Enhancing multi-bioreactor platforms for automated measurements and reactive experiment control

New small-scale, low-cost bioreactor designs provide researchers with exquisite control of environmental parameters of microbial cultures over long durations, allowing them to perform sophisticated, high-quality experiments that are particularly useful in systems biology, synthetic biology and bioengineering. However, existing setups are limited in their automated measurement capabilities, primarily because sensitive and specific measurements require bulky, expensive, stand-alone instruments (for example, most single-cell resolved measurements require a cytometer or a microscope). We present here ReacSight, a generic and flexible strategy to enhance multi-bioreactor platforms for automated measurements and reactive experiment control. We use ReacSight to assemble a platform for single-cell resolved characterization and reactive optogenetic control of parallel yeast continuous cultures. We demonstrate its usefulness by achieving parallel real-time control of gene expression with light in different bioreactors and by exploring the relationship between fitness, nutrient scarcity and cellular stress density using highlycontrolled and informative competition assays.


Introduction
Small-scale, low-cost bioreactors are emerging as powerful tools for microbial systems and synthetic biology research [1][2][3][4] . They allow tight control of cell culture parameters (e.g. temperature, cell density, media renewal rate) over long durations (many cell generations). These unique features enable researchers to perform sophisticated experiments and to achieve high experimental reproducibility. Examples include characterization of antibiotic resistance when drug selection pressures increases as resistance evolves 1 , cell-density controlled characterization of cell-cell communication synthetic circuits 2 , and genome-wide characterization of yeast fitness under dynamically changing temperature using a pooled knockout library 3 .
A weakness of existing small-scale, low-cost bioreactors is their limited automated measurement capabilities: in situ optical density measurements only inform about overall biomass concentration and its growth rate, and when available 2,4 fluorescence measurements suffer from low sensitivity and high background. To quantify other key characteristics of the cell population under study (e.g. gene expression levels, cellular stress levels, cell size and morphology, cell cycle progression, proportions of different genotypes/phenotypes), researchers usually need to manually extract, process and measure culture samples to run them through more sensitive and specialized instruments (e.g. a cytometer, a microscope, a sequencer). This is usually tedious and strongly constrains the available temporal resolution and scope (i.e. no time-points during night-time). It also precludes the dynamic adaptation of culture conditions in response to such measurements (reactive experiment control). This would be useful to either maintain a certain state of the population (external feedback control) or to maximize the value of the experiment (reactive experiment design). For example, external feedback control can be used to disentangle complex cellular couplings and signaling pathway regulations [5][6][7][8] , to steer the composition of microbial consortia 9 or to optimize industrial bioproduction 10 . Reactive experiment design can be especially useful in the context of long and uncertain experiments such as artificial evolution experiments 11 .
In principle, commercial robotic equipment and/or custom hardware can be used to couple a multibioreactor setup to a sensitive, multi-sample (typically accepting 96-well plates as input) measurement device. However, this poses tremendous challenges regarding equipment sourcing, equipment cost and software integration, and very few examples have been reported. For instance, only two groups have demonstrated automated cytometry and reactive optogenetic control of bacteria 12 and yeast 6,7 cultures, with setups limited to either a single continuous culture or multiple batch-only cultures.
Here, we present ReacSight, a generic and flexible strategy to enhance multi-bioreactor platforms for automated measurements and reactive experiment control. We first use ReacSight to assemble a platform enabling single-cell resolved characterization and reactive optogenetic control of parallel yeast continuous cultures. We then demonstrate its usefulness on two case studies. First, we achieve parallel real-time control of gene expression with light in different bioreactors. Second, we explore the relationship between fitness, nutrient scarcity and cellular stress density using highly-controlled and informative competition assays.

Measurement automation, platform software integration and reactive experiment control with ReacSight
The ReacSight strategy ( Figure 1, Text S1.1) to enhance multi-bioreactor platforms for automated measurements and reactive experiment control combines hardware and software elements. A pipetting robot is used to establish, in a generic fashion, a physical link between any multi-bioreactor setup and any plate-based measurement device (Figure 1, left). Bioreactor culture samples are sent to the pipetting robot through pump-controlled sampling lines attached to the robot arm (sampling). A key advantage of using a pipetting robot is that treatment steps can be automatically performed on culture samples before measurement (treatment). Samples are then transferred to the measurement device by the pipetting robot (loading), thanks to a positioning of the device such that upon loading tray opening, part or all of the device input plate is accessible to the robot. Partial access to the device input plate is not problematic because the robot can be used to wash input plate wells between measurements, allowing re-use of the same wells over time (washing). Importantly, if reactive experiment control based on measurements is not needed, the robot capabilities can also be used to treat and store culture samples for one-shot offline measurements at the end of an experiment, enabling automated dynamic measurements with flexible temporal resolution and scope.  ReacSight also provide a solution to several software challenges that should be addressed to unlock automated measurements and reactive experiment control of multi-bioreactors (Figure 1, right). First, programmatic control of all instruments of the platform (bioreactors, pipetting robot, measurement device) is required. Second, a single computer should communicate with all instruments to orchestrate the whole experiment. ReacSight combines the versatility and power of the python programming language with the genericity and scalability of the Flask web-application framework to address both challenges. Indeed, python is ideally suited to easily build APIs to control various instruments: there exist wellestablished, open-source libraries for the control of micro-controllers (such as arduinos), and even of GUIonly software (pyautogui) for closed-source instruments lacking APIs. Importantly, the open-source, lowcost pipetting robot OT-2 (Opentrons) is shipped with a native python API. Flask can then be used to expose all instrument APIs for simple access over the local network. The task of orchestrating the control of multiple instruments from a single computer is then essentially reduced to the simple task of sending HTTP requests, for example using the python module requests. Finally, in addition to this versatile instrument control architecture, ReacSight provides a generic object-oriented implementation of events (if this happens, do this) to facilitate reactive experiment control. Key software elements as well as source files for hardware pieces are available in the ReacSight Git repository.

Reactive optogenetic control and single-cell resolved characterization of yeast continuous cultures
Our first application of the ReacSight strategy is motivated by yeast synthetic biology applications. Being able to 1) accurately control synthetic circuits and 2) measure their output in well-defined cellular and environmental contexts and with sufficient temporal resolution and scope is critical in this context. Optogenetics provides an ideal way to control synthetic circuits, and bioreactor-enabled continuous cultures are ideal to exert tight control over growth conditions for long durations. To measure circuit output in single cells, cytometry is also ideal due to high sensitivity and throughput. We thus resorted to the ReacSight strategy to assemble a fully automated experimental platform enabling reactive optogenetic control and single-cell resolved characterization of yeast continuous cultures (Figure 2A), using a benchtop cytometer as a measurement device.
Detailed information on the platform hardware and software is provided in Text S1.2, and we discuss here only key elements. Eight reactors are connected to the pipetting robot, meaning that each timepoint fills one row of a sampling plate. While three rows of the cytometer input plate are accessible by the robot, we use only one row, washed extensively by the robot to achieve less than 0.2% carry-over as validated using beads. We typically fit two tip boxes and two sampling plates on the robot deck, therefore enabling 24 x 8 timepoints without any human intervention. To enable reactive experiment control based on cytometry data, we developed and implemented algorithms to perform automated gating and spectral deconvolution between overlapping fluorophores ( Figure 2B). We first validated the performance of the platform by performing long-term turbidostat cultures of yeast strains constitutively expressing various fluorescent proteins from chromosomally integrated transcriptional units ( Figure 2C). Distributions of fluorophores levels were unimodal and stable over time, as expected from steady growth conditions with a constitutive promoter. Distributions of mNeonGreen and mScarlet-I exactly overlapped between the single-and 3-color strains, as expected from the assumptions that expressing one or three fluorescent proteins from the strong pTDH3 promoter has negligible impact on cell physiology and that the relative positioning of transcriptional units in the 3-color strain (mCerulean first, followed by mNeonGreen and mScarlet-I) has little impact on gene expression. Measured levels of mCerulean appear slightly higher (~15%) in the 3-color strain compared to the singlecolor strain. This could be caused by residual errors in the deconvolution, exacerbated by the low brightness of mCerulean compared to autofluorescence and to mNeonGreen.
Finally, to validate the optogenetic capabilities of the platform, we built and characterized a light-inducible gene expression circuit based on the EL222 system 13 ( Figure 2D). As expected, applying different ON-OFF temporal patterns of blue light resulted in dynamic profiles of fluorophore levels covering a wide range, from near-zero levels (i.e., hardly distinguishable from auto-fluorescence) to levels exceeding those obtained with the strong constitutive promoter pTDH3. Cell-to-cell variability in expression levels at high induction is also low, with CV (coefficient of variation) values comparable to the pTDH3 promoter (0.22 vs 0.20).

Real-time control of gene expression using light
To showcase the reactive optogenetic control capabilities of the platform, we set out to dynamically adapt light stimulation so as to maintain fluorophore levels at different target setpoints. Such in-silico feedback for in-vivo regulation of gene expression is useful to dissect the functioning of endogenous circuits in the presence of complex cellular regulations and could facilitate the use of synthetic systems for biotechnological applications 6,10,14 .

Every hour, cytometry acquisition is performed, and after gating, deconvolution and FSC-normalization the data is fed to a model-predictive control (MPC) algorithm. The algorithm uses the model to search for the best sequence of duty fractions for 10 duty cycles of period 30 minutes (i.e. a horizon of 5 hours) in order to track the target level. (C) Real-time control results for four different target levels, performed in parallel in different bioreactors. Left: median of single cells (controlled value). Right: single-cell distributions over time. Note that a linear scale is used on all plots.
We first constructed and validated a simple mathematical model of light-induced gene expression ( Figure  3A). Joint fitting of the three model parameters to the characterization data of Figure 2D resulted in an excellent quantitative agreement. This gives further confidence in the ability of the platform to provide reproducible and reliable quantitative data. We then incorporated the fitted model into a modelpredictive control algorithm ( Figure 3B). Together with the ReacSight event system, this algorithm enabled accurate real-time control of fluorophore levels to different targets in different reactors in parallel ( Figure  3C).

Exploring the relationship between fitness, nutrient scarcity and cellular stress
A key advantage of using cytometry to characterize bioreactor cultures is the possibility to phenotype single cells, and even to genotype them if different strains harbor distinct fluorescent barcodes. This capability unlocks a new class of experiments: multiplexed strain characterization and competition in dynamically controlled environments ( Figure 4A). Indeed, some fluorophores can be used for genotyping and others for phenotyping. Automated cytometry (including raw data analysis) will then provide quantitative information on both the competition dynamics between the different strains and cell state distribution dynamics for each strain. Depending on the goal of the experiment, this rich information can be fed back to experiment control to adapt environmental parameters for each reactor.
As a proof of concept that such experiments can be carried out, we set out to explore the relationship between fitness, nutrient scarcity and cellular stress ( Figure 4B, top-left). Different species in microbial communities have different nutritional needs depending on their metabolic diversity or specialization, and their fitness therefore depends not only on external environmental factors but also on the community itself through nutrient consumption, metabolite release, and other inter-cellular couplings 15,16 . As opposed to competition assays in batch, continuous culture allows to control for such factors. For example, in turbidostat cultures, nutrient availability depends on both nutrient supply (i.e. nutrient levels in the input medium) and nutrient consumption by cells (which primarily depends on the OD setpoint). We used histidine auxotrophy as a model for nutrient scarcity: for his3 mutant cells, histidine is an essential nutrient. By competing his3 mutant cells with wild-type cells at different OD setpoints and different histidine concentrations in the input medium, we can measure how nutrient scarcity affects fitness ( Figure  4B, top-right). Using a stress reporter in both strains also informs about the relationship between fitness and cellular stress in the context of nutrient scarcity. We focused on the UPR (Unfolded Protein Response 17 ) stress response to investigate whether nutrient stress can lead to other, a priori unrelated types of stress, which will be indicative of global couplings in cell physiology.  At a histidine concentration of 4 µM, his3 mutant cells are strongly outcompeted by wild-type cells over the range of OD setpoints (0.1 -0.8) we considered ( Figure 4B, bottom-left). This is not the case anymore at a concentration of 20 µM: then, the growth rate advantage of wild-type cells is close to zero below an OD setpoint of 0.6 and becomes larger than 0.2 hr -1 at the largest OD setpoint of 0.8. Therefore, for this level of nutrient supply, levels of nutrient consumption by cells have a strong impact on fitness of his3 mutant cells. This qualitative change between 4 µM and 20 µM is highly consistent with the reported value of 17 µM for the Km constant of the single high-affinity transporter of histidine, HIP1 18 . Also, because the growth rate difference between wild-type and mutant cells for a histidine concentration of 4 µM is close or exceeds the typically observed growth rate of wild-type cells (between 0.3 and 0.45 hr -1 depending on the OD setpoint), we can conclude that mutant cells are growth-arrested in these conditions. Single-cell UPR phenotyping data shows little difference between mutant and wild-type cells across all OD setpoints for a histidine concentration of 20 µM but a clear activation of the UPR response in mutant cells at a histidine concentration of 4 µM ( Figure 4B, bottom-right). Therefore, as opposed to growth arrest, growth rate reduction caused by nutrient scarcity does not lead to strong UPR activation.

Discussion
We report the development of ReacSight, a strategy to enhance multi-bioreactor setups with automated measurements and reactive experiment control. ReacSight addresses an unmet need by allowing researchers to combine the recent advances in low-cost, open-hardware instruments for continuous cultures of microbes (e.g. eVOLVER, Chi.Bio 3,4 ) and multi-purpose, modular, programmable pipetting robots (e.g. Opentrons OT-2) with sensitive, but generally expensive, stand-alone instruments to build fully automated platforms that open up radically novel experimental capabilities. ReacSight is generic and easy to deploy, and should be broadly useful for the microbial systems biology and synthetic biology communities. While we deployed the ReacSight strategy for only one measurement device (a benchtop cytometer), it should be possible to position two devices on each side of the pipetting robot to enable even more advanced workflows.
As already noted by Wong and colleagues 3 , connecting a multi-bioreactor setup to a cytometer for automated measurements could enable single-cell resolved characterization of microbial cultures across time. Automated cytometry in the context of microbial systems and synthetic biology has in fact already been demonstrated years ago by a small number of labs 6,12,19 , but low throughput or reliance on expensive automation equipment likely prevented a wider adoption of this technology. Automated cytometry from continuous cultures becomes especially powerful in combination with recently developed optogenetic systems 20,21 , enabling targeted, rapid and cost-effective control over cellular processes 12 . We demonstrated the key capabilities of such platform by performing parallel, light-driven real-time control of gene expression in different bioreactors as well as fully automated, cell-state informing competition assays in tightly controlled environmental conditions. While we used ReacSight to connect our own, preexisting custom multi-bioreactor setup with a cytometer, it can instead be used in combination with the recent Chi.Bio optogenetic-ready bioreactor system to build such platform at a small time and financial cost (excluding the cost of the cytometer, which are expensive but already widespread in labs given their broad usefulness even in absence of automation). We only touched the surface of the large space of potential applications offered by such platforms. Strain barcoding can be scaled-up to 10-15 strains (for example by integrating different copy numbers of two fluorophores driven by the same promoter, their ratio providing a readout of cell genotype). Such multiplexing capabilities can be especially useful to characterize the input-output response of various candidate circuits (or the dependence of circuit behavior across a library of strain backgrounds) in parallel (using different light inductions across reactors). Immunobeads can be used for more diverse cytometry-based measurements (the robot enabling automated incubation and wash, for example using the Opentrons OT-2 magnetic module). Technologies such as surface display 22,23 can also be used to engineer biosensor strains to measure even more dimensions of the cultures with a single cytometer and at no reagent costs.
In the future we hope that many ReacSight-based platforms will be assembled and their design shared by a broad community to drastically expand our experimental capabilities, in order to shed new light on fundamental questions in microbiology or to unlock the potential of synthetic biology in biotechnological applications.

Methods
Cloning and strain construction. All plasmids and strains are constructed from the common laboratory strain BY4741 using the modular cloning framework for yeast synthetic biology Yeast Tool Kit by Lee and colleagues 24 . All strains used express the light-inducible transcription factor from the URA3 locus (transcriptional unit: pTDH3 NLS-VP16-EL222 tSSA1). Single-color constitutive expression strains ( Figure  2) also harbor a pTDH3 FP tTDH1 transcriptional unit at the LEU2 locus where FP is mCerulean, mNeonGreen or mScarlet-I. Corresponding CDS have been codon-optimized for expression in S. cerevisiae. The three-color strain harbors the same three transcriptional units in tandem (order: mCerulean, mNeonGreen, mScarlet-I) at the LEU2 locus. The autofluorescence strain harbors an empty cassette at the LEU2 locus to match auxotrophy markers between strains. For light-inducible gene expression (Figure 2 and 3), a pEL222 mNeonGreen tTDH1 transcriptional unit (where pEL222 is composed of 5 copies of the EL222 binding site followed by a truncated CYC1 promoter 13 ) is integrated at the LEU2 locus. For the histidine competition experiments, the mutant strain is simply BY4741 expressing the light-inducible transcription factor from the URA3 locus, i.e. the common parental strain for all strains used in this work, which is auxotroph for histidine, leucine and methionine. The wild-type strain was obtained from the mutant strain by integrating two identical pTDH3 mCerulean tTDH1 transcriptional units in tandem at the HO locus with HIS3 selection, thereby restoring histidine prototrophy and enabling fluorescent barcoding.
Cell culture conditions. All experiments were performed in 30 mL culture volume bioreactors (cf Text S1.2) at 30 degrees and in turbidostat mode (OD 0.5, typically corresponding to 10 7 cells/mL according to cytometry data) with synthetic complete medium (ForMedium LoFlo yeast nitrogen base CYN6510 and Formedium complete supplement mixture DCS0019) except for histidine competition experiment where histidine drop-out amino-acid mixture was used (Sigma Y1751) and complemented with desired levels of histidine (Sigma 53319).
Cytometry acquisition and raw data analysis. Gain settings of our cytometer (Guava EasyCyte 14HT, Luminex) for all channels were set once and for all prior to the study such that yeast auto-fluorescence under our typical growth conditions is detectable but at the lower end of the instrument 5-decade range.
We verified that cytometry data was reproducible week-to-week with those fixed settings. Single-color strains described above were used together with the autofluorescence control strain to obtain 'spectral' signatures of the three fluorophores mCerulean, mNeonGreen and mScarlet-I and autofluorescence levels for each channel. These signatures were also highly reproducible week-to-week. To convert raw cytometry data into fluorophore concentrations in relative promoter units (RPU 25 ), we used a pipeline described in Text S2.2. This pipeline was implemented in python and is available in the ReacSight Git repository.
Histidine competition assays. Pre-cultures were performed in synthetic complete medium. Cells were washed in the same low histidine medium as the one used for turbidostat feeding of the competition culture and mixed with an approximate ratio mutant:WT of 5:1 (to ensure good statistics for long enough even when the mutant fitness is very low) before inoculation. Cytometry was acquired automatically every 2 hours. At steady-state, the ratio between two competitors in a co-culture evolves exponentially at a rate equals to their growth rate difference. Linearity of the ratio logarithm for at least 3 timepoints was therefore used to assess when steady-state is reached. A threshold of 1 mCerulean RPU was used to assign each cell to its genotype. Size gating was performed as described in Text 2.2 (parameters: size threshold = 0.5 and doublet threshold = 0.5, less stringent than for experiments of Figure 2 and 3) to discard dead or dying cells.