors, PEI type and cell density were kept constant. With this aim, we took advantage of a CCD, centered on the most promising area of the factor space, to provide high-quality predictions by adding more points and above all by exploring also factor levels outside the reference interval investigated in the previous designs ( Figure S2).
The CCD includes star points that are fixed at a distance from the design center identified by the α parameter, set at the standard value probably also due to their cytotoxic effects.

| Modeling and validation
By combining the output data obtained from full fractional, BB and CC designs, keeping constant the optimized levels of PEI type and cell density, a final model was obtained, that we named the DoT model ( Figure 5). The DoT model is able to predict the transfection output at different PEI and DNA concentrations and to clearly identify a response surface peak corresponding to the optimized transfection setting. The resulting response surface plot (Figure 5a) and, even more so, the contour plot (Figure 5b) highlighted the best setting of PEI and DNA concentration, able to maximize the transfection efficiency. Finally, to validate the DoT model, we tested two different factor settings: (1) a suboptimal one (5 μg/ml LPEI22, 2 μg/ml DNA, 25,000 cells/cm 2 ), corresponding to the amount of DNA and PEI set in the transfection protocol optimized for primary tissue-derived cells, but using BPEI polymer (Ming Hsu & UludaĞ, 2012), and (2) the optimized conditions here identified (6.5 μg/ml LPEI22, 1 μg/ml DNA, 25,000 cells/cm 2 ). For each factor setting, three independent experiments were performed and the transfection efficiencies were measured ( Figure 5c). It is worth noting that using the optimized factor setting, we obtained a quite satisfactory MTE of 34%, whereas the suboptimal setting resulted in a mean efficiency of about 6%. In both cases, we obtained the output value predicted by the DoT model (Figure 5b), confirming its validity. These results clearly show the effectiveness of the DoT method in optimizing transfection efficiency.
We also took into account the PEI/DNA ratio, which refers to the balance between PEI negative charges and DNA positive ones, a key factor in transfection. The relationship between the transfection efficiencies obtained in the conditions tested in all the three experimental designs and their relative PEI/DNA ratios is extremely variable (Figure 5d). However, the best outputs at the different ratios are usually obtained by using as transfectant the LPEI molecule, with the exception where the PEI/DNA ratio is close to 1 (1.32), for which BPEI is preferable to LPEI. This is probably due to the higher number of negative charges on the branched conformation, which causes, albeit minimally, an excess of positive charges that seems to favor transfection at this low PEI/DNA ratios. Drawing the contour plots for transfection efficiency vs. PEI/DNA ratio and DNA or PEI, respectively, we highlighted two ranges of PEI/DNA ratio, named A and B, predicting the highest transfection output (Figure 5e,f). The amount of transfectant used per μg of DNA obviously has also economic implications, therefore, lower values, corresponding to the A interval, are preferable, especially when an expensive reagent is used. The optimal conditions identified and tested, that is 1 μg/ml of DNA and 6.5 μg/ml of PEI, correspond to a value of 6.5 for the PEI/ DNA ratio, which approximately matches the lower limit of the A interval, also allowing transfection reagent costs to be minimized.

| CONCLUSIONS
In our DoT assay, we analyzed the impact on neural progenitor transfection of four independent factors, namely PEI concentration, PEI type, DNA concentration, and cell density. Our results highlight that: (i) LPEI is associated to higher transfection efficiency than BPEI; (ii) DNA concentration and especially PEI concentration levels strongly influence transfection efficiency, and, related to this, (iii) the PEI/DNA ratio used for transfection is a good indicator of the obtainable efficiency. Finally, we identified an optimized setting of transfection conditions for neural progenitors (6.5 μg/ml LPEI22,

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available in the supplementary material of this article. Additional data are available from the corresponding author upon reasonable request.

Sara Mancinelli
http://orcid.org/0000-0002-6290-5425 Giovanna Lucia Liguori https://orcid.org/0000-0002-7239-3224 F I G U R E 6 Design of transfections (DoT) workflow. Rectangles represent processes whereas rhombuses represent flow checkpoints. OK indicates that it is possible to continue to the next step, whereas KO indicates that the checkpoint has not been overcome and the step indicated by the arrow has to be repeated. Along Path (1), an OPTIMIZATION stage, through response surface designs, allows to model transfection and, then, to identify the best factor setting that optimizes transfection efficiency (OUTPUT). Optimization can be refined until a good model describing the factor space is obtained. In some cases, the factorial design and the subsequent statistical analysis might be sufficient to identify a satisfactory output (Path 2)