High-resolution dose–response screening using droplet-based microfluidics

Abstract

A critical early step in drug discovery is the screening of a chemical library. Typically, promising compounds are identified in a primary screen and then more fully characterized in a dose–response analysis with 7–10 data points per compound. Here, we describe a robust microfluidic approach that increases the number of data points to approximately 10,000 per compound. The system exploits Taylor–Aris dispersion to create concentration gradients, which are then segmented into picoliter microreactors by droplet-based microfluidics. The large number of data points results in IC₅₀ values that are highly precise (± 2.40% at 95% confidence) and highly reproducible (CV = 2.45%, n = 16). In addition, the high resolution of the data reveals complex dose–response relationships unambiguously. We used this system to screen a chemical library of 704 compounds against protein tyrosine phosphatase 1B, a diabetes, obesity, and cancer target. We identified a number of novel inhibitors, the most potent being sodium cefsulodine, which has an IC₅₀ of 27 ± 0.83 μM.

In the early 16th century the Swiss chemist Paracelsus declared “all substances are poisons, there is none which is not a poison; only the right dose makes a substance non-poisonous.” This idea that the biological effects of a chemical compound are dependent upon its concentration was quantified by A. V. Hill in 1910 (1). However, despite the fact that compounds can display complex concentration-dependent relationships, varying in potency, efficacy, and steepness of response, usually just a single measurement at a single concentration (approximately 10 μM) is obtained for each compound in the chemical library during a primary drug screen, even when using state-of-the-art robotic microplate-based screening systems. This results in high numbers of false positives and false negatives (2), as well as the inability to identify subtle, complex pharmacology, such as partial agonism or antagonism. Even when dose–response curves are generated during a quantitative primary screen (3) or, more typically, during the follow-up of a single-point screen, the time and cost limitations mean that the curves typically contain only 7–10 data points each. With ≤ 10 nonduplicated data points, and as many as four adjustable nonlinear parameters (e.g., in the four-parameter Hill function), the results are highly sensitive to the data quality: For example, the presence of a single outlying data point can substantially alter the fit of the data, unless the outliers are identified and removed (4).

We have developed a system that uses droplet-based microfluidics to generate high-quality dose–response data during drug screening. Droplet-based microfluidics is itself a new technology for creating and manipulating picoliter-volume aqueous droplets that function as independent microreactors (for a review see ref. 5). As a result of the miniaturization inherent in this approach, our system (Fig. 1) is capable of generating dose–response curves at materially higher resolutions than existing microplate-based (3) and microfluidics-based approaches (6–14). Each dose–response curve contains approximately 10,000 data points, 1,000 times more than in conventional systems, resulting in extremely precise measurement of dose–response relationships using minimal quantities of reagents.

Fig. 1.

The microfluidic screening system. (A) Overview of the system. (B) Schematic showing how sequential injections of different compounds from the autosampler (C₁ to C_N) are transformed into smooth pulses by Taylor–Aris dispersion in the capillary. Each pulse gradually rises and then falls in concentration as it arrives at the subsequent microfluidic device. (C) Design of the microfluidic device (plan view) showing the two depths of channel: 25 μm (dark gray) and 75 μm (pale gray). Dotted black arrows show the route the droplets take through the device. (D) Schematic of the droplet production region of the device. Smoothed compound pulses from the capillary are combined with the enzyme (green) and the substrate (blue) and are then segmented into droplets by two streams of fluorinated oil (yellow). Each droplet contains a different concentration of the compound but constant concentrations of enzyme and substrate. (E) Light micrograph of one of the 10 analysis points with the position of the laser spot indicated just after the triangular droplet-respacing feature.

Results

Design of the Screening System.

The design of the system is outlined in (Fig. 1 and SI Appendix, Fig. S1). A high-performance liquid chromatography (HPLC) autosampler injects pulses of compounds from a 96- or 384-well microplate into a continuously flowing stream of buffer. The buffer passes through a capillary where Taylor–Aris dispersion (15) gradually smoothes the initially rectangular concentration profile of each compound pulse into a Gaussian profile (Fig. 1B and SI Text and Figs. S2 and S3). Previously, this phenomenon has been exploited to enhance mixing in microfluidic micromixers (16), but here we describe its use in creating concentration gradients. Following dispersion in the capillary, each smoothed pulse passes to a microfluidic device where it is combined with the assay components (the target enzyme and a fluorogenic substrate) and then segmented into approximately 140 pl droplets by two intersecting streams of fluorinated oil containing a surfactant (17). As each compound pulse is segmented, thousands of independent microreactors are generated containing fixed concentrations of the assay components but varying concentrations of the compound. This compartmentalization is necessary to prevent further dispersion of the compound after the enzyme and substrate are added. In a conventional single-phase system, the continued dispersion of the compound in the microfluidic channels would cause its concentration to vary over the course of the assay and preclude the construction of dose–response curves from single injections. The internal flow fields of the droplets also shorten the time scale for complete mixing of the assay components from seconds for diffusion-only mixing—as in single-phase systems—to milliseconds (18). After generation, the droplets pass through an on-chip delay line (19) with a series of analysis points spaced at increasing time intervals. During screening, the optical setup is focused at one analysis point, corresponding to a suitable incubation period for the enzyme, and the droplets are analyzed one by one as they stream through. By premixing each compound with a fluorescent dye before injection (the “concentration encoder”), it is possible to infer the compound concentration in a droplet from its fluorescence in the relevant channel. In parallel, the degree of enzyme inhibition in the droplet is determined from the fluorescence of the product of the enzymatic reaction in another, nonoverlapping fluorescence channel. Offline, the data for the droplets corresponding to a single injection are plotted on a graph of enzyme inhibition versus compound concentration, creating a high-resolution dose–response profile. The number of data points in this profile increases with the molecular weight of the compound, but 10,000 data points is typical for a small molecule (SI Text).

Characterization of Taylor–Aris Dispersion.

We characterized the dispersion of compounds inside the microfluidic system by injecting six compounds with different molecular weights (376 to 20,000 Da; SI Text). Because the compounds were all fluorescent, it was possible to observe the arrival of each pulse at the microfluidic device after dispersion in the capillary. The profiles obtained for a near-infrared (NIR) fluorescent dye DY-682 (SI Appendix, Fig. S4) and the five other fluorophores (SI Appendix, Fig. S5) varied in shape but were closely fitted by a model for Taylor–Aris dispersion that takes into account the injection method of the autosampler (SI Text). The fitted diffusion coefficients (D) were close to published values (SI Appendix, Table S1) and scaled with molecular weight, following a power-law with a scaling exponent (k) of −0.461 (SI Appendix, Fig. S6). Hence, under the same flow conditions, the dispersion profile of a molecule is simply a function of its D value and, thus, its molecular weight (SI Appendix, Fig. S7). Via a numerical approach, this allows the concentration of a compound in a droplet to be determined from the concentration of a coinjected fluorescent dye possessing a different D (the concentration encoder). This approach contrasts with capillary electrophoresis, high-performance, and ultra-performance liquid chromatography separation systems, which have also been integrated with microfluidic droplet production, in which the concentration gradients are strongly influenced by the chemical properties of the compounds (20–22).

High-Resolution Dose–Response Curves of β-Galactosidase Inhibition.

The complete screening system was validated using the reporter enzyme β-galactosidase and one of its known inhibitors, 2-phenylethyl β-D-thiogalactoside (PETG), as a model system. A 96-well plate was prepared with each well containing a fixed concentration of DY-682 (the concentration encoder) and one of four different concentrations of PETG (including zero). One μl was injected from each well, and the flow from the capillary was combined with β-galactosidase and the fluorogenic substrate fluorescein di-β-D-galactopyranoside (FDG) on-chip. Droplets flowed through the delay line and were analyzed by the optical setup to determine initial reaction rate (SI Appendix, Fig. S8). The raw data for each injection (Fig. 2A) were then processed, adjusting for the different dispersion profile of PETG versus the concentration encoder, and fitted with the four-parameter Hill function (Fig. 2B).

Fig. 2.

Construction of a high-resolution dose–response curve and comparison with a conventional dose-response curve. (A) The raw data for the droplets derived from an injection of PETG (well A3 in C) are plotted as photomultiplier signal versus time for the two fluorescence channels: green (proportional to β-galactosidase activity) and NIR (proportional to the concentration of the concentration encoder). The data for one second before each injection arrives at the analysis point is averaged to find a “control” value for the enzyme assay (0% inhibition). The data for the droplets in the rising phase of each dispersed compound pulse (“Dose–response droplets”) are then used to calculate the high-resolution dose–response profile. The droplets in the falling phase of the pulse are ignored because the shape of the curve typically deviates from the expected profile, especially when the passivated capillary has been used for many injections. (B) Scatter plot of percentage inhibition against PETG concentration for the same data, after correction (SI Appendix, Fig. S9). A total of 11,113 droplets (blue dots), were binned along the x axis and averaged, yielding 28 points (squares). These points were used to fit the four-parameter Hill function (black line; fit parameters are shown inset with 95% confidence intervals). The error bars for each binned point are largely the result of the dispersion in individual incubation times as the droplets pass through the delay line (19). This variation is most pronounced at the lowest PETG concentrations, but the 95% confidence interval never exceeds ± 1.62 percentage points of inhibition in this subfigure. (C) The same dose–response curve, but acquired using a conventional microplate-based assay. There were 10 replicates for each of the eight different concentrations of PETG. All precisions in the figure are the 95% confidence interval. Error bars correspond to .

The IC₅₀ and Hill slope, especially after adjustment for molecular weight (SI Appendix, Fig. S9), were found to be in good agreement with the values obtained in microplate (a mean IC₅₀ of 2.04 versus 2.87 μM; a mean Hill slope of 1.07 versus 0.967; Fig. 2 B and C and SI Appendix, Table S2) and a published IC₅₀ (3.10 μM) (23). The precision of the IC₅₀ value was, however, found to be much higher in the microfluidic system than in a conventional eight-point microplate assay: For a single injection the 95% confidence interval was, on average, ± 2.40% versus ± 62.6% in microplate. Furthermore, the results are highly reproducible (Fig. 3): The CV for the IC₅₀ was 2.45% (n = 16), compared to 28.0% in microplate (n = 10). The IC₅₀ values for the three different injection concentrations of PETG were very similar, indicating that the system can operate effectively over concentration ranges exceeding 3 orders of magnitude (Fig. 3B). The Z-factor (24) for the assay was 0.685 (SI Text), indicating that it was “excellent,” and cross-contamination was minimal: Inhibition never rose above 2.37% in the 48 intercalated control injections (Fig. 3B). Subsequent experiments with a panel of different fluorescent molecules directly measured cross-contamination to be less than 0.14% (SI Appendix, Fig. S10 and SI Text).

Fig. 3.

Validation of the high-resolution dose–response screening method with β-galactosidase. (A) High-resolution dose–response curves for injections of PETG from a 96-well microplate at four different concentrations: high (600 μM; red wells), medium (120 μM; blue wells), low (24 μM; green wells), and zero (yellow wells). Percentage inhibition (y-axis of each well; −10 to 110%) is plotted against compound concentration (logged x-axis; 0.5 to 250 μM for high, 0.1 to 50 μM for medium, 20 nM to 10 μM for low) or relative dilution for the buffer only injections (0.2 to 100%). Fitting the data with the four-parameter Hill function (not shown) reveals very similar IC₅₀ values for all injections at all three concentrations of injected PETG: The mean IC₅₀ values were 1.78 μM (high; CV = 4.52%), 2.04 μM (medium; CV = 2.45%), and 2.03 μM (low; CV = 3.28%). (B) All 96 high-resolution curves from the previous subfigure plotted together to demonstrate the superposition of the IC₅₀ points.

Identification of Novel Inhibitors of PTP1B.

A chemical library comprising 704 compounds from the Prestwick Chemical Library® (all marketed drugs with molecular weights between 113 and 1,882 Da; SI Appendix, Table S3) was then screened for inhibition of PTP1B, a target for type 2 diabetes mellitus, obesity, and cancer (25). In this case, fluorescein diphosphate (FDP) was used as the fluorogenic substrate. Compounds were injected automatically from two 384-well plates, and each row of 24 wells terminated with a positive control well containing sodium suramin, a known inhibitor of PTP1B (26), followed by a negative control well containing injection buffer alone. The Z-factor for the assay was 0.657 (SI Text), indicating that it was “excellent.” The mean IC₅₀ for the 32 injections of sodium suramin was 17.7 μM (CV = 11.5%).

The screen revealed that seven compounds in the chemical library were PTP1B inhibitors with IC₂₀ values less than 20 μM (Fig. 4 and SI Appendix, Fig. S11 and Table S4). One of these compounds, sodium cefsulodine, exhibited an IC₂₀ of 8.51 μM and an IC₅₀ inside the concentration range tested: 27.0 μM (Fig. 4C). Its activity was confirmed in microplate, as was the activity of one of the novel weak inhibitors, methimazole (Fig. 4D and SI Appendix, Fig. S12). In the case of this weak inhibitor, the microplate profile deviated somewhat from the microfluidic profile, but such weak effects are challenging to observe accurately in microplates. Interestingly, the high-resolution dose–response curve for the known inhibitor sodium suramin revealed that with increasing concentration, it first activates PTP1B (< 15 μM) and then steeply climbs in inhibition (Fig. 4B and SI Appendix, Fig. S12A). This biphasic aspect, as seen in hormesis (27), is an example of a complex dose–response relationship that would have been ambiguous in a 7–10-point dose–response study and completely absent from a single-point screen. Being able to clearly observe this kind of behavior, including the pathologically steep Hill slope (28), will be an aid to maximizing the quality of lead compounds coming out of a screen.

Fig. 4.

High-resolution dose–response screening of a chemical library against PTP1B. Of the 704 compounds injected, 701 were successfully analyzed. (A) Superimposed high-resolution dose–response profiles of the 701 compounds, plus sodium suramin (the control inhibitor) and buffer alone. High-resolution dose–response profiles (and chemical structures) of (B) sodium suramin, (C) the novel inhibitor sodium cefsulodine, and (D) methimazole, one of the novel weak inhibitors. The black lines in C and D are the fitted four-parameter Hill function. In C the fit parameters are shown inset because an IC₅₀ could be meaningfully extracted. In the remaining plots the black line is merely an aid to the eye. The IC₅₀ and “Slope at IC₅₀” values in B were the x value of the crossing point of the line at y = 50% and its gradient at that point, respectively. The IC₂₀ value in D was determined by finding the crossing point of the fit at y = 20%. All precisions in the figure are the 95% confidence interval. Error bars correspond to .

Discussion

In the past, small molecule inhibitors of PTP1B have been identified by a number of approaches, including high-throughput screening (HTS; reviewed in refs. 29–31). The most potent and selective of these was isolated by screening a focused library containing a biasing phosphotyrosine (pTyr) to ensure association with the active site (32). However, inhibitors have also been identified by HTS of nonfocused compound libraries (as used in this study). HTS of corporate compound collections has, notably, allowed the identification of several pTyr mimetics (30, 31), including a number of sulfamic acid derivatives (33). Sodium cefsulodine, the most potent PTP1B inhibitor identified using the droplet-based microfluidic system, contains a sulfonic acid moiety linked to a cephalosporin scaffold; this molecule is at least 10-fold more potent (IC₅₀ = 27.0 μM; Fig. 4C) than the best hit from the earlier HTS study, a sulfamic acid (IC₅₀ = 322.5 μM) (33). It seems likely that the sulfonic acid moiety in sodium cefsulodine plays a similar role as a pTyr mimetic to the sulfamic acid moiety and the 1,2,5-thiadiazolidin-3-one 1,1-dioxone (TDZ) moiety found in other PTP1B inhibitors (31). Although competitive inhibitors of PTP1B with nonfluorinated methylenesulfonic acids have been synthesized with IC₅₀ values similar to that of sodium cefsulodine (19 μM and 44 μM) (34), we believe that the latter molecule is an interesting candidate for further drug development because it is already marketed as a pharmaceutical (a broad-spectrum antibacterial).

As a technology for studying enzyme kinetics, droplet-based microfluidics has several advantages over conventional single-phase microfluidics: On-chip dispersion of reagents can be controlled and individual microreactors can be created at high speed and without resorting to complex arrays of on-chip valves. Laboratories, including our own, have described techniques that exploit these features to measure concentration-dependent effects at the microscale. Several of these techniques generate droplets containing different concentrations of a compound by varying the relative flow rates of two or more aqueous streams (13, 18). However, it is difficult to maintain stable coflows that exceed a 1∶9 ratio, and hence to achieve dilutions of greater than 10-fold, which makes this method poorly adapted for studying phenomena such as inhibition where concentrations must be varied logarithmically. Prediluting compounds in microplates (6, 14), composing droplets one at a time in a glass capillary (35), or employing on-chip diffusive mixing (36, 37) permits a wider concentration range to be explored, but the sampling of this range becomes sparse, as with microplate-based methods (≤ 10 concentrations). Better sampling (23 concentrations) can be achieved by manipulating a single nanoliter-volume droplet trapped in a microfluidic channel by splitting and merging buffer droplets (38), but this method has not been applied to screening. In contrast with these existing techniques, the droplet-based microfluidic system described here is capable of processing entire chemical libraries and rapidly generating dose–response curves that cover wide concentration ranges (approximately 3 orders of magnitude) with very large numbers of dose–response points (approximately 10,000 for a typical small molecule). This cloud of data, once fitted, yields clear dose–response information at precisions much greater than conventional microplate-based formats (approximately 26-fold; SI Appendix, Table S5). We believe that this kind of data makes the approach ideal for focused or iterative drug screening: new approaches to drug discovery that rely on intelligent selection and refinement of chemical libraries (39). These methods are dependent on data quality and will benefit from the precision and reliability of high-resolution dose–response data where subtle, complex pharmacologies are revealed.

It is worth noting that in delay lines of the sort used here, in which droplets are not maintained in single file, a dispersion of droplet incubation times (t) equal to approximately 5% of the delay time is observed (19). In order to prevent droplets containing one compound arriving at the same time as droplets containing the subsequent compound, the period separating dispersed compound pulses should be at least 0.1t. The maximum delay currently available in the delay line is 3.5 min, but slower enzymes could be accommodated by increasing enzyme concentration or by increasing the length of the delay line.

Throughput currently stands at 1 compound every 157 s, with the system consuming 17.5 μl of assay reagents during this period (SI Appendix, Table S5). Compared to an 8-point microplate assay, this represents an approximately 25,000-fold reduction in reagent consumption per dose–response data point and an approximately 18-fold reduction per dose–response curve. In the future, we believe that it should be possible to increase throughput to 1 compound every 10 s by using a faster autosampler and a higher flow velocity in the capillary, without significantly reducing data quality (SI Appendix, Table S5). This rate would approach microplate-based quantitative high-throughput screening in terms of speed (3, 40), while consuming approximately 1 μl of assay reagents (approximatley 2 orders of magnitude less) and yielding, on average, 700 data points per curve (approximately 2 orders of magnitude more). Although Taylor–Aris dispersion can be used to generate any concentration from zero to approximately 100% of the starting concentration (see SI Appendix), in practice the concentration range of each dose–response profile is limited by the dynamic range of the optical detection system in quantifying the concentration encoder. This range, currently spanning approximately 3 orders of magnitude, could be extended to 4 or 5 orders of magnitude using a high dynamic range fluorescence detector.

The system is compatible with any rapid fluorescence-based assay, as long as the emission spectrum does not overlap the concentration encoder channel. Recently, microfluidic devices have been used to assay mammalian and insect cells in droplets (13, 14, 41, 42), so it should be possible to adapt this system to perform cell-based assays with rapid fluorescent readouts, such as intracellular calcium flux assays for identifying agonists and antagonists of G-protein-coupled receptors (GPCRs) and ligand-gated ion channels (43, 44). Furthermore, the abstemiousness of this system—and other microfluidic approaches—with regard to reagent consumption presents the exciting opportunity of using primary cells in drug screens, dramatically increasing the relevance of the screening data and helping to forestall clinical dead ends (45).

Materials and Methods

Microfluidic System.

A capillary with an internal diameter of 75 μm was passivated with a fluorosilane (SI Text) and used to connect the “Column” port of an WPS-3000 HPLC autosampler (Dionex Corp.) to a microfluidic device. The microfluidic device was fabricated from poly(dimethylsiloxane) (PDMS) and glass by soft lithography (SI Text). During operation, “running buffer” was pumped to the autosampler at a rate of 200 μL/hour. The flow passed through the autosampler’s injection valve, then the capillary, and arrived at the microfluidic device. On-chip, this stream was combined with two other aqueous streams, each flowing at 100 μL/hour. The combined aqueous flow was segmented into approximately 140-pl droplets at a nozzle by two streams of the fluorinated oil HFE-7500 (3M), each flowing at 200 μL/hour. The oil contained 0.5% (w/w) EA-surfactant (RainDance Technologies, Inc.), a biocompatible PEG-PFPE amphiphilic block copolymer (46), to stabilize the droplets against fusion. After production, the droplets flowed into a deep, wide delay line where their mean velocity decreased, allowing incubation times of up to 210 s on-chip. At several points along the delay line, the deep serpentine channel constricted, allowing the droplets to be examined one at a time by the laser-based optical setup (SI Text). The peak green (529 nm) and NIR (794 nm) fluorescence of each droplet was measured by optical detectors and recorded.

High-Resolution Dose–Response Curves of β-Galactosidase Inhibition.

PBS was the running buffer. The final concentrations in the droplets were: 5 U/mL E. coli Grade VIII β-galactosidase (the enzyme), 60 μM FDG (the substrate), 100 nM sodium fluorescein, and 1 g/L bovine serum albumin (BSA) (all from Sigma-Aldrich Co.). The optical setup was positioned just before the delay line, and individual droplets were discriminated by green fluorescence. The measurement at this point provided a “pseudo blank” (equivalent to 100% inhibition). The optical setup was repositioned to the 30-second measurement point in the delay line in order to observe initial reaction rates in the droplets (SI Appendix, Fig. S8 and SI Text). Subsequently, the autosampler was used to inject 1 μL from each well of a 96-well plate. Each well contained 20 μL of 100 μM DY-682 in PBS, plus one of four different concentrations of the inhibitor PETG (Invitrogen Corp.): 600 μM (“high”), 120 μM (“medium”), 24 μM (“low”), or zero. As each dispersed pulse of DY-682 and PETG, mixed with the reaction components and segmented into droplets, arrived at the optical detector, a dose–response profile was recorded. For each injection, occurring with a period of 157 s, 17.5 μL of reagents were consumed (SI Appendix, Table S5). The entire screen, which lasted 4.2 h, was performed without manual intervention using a single microfluidic device.

Offline, each profile was fitted with the four-parameter Hill function to determine the IC₅₀ and Hill slope (SI Appendix, Fig. S9 and SI Text).

High-Resolution Dose–Response Screening of Chemical Library for PTP1B Inhibition.

This was performed in a similar manner to the analysis of β-galactosidase inhibition. Fifty mM 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid (HEPES) pH 7.2 (Sigma-Aldrich Co.) was the running buffer. The final concentrations in the droplets were: 134 nM PTP1B (the enzyme; EMD Biosciences, Inc.), 17 μM fluorescein diphosphate (FDP; the substrate), 1 mM dithiothreitol (DTT), 1 mM ethylenediaminetetraacetic acid (EDTA), and 1 g/L BSA (all except the enzyme from Sigma-Aldrich Co.). Two 384-well plates containing 704 compounds comprising a subset of the Prestwick Chemical Library®, were screened (SI Appendix, Table S3). The compounds were at 120-μM concentration in the “injection buffer”: 50 mM HEPES pH 7.2 containing 100 μM DY-682 and 1.2% (v/v) dimethyl sulphoxide (DMSO; Sigma-Aldrich Co.). Each row of 24 wells terminated with a control well containing 120 μM sodium suramin, a known PTP1B inhibitor, in the injection buffer, followed by a well containing injection buffer alone. The measurement point was at 210 s delay (SI Appendix, Fig. S8). A single microfluidic device was used for the entire screen, which was broken into four runs totaling 33.6 h. No manual intervention was required other than to refill reagent reservoirs between runs.