Towards sensitive, high-throughput, biomolecular assays based on fluorescence lifetime

In this work water-in-oil microdroplets are used as small-volume reservoirs that contain the samples of interest and can be manipulated at very high throughput in microfluidic channels. As previously reported [36], we use the conventional 'rapid prototyping' technique [47] to manufacture polydimethylsiloxane microfluidic chips. A so-called T-junction is used for droplet generation: three aqueous inlets merge into one channel which intersects orthogonally a fourth channel flowed with immiscible oil (10:1 mixture of perfluorodecalin with 1H,1H,2H,2H-perfluorodecyltrichlorosilane). The mixing of the three aqueous solutions is initiated by the droplet formation and is further accelerated by chaotic advection induced within the droplets by the wiggling shape of the channel after the T-junction [48]. The velocity, size and concentration of the droplets are precisely controlled by regulating the flow rates in the aqueous and oil channels using a commercial flow control system (Elveflow, Elvesys, France). Droplets of less than 200 pl volume and 1 μM fluorophore or substrate concentration were used in the present work. This represents no more than 200 attomol (10⁻¹⁸ mol) of solute per droplet, i.e. per sample.

To illustrate the efficiency of the proposed FLT detection technology for HTS, we demonstrate the feasibility of an enzymatic activity assay relying on FLT detection. The assay [15, 18, 19] is designed to evidence the activity of the protease human pancreatic trypsin (from autolyzed human pancreas, Elastin Product Company, Inc., Owensville, MO, USA, catalog number TR127) on a peptide substrate tagged with the fluorescent probe PT14. Due to the structural flexibility of the peptide the probe undergoes dynamic quenching resulting in a FLT of 7 ns. When the enzyme cleaves the substrate, the probe is released, and its FLT of 14 ns is restored. Figure 1 displays the kinetics of FLT recovery after mixing 1 μM substrate with 1 or 10 nM enzyme in PBS. In this work, the fluorescently-labeled substrate was Biosyntan 9385.1 C(PT14)-GFKY-NH₂, purchased from Biosyntan (Berlin, Germany). It was, delivered as dry powder, dissolved in DMSO and kept at −20 °C. For all experiments it was dissolved in PBS (phosphate buffered saline, pH = 7.4) and 0.02% v/v TWEEN (polyethylene glycol sorbitan monolaurate) was added. In the microfluidic droplets, the concentrations of substrate and enzyme were adjusted to 1 μM and 1 nM, respectively.

Zoom In Zoom Out Reset image size

Figure 2 displays a sketch of the experimental set-up. A home-made fluorescence microscope is assembled to detect the fluorescence of water-in-oil droplets circulating in the microfluidic chip. A dichroic mirror (Semrock, BLP01-442R) is used to reflect the excitation light and transmit the fluorescence emission. A microscope objective (Nikon, x50, NA = 0.6) is used to focus the excitation beam to a ∼5 μm spot inside the droplets circulating in the main microfluidic channel, and to collect and collimate the fluorescence emitted by the droplets. A second objective (Nikon, x20, NA = 0.4) is used to focus the fluorescence light on a single photon avalanche diode (SPAD, ID100, Id-quantique, Switzerland). The excitation source is a diode laser (DL-4146-101S, Roithner lasertechnik, Austria) emitting at 405 nm and operated in a pulsed mode, by a home-made, cost-effective driver consisting of a current pulse generator described previously [49]. The laser pulse repetition rate is fixed at 20 MHz with 80 ps duration FWHM. The laser average power is 100 μW, attenuated to 50 μW in the microfluidic experiments performed with the PT14 probe, which is however characterized by a relatively low extinction coefficient (ε_max ∼ 7600 M⁻¹ cm⁻¹ at 394 nm) [15].

Zoom In Zoom Out Reset image size

According to conventional TCSPC experiments, a TDC is used to measure the time laps between each single photon detection event delivered by the SPAD and the previous laser pulse. As a TDC, we use here a Field-Programmable Gate Array (FPGA, Altera^® Cyclone EP4CE55) board configured to realize the function of a TDC. The FPGA board actually generates the 20 MHz clock signal that triggers the laser diode driver, acquires the signal delivered by the SPAD, and measures the time delay between trigger and SPAD signals as a conventional TDC. The design of this home-made, cost-effective, accurate FPGA-based TDC was described in detail elsewhere [50, 51]. It is characterized by a 42 ps time resolution, 1280 ns full scale, and 20 ns dead time. As an illustration, figure 3 displays the fluorescence decay of fluorescein dissolved to 10 μM in PBS, acquired in a conventional spectroscopy cuvette with this FPGA-based TDC, and showing the expected monoexponential decay kinetics with a FLT of 4.2 ns.

Zoom In Zoom Out Reset image size

The above FPGA-based TDC is used for TCSPC FLT detection in microfluidic droplets as illustrated in figure 2. A typical experimental run consists of circulating droplets of given size at a controlled flow rate and concentration, while detecting and time-stamping single photons continuously over seconds to minutes. Typical data are displayed in figure 4. Data processing is performed subsequently as follows. First the analysis of the instantaneous photon counting rate (computed as the number of photons detected over a 40 μs long sliding acquisition window) allows us to detect the passage of successive droplets across the excitation laser, with a higher counting rate (i.e. above a fixed threshold) inside droplets, and lower (below threshold) in between two droplets. Second, all the time-correlated, single photon detection events corresponding to a given droplet are assembled in one histogram representing the fluorescence decay kinetics of this very droplet. At a typical MHz photon counting rate inside the droplets, fluorescence decay curves consisting of a few thousand photons in total may be collected per droplet, at a droplet circulation rate exceeding 300 per second. An automatic fitting procedure is programmed to extract the FLT of all single-droplet histograms, by performing a single parameter monoexponential fit (nonlinear least-square minimizing routine implemented with Matlab). More precisely, the fitting function is an exponential decay ($N/\tau \times \exp (t/\tau ),$ with N the total number of photons in the histogram, and τ the FLT), convolved with the instrument response function, determined separately and modeled analytically by a gaussian function of 80 ps full width at half maximum. An example of such a fit is displayed in figure 4(c).

Zoom In Zoom Out Reset image size

Previous investigation of the physical limitations of the photon counting technique in these very high throughput conditions allowed us to make several conclusions [36]. First, as a very general property of single photon counting, the measured FLT is accurate as long as the probability that more than one photon hitting simultaneously the SPAD remains low. If this condition is not fulfilled, the measured fluorescence decay histogram is distorted, and the FLT underestimated. This is known as the pile-up effect [52, 53]. In the present conditions, it can be shown to remain negligible [36] at photon detection rates lower than one tenth of the laser repetition rate typically, i.e. 2 MHz here. Second, due to this fundamental limitation in photon counting rate, the total number of photons detected during a typical 3 ms exposure time remains relatively small (few thousands), and consequently the photon detection noise (so-called 'shot noise') is the origin of the weak signal-to-noise ratio observed in the histograms of individual droplets (figure 4(b)). This is therefore also the dominant source of imprecision on the FLT, determined by the monoexponential fit of the noisy histograms. Third and most importantly, this shot noise limits the relative precision on the FLT value to $1/\sqrt{N}$ at best, with N the total number of photons in the histogram [54, 55]. This is a well-known consequence of the Poissonian statistics which characterizes the shot noise inherent to single photon detection. This implies in particular that for a monoexponential decay, measuring a FLT is in principle not less precise and does not require more photons than measuring a FLINT, since the latter is also characterized by the same $1/\sqrt{N}$ imprecision at best. With the present experimental set-up, we have shown that the actual relative precision on the measured FLT approaches very closely the ultimate value of $1/\sqrt{N},$ to within a factor of 1.2–1.5 [36]. Therefore accurate discrimination between two distinct FLT's corresponding to positive and negative control samples is achievable in a screening experiment, even at a very high throughput implying the detection of a relatively limited number of photons per sample. This is what we will demonstrate below in a proof of principle experiment based on a biologically relevant, enzyme activity assay.

The enzyme activity assay was developed and validated previously [15, 18, 19]. It is based on the increase in the FLT of the PT14 probe which is initially quenched by collisions with the flexible tryrosine residue in the substrate. After trypsin has cleaved the substrate, the tyrosine is separated from PT14 and the probe's FLT is restored. Hence the increase of the FLT is the signature of the enzyme activity. As a test to demonstrate the feasibility of this assay by implementing TCSPC in the conditions illustrated in figures 3 and 4, we performed two successive experiments, where we measured the FLT of microdroplets circulating at a flow rate of about 300 per second. In the first one, the droplets were produced to contain the substrate solution at 1 μM concentration (see Material and Methods), by injecting the same solution in the three aqueous inlets. In the second experiment, a solution of 1 μM substrate and 1 nM enzyme was prepared about half an hour in advance such that the enzymatic reaction was initiated (see figure 1) before injection in the three aqueous inlets of the microfluidic chip. Hence, in these two experiments, each individual droplet is a new control sample to evaluate the FLT of the negative (first experiment) or positive (second experiment) control samples. Each acquisition run lasts for a few seconds only. By analyzing statistically the result of the FLT measurement in each individual droplet (sample), we quantify the precision with which the FLT of the negative and positive control samples may be determined in these very high throughput conditions, and very low sample volumes.

The fluorescence decay histograms obtained for individual droplets were composed of approximately 2300 photons and were fitted automatically as described above, with the FLT τ as a unique fitting parameter. Figure 5(a) displays the distribution of the FLT's measured on thousands of droplets in both experiments. More precisely, the FLT of the negative control sample (first experiment) is determined to be ${\tau }^{-}=6.9$ ns with a RMS standard deviation of ${\sigma }^{-}=0.25$ ns (green distribution), while that of the positive control (second experiment) is ${\tau }^{+}=11.4$ ns with ${\sigma }^{+}=0.42$ ns (red distribution). The corresponding coefficients of variations (CVs) are thus no more than ∼3.6% in both cases. The feasibility of the enzymatic assay in these very high throughput conditions may be quantified by the so-called Z' factor [56]. In the present case of an assay based on a FLT measurement it writes: $Z^{\prime} =1-\tfrac{3{\sigma }^{+}+3{\sigma }^{-}}{{\tau }^{+}-{\tau }^{-}},$ with the notations introduced above. With as little as 3 ms exposure time per sample we already achieve Z' = 0.55, which exceeds 0.5 and therefore demonstrates the feasibility of the assay in these conditions.

Zoom In Zoom Out Reset image size

Finally, we note that since the precision on the FLT determination is limited by Poisonian photon shot noise [36], a 10 fold increase in the sample exposure time should improve by $\sqrt{10}$ the CVs. This is easily confirmed by summing the histograms of ten successive droplets (all droplets are identical in a given experiment), before performing the automatic fit of the resulting histograms. The results are displayed in figure 5(b) and confirm that the CVs decrease to 1.2% as expected. By this analysis, we show that extending the exposure time to 30 ms, i.e. reducing the throughput to 1800 samples per minute which is still very high, the assay can be performed with a Z' factor as large as Z' = 0.85.