Machine vision-enabled OD₆₀₀ measurement and mass spectrometric determination of metabolites of microbial cell-containing microfluidic droplets in real-time

Abstract

Droplet microfluidics has become a versatile tool for high-throughput phenotypic analysis of cells, employed in enzyme engineering, drug screening, and antibiotic assays. An important metric in such bioanalyses is the optical density (OD₆₀₀) of the culture, but current methods for measuring OD₆₀₀ within cultured droplets are limited. Herein, we present a machine-vision method for the measurement of OD₆₀₀ between 0.1 and 4.3 in flowing microdroplets from 0.5–50 nL in volume. This system can make real-time decisions based on cell density and volume measurements and was demonstrated to remove vacant droplets from an emulsion by dielectrophoretic sorting at up to 10 droplets per s based on cell density with an error rate below 10%. We show that droplets can also be subjected to subsequent on-line analysis such as mass spectrometry (MS) and their growth levels correlated to metabolite abundances. We show that growth level is correlated with lysine abundance in droplets incubated with engineered E. coli cells.

Machine vision camera method for label-free measurement of OD₆₀₀ of microbes growing in flowing nanoliter droplets at 10 samples per s is presented.

1. Introduction

Droplet microfluidics has transformed the ability to perform miniature-scale analyses and screens. Although years of benchtop research and optimization have fine-tuned techniques for performing macroscopic measurements essential to several applications, advancements for the precise analysis of droplets on the micro-scale are still emerging. An increasingly important use of droplet microfluidics is the encapsulation and culture of cells. This capability is used in a variety of applications including biocatalyst development,^1–3 antibiotic testing,^4–7 drug screening, tumor viability studies,⁸ and stem cell analysis.^9,10 Often these applications use isogenic microbial colonies grown in droplets that originate from single cells encapsulated into droplets according to the Poisson distribution.¹¹ When evaluating cell growth, an important metric is cell density or concentration. Here, we report a method to measure bacterial concentration within pL–nL droplets in flow, which is compatible with downstream methods like mass spectrometry (MS).

Measurement of cell growth and concentration is commonly required in microbiology applications. For instance, the metabolic state and growth kinetics of cell cultures are related to their proliferation in a medium (i.e. defining lag, exponential and stationary phases of growth).¹² Such measurements are applied in determining when to induce gene expression of a culture in the preparation of recombinant proteins.¹³ Additionally, in metabolic engineering it is essential to control the concentration and growth of cells for the optimized production of a biomolecule or metabolite of interest.¹⁴ Cell concentration is known to be proportional to the turbidity of a suspension, where an increased concentrations of cells increases the opacity of a suspension due to scattering of incident light. On a macroscopic scale, this turbidity is routinely measured in a 1 cm-pathlength cuvette at a wavelength of 600 nm and is characterized by the negative log of transmittance, known as the optical density (OD₆₀₀).¹⁵ At low cell densities (OD₆₀₀ < 1.0) the optical density value is directly proportional to the concentration of cells in a solution, often expressed as colony forming units (CFU), in a pseudo-Beer–Lambert relationship. The optical density is measured at 600 nm due to the minimal interference by the yellow-brown hued media and low interactions with other biomolecules in the medium.¹⁶ At 600 nm the reduction in transmission of incident light in cell suspensions is primarily due to scattering as opposed to traditional molecular absorbance.¹⁷

Although the measurement of OD₆₀₀ in cuvettes is a routine procedure done daily in biological laboratories, measurement of bacterial concentration in microfluidic droplets remains a difficult task. The miniaturized hardware of microfluidics presents challenges in both integrating external measurement hardware, as well as challenges in sensitivity of optical measurements due to path lengths that are ∼10³-times shorter compared to the typical cuvette measurement. Embedded photodiodes and LEDs have been used to correlate transmittance with OD₆₀₀ in microfluidic devices and slugs from 0.4–2.5 µL, providing real-time measurement of bacterial concentration in drops, but not on the pL–nL-scale typically employed in microdroplet emulsions.^6,18 In pL-scale droplets, the measurement of E. coli growth has been accomplished indirectly by injecting a colorimetric cell stain to droplets after growth, and correlating cell density to the molecular absorbance of droplets in a microfluidic chip.¹⁹ Fluorescence reporters as a surrogate for cell density have been used to identify cell growth in small droplets on the pL-scale. For example, a fluorescence sensor was used to detect droplets displaying E. coli growth with a Python-based image processing algorithm detecting bright droplets within a static fluorescence micrograph.²⁰ In flow, droplet shrinkage has been associated with cell growth, which could be measured by increased intensity of a spiked fluorophore due to the decreased volume and proportionally increased fluorophore concentration.²¹ Rather than relying on exogenous fluorophores, genetically-encoded reporters have been utilized to track cell growth by measuring the fluorescence of E. coli expressing red fluorescent protein.⁵

Instead of relying on fluorescent or colorimetric markers, sorting low growth from high growth cells based purely on light scattering has been accomplished by using embedded optical fibers, where high-growth droplets provide greater scattering, and thus, higher intensity radiation measured perpendicularly to the incident light. Using this differentiation, antibiotic-resistant mutant cells were isolated based on their higher observed density.⁷ Though these studies accomplished sorting at up to 300 droplets per s, the classification of droplets was limited to binary assignments, and quantitative growth information for each droplet was not gathered, such as would be provided by an optical density measurement. Specific growth levels within a microdroplet can be measured through individual cell counting, either with cells or cell clusters that are relatively large compared to the droplet,⁴ or with fluorescent reporters such as green fluorescent protein in E. coli containing droplets.²² However, these counting methods rely on fluorescent labeling and either manual or computationally intense image processing algorithms to identify and count individual cells. In eukaryotic systems, cells of different types can be detected in droplets and differentiated by machine-learning algorithms in order to isolate them by sorting,^23,24 but these machine-learning algorithms are likewise computationally intense and therefore these image-based methods are generally limited to small droplets in the pL-range with relatively small total numbers of cells.

Recently, image-based turbidity measurements relying on greyscale intensity of microdroplets have been proposed as a simple alternative to traditional cell/particle counting or embedded hardware methods.^7,25–28 Greyscale intensities of droplets containing E. coli have been successfully correlated to bulk OD₆₀₀ by using triggered darkfield microscopic imaging,²⁷ where analysis involved offline measurement of the average grey value extracted from a region within the center of droplets using a custom algorithm. However, as an offline method, it does not allow for droplet manipulation, such as sorting based on growth, or for other modes of analysis without losing the connection between a droplet and its corresponding data.

Machine vision is a versatile technology that is receiving increased attention in the field of droplet microfluidics. By combining real-time, user-defined image processing algorithms with integrated logic and output, machine vision systems can record and make decisions based on a variety of optical information. Machine vision algorithms have been developed for size measurement,²⁹ fluorescence analysis,^30,31 and colorimetric assays in droplet microfluidics.³² In this work, we develop a machine-vision-based microfluidic platform to characterize cell growth within droplets by real-time measurement of greyscale darkness in addition to sorting and collection of droplet fractions based on measured values. The on-line measurement allows for multi-modal analysis of droplet content with subsequent methods such as droplet electrospray ionization (ESI) MS.^5,33–38 The utility of this system is demonstrated for measuring metabolite concentrations and their correlation to cell density within droplets containing DapA engineered E. coli. (dapA is the gene that encodes dihydrodipicolinate synthase which is part of the diaminopimelate pathway.)

2. Materials and methods

2.1. Chemicals and reagents

All chemicals were purchased from Sigma-Aldrich (St Louis, MO) unless otherwise stated. Lysine-deficient Neidhardt “EZ” media (Teknova, Hollister, CA) was prepared according to the manufacturer's suggestions and used for all bacterial cultures. The continuous phase for droplet generation and reinjection was a 2 wt% solution of 008-Fluorosurfactant (RAN Biotechnologies, Beverly, MA) in Novec HFE-7500 (Oakwood Chemicals, Estill, SC). Blue #1 food dye was purchased from Kroma Kolors (Torrance, CA). The cell strain used in this work, an engineered E. coli strain dapA E84T + dapA J23100 + glnDJ23100, referred to as simply DapA, was obtained from Manus Bio (Waltham, MA).³⁹

2.2. Microfluidic device fabrication

Microfluidic devices were fabricated according to standard soft lithography protocols.⁴⁰ The designs were drawn in AutoCAD (Autodesk Inc., San Rafael, CA) and plotted on mylar photomasks (FineLine Imaging, Colorado Springs, CO). Master molds were fabricated with SU-8 photoresist (Kayaku Advanced Materials, Westborough, MA) spun to a height of 200 µm or 50 µm according to manufacturer's directions on 3-inch silicon wafers (University Wafer, Boston, MA). Polydimethylsiloxane (RTV615, Momentive, Niskayuna, NY) was mixed according to manufacturer's directions and degassed under vacuum. Devices were cast over the masters, cured overnight at 70 °C, removed from the molds, and inlets and outlets were formed with 20-gauge and 22-gauge biopsy punches for standard inlets/outlets and for MS-interfaced outlets, respectively. For dye-doped PDMS devices, a 30 mg mL⁻¹ solution of Sudan I dye in toluene was added to the PDMS to a final concentration of 1 mg mL⁻¹ prior to curing. The punched devices were cleaned with water and isopropanol (Fisher Chemical, Waltham, MA) and plasma bonded to glass microscope slides (VWR, Radnor, PA) using a handheld corona treater (Electro-Technic Products, Chicago, IL). After bonding, devices were flushed with 2% v/v tridecafluoro-1,1,2,2-tetrahydrooctyl trichlorosilane in HFE-7500 to derivatize the channels. PTFE tubing (300 µm i.d., 760 µm o.d., MasterFlex, Radnor, PA) was used to connect to device inlets, while PFA tubing (150 µm i.d., 360 µm o.d., IDEX Corporation, Northbrook, IL) was used to connect device outlets to the mass spectrometer via a tri-coaxial electrospray sheath sprayer.

2.3. Bulk cell culture

For cells grown in bulk, an inoculum of 3× Lysine DapA E. coli was removed from a glycerol stock of cells stored at −80 °C and placed into Luria Broth medium containing 25 µg mL⁻¹ of chloramphenicol (Gold Biosciences, Boulder, CO) dissolved in ethanol which was prepared at a 1000× stock solution. Typically, cells were grown in 2–5 mL of medium in a 14 mL aeration cell culture tube (Thermo Fisher Scientific, Waltham, MA) overnight at 37 °C with shaking at 200 rpm. The bulk solution optical density at 600 nm was measured with a Implen OD600 ® (Implen GmbH, Munich, Germany) at a 10× dilution such that the measured value was between 0.1–1 using standard 1.5 mL 340 to 750 nm polystyrene disposable cuvettes with a semi-microcell measurement region (Fischer Scientific, Waltham, MA). Dilutions were performed with fresh media containing antibiotics as needed. Between optical density measurement and use, cells were stored on ice to slow division to a negligible rate and maintain a stable known concentration of cells.

For cells grown in EZ – lysine deficient medium, the above general procedures were adhered to with the exception of cultures being 1–3 mL in volume in a 14 mL aeration culture tube. For optical density measurements, the spectrophotometer was blanked and cells were diluted with 18 MΩ cm ultrapure water as there was no appreciable difference in the absorbance at 600 nm between the EZ medium and water.

2.4. Instrumental and device operation

Droplets were generated by using microfluidic devices with flow-focusing geometry (Fig. S1a and b), with continuous and dispersed phases controlled by independent-driven gas-tight syringes (Hamilton Co., Reno, NV) in syringe pumps (Fusion 200, Chemyx Inc., Stafford, TX). Observation and optical analysis of droplets was enabled by a machine vision camera (In Sight 7600C, Cognex, Natick, MA) coupled to the C-mount port of an inverted microscope (Eclipse Ts2, Nikon, Tokyo, Japan) with 1× and 4× magnification objective lenses (Plan Achromat, Nikon, Tokyo, Japan). To achieve single-cell encapsulation of droplets, bulk cultures were diluted such that droplets of the target volume were ∼14% singly encapsulated and ∼1% multiply encapsulated. Encapsulated droplets were collected in a 2 mL polypropylene microcentrifuge tube and incubated at 30 °C for 24 h to allow for cell proliferation. Reinjection of incubated droplets onto a microfluidic device was accomplished by transferring droplets to a glass pipette with silicone caps.² HFE-7500 oil with 2 wt% 008-Fluorosurfactant was introduced from the bottom of the pipette by syringe pump which displaced droplets through the top tubing and into microfluidic respacer device that created a flow of droplets segmented by carrier phase (Fig. S1c). Droplet ESI-MS was accomplished by flowing reinjected droplets into an Agilent 6135B single-quadrupole mass spectrometer using a tri-coaxial sheath-flow ESI nebulizer (G1607A, Agilent Technologies, Inc., Santa Clara, CA).³⁸ Droplets were reinjected at 2 µL min⁻¹ and respaced with neat HFE-7500 continuous phase at 5 µL min⁻¹ prior to introduction to the electrospray source. The spray voltage was set to 3000 V and the nebulizer pressure to 20 psig. Sheath liquid consisted of 30% v/v methanol in water with 0.1% formic acid and was pumped with a syringe pump (Harvard Apparatus, Holliston, MA) at a flow rate of 35 µL min⁻¹. Lysine was measured in positive mode with 1 ms dwell times at 147.1 m/z. A fragmentor voltage of 100 V was used for all compounds. A photograph of the setup with major components labeled is shown in Fig. S2.

2.5. Machine vision algorithm

A machine vision algorithm was developed using In-Sight Explorer software (Cognex, Natick, MA) to perform optical analysis of droplets in real time. The algorithm monitors a user-defined detection region for droplets, identified by an edge detection function in the longitudinal axis. Upon droplet detection, the length of the edge-to-edge droplet is measured and used to calculate the volume with one of three formulas depending on the nominal shape of the droplet—slug, pancake, or sphere. At the same time, a smaller region is automatically positioned inside the droplet, and the color (red, green, and blue pixel intensities) and average greyscale intensities are measured within this region at the droplet center. Color matching is performed to determine if a droplet matches a pre-defined marker color, and the greyscale intensity is converted to darkness by inverting the 8-bit value. Finally, the droplet information is displayed in a live graphical readout, as well as being exported over a serial communication line, where it is logged in a text file using PuTTY. Optionally, a decision can be made by the machine vision program based on droplet darkness, volume, and color, which triggers a 20 ms contact closure signal which may be used to interface with a droplet sorting device or other external instrumentation. The code for the machine vision algorithm described here can be accessed in the Zenodo repository (see Data availability statement).

2.6. Sorting droplets by growth and size

Droplet sorting by cell growth level was accomplished by a reinjection device where respaced droplets encounter a bifurcation with opposing dielectrophoresis electrodes (Fig. S1d). Droplets were reinjected to the sorting device at a flow rate of 2 µL min⁻¹ and were respaced with HFE-7500 continuous phase at a flow rate of 110 µL min⁻¹. A signal provided by the machine vision camera triggers a 20 ms high-voltage pulse (1.5 kVAC inverter board) across the opposing saline electrodes that directs selected droplets to the positive channel by dielectrophoresis (DEP), while unselected droplets flow passively to the negative channel.⁴¹

2.7. Synchronization of optical and downstream mass spectrometric data

Optical droplet information obtained from the microfluidic chip and MS data from the mass spectrometer are offset from one another by a significant time delay due to the length of tubing required to flow droplets from the optical detection point to the ESI source. To synchronize signals, marker droplets were generated and mixed with the droplet emulsion to be analyzed at a 1 : 5 ratio of marker to sample droplets. Marker droplets were composed of cell growth medium doped with 1% blue food dye to act as an optical marker, and 20 µM neostigmine to act as a MS marker. The offset could be determined by aligning the marker pattern between the optical and MS trace. Occasional misalignments and desynchronizations were corrected for by the random distribution fingerprint of marker droplets across the entire trace, allowing for quick realignment at the next marker sequence.

2.8. Determination of an experimental conversion factor between OD₆₀₀ and cell density

An overnight culture of E. coli was diluted into fresh LB media and grown for 4 h at 37 °C with shaking in an incubator. After 4 h, the cells were removed and diluted into fresh LB media at 5×, 10×, 20×, and 40× dilutions to generate a calibration curve of OD₆₀₀versus dilution factor and ensure linearity of the measured range. OD₆₀₀ measurements of the diluted cultures were taken with an Implen OD600 ® meter (Implen GmbH, Munich, Germany) in 1 cm cuvettes. The OD₆₀₀ of the stock culture was extrapolated from this trend line. In triplicate, 5 µL of the culture was combined with 995 µL of fresh LB followed by two serial dilutions of 10 µL into 990 µL. 10 µL of the first 100× serial dilution and 20 µL of the second 100× serial dilution was plated for each of the technical replicates resulting in an observed 2 × 10⁶ and 1 × 10⁸ dilution from the parent stock. Plates were grown at 37 °C overnight and single colonies were counted to determine the CFU for each dilution. The replicates were averaged for each dilution and the relation of CFU/OD₆₀₀ was calculated. Data and results are available in Fig. S3.

3. Results and discussion

3.1. Measurement of cell concentration in droplets

The machine vision algorithm was successfully able to detect droplets of a wide range of volumes from 0.5–50 nL and determine their OD₆₀₀ as they flowed through a microfluidic channel. User-selected regions for detection and measurement, depicted in Fig. 1a, were defined using the program's graphical interface. Centering the detection region about the channel and sizing it to be approximately 70% longer than maximum length of the droplets to be measured ensured consistent detection. The internal measurement region is automatically positioned in the center of the detected droplet and is sized such that the edges of the droplet (where the curvature effects the path length and causes lensing) are avoided, and only the uniform center is measured. A flow chart representation of the machine vision algorithm for the measurement and sorting of droplets by OD₆₀₀ is shown in Fig. 1b. A video of the algorithm running in real time is shown in SI Movie S1. To determine the relationship between greyscale darkness of droplets and their bulk optical densities, a calibration curve (Fig. 1c) was constructed with serial dilutions of an E. coli solution with a stock optical density of 4.3 measured in a cuvette at a 1 : 10 dilution. Droplets were generated at a rate of ∼2 droplets per s on a device of 200 µm depth for each of these dilutions and their darknesses were measured by the machine vision program under 1× magnification and recorded (N = 300 droplets of each concentration). A strong linear relationship (R² = 0.9993) is seen between the bulk optical density of cell suspensions and the greyscale darkness of their generated droplets, enabling the label-free determination of OD₆₀₀ in individual microfluidic droplets.

Fig. 1. (a) A micrograph of a droplet generation device as seen by the machine vision program during droplet measurement with key regions overlain and annotated. (b) Flow chart of the machine vison droplet analysis algorithm. (c) A linear least-squares regression of the relationship of droplet darkness (arbitrary units; a.u.) measured by the machine vision program to bulk optical density of the dispersed phase. Error bars represent one standard deviation.

Additional calibration curves were made to examine day-to-day stability of the method as well as the possible range of droplet sizes for which cell concentrations can be measured. The analytical figures of merit for these curves are listed in Table 1. The calibration curves were found to be stable, with overlapping 95% confidence intervals for both slopes and intercepts from day-to-day and maintaining high linear correlation when using the same microfluidic device and with the incident light adjusted to the same intensity. Curves were found to be less repeatable between devices, possibly due to slight differences in device depth caused by imperfections in the photolithography process as well as occlusions or surface imperfections such as trapped dust and microbubbles in the PDMS, optical occlusions within the glass slide, and smudging on the outer surfaces of the chip. The variability suggests that each device should be calibrated before use. However, individual devices may be used dozens of times if care is used to prevent clogging. The absolute limits of detection (the number of cells in a single droplet) shown in the table were estimated by multiplying the concentration limit of detection by the measured volume of the droplet, using the conversion for E. coli of 8 × 10⁸ CFU mL⁻¹ at an OD₆₀₀ of 1.0 which was determined from plating experiments shown in Fig. S3, and matched previous studies.

Table 1. Analytical figures of merit for three calibration curves demonstrating statistically equivalent values when measured on two different days using the same device, as well as on a device of smaller dimensions, where picoliter-scale droplets are measured. Error on slopes and intercepts represents the 95% confidence interval.

Metric	Large droplets day 1	Large droplets day 2	Small droplets
Droplet volume (nL)	42 (6% RSD)	41 (5% RSD)	0.5 (13% RSD)
Slope (a.u.)a	9.0 ± 0.4	9.1 ± 0.3	1.80 ± 0.05
Intercept (a.u.)a	55.4 ± 0.8	56.0 ± 0.8	46.3 ± 0.1
R 2	0.9991	0.9993	0.9975
LOD (OD600)b	0.1	0.1	1.5
RSDc	0.4%	0.4%	2%
ALOD (cells per droplet)d	4000	4000	600

^aUncertainties reflect 95% confidence intervals. a.u. = arbitrary units.

^bLimit of detection determined as three times the standard deviation of the blank over the slope. LOD = 3S_b/m.

^cRelative standard deviation determined at the centroid (N = 300).

^dAbsolute limit of detection determined by conversion between OD₆₀₀ and cells/droplet using conversion factor of 8 × 10⁸ cells per mL at OD₆₀₀ = 1.

Smaller droplets down to 500 pL in volume were measured on a device of 50 µm depth. The shorter pathlength of the device resulted in an approximately proportional five-fold decrease in sensitivity. Increased variability was also observed, possibly due to the greater polydispersity in droplet volume. Additionally, as darkness is only measured in a single-frame snapshot, with a lower absolute number of cells in a droplet, darkness variability may be higher due to higher variability in their possible positional distributions within a droplet. The combination of lowered sensitivity and raised variability led to a 15×-higher limit of detection compared to the 200 µm calibration. However, due to the hundred-fold lower droplet volume, the absolute limit of detection decreased by a factor of 7 to 600 cells (Table 1). The bulk cuvette system used in this work has a linear dynamic range of approximately OD₆₀₀ 0.01–0.6. The droplet method therefore provides a method with higher limits of detection, but maintains linearity at higher cell densities, avoiding the need for the dilutions commonly used in bulk measurement. Additionally, the absolute limit of detection for the bulk method is approximately 8 × 10⁶ cells in a standard 1 mL cuvette, meaning the droplet system achieves lower absolute limits of detection by over four orders of magnitude.

A limitation of using greyscale darkness values to measure cell concentration is that transmission is not measured at a single wavelength, as in OD₆₀₀, but intensity is measured across the entire visible spectrum. Whereas wavelength selectivity is not needed to achieve high correlation in turbidity measurements, as the Mie scattering of E. coli cells (1–2 µm) is largely independent of wavelength, the method is not resistant to interference from molecular absorbers such as those found in “Luria” (LB) and “terrific” (TB) broth cell growth media. To better represent OD₆₀₀, multiple methods of filtering were employed and characterized by their response to yeast extract, which is the most abundant colored component in LB and TB media. The response curves of various filtering methods are compared to the unfiltered greyscale response shown in Fig. 2a.

Fig. 2. Instrumental diagram and response to varying concentrations of yeast extract within bacterial growth medium with (a) no wavelength filtering, (b) an in-line 580 nm bandpass filter, (c) isolated signal from red sensors in camera pixels, and (d) Sudan I yellow dye incorporated into the PDMS microfluidic device. Low slopes indicate low interference from background absorbance on a darkness measurement. Uncertainties represent 95% confidence intervals. Error bars represent one standard deviation. a.u. = arbitrary units.

The first method of wavelength filtering investigated was filtering the incident light with an in-line bandpass filter (580 nm/30 nm FWHM), as illustrated in Fig. 2b. This method results in a monochromatic image with significantly lower response to yeast extract, as shown by the slope of darkness to yeast extract concentration being reduced to 0.040 from 0.104 without filtering. The second method (Fig. 2c) involves isolation of the red pixel intensities within the camera sensor, which provides a broad response centered around approximately 600 nm, and avoids major interference from molecular absorbers in the blue region of the spectrum. While not as effective at removing interference from yeast extract as filtering incident light (seen by the slightly higher slope with increasing concentration of yeast extract), this method still removes most of the interference while maintaining color imaging which is essential when colored optical marker droplets are used for the alignment of optical droplet information to other modes of analysis such as MS.

The final method of removing molecular interference is by filtering the incident light, like the first method, but by doping the PDMS used to cast the device with dye. This method, shown in Fig. 2d, was most effective at reducing response to yeast extract, and avoids the need for costly optical filters and microscopes with filter options, but is limited to monochromatic imaging, and the filtering is a permanent feature of each chip with the formulated PDMS. An example device cast with the dye-doped PDMS is shown in Fig. S4. Calibration curves for determining OD₆₀₀ from droplet darkness from systems with both filtered incident light and with red pixel isolation showed identical limits of detection to the unfiltered greyscale method, as well as other similar analytical figures of merit (Table S1). In summary, filtering the light or isolating only red pixels from the camera can reduce the impact of other wavelengths on the measurement and provide different advantages.

It may be possible to achieve lower limits of detection by further improvements in the system. Devices of greater depth could increase the optical pathlength of droplets. A more sensitive and lower-noise image detector could be employed. It would be possible to obtain many measurements on each droplet and their values averaged to reduce the effect of measuring droplets at a slightly different position in the image. Specialized optics may be used to provide more uniform lighting and a reference measurement to correct for intensity variation of incident light. While such improvements are feasible, the current system is relatively simple and adequate for many studies that evaluate microbe growth.

3.2. Sorting droplets by growth and volume

One potential use of the real time measurement of OD₆₀₀ is to collect droplets that contain cells in a mixed population emulsion, such as generated from encapsulation of single cells that result in a Poisson distribution of cell colony containing droplets. Typically, when performing single-cell encapsulation, approximately 80% of droplets generated end up vacant of any cells, while the other 20% mostly contain single cells that are incubated and multiply to hundreds or thousands of cells.^10,42 To enrich droplet emulsions encapsulated with cells, the machine vision system can be used to analyze droplets based on cell growth. The camera also allows sorting based on volume and color of droplets. Sorting by volume allows for the removal of merged or broken droplets, which are a common source of noise in droplet microfluidics screens. A video of volume-based sorting is shown in SI Movie S2, where a mixed population of 5 nL and 10 nL droplets are sorted into two separate populations. Sorting by color allows for the capture of marker droplets optically identified by an included dye. Integrated electronic outputs of the camera are used to trigger a high voltage pulse for DEP sorting, positioned downstream of the camera detection, based on user-defined selection criteria.

To demonstrate such sorting, single cells were encapsulated in droplets of approximately 5 nL volume at a rate of 15% according to the Poisson distribution, and the droplet emulsion was incubated for 24 h at 30 °C. When visualized under a microscope, droplets containing cells are observed to be darker, due to increased turbidity, and merged droplets are seen around the edge of the micrograph (Fig. 3e). The incubated droplets were introduced into a binary sorting device, and the droplet detection region was placed approximately 1.8 mm before the opening of the bifurcation (Fig. 3a). Logic was programmed into the machine vision algorithm to provide a positive sorting signal for droplets that both contained cells and were 3–7 nL in volume. Droplets were introduced and sorted at ∼5 droplets per s. In this way, vacant droplets containing no cells, as well as merged, broken, and satellite droplets are directed to the negative sorting channel and removed from the emulsion, while monodisperse cell containing droplets are steered to the positive channel, resulting in an enriched fraction of droplets collected from the positive outlet. A video of the sorting is shown in SI Movie S3.

Fig. 3. (a) An annotated diagram of the sorting device used showing the placement of the droplet detection region of interest and electrode placement for DEP sorting. (b–d) Micrographs of the sorting of an emulsion of droplets incubated from single encapsulated cells such that approximately 15% of droplets are occupied showing (b) the measurement of a vacant droplet entering the sorting region (c) the measurement of a subsequent occupied droplet as the vacant droplet is sorted to the negative channel (d) the sorting of the occupied droplet to the positive channel. (e) Image of emulsions shown prior to sorting. (f) Droplets sorted to the negative channel where red arrows indicate false negatives. (g) Droplets sorted to the positive channel where green arrows indicate false positives.

An initial droplet (α) that does not contain cells can be seen entering the detection region in Fig. 3b. Upon determination of cell contents by darkness measurement, the droplet is seen being passively directed to the negative (lower) channel in Fig. 3c. At the same time, a second droplet (β) is observed entering the detection region. The darkness of the second droplet is indicative of cell presence, and it is appropriately sorted to the positive (upper) channel (Fig. 3d). A collected fraction of droplets from the negative outlet is shown in Fig. 3f, showing mostly vacant droplets, as well as several merged and broken or satellite droplets. A fraction of positive droplets is shown in Fig. 3g, showing a monodisperse population of droplets containing cells. The false positive and false negative rates of this sort at ∼5 droplets per s were determined from these micrographs to be 5% and 6%, respectively (total positive N = 300, total negative N = 84). These error rates a similar to many other droplet sorting methods and are suitable for many applications;^{3,23,24,28,43} however, for scenarios where higher sorting accuracy is required, sorting at lower rates and allowing for more space in between droplets has been shown to improve accuracy.¹ The maximum sorting rate of the device was determined to be ∼10 droplets per s for this system. This limitation is due to the combination of several steps within the hardware and software of the system that contribute to the total program run-time including exposure, image transfer, algorithm execution, and readout. Image transfer time (the amount of time required to transfer the image information from the sensor to the processor) can be minimized by reducing the height of the captured image. This reduction can increase the frame rate of the method, but it becomes limited by algorithm execution and readout times. A detailed breakdown of each step by the amount of time it takes is shown in Table S2. A machine-vision camera with more powerful image-processing and readout hardware would decrease the total run-time of the method by reducing these steps, which would increase the maximum possible sorting rate.

The sorting of emulsions by cell growth may have several applications including to isolate highly proliferative cells (tumor analysis) and droplet-based antibiotic screening.^6,7 Another example involves improving throughput when analyzing droplet contents by methods like MS. Removing empty droplets prior to a downstream screening method will increase throughput for any droplet screen with a lower screening rate than the sorting rate according to the equation k = F_p/(F_m + αF_p) where F_p is the pre-sort rate, F_m is the screening rate, α is the fraction of occupied droplets, and k is the throughput-increase factor. Although optical detection methods can sort rapidly, mass-activated droplet sorting (MADS)^2,43 is relatively slow and analysis of empty droplets can extend screening times. In MADS, droplets are split into two droplets, one of which is introduced to a mass spectrometer, and the other flows through a delay line towards a DEP sorter. The paired droplets reach the mass spectrometer and sorter at synchronized times, and droplets are sorted based on their MS abundance of a target analyte at rates of ∼0.5 Hz. If pre-sorting were implemented in a MADS screen with singly encapsulated cells (λ = 0.2), a pre-sort rate of 8 droplets per s, and a screening rate of 0.5 droplets per s, the analytical throughput would be increased 4-fold compared to without the pre-sort.

3.3. Correlation of growth to E. coli metabolism in droplets

An important application of droplet microfluidics is to enable rapid screening of cell phenotypes for enzyme engineering or synthetic biology. In such applications, a desired chemical product may be detected and used to sort^1–3 or collect cells.^44–46 A confound in such measurements may be that amount of product formed may also relate to number of cells that have grown in the colony which may be variable. For example, it has been observed in previous studies that cell-containing droplets that had been incubated from single cells showed high variability in lysine metabolic production compared to droplets of the same cell line generated from a bulk suspension of cells.² Growing from single cells in a droplet emulsion, such growth variability may arise from several factors including slight differences in conditions experienced during incubation or individual variability in lag phase length between cells.^19,27,47 The ability to directly measure cell density via OD₆₀₀ and correlate it to a phenotype measurement, such as amount of a product formed, could help better characterize a cell culture and reduce variability in measurements. We tested this idea by measuring OD₆₀₀ and MS abundance for lysine in a collection of DapA engineered E. coli cells.

We initially performed a control experiment in which we encapsulated cells grown in bulk at different concentrations into droplets and then measured the OD₆₀₀ and lysine production. For this experiment, a suspension of lysine-producing E. coli was diluted with fresh EZ media on a two-inlet droplet generator device. Independent syringe pumps were used to pump media and E. coli suspension at 0.05–1.25 µL min⁻¹ in oscillating triangle wave patterns with a 2 min period, 180° out of phase from each other. This procedure yielded a sequence of droplets that had different concentrations in a sawtooth temporal pattern. A micrograph of the two-inlet gradient generator device operating with blue dye is shown in Fig. S5. The darkness of each droplet was measured immediately following their generation and then subsequently introduced to the mass spectrometer where the abundance of lysine was measured as the maximum intensity of each individual droplet. Synchronized traces of droplet darkness and lysine abundance are shown in Fig. 4a, displaying the oscillating gradient pattern. The darkness and lysine abundance of each droplet can be plotted against each other, which is shown in Fig. 4b, to demonstrate extremely strong correlation between cell and lysine concentration with droplets generated from the bulk. This correlation demonstrates that OD₆₀₀ and MS signals are related indicating that the number of cells (grown under identical conditions) governs the amount of lysine produced.

Fig. 4. Correlation of cell density of 3× DapA E. coli in 30 nL droplets, measured by machine vision (arbitrary units; a.u.), and lysine production measured by MS. Droplets generated on a two-inlet device where flowrates of a suspension of lysine-producing cells are diluted in a cyclic gradient with growth medium. (a) Traces plotted over time, showing the correlated gradient pattern can be observed in cell density and lysine abundance. (b) The abundance plotted vs. darkness fit to a linear regression, where each point represents an individual droplet and r is the Pearson correlation coefficient.

In experiments where droplets containing different concentrations of cells are randomly distributed it will be necessary to match each mass spectrometric signal to optical signal for the same droplet. Since the optical measurement is necessarily done prior to droplet-ESI-MS there is a delay that requires aligning or registering of the droplets for each signal. In principle this alignment is easily done with a fixed time interval; however, variations such as flow fluctuation and droplet fission during travel could confound such simple tracking. To more accurately match droplets, we doped mixtures of cell-containing droplets with marker droplets at a 5-to-1 ratio. The marker droplets contain a blue dye, for an optical signal, and neostigmine for a mass spectrometric signal. Detection of these droplets are used as time points for alignment and ensure registry of droplet OD₆₀₀ and MS signal. The measurement of a cell-containing droplet by the machine vision program is shown in Fig. 5a. It is identified as not being a marker droplet because of the lack of blue color (marker value of 0). Fig. 5b shows a marker droplet identified by its blue color (marker value of 1). The marker droplet trace from the machine vision system can be aligned with the neostigmine trace from the mass spectrometer by the pattern created by the stochastic distribution of marker droplets as seen in Fig. 5b. Plotting the observation time of individual marker droplets on both systems against each other reveals that there is minimal drift between the two systems, characterized by a slope of approximate unity, as seen in Fig. 5c, and a simple global offset is used to synchronize them. The residuals plot of the synchronized markers shown in Fig. 5d shows that there is never more than 3 s of desynchronization between the two measurement systems. Desynchronization is likely due primarily to flow rate noise between the point of optical measurement and the point of mass measurement.⁴⁴

Fig. 5. The synchronization of optical droplet data obtained by the machine vision algorithm to mass spectrometric data achieved by marker droplets. Marker droplets containing blue dye and neostigmine are mixed into the droplet emulsion at about a 15% proportion and are identified by the machine vision algorithm through color matching where analyte droplets receive a marker value of 0 (a) and marker droplets a value of 1 (b). These marker droplets are (c) matched to the mass spectrometer by the pattern of markers seen in the neostigmine spectrometer chronogram. Synchronization (d) is performed by a fitting a linear regression of time points for marker droplet detection between systems.

With the ability to align the OD₆₀₀ and MS signals, we investigated the effect of growing cell colonies in droplets, starting from a small (<10⁶ CFU mL⁻¹) number of cells in the droplet, on metabolite abundance. For this experiment, droplets were generated from a dilute 3× DapA cell suspension such that >99% of droplets contained 10–40 cells initially. The droplets were allowed to incubate in lysine deficient EZ-media at 30 °C for 24 h in a vial and the droplets were mixed by inversion every 8 h. Droplets were then reinjected into a respacer chip, where their darkness was measured online prior to introducing them to the mass spectrometer. Subsequent OD₆₀₀ and MS signals were aligned using marker droplets.

This experiment revealed a biphasic distribution of lysine abundances (Fig. 6) where droplets with OD₆₀₀ < 0.5 showed slight negative correlation (r = −0.32) between optical density and lysine signal and droplets with OD₆₀₀ > 0.5 showed moderate positive correlation (r = +0.55). The biphasic nature of the relationship can be rationalized by the growth dynamics of E. coli. At low densities, droplets are undergoing a cell-growth phase, where residual lysine in the media is consumed to support bacterial proliferation, explaining the initial negative correlation. However, after a certain density is reached (OD₆₀₀ ∼ 0.5 or ∼10 000 cells per droplet), the production of lysine via the DapA pathway can be sustained and the cells produce lysine at a faster rate than they consume it leading to an observed increase in lysine abundance at higher optical densities. Some of the initial negative correlation may also be explained by consumption of glutamine, which is isobaric to lysine and present at low concentration in the media, although lysine signals are much higher.²

Fig. 6. The correlation of optical density to mass spectrometric abundance of lysine in droplets incubated with 3× DapA E. coli. Droplets were incubated for 24 h following initial encapsulation of 10–40 cells per droplet. Independent linear regressions are fit for droplets with an OD₆₀₀ < 0.5, and droplets with OD₆₀₀ > 0.5, where r is the Pearson correlation coefficient. In droplets with higher growth levels, increased levels of lysine are observed, which allows for correction of the abundance signal by optical density.

Overall, a reasonable correlation is seen between OD₆₀₀ and lysine abundance (r = +0.71 for all droplets), but the graphs show much more noise than that seen for cells grown in bulk and diluted (Fig. 4b), suggesting that factors other than cell density play a significant role in the lysine production by colonies grown within droplets as detected by ESI-MS. Nevertheless, the use of OD₆₀₀ does reduce variability in the lysine signal (i.e., when the lysine signal is normalized to the amount of cells as measured by OD₆₀₀, the RSD for the lysine signal in all droplets was 54%, compared to 84% without normalization). Thus, measuring OD₆₀₀ can account for some variability in metabolite production in this system.

4. Conclusion

We demonstrate a label-free system for the measurement of cell concentration in microfluidic droplets at 10 droplets per s rates. The machine vision platform avoids the need for fluorescent or colorimetric reporters, embedded optical fibers, or complex droplet microscopy employing slits.^5,7,19–21 While other imaging systems have succeeded in sorting vacant droplets from cell-laden droplets, the quantitation of bacteria and measurement of size allows for isolation of specific growth levels and volumes.^4,23,24,28 Compared to bulk measurement, the minimum detectable number of cells was decreased by four orders of magnitude, and the linear dynamic range allows for higher density measurement without dilution; however, the lowest concentration of cells detectable was higher. Droplets were successfully sorted by volume and optical density, allowing for enrichment of microfluidic emulsions by removing vacant, merged, and broken droplets, increasing throughput in screening campaigns. Additionally, density measurement was coupled with mass spectrometric detection and metabolic production of E. coli was correlated to cell growth. Whereas our main goal is to use this technology for enabling directed evolution of biocatalysts, other applications are possible. Such potential applications include detection and isolation of highly proliferative tumor cells, which could be differentiated based on their growth level,^48,49 or the measurement of cell density in body fluids (e.g. urine, plasma, or cerebrospinal fluids) using only nanoliters of sample.^50–52 Additionally, the sorting system could be used in antibiotic susceptibility assays to separate cells with specific antibiotic resistance.^6,7

Conflicts of interest

There are no conflicts to declare.

Data availability

Key raw data sets collected in this work as well as the code for the machine vision algorithm used in this study can be found in the Zenodo repository at https://doi.org/10.5281/zenodo.22881518. The raw data was processed in Microsoft Excel and the code employed for this study was written and executed on Cognex In-Sight Explorer version 6.5.0.

Supplementary information (SI): annotated drawings of the microfluidics devices used in this study, photographs of the instrumental setup and formulated microfluidic chip, micrographs of gradient droplet generation, analytical figures of merit for filtered calibration curves, experimental information for determining the conversion factor between OD₆₀₀ and CFU mL⁻¹, run-times of the image-processing algorithm, and videos of droplet detection, measurement, and sorting. See DOI: https://doi.org/10.1039/d6an00623j.