Combinatorial nanodroplet platform for screening antibiotic combinations

Abstract

The emergence and spread of multidrug resistant bacterial strains and concomitant dwindling of effective antibiotics pose worldwide healthcare challenges. To address these challenges, advanced engineering tools are developed to personalize antibiotic treatments by speeding up the diagnostics that is critical to prevent antibiotic misuse and overuse and make full use of existing antibiotics. Meanwhile, it is necessary to investigate novel antibiotic strategies. Recently, repurposing mono antibiotics into combinatorial antibiotic therapies has shown great potential for treatment of bacterial infections. However, widespread adoption of drug combinations has been hindered by the complexity of screening techniques and the cost of reagent consumptions in practice. In this study, we developed a combinatorial nanodroplet platform for automated and high-throughput screening of antibiotic combinations while consuming orders of magnitude lower reagents than the standard microtiter-based screening method. In particular, the proposed platform is capable of creating nanoliter droplets with multiple reagents in an automatic manner, tuning concentrations of each component, performing biochemical assays with high flexibility (e.g., temperature and duration), and achieving detection with high sensitivity. A biochemical assay, based on the reduction of resazurin by the metabolism of bacteria, has been characterized and employed to evaluate the combinatorial effects of the antibiotics of interest. In a pilot study, we successfully screened pairwise combinations between 4 antibiotics for a model Escherichia coli strain.

Graphical Abstract

We have developed a combinatorial nanodroplet platform for screening antibiotic combinations and successfully screened drug response of pairwise antibiotic combinations from selected antibiotics using the platform.

Introduction

Bacterial infection is a worldwide healthcare concern and among the leading causes for morbidity and mortality ^1–3. It results in at least 2.8 million antibiotic-resistant infections, among which 35,000 deaths, and as high as $20 billion healthcare costs in the United States each year ⁴. The trend continues and is getting worse as multidrug resistant bacteria that typically result from the misuse and overuse of antibiotics are widely spreading while we are running out of effective antibiotic candidates to treat these superbugs, which raises genuine concerns for the so-called resistance era and narrow-spectrum era ^5–7. Tremendous efforts have been made to address these challenges. For instance, advanced diagnostic strategies have been developed to detect the presence of bacteria, identify the pathogenic strains, and conduct antimicrobial susceptibility testing in a timely manner, which facilitate precise and systematic management of bacterial infections and assist physicians in personalizing antibiotic prescriptions based on evidence instead of empirical experiences ^8–37. Similarly, although the development of novel antibiotic agents lags behind due to the lengthy and costly procedures involved and the rapid development of bacterial resistance ^{5, 6, 38}, combinatorial antibiotic therapies that combine various antibiotics as a treatment strategy are able to empower existing antibiotics with exceptional bacteriostatic/bactericidal performance and fuel to combat for the threat of bacterial infections ^39–44. For example, combinatorial therapies are the predominant treatment for Mycobacterium tuberculosis, and the treatment of multidrug-resistant Gram-negative infections typically requires a combination of up to 3 antibiotics ⁴⁵. Therefore, there is a critical need to screen combinations of existing antibiotics to prolong the effectiveness of the antibiotics and offer an additional tool for combating antibiotic resistance.

The screening for antibiotic combinations is typically performed using microtiter plates in current practice ^46–48. These approaches may require labor-intensive procedures and most notably high reagent consumption as the number of antibiotic combinations increases exponentially with the number of antibiotics to test. For instance, screening of pairwise combinations from a library of 1000 drugs needs nearly 8 million tests, assuming only 4 concentrations are tested for each drug and no experimental replicate is yet considered. Screenings at this scale may easily consume a whole chemical inventory since the minimum reaction volume for each microtiter plate test is at the microliter level. To this end, novel platforms are investigated to achieve high throughput screening with a reduced amount of reagent consumption. For instance, delicate microfluidic designs are demonstrated to conduct combinational drug screening in microreactors and hold capability for multiplex reactions and in situ analysis ^49–55. By incorporating advances in liquid handling techniques, such as inkjet printing, contact printing, and acoustic ejection, drug screening is realized by combining reagents in a step-by-step manner that enables a high degree of freedom for reagent combinations and allows for multistep screening reactions ^56–64. Alternatively, nano/pico-liter droplet techniques hold great potentials for combinatorial screening since multiple reagents can be readily configured in a droplet and droplet systems typically allow high scalability testing ^65–72. However, these techniques do not fully meet the needs for combinatorial antibiotic screening, which requires on-demand combination conditions with little reagent consumption, automated screening procedures at high throughput, and easy data analysis and interpretation.

In this study, we present a combinatorial nanodroplet platform for screening antibiotic combinations, which allows programmable generation of droplets with reagents on demand, automates high-throughput screening procedures, and supports straightforward readout and data analysis. In particular, the droplets are created by combining desired reagents, such as bacteria, culture medium, fluorogenic bacterial viability indicator dye, and antibiotics of interest, and the ratio of involved reagents can be tuned in each droplet by the volume of respective reagents. Importantly, this droplet generation process is automated via a MATLAB program. Then, the droplets are orderly transferred into a tubing where fluorescence biochemical reactions can be performed within all droplets in a parallelized and high-throughput manner.

Since the droplets are spatially coded in the tubing, they can be subsequently ejected through a detection channel in order and sequentially detected by a laser-induced fluorescence detector to quantitatively evaluate the drug response. This approach directly correlates the fluorescence signal with respective combinatorial conditions in droplets, which thereby enables simple data analysis and interpretation. To demonstrate the feasibility of the platform, we screened all pairwise combinations between 4 antibiotics. The results are consistent with benchtop results using standard multiwell plates and are correlated with findings in other groups.

Results

Design of the combinatorial nanodroplet platform for screening antibiotic combinations

We developed a combinatorial nanodroplet platform for screening antibiotic combinations, which consisted of a microdevice for on-demand droplet generation, a segment of tubing for high throughput assays in the droplets, and a detection system for assay readout (Figure 1a). The microdevice was designed with a reagent loading layer, a valve control layer, and a glass substrate (Figure 1a, step 1). The device was fabricated using a standard soft lithography process (Supplementary Figure S1). The reagent loading layer contained microchannels that served as the inlets for various reagents. Correspondingly, there were individual pneumatic valves on the valve control layer, which could independently open or close the sample loading channels to create droplets with on-demand contents at the nanoliter level (Figure 1b, step 1, the horizontal gray and black blocks). These highly packed reagents are coalesced into a single droplet in the microchannel. A segment of oil was loaded between adjacent droplets to prevent droplet merging and served as the continuous phase in the system. This droplet generation process was automated and systematically managed using a MATLAB program. Then, the droplets were transferred into a tubing where fluorescence biochemical assays were implemented to examine the effect of antibiotic combinations (Figure 1a, step 2 and Figure 1b, step 2). Of note, the tubing enabled flexible assay conditions, such as the temperature and assaying time, and allowed high throughput testing in parallel. Lastly, the droplets were subsequently loaded to a detection microchannel, and the fluorescence intensity of the droplets was orderly captured using a laser-induced fluorescence detection system when they passed through the laser beam (Figure 1a, step 3). As the drug response was correlated with the fluorescence signal in this study, the data set of the droplet array was analyzed to study the drug response of antibiotic combinations (Figure 1b, step 3).

Figure 1. Schematic for the combinatorial nanodroplet platform for high-throughput screening of antibiotic combinations.

(a) The combinatorial nanodroplet platform consists of a device for droplet generation (step 1), a tubing-based droplet assaying (step 2), and a laser-induced fluorescence detection system for assay readout (step 3). (b) Schematic for the drug combinations in droplets (step 1), assaying in independent droplets (step 2), and the evaluation of the drug response by the fluorescence signal of the droplets (step 3).

Droplet generation with on-demand size and content

The droplet generation was implemented in a reagent loading layer via pneumatic microvalves in a valve control layer, which were represented by filling with blue and red food dyes, respectively (Figure 2a). The reagent loading layer consisted of 14 independent microchannels with reagent supplies, which were individually opened or closed by pneumatic microvalves in the valve control layer. The droplet size/volume was tuned by the opening time of the microchannels. In particular, we captured the droplet size with the valve opening time ranging from 0.1 s to 4 s, which was the region of our interest in this study (Figure 2b). As the cross-section of the microchannel was fixed, the droplet size was indicated by the length of the droplets. To quantitatively analyze the droplet size, a series of droplets were generated with fluorescence dye and the length of the droplets was measured by detecting the fluorescence signal of the droplet when it passed through the laser-induced fluorescence detector (Supplementary Figure S2). The droplet size linearly increased over the valve opening time, suggesting that the droplet size can be precisely controlled via the valve opening time (Figure 2c).

Figure 2. Characterization of the combinatorial nanodroplet system.

(a) The photography of the device. The reagent loading layer and the valve control layer were represented by filling with blue and red food dyes, respectively. (b-c) Calibration of the droplet size with the valve opening time. For quantitative analysis, the droplet size was normalized to the internal reference points where the valve opening time was 2 s (n = 9) (c). (d) Generation of droplets with combinations of reagents. The opening microchannels were highlighted during droplet generation. (e) Droplets combined with various volume ratios between water and a green food dye were captured to visualize the combination results. (f) The combination was quantitatively analyzed by tuning the volume ratios between the fluorescein isothiocyanate (FITC) dye and DI water. The internal reference points were the pure fluorescence dye (n = 11). (g) The outlet of the microchannel was designed with gradient height (i.e., 3 height steps) to ensure smooth transition of the droplets into tubing. Time-lapse images of the droplet transition were shown in the close-up view. (h) The droplets were orderly transferred into tubing for parallelized and high-throughput biochemical assays. Close-up view of a droplet was represented on the right. (i) The droplets were subsequently delivered to a microchannel and sequentially detected using a laser-induced fluorescence detector. Each detection chip consisted of a pair of microchannels. The ends of the tubing were interfaced across the microchannels; whereas the left microchannel was connected with a pump to push the fluid system and the other served as the detection channel. The flow direction, droplets, and detection system are depicted in blue, yellow, and red, respectively; the ends of the tubing and the outlet of the pump are presented. Data represent mean ± SEM. Scale bar: a and i, 5 mm; b, d, e, and g, 1 mm.

Combinations of reagents within a single droplet were achieved using the droplet generation device. Desired reagents were orderly loaded into a microchannel, which was pre-treated with Novec HFE-7500 fluorinated oil to facilitate coalescing of reagents. Using water and two food dyes as reagents and programmable opening the respective microchannels, we demonstrated the generation of droplets that contain 1, 2, and 3 reagents (Figure 2d, Supplementary Movie S1). Droplet coalescence occurred upon reagents contact, which obviated engineering assistance, such as electrocoalescence, physical confinement, and synchronized design ^{68, 70, 73, 74}. Two reagents (e.g., water and green food dye) were combined at various volume ratios within a single droplet to visually demonstrate on-demand droplet generation with varied concentrations. As the volume ratio of the green food dye increased – controlled by the opening time ratio of its microchannel – the green hue of the droplet became visibly darker (Figure 2e). Moreover, the combination was quantitatively analyzed by tuning the volume ratio between the fluorescein isothiocyanate (FITC) dye and DI water, where the concentration of FITC was linearly related with the fluorescence intensity (Supplementary Figure S3). The fluorescence intensity of droplets was measured using FITC droplets as internal references to ensure consistency of detection across experiments. The fluorescence intensity of the droplets linearly increased with the valve opening time ratio, demonstrating the accuracy of on-demand combinations of reagents (Figure. 2f). The near-zero background between fluorescence peaks of droplets was ascribed to minimal cross-contamination across droplets (Figure 2f, Supplementary Figure S4a, S4b). Of note, our system was capable of maintaining uniform droplet size when the total opening time for droplet generation remained constant, which can be critical for applications that require uniform droplets for assaying and/or analysis (Supplementary Figure S4c).

Tubing interface for biomedical assays in droplets

The droplets were transferred into tubing to perform fluorescence biochemical assays to study the drug response of antibiotic combinations. We engineered the outlet of the microchannel with gradient height (e.g., 3 height steps in the device) to avoid breakage of the droplets and enable smooth transition of the droplets into the tubing ^{75, 76} (Figure 2g, Supplementary Movie S1). The length of the droplet decreased as the channel height increased. Moreover, this design assisted mixing of reagents. For instance, the food dye and DI water were well mixed in the droplet before it reached the tubing interface. The droplets were transferred into the tubing in sequence, which allowed spatially coded assaying conditions (Figure 2h). It avoided the efforts to identify the droplets, such as barcoding with fluorescence dyes ^{66, 68} and coding with colored beads ⁶⁵, which raised another layer of data analysis and interpretation. Meanwhile, the tubing interface separated the droplet generation and incubation modules, which could facilitate the repeated usage of the droplet generation device. Moreover, the tubing was made of Teflon that is compatible with a variety of biochemical assay conditions (e.g., temperature and duration). Importantly, using the gas impermeable Teflon tubing with two ends sealed, we designed an airtight environment for the droplets to minimize evaporation during the incubation (e.g., negligible evaporation after incubation at 37 °C for 24 h (< 2% change of the droplet size, data not shown)). After proper biochemical assays, the droplets were delivered into a detection channel, where the laser-induced fluorescence detector was applied underneath to measure the fluorescence signal (Figure 2i, Supplementary Movie S1). As the droplets passed the laser beam, the average fluorescence signal of each droplet was recorded in a LabVIEW program and analyzed in MATLAB to evaluate the drug response.

Characterization of a fluorescence biochemical assay to evaluate drug response

A fluorescence biochemical reaction, where resazurin can be converted to resorufin by bacterial metabolism, has been employed to evaluate the drug response ^{18, 25, 70, 71, 77}. The resorufin was quantitatively detected and analyzed in fluorescence mode to measure the bacterial activity and the drug efficiency (i.e., the more resorufin produced in an assay, the lower drug efficiency). Escherichia coli (E. coli) was investigated as the model strain in this study. We optimized the assay conditions to evaluate the drug response. Firstly, various bacterial concentrations were tested to study the bacterial response toward drug treatment (Figure 3a). For the drug control groups, bacteria were treated with cefsulodin at the minimum inhibitory concentration (Supplementary Figure S5). We found that the bacterial concentration of 5×10⁶ cfu/mL enabled the most distinct differentiation between the groups in the absence/presence of the antibiotic and a high signal-to-noise ratio was realized after 5 h incubation (Figure 3a, middle panel). Secondly, the resazurin concentration was optimized in the droplet system. We measured the signal change over bacterial growth (5×10⁶ cfu/mL; 37 °C for 5 h) in the absence of antibiotics at three resazurin concentrations (Figure 3b). The results suggested that the highest signal change was achieved at resazurin concentration of 25 μM, while further increasing the resazurin concentration would raise the background signal and saturate our fluorescence detector after bacterial growth. Moreover, under these optimized conditions, there was no bacterial contamination detected during the incubation in the droplet system (Supplementary Figure S6).

Figure 3. Characterization of a fluorescence biochemical assay for evaluation of the drug response.

(a) The fluorescence signal change of bacterial response in the absence/presence of antibiotics at different inoculum concentrations in 96-well plates (n = 3–4). (b) The fluorescence signal change over bacterial growth at various resazurin concentrations in droplets (n = 4). Data represent mean ± SEM.

Screening for pairwise antibiotic combinations

We piloted a study to screen pairwise antibiotic combinations from four antibiotics using the proposed system, including Penicillin G (PEN), oxacillin (OXA), piperacillin (PIP), and cefsulodin (CEF). The antibiotics were in the same category (i.e., β-lactams targeting bacterial cell wall), which were selected due to the high potential of the synergistic effect that is the major target of interest ^{46, 78–80}. An array of antibiotic pairs consisting of various antibiotic concentrations, which was set at the MIC, half MIC, a quarter of the MIC, and 0, was established to evaluate the effect of drug combinations (Figure 4a). The assays were orderly created in droplets using the device in an automated manner. For instance, the final concentration of antibiotics A and B were MIC/2 and MIC/4 in the #7 droplet. The droplet array was transferred in tubing, incubated at 37 °C for 5 h, and delivered to the detection device to measure the fluorescence signal (Figure 4b). The signal was normalized between the maximum and minimum points and allocated to 8 subgroups to evaluate the combination effect of the drug pairs. Each subgroup consisted of 4 signal points, where one antibiotic dose was fixed and the other was variable, such as the individual rows or columns highlighted in Figure 4a. The subgroups were fitted using the dose-response curve model in GraphPad Prism 8 (Figure 4c). For instance, the blue curve represented the drug response as PEN dose increased in the absence of CEF; the red curve indicated the drug response when PEN concentration increased and CEF dose was fixed at the MIC/4. The IC₈₀s were determined for each subgroup (i.e., the cross-points between the dot line and the fitting curves). Then, the IC₈₀s were plotted in the x-y plane (Figure 4d). Each axis represented the normalized concentration range of respective drug, from 0 to 100% of the MIC. The IC₈₀s of all subgroups were mapped into the figure. For instance, the orange and blue points were the IC₈₀s of CEF and PEN independently; the light green and red points were the IC₈₀s of the subgroups where the IC₈₀s of PEN were calculated with a fixed CEF concentration at the MIC/2 and MIC/4 respectively. The combination effect was studied using the standard Loewe additivity model ^{46, 50, 70, 81, 82}, which compares the connection line of all IC₈₀s with the straight line between the IC₈₀s of independent drugs (Figure 4d, solid black line and dotted red line respectively). A concave connection line indicates a synergistic effect while a convex connection line suggests an antagonistic effect. In particular, our results revealed that the combination between CEF and PEN was a synergistic pair for the model E. coli (Figure 4d, highlighted in red). To visualize the signal of the whole group, the fluorescence intensity of the droplet array was normalized in a heatmap and was set underneath the plot of IC₈₀s. Moreover, our results suggested that all combinatorial pairs between the selected antibiotics were synergistic to treat the model E. coli in this study (Figure 4e).

Figure 4. High throughput screening for antibiotic combinations.

(a) Design of the drug combinations and sequence for evaluating the combinatorial effect. The points represent the pair-wise combinations at respective antibiotic concentrations. The rows and columns are subgroups for data analysis in (c). (b) Representative fluorescence signal in a set of droplets in the same order of (a). (c) The fluorescence signal was allocated to 8 subgroups (e.g., the rows and columns in (a)) and the IC₈₀s were determined for the subgroups (i.e., the cross-points between the dot line and the fitting curves). (d) The IC₈₀s were plotted in the x-y plane, where each axis represented the normalized concentration range of respective drug, from 0 to 100% of the MIC, and the IC₈₀s of all subgroups were correlated in color with the cross-points in (c). The difference between the connection line of all IC₈₀s and the straight line of the IC₈₀s of independent drugs is highlighted in red. The heatmap underneath the plot of IC₈₀s represented the normalized fluorescence intensity of the droplet array. Deeper color indicated higher intensity. (e-f) Evaluation of combinatorial effects using the nanodroplet platform (e; n = 4) and multiwell plates (f; n = 2) for antibiotic pairs from four antibiotics. The synergistic effects are highlighted in red.

Moreover, we investigated the combinatorial effect of the antibiotic combinations between involved antibiotics using 96-well plates in benchtop experiments. In particular, the bacteria was cultured in the presence of antibiotic combinations, where the antibiotic concentration was designed following the pattern in figure 4a. The drug response was studied by bacterial growth, measured by optical density, in each condition. Using a similar analysis approach applied in the nanodroplet system, we found that the connection lines of IC₈₀s were in concave shapes compared with the straight lines between IC₈₀s of independent drugs in all the pairwise combinations between selected antibiotics, indicating synergistic effects of these combinations (Figure 4f). The results tested by our nanodroplet platform were consistent with that obtained from the standard method (i.e., microtiter plate) and correlated with the findings in the literature ^{46, 50}.

Conclusion and discussion

In this study, we demonstrate a combinatorial nanodroplet platform for screening antibiotic combinations. This platform is able to create nanoliter droplets with desired reagents in an automated manner. The size of the droplets reduces the reagent consumption by 3 orders compared to the traditional microtiter plate methods. The integration of the tubing interface supports parallelized assays with flexible assay conditions. These features collectively support the implementation of high throughput drug screening. Moreover, we characterized the resazurin-based assay to evaluate the drug response for antibiotic combinations. The feasibility for low-volume and high-throughput drug screening was demonstrated by screening pairwise antibiotic combinations to combat for a E. coli strain using the proposed platform.

Although combinatorial antibiotic therapies have become promising strategies for the treatment of bacterial infections, the widespread adoption of drug combinations has been hindered by the complexity of the screening techniques and cost of reagent consumptions. The microtiter plate is currently the most widely used tool for drug screening in current practice, which can be labor intensive and costly for large scale applications. Although significant progress has been made in the field, these platforms require delicate engineering designs (e.g., inkjet printing, contact printing, or complex microfluidic device design), extra efforts for assay identification (e.g., barcoding), or complicated data analysis (e.g., decouple the signal from the ID information). Our platform directly addresses these needs for drug screening. In particular, the flexibility of creating nanodroplets with on-demand contents supports to screen antibiotic combinations with low reagent consumption and the automation of the system enables the scalability for high-throughput screening.

The proposed platform represents a promising engineering tool for high throughput screening of antibiotic combinations. In the future, screening high order antibiotic combinations using various antibiotics at fine concentrations for different bacteria strains can be implemented using the proposed platform to improve our understanding of combinatorial effects (e.g., antagonistic and synergistic effects) and strenghten the capability of developing novel antibiotic treatment strategies. Meanwhile, the number of microchannels for sample loading should be increased to expand drug candidates for combinations. To optimize the potentials for extremely high-throughput screening applications, the capability for manipulating devices in parallel should be improved. Integration with other engineering designs, such as in-situ analysis of droplets and automated analysis algorism, can facilitate the platform with higher degree of simplicity and enable in parallel data analysis and interpretation.

Experimental Section

Device fabrication

The droplet generation device consisted of a reagent loading layer, a valve control layer, and a glass substrate (Supplementary Figure S1). The device was fabricated using a soft lithography process. In particular, the reagent loading layer mold was created with one layer of positive photoresist (SPR220–7.0) and four layers of SU-8 (SU-8 3050). The first layer was the SPR220–7.0, which was spin-coated on a silicon wafer at 1600 rpm for 60 s. This layer was used at the intersection of the valve control region and fluidic channel region to ensure the control of the valves ⁸³. The four layers of SU-8 were subsequently spin-coated on the wafer at 1500 rpm for 60 s. The valve control mold was fabricated by spin-coating a SU-8 layer on a silicon wafer at 2500 rpm for 60 s. Premixed polydimethylsiloxane (PDMS, from Sylgard 184), at a ratio of 10:1 between prepolymer and crosslinker, was poured on the reagent loading layer mold and cured at 80 °C. To fabricate a thin layer of the valve control channels, PDMS was spin-coated on the valve control mold at 1000 rpm for 1 min and cured at 80 °C. Then, the valve control layer was peeled off from the mold with the help of a sacrificed PDMS sheet on top and bonded with a glass slide after oxygen plasma treatment for 1 min. The reagent loading layer was bonded with the valve control layer after oxygen plasma treatment for 1 min. All devices were stored at 80 °C in an oven (Isotemp Oven 637F, Fisher Scientific) to avoid contamination.

Droplet generation and characterization

The droplet was generated and characterized in the device. First, reagents were pressurized in the inlets. Sample loading was systematically managed via a MATLAB program, which is able to open or close each individual pneumatic microvalve and tune the opening time. The reagents were coalesced when they encountered in the microchannel, which was pre-washed with Novec HFE-7500 fluorinated oil (Sigma). Droplets with on-demand contents were generated by tuning the volume/ratio of involved reagents. A segment of oil, containing a nonionic fluorous-soluble surfactant, 1H, 1H, 2H, 2H-Perfluoro-1-octanol (PFO, Sigma-Aldrich), in the HFE-7500 fluorinated oil (final concentration was 1.2%), was loaded between adjacent droplets and served as the continuous phase oil in the system. The process was monitored using a microscope (Olympus, IX71) equipped with a digital camera (Canon, EOS 60D).

Fluorescein isothiocyanate (FITC) dye was loaded in the droplets to characterize the droplet size and content. The droplet size was tested by measuring the length of the droplet when it passed through a fixed laser beam. The fluorescence signal was tested at excitation/emission wavelength of 488 nm/525 nm using the laser-induced fluorescence detector. The concentration of the involved reagents in each droplet was calibrated using the fluorescence intensity of the droplets, which was linearly correlated with the concentration of FITC dye (Supplementary Figure S3). The droplets were generated by tuning the volume ratios between DI water and FITC, where the FITC concentration ranged from 0% to 100% and also covered representative concentrations of 10%, 25%, and 50% (meanwhile, the counterpart DI water was at 90%, 75%, and 50%, respectively). The droplets were metered using the same setting in the detection system.

Droplet transportation into tubing

The droplet was transported from the droplet generation device to tubing to perform a fluorescence biochemical assay (Figure 1 and Supplementary Movie S1). The outlet of the droplet generation device was created by punching a hole with a 21G probe needle. The tubing was purchased from Cole-Parmer Instrument Company. It is made of Teflon. The inner and outer diameter were 0.31 mm and 0.76 mm, respectively. The length of the tubing was ~80 cm in our experiments. The tubing was washed with bleach (10% v/v in DI water) for 1 min and ethanol (70% v/v in DI water) for 1 min and stored at 80 °C in the oven until use. The tubing was directly plugged in the outlet and the droplet was smoothly transferred from the device to the tubing. The device part at the interface between the device and tubing was monitored using the microscope and the digital camera mentioned above. Meanwhile, the droplet motion in the tubing was captured using a cellphone camera (Redmi, Xiaomi) and/or a digital camera (T3i, Canon).

Characterization of a biochemical assay to study the drug response

The four antibiotics used in this study, including Penicillin G (PEN), oxacillin (OXA), piperacillin (PIP), and cefsulodin (CEF), were from Sigma-Aldrich. Resazurin, which was used to measure the drug response, was from Sigma-Aldrich. Escherichia coli (E. coli) (ATCC 25922) was used as the model strain in this study. The bacteria were cultured to 1×10⁹ cfu/ml in Mueller Hinton Broth (MHB) in a shake incubator at 37 °C, washed three time with fresh MHB, and stored at −80 °C with glycerol for further experiments. The minimum inhibitory concentration (MIC) of each individual antibiotic was tested using the standard microdilution method recommended by the Clinical and Laboratory Standards Institute (CLSI) guidelines (Supplementary Figure S5). In brief, the bacteria (initial concentration at 5×10⁵ cfu/ml) was cultured with antibiotics at various concentrations at 37 °C for 16–20 h. The optical density at 600 nm (OD₆₀₀) was tested in a plate reader (Synergy H1, BioTek) to examine the bacterial growth and determine MICs.

To study the fluorescence signal change of bacterial response toward antibiotic treatment, the initial bacterial concentration was set from 5×10⁵ to 5×10⁷ cfu/ml, the resazurin concentration was at 500 μM, and the antibiotic was cefsulodin (CEF) with concentration at the MIC. The fluorescence intensity was monitored over time in a 96-well plate using the plate reader. The readout was achieved at excitation/emission wavelength of 530 nm/ 585 nm. The signal was normalized to the initial stage of each assay.

To test the signal change over bacterial growth at different resazurin concentrations in the absense of antibiotics, droplets were generated using samples before and after the incubation process (37 °C for 5 h). The bacterial concentration was at 5×10⁶ cfu/ml and the resazurin concentration was at 10, 25, and 50 μM. The signal was tested using the laser-induced fluorescence detector and the excitation/emission wavelength was 552 nm/ 575 nm. Relatively low resazurin concentration was used to fit the dynamic range of the laser-induced fluorescence detector.

To evaluate the bacterial contamination in the droplet system (Supplementary Figure S6). There were two groups of droplets generated using the proposed device, including the droplets consisting of bacteria, drug (CEF), and resazurin, and the droplets consisting of resazurin only (i.e., sterility control). The concentration of bacteria, drug, and resazurin was 5 × 10⁶ cfu/ml, MIC, and 25 μM, respectively. These droplets were incubated at 37 °C for 5 h in the tubing and subsequently detected by the laser-induced fluorescence detector. The fluorescence signal was normalized to the sterility droplets.

Drug response of antibiotic combinations

The drug response of antibiotic combinations was studied using both the proposed droplet platform and the standard 96-well plates. The droplet platform automated the antibiotic combinations by creating addressable droplets that contain appropriate contents of the reagents, as shown in Figure 4a. Herein, each column/row represented a concentration of corresponding antibiotic ranging from 0 to the MIC of the antibiotic with a two-fold increase. The final concentration of bacteria and resazurin were 5×10⁶ cfu/ml and 25 μM, respectively. These conditions were achieved by programable control of the opening time of respective valves for each droplet. In particular, the total valve opening time was 0.6 s and the pressure applied on the reagent-supply microchannels was 2.5 psi. The droplet volume was ~50 nL. Given the bacterial concentration of 5×10⁶ cfu/mL, there were ~250 bacterial cells in each droplet. In addition, the reagents and materials were freshly prepared for use in all experiments. The droplet array was subsequently transferred into tubing in sequence. In our experiments, there were ~80 droplets in each tubing. Then, the tubings were sealed and incubated at 37 °C for 5 h in an incubator. Lastly, the droplets were pumped into a microchannel in a detection device at 20 μl/min using a precise pump (Harvard Apparatus) (Figure 2i). The device was simply fabricated by spin-coating a SU-8 layer on silicon wafer and molded using PDMS (at a ratio of 10:1 between prepolymer and crosslinker) at 80 °C for 1 h. The fluorescence intensity of the droplets was orderly captured at excitation/emission wavelength of 552 nm/ 575 nm using the laser-induced fluorescence detector.

We examined the drug response of antibiotic combinations using 96-well plates in parallel. Each antibiotic pair was tested using 16 assays that contained combinations of involved antibiotics at various concentration (i.e., from 0 to the MIC of the antibiotic). The reagents and materials were freshly prepared for use in all experiments. The final concentration of bacteria was 5×10⁵ cfu/ml. The assays were incubated at 37 °C for 16–20 h. The OD₆₀₀ was tested using the plate reader before and after the culture process to examine the growth in each condition.

Dose-response curve model was utilized to analyze the IC₈₀s of the growth curve in GraphPad Prism 8. Briefly, the concentration of the drug was transformed to log phase and the fluorescence intensity or OD₆₀₀ values were normalized in the system. Then, the relationship between the drug and detected signal was fitted using the “Dose-response – Special” model to determine the IC₈₀s (Figure 4c). These IC₈₀s were plotted in the x-y plane to determine the combinatorial effect. Moreover, the fluorescence intensity of the droplet array or the OD₆₀₀ values were normalized in a heatmap and set underneath the plot of IC₈₀s (Figure 4d–f). Deeper color indicated higher intensity or higher OD₆₀₀ value. This approach helps to visualize the signal of the whole group in each drug combination.