Rapid phenotypic antimicrobial susceptibility testing using nanoliter arrays

Description of the SNDA–AST System

The method uses the SNDA (30) as a base platform for simple well loading/stationary droplet formation. The microfluidic device can interface with nearly any substrate, such as a conventional microscope glass slide, and can be loaded with a conventional 10-µL pipette. Each array consists of 200 wells of 8 nL, branching off a main channel (Fig. 1A). Well dimensions are 200 µm × 400 µm × 100 µm (width × length × height) and the main channel is 300 µm wide. Restrictions are 2–5 µm wide and serve as unique structures that are specifically designed to allow air in the wells to escape to surrounding channels, thus facilitating simple capillary-based filling. The system uses the standard AST cell concentration as well as growth medium (Mueller Hinton II, or MHII) to improve clinical translatability and interpretability using known standard breakpoints produced by organizations such as the European Committee on Antimicrobial Susceptibility Testing (EUCAST) and the Clinical & Laboratory Standards Institute. The well volume was set specifically so that the standard AST cell concentration (5×105 cfu/mL) would produce an average of 4 cfu per well. This number was chosen specifically to balance three design considerations: (i) to use the minimal amount of bacteria possible to aid in the compatibility of future novel sample preparation schemes, (ii) to have large enough wells so that the standard AST testing concentration would result in a sufficient amount of bacteria per well to produce a predictable distribution of bacterial loading as governed by the Poisson distribution, and (iii) to maximize sample dispersion. Dispersion of the sample into hundreds of smaller nanoliter volumes serves to capture more cell–cell heterogeneity while lowering sample volumes. Because each well holds only 8 nL, each testing treatment requires <2 µL. In addition, sample dispersion increases the surface area-to-volume ratio of the cultures by an order of magnitude, thereby promoting increased gas transport and growth rates and faster AST times. Finally, sample dispersion can theoretically lower the rate of false-positive results by quarantining unrelated pathogens or contaminants to only a few number of wells.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

Device design and principle of operation. (A) Illustration of SNDA–AST device on a conventional microscope slide being loaded with a conventional 10-µL pipette. The SNDA device consists of two rows of 8-nL wells connected by a main delivery channel. Each well contains ∼3-µm restrictions that allow the air to escape to two surrounding channels, thus facilitating simple capillary based filling. (Inset) 1, 200 µm; 2, 400 µm; 3, 100 µm; and 4, 2–5 µm. (B) Step 1: The SNDA-based AST device is loaded with a single-step injection of a two-plug formulation of the bacterial suspension with 10% resazurin followed by a plug of FC-40 oil. The purpose of the oil is to discretize the sample by isolating the wells while delivering oxygen and preventing evaporation. Step 2: The well fluorescence, indicating the level of metabolic activity occurring in the culture, is measured every 30 min and is proportional to the amount of bacteria/metabolism in the well. Bacteria are not drawn to scale.

Principle of Operation

The sample is loaded with a single-step injection of a two-plug solution using a conventional laboratory micropipette (Fig. 1A). The first, or bottommost, plug is a ∼1.6 µL bacterial suspension of 5×105 cfu/mL supplemented with 10% resazurin. The second, or topmost, plug contains ∼3 µL of FC-40 oil. The two-plug solution is achieved simply by aspirating the respective liquids sequentially. Low-pressure loading is enabled by restrictions in the well, allowing the volume of air in the wells to escape and be replaced by that of the liquid, making loading by hand possible using this system (Movie S1). After the first plug containing the bacteria flows in and fills the cavities of the reservoirs the oil plug follows and fills the main channel, thereby separating the wells with an immiscible barrier, effectively discretizing the sample (Fig. 1B, step 1). We specifically use fluorinated oil because it allows delivery of dissolved oxygen to the culture chambers while preventing evaporation of the well contents. In this work, an array is loaded for each testing treatment, and for each experiment a positive and negative control are included. We defined the positive control to be the bacterial testing suspension without the addition of any antibiotics and the negative control to be the bacterial testing suspension with the addition of sufficiently high amounts of antibiotics (one to two orders of magnitude higher in concentration to that of the testing concentration). We specifically chose this model for a negative control to account for any interactions between dead or nonproliferating bacteria in the system with resazurin that would not be present in a negative control composed simply of sterile growth medium with resazurin. For all experiments 100 wells were analyzed for each treatment (e.g., positive control, negative control, and ampicillin 8 mg/L) to account for nonuniform cell loading according to the Poisson distribution with λ = 4 cells per well, a 95% confidence level, and a 10% acceptable error (Fig. S1). Although only 100 wells are needed for analysis, in this work we used arrays containing 200 wells due to fabrication yield and loading aspects that seldom reduce the number of usable chambers. The fluorescence buildup within the wells indicates the level of metabolic activity occurring in the culture and is measured every 30 min for different antibiotic conditions to obtain results (Fig. 1B, step 2).

• Download figure
• Open in new tab
• Download powerpoint

Fig. S1.

A table summarizing calculations made using the Poisson distribution for different levels of acceptable error and confidence and a mean of λ=4 bacteria per well. For our system, we chose to settle at an acceptable error of 10% in the concentration and be 95% confident that we have that concentration. This leads us to the conclusion that we have to average the results of ∼100 wells per treatment to achieve the desired level of accuracy.

Time to Results and Data Analysis

For automatable S/R determination we develop an algorithm that analyzes the kinetic growth curves of each well to make rapid determinations. After subtracting the data by the average rise profile of the negative control (Fig. 2A) we fit the rise profile of each well (100 wells per treatment) to both a linear and exponential model. Then, by comparing the goodness of the fit through the rms error, we keep only the best-fitting model for each specific well. The instantaneous slope of each fit is calculated analytically from the model equation and the slopes of the fits are averaged for each treatment at the last time point along the curve (Fig. 2B). This process is then repeated, and new fits are generated for each data point that is collected. For each experiment we used the difference in the average slopes of the positive and negative control to define a working range, Δ, for which we compare the slopes of the treatments for the S/R determinations. Treatments with slopes that were ≤30% of Δ were considered susceptible. Treatments with slopes that were ≥60% of Δ were considered resistant. These thresholds were optimized based on an existing dataset (Methods for more information). Slope comparisons between the treatments and the thresholds of the working range were performed with a one-tailed t test with significance level of α=0.95, thus allowing the method to use both the effect size (31) and P value for making determinations. In addition, to help avoid nonmeaningful determinations in the early stages of the growth curves, a logistical restriction was set so that determinations could only be made once there was a significant difference between the slopes of the positive and negative controls. We then visualize the results in the form of S/R determinations over time (Fig. 2 C and D) and consider the correct determination to be the one that the algorithm converges for sufficiently long times. The time to result is then defined as the time at which the algorithm made a decisive decision, that is, one that did not change for the remainder of the analysis period. Fig. 2 C and D and present the determination plots for treatments 1 and 2 from the dataset displayed in Fig. 2 A and B that produced resistant and susceptible determinations, respectively. We applied our determination assay and algorithm based on a “crude” minimum inhibitory concentration (MIC) of a single breakpoint concentration to a wide range of different microorganisms and antibiotic combinations and assessed the time to result. We then aggregated the result times from the different experiments to determine the characteristic time to results (Fig. 2E). Using this system, designations are evident in less than 5.5 h, with >80% of designations evident in 4.5 h or less, whereas the results on the same samples using the VITEK 2 took longer than the typical 8-h workday. We also found that our S/R determinations matched the determinations obtained in the clinic using the VITEK 2, as interpreted by EUCAST guidelines (32). Table 1 lists the 12 bacteria–antibiotic combinations that we tested and used for the time analysis. We designed our experiments in a manner that allowed us to test both gram-positive and -negative bacteria from different families as well as antibiotics from different mechanistic classes. In addition, we aimed to include both resistant and susceptible determinations in the analysis.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Time to S/R determination. (A) First, the fluorescent profile of each well is normalized to the average profile of the negative control by subtraction. Here, data for Klebsiella pneumoniae are used where treatments 1 and 2 were assigned resistant and susceptible determinations, respectively. Averages of each treatment are shown. Next, the normalized profile of each well is fit to both a linear or exponential model with an offset however only the ‟best” model is kept, using the root mean square error as a determiner. (B) Slope distributions are then obtained from the fits created for every time point. This is done iteratively after every data point collected using all data points collected up to that time point for the fits. Here the slopes of the fits created at the last time point are shown. (C and D) Finally, determinations are made for each treatment by categorizing the average slopes of the treatments using fixed fractional thresholds of that of the normalized positive control. We define the time to results as the time point at which the algorithm has presented a decisive determination (one that does not change for the remainder of the experiment). (E) A summary of AST times obtained for the clinical isolates studied in this work using the SNDA–AST system compared with the time to results obtained in the clinic for the same isolate sample using the VITEK 2 AST system. Blue circles represent susceptible determinations, green circles represent resistant determinations, and orange circles marked with a “V” present the results obtained by the VITEK 2 AST system in the clinic. A red dashed line illustrates the ending of a typical 8-h work day.

Antibiotic Incorporation via lyophilization and Multiplexing

We developed a lyophilization scheme to incorporate freeze-dried antibiotics within the wells of the SNDA before sample loading to improve ease of use, practicality, and multiplexing. To lyophilize antibiotics within the SNDA, before the AST experiment, we loaded liquid antibiotics at the target concentration in the SNDA and separated the wells with air instead of oil. Next, we froze the arrays at −80 °C for 40 min and subsequently placed them into vacuum chambers for overnight lyophilization in a lyophilizer machine (Fig. 3A). After lyophilization, these devices had dried antibiotics within the wells and were then ready to be used for AST experiments and loaded with a bacterial suspension using the normal loading procedure mentioned previously. Using this method we observed similar kinetic growth profiles with freeze-dried antibiotics compared with those obtained using the equivalent “wet” antibiotics that we supplemented before loading (Fig. 3B).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Antibiotic incorporation via lyophilization and multiplexing. (A) Workflow illustration for antibiotic loading and lyophilization. (B) Comparison of the standard “wet” and dried gentamicin at the breakpoint concentration using E. coli with 8 mg/L ampicillin showing that the dried antibiotic has efficacy similar to that of the standard “wet” counterpart. (C) Schematic of parallel SNDA–AST device used for multiplexing. A bacterial sample can be loaded into the device and tested against two different types of antibiotics simultaneously. Inlet/outlet reservoirs are indicated by gray and yellow circles. Yellow circles specifically indicate oil reservoirs. (D) Demonstration of sensitivity testing of two different dried antibiotics simultaneously at breakpoint concentrations using the parallel SNDA–AST device, confirming that this isolate of E. coli is resistant to ampicillin and susceptible to gentamicin. Error bars in all graphs present 95% CIs on the mean.

To demonstrate the concept of multiplexing through parallelization we designed a device of two parallel and interconnected arrays for which two different antibiotics can be lyophilized within the wells of each array (Fig. 3C). Using this system, we could investigate the susceptibility of Escherichia coli to gentamicin and ampicillin simultaneously and confirm that our determinations matched the expected determinations (Fig. 3D) that were produced by the VITEK 2 for the same strain.

Same-Day Detection and AST for UTIs

UTIs are among the most common type of bacterial infections dealt with in the clinic and pose a significant healthcare burden (33). After a urine sample is collected, current clinical protocols dictate the need to culture the sample on an agar-growth medium to first confirm the presence of bacteria in the sample, isolate the bacteria from the urine, and identify the purity of the culture (34, 35). This step is necessary because most urine cultures for suspected UTIs yield negative results (do not contain a significant concentration of bacteria). However, urine cultures require 18–24 h of incubation, and only after that can AST testing begin, thereby greatly lengthening the diagnostic time. Here, we investigated the ability to perform rapid AST directly on bacteria collected from fresh urine samples to skip the plating and incubation steps and provide same-day AST results. Because current plating and AST methods only yield results the next working day, we hypothesized that using the SNDA–AST system directly on bacteria harvested from urine samples can save ∼2 d of clinical diagnostic time (Fig. 4A).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 4.

Same-day AST for UTIs. (A) Timeline comparing the current clinical workflow (Top) to the SNDA–AST workflow (Bottom). In the current workflow, lengthy plating and AST steps cause the next step to be initiated only the next working day (shaded regions), causing AST results (asterisk) to be delivered in 2 d. With the SNDA–AST setup, the entire workflow can be performed and results (asterisk) can be delivered the same day. (B) Illustrated schematic of bacteria isolation protocol developed for clinical urine samples. (1) The urine sample is directed through a 5-µm filter to trap WBC and large debris followed by a 0.22-µm filter to trap bacteria. (2) Growth medium is directed through the 0.22-µm filter in the opposite direction to flush out trapped bacteria and the suspension is collected in a separate vial. (3) The bacterial concentration is adjusted via microscope observation. (4) The sample is then supplemented with 10% resazurin and loaded into the SNDA–AST device for analysis. (C) A summary of AST times obtained for the clinical urine samples studied in this work using the SNDA–AST system compared with the time to results obtained in the clinic for the same isolate sample using the VITEK 2 AST system. Blue circles represent susceptible determinations, green circles represent resistant determinations, and orange circles marked with a “V” present the results obtained by the VITEK 2 AST system in the hospital. A red dashed line indicates the assumed 8-h work day, showing that, unlike results from the VITEK 2, results from the SNDA–AST system can be used the same day.

Performing AST directly on urine samples requires that the urine be used as a growth medium. This can alter growth kinetics and thus affect determinations and their compatibility with breakpoint standards for interpretation. In addition, extraneous components in the urine sample could cause inaccurate testing. Instead, we opted to use a series of carefully designed filtering steps to isolate the bacteria directly from clinical urine samples into our growth medium of choice. First, the sample is filtered sequentially through a series of a 5-µm filter followed by a 0.22-µm filter (Fig. 4B and Fig. S2A). The 5-µm filter traps WBC and large debris that may be present in the sample and the 0.22-µm filter traps the bacteria. Second, the standard AST growth medium is passed through the opposite direction of the 0.22-µm filter only, consequently flushing out the bacteria that were trapped during the first filtering step. The bacteria that are removed during this process are kept for use in AST. The ratio of volumes of the urine sample to the volume of the growth medium used is roughly 5:1, which consequently concentrates the bacteria five times in the growth medium compared with the load in the original urine sample. Because positive urine samples of patients with UTI generally contain more than 10⁵ cfu/mL bacteria, this ensures that there will be a high-enough concentration to perform AST accurately when taking into account the estimated efficiency of the filtering scheme (Fig. S3). Also, the SNDA–AST system requires only a few tens to hundreds of microliters of a 5×105cfu/mL solution and urine samples are generally collected in volume of ∼9 mL, and thus sufficient sample volume is most likely not a concern with this system. Next, because AST requires a specific bacterial concentration, the sample is adjusted to the correct concentration by manually probing the average number of bacteria per well (Fig. S2B) in the SNDA and diluting accordingly. Finally, AST is performed using the SNDA–AST using the same methodology mentioned earlier. If developed, perhaps an image analysis program can perform counting before AST automatically and also give insight into the gram classification of the bacteria with some level of confidence.

• Download figure
• Open in new tab
• Download powerpoint

Fig. S2.

(A) An image of the filtration setup used to isolate the bacteria from clinical urine samples where the sample is drawn through the needle and through the series of filters from left to right. (B) A microscope image of bacteria in a single well of the SNDA as seen during the counting step required for adjusting the cell concentration for subsequent AST.

• Download figure
• Open in new tab
• Download powerpoint

Fig. S3.

Retrieval efficiency of the filtration process. Here we quantify the retrieval efficiency of the filtration process used to produce the data in Fig. 4 on clinical urine samples. A laboratory-grown culture of E. coli was used at an initial concentration of 5 × 10⁵ cells per mL as determined by optical density and a standard curve. n = 4 repetitions for each measurement. Error bars represent 95% CIs on the mean. Results were quantified by plating the bacterial suspensions before and after the respective filtering steps using the Drigalski technique. Percent values are calculated with respect to the bacterial concentration before each specific step.

We next aggregated the time to results of 10 bacteria–antibiotic combinations from five clinical urine samples using this system (Fig. 4C). We observed times similar to those found in our time analysis with the clinical isolates shown in Fig. 2B. This allowed us to validate our hypothesis that this method is capable of shaving up to 2 d off the clinical diagnostic timeline for UTIs. Table 2 summarizes the experiments performed and results obtained on clinical urine samples for the time analysis of AST for UTIs. We found our determinations to match the determinations obtained using the VITEK 2 in the clinic, with the exception of two cases (single determinations in Exp. 2 and 4). For these discrepancies, the results obtained by a bulk control ran in parallel did match the VITEK results (Fig. S4), indicating that there was no problem regarding the accuracy of the resazurin assay itself, but rather that a delayed proliferation cycle was present. For these experiments, we hypothesize that this delayed growth cycle is most likely attributed to a pronounced lag phase associated with the bacteria coming directly from patient urine samples which, for the work described in this paper, were stored in a refrigerator overnight before use.

• Download figure
• Open in new tab
• Download powerpoint

Fig. S4.

Bulk colorimetric assay performed in parallel with the same sample used for AST with the SNDA–AST. This example is from the E. coli tested in Table 2 in Exp. 4. From left to right: Positive control, negative control, ampicillin 8 mg/L, and ciprofloxacin 0.5 mg/L. The result with ampicillin determined using this bulk colorimetric assay matches that of the VITEK 2 on the same isolate performed in the clinic, although does not match the determination produced by the SNDA–AST. We hypothesize that this is most likely due to a pronounced lag phase associated with the bacteria that were extracted directly from clinical urine samples that were refrigerated overnight before testing.