3.2.1. Media composition

To maintain long-term precision in HTP processes, screening groups often select highly defined culture media with minimal lot-to-lot variability. Some basal media have consisted of defined mixtures of amino acids along with essential enzyme cofactors and trace metals (Postma et al., 1989); other media simply leave the strain to produce its own amino acids from a nitrogen source such as ammonium or urea (Ponomarova et al., 2017). If amino acid biosynthesis is heavily challenged in a manufacturing process due to minimal media or high nitrogen flux to product, it is advisable to consider these factors during screening media selection. It is also wise to consider the scalability of media components while designing a plate model. If strains are iteratively engineered in media containing a reagent that is expensive, environmentally unfriendly, or otherwise not scalable, the performance of the resulting strain could depend critically on a process that cannot be commercialized.

Maintenance of pH is critical to ensure consistency in the chemical environment of cultures; large drifts in pH can affect the performance of engineered strains as well as the stability and solubility of products. In contrast to lab-scale bioreactors, it is not feasible to control pH in individual wells by way of an acidic or basic feed, as the measurement of pH in each well and reagent additions necessary for pH adjustment pose major challenges. Many plate models therefore take advantage of buffering media components, such as succinate (Westfall et al., 2012), 2-(N-morpholino)ethanesulfonic acid (MES) (Nandy et al., 2015), and (3-(N-morpholino)propanesulfonic acid) (MOPS) (Neidhardt et al., 1974). A caveat with the use of succinate is that it may itself serve as a carbon source for some microbes, which can impact the physiology of the strain and deplete the buffer over time. Buffers can be added in relatively high molar concentrations, on the same order of magnitude as the carbon source, to reduce the risk of acid or base generation in excess of the buffering capacity. Alternatively, slow-release polymer discs containing alkaline compounds have also been used to control pH in some shake-flask fermentations, allowing a significant reduction in the buffer required (Scheidle et al., 2011). A similar strategy could theoretically be applied in a plate model.

3.2.2. Subculturing for multistage process simulation and biomass normalization

“Switch”-controlled fermentation processes used in manufacturing and in lab-scale models can significantly impact the strain metabolic state, potentially interacting with designed or random genetic modifications in engineered strains. However, this process typically involves media exchange, which is not easily accomplished in plate models. To induce physiological shifts that recapitulate the transition made during switch activation at manufacturing scale, strains in plate models can be subcultured, diluting into secondary media that triggers the switch (Chua et al., 2016).

Subculturing also provides an opportunity to standardize the inoculum for each well of the microplates. Although automation can enable picking of up to 20,000 colonies per day (Anzai et al., 2017), even the most advanced robotics will inoculate a variable number of viable cells, which can negatively impact precision among biological replicates. To address this, relatively uniform cell densities can be achieved in parallel “precultures” using common practices for protein overexpression: cells are grown in batch culture to stationary phase, which then allows precise inoculation of production culture from an aliquot of the preculture (Berrow et al., 2006). More sophisticated fed-batch precultures have been shown to give even more uniform performance by eliminating stationary-phase incubation and the associated growth lag following subculture (Keil et al., 2019b).

3.2.3. Feeding strategies

Because of its operational simplicity, batch mode is the most used feeding scheme for plate models. As discussed in Section 2.1, carbon sources, nutrients, and buffers available in the culture are added to the growth medium before inoculation (Jones and Kompala, 1999). A consequence of running cultures in batch mode is the risk of activating overflow metabolism, which occurs for many microbes when high concentrations of feedstock are readily available (Xu et al., 1999; Keil et al., 2019a). Batch cultures may also lead to oxygen limitation (Vasala et al., 2006; see also Section 3.2.4) or substrate inhibition (Minihane and Brown, 1986) that are not observed in lab- or manufacturing-scale fermentation process. These phenomena can cause strain behavior in fed-batch or lab-scale models to sometimes be poorly modeled by plate models that are run in batch mode.

Some limitations of batch feeding can be circumvented with “pseudo-fed-batch” systems that use controlled-release technologies (Schultz and Gerhardt, 1969; Lübbe et al., 1985). In these strategies, all media components are still added at the outset, but a limiting nutrient is released over time from within the culture. This allows for fed-batch-like conditions in standard microplates incubated in common shakers.

There are two main approaches used in pseudo-fed-batch plate models: diffusion and enzymatic-release systems. The former embeds a nutrient inside a polymer matrix with the polymer either suspended within the culture or affixed to the bottom of the vessel (Jeude et al., 2006; Huber et al., 2009). The nutrient then diffuses into the bulk medium over the course of the culture. Commercial versions of such diffusion feeding systems are available from Kuhner AG (Switzerland). The enzymatic approach uses an inaccessible form of a substrate, usually a sugar polymer, which is converted into a consumable form by a supplemented enzyme (Panula-Perälä et al., 2008; Krause et al., 2010). A glucose-releasing product is available from Enpresso GmbH (Germany) and a similar system can be achieved with maltodextrin and amyloglucosidase (Nagy et al., 1992). Both pseudo-batch-fed methods can have their release rates tuned to some extent by modifying substrate and enzyme loading, respectively.

3.2.4. Oxygenation control

As previously emphasized in Section 2.2, oxygen availability is one of the most important considerations in aerobic culture. It has been shown that when lab- and pilot-scale E. coli fermentations are adjusted such that their measured kLa values match those observed in plate models, near-equivalent biomass and protein production can be achieved (Islam et al., 2008). Unfortunately, kLa for an optimal manufacturing process is often too high to be achievable in microplates (Doig et al., 2005). Despite this limitation, a plate model can still attempt to match the physiology of the strain to that in a lab-scale process by targeting the same cell-specific OUR, sometimes called qO2, with units of mmol O2 consumed per gram of dry cell weight per hour. Because the cell density reached in microplates can sometimes be up to 10-fold lower than in larger bioreactors, equivalent qO2 may be attainable even with a smaller maximum kLa. Nevertheless, it is usually desirable to maximize OTRs in microplates to avoid oxygen limitation of flux through overexpressed pathways unless it is expected that this limitation occurs at manufacturing scale.

Incubator design is a key factor in attaining a desired OTR and in maintaining consistent OTRs across plates. Incubator chambers should promote circulation of gases to distribute oxygen evenly and will ideally allow feedback regulation of CO2 and O2 content. Monitoring dissolved O2 within plate media can also be helpful in determining whether proper oxygenation is achieved in a plate model (A. Kumar 2004). Several sensors have been published based on oxygen-sensitive phosphorescent or fluorescent materials (O'Riordan et al., 2000; Stitt et al., 2002; John et al., 2003; Deshpande and Heinzle, 2004). Changes in dissolved oxygen can then be tracked to obtain estimates of OTR and OUR.

Increasing air-liquid interfacial area within a microplate well is the most practical and effective way to increase kLa, and therefore OTR (Duetz et al., 2000). This can be achieved by shaking at high frequencies (i.e. ∼1000 rpm) and with large throw (diameter of the circle described by the motion of the shaker), which cause the liquid to spread vertically up the well wall (Hermann et al., 2003; Islam et al., 2007). In practice, most microplates are shaken with a smaller throw because it allows higher shaking frequency without culture broth spilling out of the plate or wetting the plate seal; the liquid spreads more radially on the well wall with a smaller throw (Kensy et al., 2005). Air-permeable seals or covers can be used to avoid spillage and cross-contamination but do affect oxygen availability to some extent. Despite this, it has been demonstrated that most commercially available seals are not rate limiting in oxygen transfer to cultures (Sieben et al., 2016).

Microplate well geometry can also significantly affect oxygen transfer, especially at lower shaking speeds (Hermann et al., 2003; Funke et al., 2009). Angular well edges can act as baffles, increasing turbulence and leading to higher air-liquid interfacial area (Lattermann et al., 2014). However, this effect is less pronounced relative to round wells at very high shaking speeds. For plates with round wells, lowering the surface tension also increases liquid spreading, leading to higher OTR (Hermann et al., 2003). It has also been shown that surface tension can be decreased with more hydrophilic plate material (Doig et al., 2005), which circumvents the need to adjust the composition of the culture. The authors are unaware of commercially available well geometries in an ANSI/SBS plate format besides round and square wells.

A final factor influencing OTR is the fill volume of the well; lower liquid volumes lead to higher OTRs because the air-liquid interface (and therefore mass of oxygen transferred over time) stays roughly constant as the volume is changed, but OTR is a volume-normalized metric (Hermann et al., 2003). It is important to note that, for most conditions, the evaporation rate also increases along with higher air-liquid interfacial area (Doig et al., 2005); the starting culture volume should therefore be selected with this eventual liquid loss in mind.

3.2.5. Temperature and humidity control

Because of the exothermic nature of aerobic fermentation processes and the low surface area-to-volume ratio of lab-scale bioreactors, temperature must often be feedback-regulated with a surrounding water jacket or immersed heat-transfer element (Table 2). Fortunately, plate models can achieve more efficient heat transfer in the absence of these control elements, especially when vigorous shaking is used (Schäpper et al., 2009). It is therefore feasible to globally control temperature and reduce evaporation in minibioreactors or microplates within large shaker-incubators such as Infors AG (Switzerland) laboratory shakers or humidity-regulated rooms (Sieben et al., 2016). These approaches can allow much higher throughput than smaller benchtop heater-shakers.

Vigorous shaking to increase OTR can result in high rates of evaporation, even when incubators are humidity controlled. Significant evaporation can increase concentrations of broth and cause osmotic stress (Silk et al., 2010), and generally introduces an uncontrolled variable into the culture model. In one study, spongy silicone mats with 1.5-mm venting holes over each well were able to reduce evaporation to less than 10 μL per well per day at all volumes tested (Duetz et al., 2000). There are also numerous commercially available breathable seals for microplates, the permeabilities of which have been compared under conditions relevant to fermentative culture (Sieben et al., 2016). Ultimately, the results revealed an unresolved tradeoff between oxygen availability and evaporation rate (Sieben et al., 2016). As a next-best approach to eliminating evaporative loss, Amyris has investigated the impacts of culture conditions and liquid volumes on evaporation in these systems over time, yielding useful correction factors when time-course comparisons are required (Schmidt, unpublished results).

3.2.6. Product capture

Subcellular localization of products and intermediates is often the subject of engineering efforts, as intracellular accumulation can be detrimental to production and create challenges for product recovery at scale. Localization can be driven by factors including (1) the localization of pathway enzymes, (2) the chemical properties and size of the product or intermediate, and (3) mechanisms of membrane transport. In designing a high-throughput screen for strain improvement, it is essential to consider these factors along with the immediate goals of engineering. For instance, potential improvements in pathway flux by an enzyme engineering effort could be obscured in the absence of expression of a transporter protein if the enzyme suffers inhibition from accumulation of a downstream product (Alberstein et al., 2012) or if the product results in cytotoxicity (Mukhopadhyay, 2015). It can therefore be advantageous to adjust culture models in a way that influences metabolite and/or product localization.

As a common example, additives to the culture can help to draw products out of the cell in the absence of strong active transport. In some cases, these may be water-soluble (e.g., Triton X, Tween-80, or cyclodextrins) and aid in solubilization or dissociation from cells (Dufosse et al., 1999; Dhamole et al., 2012). Other additives are immiscible with water, forming a separate phase (a layer or emulsion) that can sequester and better dissolve or suspend the product (Gu et al., 1999). Note that while the latter usage has been very common for extractive fermentations in lab-scale and larger bioreactors, there are few literature examples of use in HTP plate models. Several industrial patents and applications do, however, outline claims over their use in plate-based screening workflows, with commonly used organic additives including dodecane, isopropyl myristate, and methyl oleate (Tsuruta et al., 2009; Ridley et al., 2019). Amyris has also reported the use of isopropyl myristate and mineral oil in lower-throughput shake-flask-based screens for artemisinin precursor production in S. cerevisiae (Westfall et al., 2012; Paddon et al., 2013).

3.3.1. Product localization and analyte extraction

Products synthesized within a cell partition between the extracellular and intracellular spaces over time, depending on the properties of the molecule. Passive diffusion is a common route of membrane transport for smaller and more hydrophobic molecules but occurs only to a very small extent for larger, uncharged, polar molecules (N. J. Yang and Hinner, 2015). Complete extraction of the whole-cell broth and subsequent sample analysis will reflect the sum of the extracellular and intracellular product concentrations. This is the standard method of analysis and is often the least technically challenging. However, when considering optimization of manufacturing potential, independent analysis of these two fractions can give insight into whether intracellular product recovery (which can be quite costly at manufacturing scale) will be necessary. As an example, recovery of fatty acids from intracellular pools can increase manufacturing cost by 40–80% (Ledesma-Amaro et al., 2016). This type of analysis is particularly useful when screening for strains with improved product export (Wagner et al., 2018).

Cell membrane permeabilization or disruption is often necessary to quantify intracellular product. Many techniques can be used to mechanically disrupt cell membranes, including bead-beating, ultrasonication (Stipetic et al., 2016), lyophilization (Jayatilaka et al., 2011), and freeze/thaw cycling (Canelas et al., 2009); these are semi-automated at best and are a poor match for HTP-sample preparation (Table 4). However these methods can serve to generate a reliable reference point for the total content of a sample. Chemical permeabilization methods are a better match for HTP workflows, since they are more compatible with automated liquid handlers. Chemical permeabilization can be achieved via enzymatic or surfactant lysis buffers, or via solvent treatment. For example, lysozyme and two other commercial enzymatic lysis agents have been compared against sonication as an automation-compatible method for releasing proteins from dried cell pellets of E. coli (Listwan et al., 2010) and found to be comparable in efficiency. In another study, combinations of lysis enzymes, surfactants, detergents, chaotropic agents, antibiotics, and chelating agents were tested in a design-of-experiment study to release proteins from E. coli that laid out a framework for selecting effective cell lysis buffers (Glauche et al., 2017). This approach resulted in a cell lysis buffer whose performance was equivalent to a commercially available product. Methanol, ethanol, and chloroform are widely cited as permeabilizing for some bacteria (especially E. coli) and yeast (Faijes et al., 2007; Mashego et al., 2007). In particular, cold methanol has been described as an ideal sample treatment for global analysis of metabolites in E. coli at native concentrations (Maharjan and Ferenci, 2003); extraction procedures for targeted quantitation of a particular metabolite should, however, be optimized for that purpose. Combining multiple methods can also be advantageous. For example, amorpha-4,11-diene has been extracted from S. cerevisiae fermentation broth using a combination of strong acid, a commercially available detergent-based protein extraction reagent, and ethyl acetate (Westfall et al., 2012).

Table 4. Sample preparation techniques for microplate culture models.

Once released from the cell, traditional liquid-liquid extraction or solubilization methods can be used to isolate analytes from the surrounding matrix and ensure that they are amenable for analysis. As previously mentioned, pipetting, agitation, and centrifugation are all commonly employed on robotics platforms and can be readily used in HTP sample preparation (Ma et al., 2008). Organic solvents are commonly used and, depending on the chemical characteristics of the target compounds and downstream analytical method, may be miscible or immiscible with the culture media. Predictions for effective extraction and solubilization systems may be made in advance based on physical characteristics of the target compound(s) such as hydrogen bonding, acidity/basicity, and octanol-water partitioning coefficient; however, due to the complex makeup of biological samples, several options should always be empirically tested on well-studied strains to optimize the sample preparation (Tshepelevitsh et al., 2017). Multi-component systems with a mixture of solvents or salt additives may be helpful, especially in extracting polar analytes from aqueous matrices (Hyde et al., 2017). As an example of a sample preparation that would be straightforward from both an automation and extraction perspective, antibacterial metabolites have been extracted from aqueous supernatant of marine actinomycete cultures using ethyl acetate (Rajan and Kannabiran, 2014).

Solid-phase extraction (SPE) has also been used with commercially available SPE cartridges in 96-well plates. These are most used for cleanup of tissue or fluid samples for bioanalysis, and integrated systems are available to automate or semi-automate the extractions in 96-well formats. As a typical example, carboxylic acid-based metalloprotease inhibitor compounds have been extracted from mammalian plasma samples using diatomaceous earth-packed extraction plates, and washed with organic solvents before elution (Peng et al., 2001). Application to microbial sample preparation has also been reported; in one study, secondary metabolites were separated from Streptomyces coelicolor culture supernatant using a hydrophilic-lipophilic balanced sorbent SPE packing (Čihák et al., 2017). Filter plates are also commonly used for sample cleanup, whether in the form of molecular weight cut-off filters or membrane-based filters. Even in 96-well plate format and with automation compatibility, the time needed for SPE or filter plate sample preparation pushes these sample techniques into the medium throughput range (∼103-104 samples per week) and they are thus generally not suitable for large libraries. Regardless of the process used, attention should be paid to the resulting sample matrix to ensure that it is compatible with the analysis platform and optimized for accuracy, precision, selectivity, and robustness.

3.3.2. Detection of products, pathway intermediates, and by-products

Analytical instrumentation has traditionally been developed to support the chemical and pharmaceutical industries, and more recently the biopharmaceutical industry. The metabolic engineering (synthetic biology) industry is relatively new; thus, instruments designed for the specific challenges related to desired throughputs are still being developed. Instrumentation development in the last decade has been especially rapid in the field of mass spectrometry (MS) and this pace of development is likely to continue into the next 10 years as well. Large step-changes in sensitivity and selectivity of commercial instrumentation are also expected for Raman-based spectroscopic techniques, which will be discussed in the context of microfluidic platforms in Section 4.3.4. This section focuses on analytical technologies that are most practical for analysis of plate model samples (Table 5).

Table 5. Analytical technologies used for analysis of microplate culture model samples.

3.3.3. Alternative response variables

Although endpoint titer measurements of products and intermediates are the most typical benchmarks for strain improvement in an HTP screen, there are occasionally other readouts that can be exploited, such as measurements of cell health, metabolic state, growth, protein expression, and kinetics of product formation. Many of these readouts are generalizable across different strain development efforts, making their implementation cost effective.

Cell density measurements are ubiquitously used to obtain relative biomass yields and monitor growth. Growth rates, which can be determined from repeated biomass measurements, are often used as indicators of strain metabolic activity and cell health. Absolute cell density can be acquired using flow cytometry or simplified automatic cell counters. However, for strains with similar cell morphologies and sizes, biomass and cell density are roughly proportional and can be more quickly estimated with absorbance measurements, usually at 600 nm. Due to the narrow linear range of the absorbance assay measurements, dilution of culture can often be required in such assays. A 96-well adaptation of the BioLector has demonstrated that measurement of light scattered from the culture (rather than that transmitted) can give a broader linear range without dilution (Kensy et al., 2009), although this is not a readout available for most plate readers.

In some cases, a strain can be utilized as a whole-cell biosensor and included in samples to allow simple OD measurements to be used as a surrogate for direct detection of product. For instance, some groups have engineered sensor strains that are auxotrophic for the desired product; the result of this can be that the sensor strain only grows in the presence of an engineered strain that provides the product at the necessary level (Saleski et al., 2019).

Direct measurements of membrane integrity (Edwards et al., 2007) and redox-equivalent generation (Nandy et al., 2015) have also been widely used for assessing product tolerance and metabolic state and can give a more nuanced illustration of strain health. While flow cytometry readouts can give the most data-rich analysis of culture samples, they are somewhat throughput-limited (∼30 s per sample). Cruder but still broadly informative alternatives have been demonstrated using spectrophotometers, providing throughputs of more than one sample per second (Kwolek-Mirek and Zadrag-Tecza, 2014).

Because sugars are common feed sources, it can be valuable to develop a high-throughput readout that can assess the residual sugar content in a plate model. Such assays can help with the development of culture models when rate of sugar consumption is a concern and can also be used in estimates of yield when sugar is only partially consumed. Absorbance assays based on the detection of NADH production through coupled enzymatic reactions that consume sugar have widely been used for this purpose (Badotti et al., 2008).

Yield is a very important factor in the final cost of production at manufacturing scale when feedstock is the major contributor to cost. When all carbon sources in a plate model are completely consumed, yield can usually be assumed proportional to endpoint titer, making relative titer the most convenient metric for strain ranking. However, if time-dependent factors like facility rental or labor dominate cost models, volumetric productivity can become a critical factor as well (Benjamin et al., 2016). A challenge in determination of volumetric productivity in plate models is that multiple measurements of titer for each sample are required to allow an estimate to be calculated. Methods to accomplish this have been used at Amyris (Sandoval et al., 2014) and elsewhere but are typically not used as primary screens because of increased operational burden. Moreover, the limited sample volume in plate models imposes a constraint on the number of assays that can be run on each plate; this means that several replicate plates must sometimes be cultured to acquire productivity data, which introduces additional error into the measurement. An interesting example of how a microfluidic screening platform can circumvent the need for multiple measurements is introduced in Section 4.2.2.

There are many literature examples demonstrating the power of metabolomic, transcriptomic, and proteomic analysis to understand the bottlenecks and feedback mechanisms in cell factories (Jewett et al., 2005; Redding-Johanson et al., 2011; Petzold et al., 2015). These techniques generate complex multivariate datasets that, together with other readouts described above, may be integrated to improve prediction of performance at lab scale. Machine learning algorithms have been successfully implemented to analyze similar datasets for applications in oncology research, personalized medicine, and neuroscience (Grapov et al., 2018) and are now beginning to demonstrate their promise in metabolic engineering as well (Serber et al., 2017; Opgenorth et al., 2019; Presnell and Alper, 2019). There is little precedent for application of multi-omics to HTP mL-scale strain screening as it is defined in this review (analysis of >104-105 cultures per week). However, there is strong potential especially for targeted proteomics and metabolomics accelerated by HTP MS, and they should be considered as possible readouts for engineering libraries.