DNA Fragment Sizing by Single Molecule Detection in Submicrometer

E. R. Brown , Edgar. ..... Michael V. Gallo , Kimberly A. Gillis , Lori A. Neely , Kevin M. Phillips , Gordon G. Wong , Steven R. Gullans , Rudolf Gil...
0 downloads 0 Views 395KB Size
Download PDF
Anal. Chem. 2002, 74, 1415-1422

DNA Fragment Sizing by Single Molecule Detection in Submicrometer-Sized Closed Fluidic Channels Mathieu Foquet,† Jonas Korlach,†,‡ Warren Zipfel,† Watt W. Webb,† and Harold G. Craighead*,†

School of Applied & Engineering Physics, and Field of Biochemistry, Molecular and Cell Biology, Cornell University, Ithaca, New York 14853

The fabrication of fluidic channels with dimensions smaller than 1 µm is described and characterized in respect to their use for detection of individual DNA molecules. The sacrificial layer technique is used to fabricate these devices as it provides CMOS-compatible materials exhibiting low fluorescence background. It also allows creating microfluidics circuitry of submicrometer dimensions with great control. The small dimensions facilitate single molecule detection and minimize events of simultaneous passage of more than one molecule through the measurement volume. The behavior of DNA molecules inside these channels under an applied electrical field was first studied by fluorescence correlation spectroscopy using M13 double-stranded DNA. A linear relationship between the flow speed and applied electric field across the channel was observed. Speeds as high as 5 mm/s were reached, corresponding to only a few milliseconds of analysis time per molecule. The channels were then used to characterize a mixture of nine DNA fragments. Both the distribution and relative proportions of the individual fragments, as well as the overall concentration of the DNA sample, can be deduced from a single experiment. The amount of sample required for the analysis was ∼10 000 molecules, or 76 fg. Other potential applications of these submicrometer structures for DNA analysis are discussed. As biotechnology progresses, there is an increased demand on high-throughput analysis systems. While the tools currently used by molecular biologists have allowed for outstanding achievements, the next steps of furthering our understanding of biological systems will require orders of magnitude more information. More efficient solutions are required to improve on the current abilities of existing technologies. Microfluidics technology1,2 allows for such novel developments. Microfluidic circuitry can be mass-produced the way electronic chips are manufactured, making them inexpensive and accessible. * To whom correspondence should be addressed: (e-mail) [email protected]; (telephone) 607-255-8707; (fax) 607-255-7658. † School of Applied & Engineering Physics. ‡ Field of Biochemistry, Molecular and Cell Biology. (1) Harrison, D. J.; Wang, C.; Thibeault, P.; Ouchen, F.; Cheng, S. B. In Micro Total Analysis Systems 2000; Van den Berg, A., Olthuis, W., Bergveld, P., Eds.; Kluwer Academic Publishers: Dordrecht, The Netherlands, 2000; pp 195-204. (2) Craighead, H. G. Science 2000, 290, 1532-1535. 10.1021/ac011076w CCC: $22.00 Published on Web 02/16/2002

© 2002 American Chemical Society

The reduction of size greatly reduces the analysis time, as has been demonstrated in several cases.3-5 Another benefit of microsystems is the reduction in sample size needed.6 Microfluidics devices have already appeared in fields as diverse as DNA analysis, single fluorophore detection,7,8 drug screening, cell culture, and chemical processing.9-11 Phenomena typical of the microscopic scale range can be exploited to develop entirely new analysis methods.12,13 In respect to DNA fragment sizing analysis, single moleculebased fragment sizing approaches have initially been developed in devices involving small capillaries and flow cytometry.14-16 While powerful, this format does not easily allow integration into labon-a-chip devices. Recently, single molecule DNA fragment sizing was demonstrated in a channel microfabricated from a silicone elastomer.17 In both cases, the channel dimensions were still fairly large (on the order of 3 × 5 µm or larger). This imposes restrictions in respect to the optimal resolution that can be obtained. First, the large channel dimensions require relatively large observation windows for uniform illumination of the entire channel width. Therefore, only slow flow speeds, or low sample concentrations, can be employed to avoid multiple molecular occupancies. Second, (3) Mathies, R. A.; Huang, X. C. Nature 1992, 359, 167-169. (4) Han, J.; Turner, S. W.; Craighead, H. G. Phys. Rev. Lett. 1999, 83, 16881691. (5) Lazar, I. M.; Ramsey, R. S.; Ramsey, J. M. In Micro Total Analysis Systems 2000; Van den Berg, A., Olthuis, W., Bergveld, P., Eds.; Kluwer Academic Publishers: Dordrecht, The Netherlands, 2000; pp 379-382. (6) Wooley, A. T.; Mathies, R. A. Proc. Natl. Acad. Sci. U.S.A. 1994, 91, 1134811352. (7) Zander, C.; Drexhage, K. H.; Han, K.-T.; Wolfrum, J.; Sauer, M. Chem. Phys. Lett. 1998, 286, 457-465. (8) Lyon, W. A.; Nie, S. Anal. Chem. 1997, 69, 3400-3405. (9) James, C. D.; Davis, R. C.; Kam, L.; Craighead, H. G.; Issacson, M.; Turner, J. N.; Shain, W. Langmuir 1998, 14, 741-744. (10) Yun, J.; Jun, G.; Jay, X.; Cheng, S. L. In Micro Total Analysis Systems 2000; Van den Berg, A., Olthuis, W., Bergveld, P., Eds.; Kluwer Academic Publishers: Dordrecht, The Netherlands, 2000; pp 485-488. (11) Eherfeld, W.; Hartman, H.-J.; Hessel, V.; Kiesewalter, S.; Lowe, H. In Micro Total Analysis Systems 2000; Van den Berg, A., Olthuis, W., Bergveld, P., Eds.; Kluwer Academic Publishers: Dordrecht, The Netherlands, 2000; pp 33-37. (12) Duke, T. A. J.; Austin, R. H. Phys. Rev. Lett. 1998, 80, 1552-1555. (13) Han, J.; Craighead, H. G. Science 2000, May 12; 288: 1026-1029. (14) Castro, A.; Fairfield, F. R.; Shera, E. B. Anal. Chem. 1993, 65, 849-852. (15) Goodwin, P. M.; Johnson, M. E.; Martin, J. C.; Ambrose, W. P.; Marrone, B. L.; Jett, J. H.; Keller, R. A. Nucleic Acids Res. 1993, 21, 803-806. (16) Van Orden, A.; Keller, R. A.; Ambrose, W. P. Anal. Chem. 2000, 72, 3741. (17) Chou, H. P.; Spence, C.; Scherer, A.; Quake S. Proc. Natl. Acad. Sci. U.S.A. 1999, 96, 11-13.

Analytical Chemistry, Vol. 74, No. 6, March 15, 2002 1415

Figure 1. Schematics of the process of a sacrificial layer fabrication. The details of every step are outlined in the text.

the larger the channel the greater noise contributions become from buffer solutions. With smaller channel dimensions, it can be ascertained that all of the DNA molecules passing the detection region will be analyzed rapidly with high signal-to-noise ratios and minimized multiple occupancies. In this paper, a device consisting of submicrometer-sized fluidic channels, fabricated by the sacrificial layer technique,18 is described and is evaluated in respect to its performance for DNA fragment sizing. A software algorithm analyzing the photon burst data is described and shown to accurately determine the composition of the sample, molar fraction, and mobility, as well as the overall concentration of DNA in the sample. In the first part, the fabrication process for the device is outlined and the strong point of the sacrificial layer fabrication technique will be highlighted. The sacrificial layer technique provides the ability to create ultrasmall microfluidics systems with very accurate dimensional control. In the second part, experiments and technical aspects in respect to DNA fragment sizing by single molecule detection using these devices will be described. EXPERIMENTAL SECTION 1. Fabrication of Submicrometer-Sized Channels. The fabrication process is derived from the sacrificial layer method.18 It allows fabrication of micro/nanostructured fluidics systems that are both extremely complex and have tight dimensional tolerance.19 The device in this paper consists of a closed rectangular channel, with several constrictions having submicrometer dimensions. These constrictions provide the femtoliter probe volumes used for DNA fragment sizing analysis. (18) Turner, S. W.; Perez, A. M.; Lopez, A.; Craighead, H. G. J. Vac. Sci., Technol. B 1998, 16, 3835-3540.

1416

Analytical Chemistry, Vol. 74, No. 6, March 15, 2002

The sacrificial layer technology relies on the use of three different materials: a sacrificial layer patterned as the internal space of the planned microfluidic device, a base/coating layer that will surround the sacrificial layer and ultimately form the walls of the microfluidic device, and a chemical etchant used to dissolve/ resorb the sacrificial layer, while leaving the wall layer intact. Many different sets of materials can be used to create sacrificial layerbased devices, ranging from polymer coatings and organic solvents to CMOS-compatible materials. Because of the requirement on the system to be optically transparent, to have low-level background fluorescence, and to be compatible with a wide range of potential chemistries, glass appeared as an ideal candidate for the wall material. Therefore, a combination of fused silica (wall material), polycrystalline silicon (sacrificial layer), and tetramethylammonium hydroxide (TMAH, MF 312 Microposit photoresist developer, Shipley Inc., Marlborough, MA, wet etchant) was used for this application. TMAH offers a selectivity of several hundreds of thousands to one in the etching of polysilicon to fused silica. Moreover, it is a gaseous species at room temperature and thus does not leave residues inside the channels upon drying. The fabrication proceeds as illustrated in Figure 1. First, a 300nm layer of polysilicon is deposited using a low-pressure chemical vapor deposition (LPCVD) furnace (Silicon Valley Group, Scotts Valley, CA), on a fused-silica wafer substrate (step 1). The top 100-nm layer of this polysilicon is then thermally oxidized at 1000 °C. Optical lithography is used to pattern the sacrificial layer (step 2). The stepper used is an i-line stepper (GCA, Andover, MA) with 0.42 numerical aperture and minimal resolution of 0.5 µm. A (19) Foquet, M. E.; Turner, S. W.; Korlach, J.; Webb, W. W.; Craighead, H. G. In Micro Total Analysis Systems 2000; Van den Berg, A., Olthuis, W., Bergveld, P., Eds.; Kluwer Academic Publishers: Dordrecht, The Netherlands, 2000; pp 549-552.

Figure 2. (A) Reflection micrograph showing a top view of a narrow region of the microchannel. The regions appearing blue are the areas filled with fluid. (B) Scanning electron micrograph showing the cross section of a wide region (10-µm width, 270-nm height) of the channel. The width appears smaller in this image because the channel is viewed from a side angle. The ridges visible in the ceiling are an artifact due to cutting of the microstructure prior to SEM imaging.

layer of positive-tone photoresist (OCG620-7i, Shipley Inc.) is used to transfer the exposed pattern into the top thermal oxide layer. The transfer is done using a trifluoromethane/oxygen plasma. Once the oxide cap has been patterned, the photoresist is removed in a barrel etcher with a high-pressure oxygen plasma. This step is necessary because the photoresist would otherwise sputter away during the transfer of the pattern from the oxide to the silicon. This would affect the final devices by creating large amounts of unwanted defects in the walls of the microfluidic structure. The pattern is then transferred to the silicon using a chlorine-based plasma etch. The entire wafer is now capped with a 1-µm-thick layer of doped fused silica (step 3) deposited in a plasma-enhanced chemical vapor deposition (PECVD) system. This layer encloses the polysilicon and can also be used as a guiding medium to form optical waveguides.20 At this point, a second level of lithography is performed to etch the irrigation holes in the cover oxide layer (step 4), thus providing access to the polysilicon sacrificial layer. The pattern is transferred from the resist to the capping oxide using a CHF3/O2 plasma etch. The sample is then immersed in a solution of 5% TMAH (v/v) and heated at 75 °C for 4 h. At the end of the etch, all the polysilicon has been removed and a microfluidic channel has been created (step 5). Next, the irrigation holes have to be reclosed. This will completely seal the cavity left by the removal of the polysilicon. The closing layer needs to be partially conformal in order to efficiently seal the system (step 6). A very low-temperature oxide deposition was found to be ideal for this step. A layer between 1.5- and 2.5-µm thickness is deposited on top of the PECVD oxide, and the seal is verified by immersing the samples in water. At this point, the full microfluidic device has been fabricated. However, it is hermetically sealed. A final step of lithography allows opening of large holes at the extremities of the device (step 7) used to fill the channels with the sample solution. With this last step, the microfluidic portion of the device is finished (Figure (20) Foquet, M. E.; Han, J.; Lopez, A.; Wright, W.; Craighead, H. G. SPIE Proc. 1998, 3258, 141-146 (Proceedings of Micro-and Nanofabricated structures for Biomedical Environmental Applications).

2). Fluid reservoirs are added and gold electrodes inserted to induce the electroosmotic flow. 2. Experimental Setup. The reservoirs were prefilled with a buffer containing 450 mM Tris-borate, pH 8.0, 10 mM EDTA (5× TBE, Sigma, St. Louis, MO), and 0.02% (v/v) Nonidet P40 (USB Corp., Cleveland, OH). Electroosmotic flow was established by applying 100 V for 1 h. This prefilling was used both to condition the walls of the channels and to remove any air bubbles. DNA was prestained with YOYO-1 (Molecular Probes, Eugene, OR) at a dye-to-base pair ratio of 1:10, in the buffer used to prefill the channels, supplemented with 3% (v/v) β-mercaptoethanol as a reducing agent to limit photobleaching. For double-stranded M13 DNA (Sigma), the final concentration was 64 pg/µL, for the DNA mixture (molecular weight marker II (Roche Molecular Biochemicals, Indianapolis, IN), 1 ng/µL. The fragment sizes are given in Table 1. All fragments of this HindIII enzymatic digest of λ DNA are present in equal amounts; however, the 4.4- and 23.1-kb fragments partially anneal to each other. This was purposely used to create a smaller fraction of the 4.4-kb fragment and another fragment of 27.5-kb length. The DNA was allowed to stain for 1 h at 30 °C. Where not otherwise indicated, measurements were performed at an estimated electric field gradient of 312 kV/m across the narrow channel region. The samples were analyzed on an upright microscope (BX50WI, Olympus, Melville, NY), providing a combination of wide-field imaging capabilities using an intensified CCD camera (Stanford Photonics, Palo Alto, CA) and a setup for carrying out fluorescence correlation spectroscopy (FCS), described below. For the time series showing DNA movement in the channels (Figure 3), mercury arc lamp illumination and a standard filter cube (UMWIB, Olympus) were used. Frames were acquired and processed using software by Epix Inc. (Buffalo, IL). For photon histogram and FCS measurements, the 488-nm line of an argon ion laser was used (American Laser Corp., Salt Lake City, UT). A λ/4 wave plate provided circular polarization. The beam was focused onto the channel by a 40 ×/0.8 NA objective lens (Olympus). Unless noted otherwise, the power at the sample Analytical Chemistry, Vol. 74, No. 6, March 15, 2002

1417

Figure 3. (A) Dependence of flow speed of DNA molecules in the microchannel on the applied electric field, using M13 dsDNA (7.3 kb). The inset shows the corresponding normalized autocorrelation curves at various voltages from which flow speeds were derived. The red curve is the fit to the FCS curve at 1000 V (625 kV/m), using a model assuming only one-dimensional translational flow (eq 2). The dashed FCS curve corresponds to conditions where the voltage was switched off. It is particularly noisy due to the rareness of single molecule event in the case of free diffusion (see text). (B) The image presents a picture taken with an Intensified CCD camera of a device under use. The outline of the capillary has been added for clarity. Time series of the motion of single molecules of DNA under the influence of an electric field can be downloaded at http://www.hgc.cornell.edu/Microfluidics/Movies.html (see text).

was 100 µW. Underfilling of the lens back aperture created a relatively large axial focal volume radius of 1.9 µm (where the intensity had decreased by a factor of e-2, obtained from FCS measurements in the channels, using Alexa488-dUTP (Molecular Probes) at 50 nM concentration in the same buffer as used for the DNA fragment sizing experiments). The detection volume thus created was ∼1 fL. Long-term stability of the position of the focal volume with respect to the channel was ensured using a MS2000 microscope stage (Applied Scientific Instrumentation, Eugene, OR). Fluorescence light was collected through the same objective lens, separated by a dichroic beam splitter and band-pass filter (DCLP500 and D575/150, Chroma Corp., Brattleboro, VT), imaged onto a 200-µm fiber (Oz Optics, Carp, ON, Canada), which was coupled to an avalanche photodiode (SPCM-AQR-14-FC, PerkinElmer Optoelectronics, Fremont, CA). The output of the APD was buffered and simultaneously sent to a computer equipped with a correlator card for FCS (Flex410R, Correlator .com, Bridgewater, NJ) and to a second computer containing a custom-build high-throughput photon counting card (United Electronics Industries, Watertown, MA) where the photon count history was continuously recorded in time bins of 20- or 50-µs duration. FCS can be used for measuring the flow speed of molecules in a channel.21,22 In FCS, emitted fluorescence from molecules passing through an excitation volume element is measured and analyzed by calculating the intensity autocorrelation function. FCS curves were fitted using a model of combined two-dimensional (21) Magde, D.; Elson, E. L.; Webb, W. W. Biopolymers 1978, 17, 361-376. (22) Gosch, M.; Blom, H.; Holm, J.; Heino, T.; Rigler, R. Anal. Chem. 2000, 72, 3260-3265.

1418

Analytical Chemistry, Vol. 74, No. 6, March 15, 2002

diffusion and uniform flow in one dimension:21

G(τ) )

(

)

1 1 τ2 exp - 2 N 1 + τ/τD τF (1 + τ/τD)

(1)

where τD is the diffusion time, τF the flow time of a molecule through the focal volume, and N the average number of molecules in the focal volume. Since τF is a measure of the average residence time of a molecule induced by the translational flow, it is inversely proportional to the flow speed. The flow speed v of the DNA molecules was then calculated using the focal volume radius w, by v ) w/τF. For fast flow speeds, the diffusion component could be neglected, and a model of one-dimensional flow could fit FCS curves:

G(τ) )

( )

1 τ2 exp - 2 N τF

(2)

For histogram analysis, an algorithm (written in C++) was designed that operated on the data from the counter card. Briefly, a burst is detected by the use of two parameters, the background level and the threshold level, which is the number of counts at which a molecule is considered to be passing through the focused beam. When a bin reaches the threshold level, the algorithm backtracks until it reaches the background level and assigns this time point as the beginning of the burst. The algorithm then goes forward in time until it reaches a bin containing the same background value to define the entire burst. The (backgroundsubtracted) maximum burst height, total burst size, burst duration, and time interval between the previous burst are calculated and

Figure 4. Sample output of the software algorithm used to analyze photon burst traces for DNA fragment sizing analysis. The top window shows a short portion of the intensity time trace. The middle left window shows the result of photon burst size analysis, in this case for M13 DNA (7.3 kb). The lower left window is a histogram analyzing the width of the photon bursts. The lower right window displays the histogram of intervals between successive bursts, from which the total DNA concentration of the sample can be calculated (see text).

plotted as histograms. The photon bursts were corrected for the saturation curve of the detector as given by the manufacturer. The algorithm also eliminates photon bursts exceeding a certain duration. This was implemented to prevent the inclusion of cases where multiple DNA molecules would simultaneously be present in the interaction volume. However, the fact that this subroutine did not alter the histograms demonstrates that the probability of multiple occupations is rare. Estimates based on a Poisson distribution show a probability of double occupancy lower than 0.5%, and analysis of the burst data was consistent with that number. The intervals between successive bursts were used for calculating the total concentration of DNA in the sample. The fit was made according to the n ) 1 term of a Poisson distribution, with the form of the fitting function being

f(x) ) FAvCxe-AvCx

(3)

where F is a scaling factor, A the cross-sectional area of the channel (0.27 µm2 in this case), v the flow speed of the molecules, and C the concentration of molecules. Burst data can be combined to create two-dimensional histograms (not shown). This is useful when species with different mobilities, but similar overall burst area, are to be differentiated. Data processing was essentially instantaneous on a Pentium III computer. A typical output of this software is shown in Figure 4. Histograms were fit using a leastsquares fit provided by the software Origin (OriginLab Corp., Northampton, MA).

For the comparison of the separation by gel electrophoresis, 50 ng of the DNA ladder was loaded onto a 0.6% agarose gel, run for 1 h at 150 V, and stained with Sybr Gold (Molecular Probes) for 30 min. RESULTS Using the sacrificial layer technique, fluidic channels can be fabricated with dimensions smaller than 1 µm (Figure 2). Figure 2A shows the top view of a portion of the device where the channel width is constricted from 10 to 1 µm. Figure 2B depicts an SEM image of a side view of the 10-µm portion of the channel. For DNA fragment sizing, channel regions with a height of 270 nm and width of 1 µm were used. The channels can be readily filled with fluid containing DNA molecules, as described in the Experimental Section. FCS can be used to analyze the electrophoretic behavior of the DNA as a function of electric field.23 In addition, it allows us to perform a quick diagnostic on accurate alignment of the focal volume to the channel and to calibrate the system. A set of FCS measurements at different voltages shows that the DNA speed varies linearly with the applied electric field (Figure 3A). It is notable that even at low fields the correlation curves are dominated by the flow, and diffusion becomes negligible. At very low electric fields, the error on the measurements becomes increasingly larger due to the smaller number of DNA molecules passing through the focal volume per unit time and, therefore, increasing lack of (23) Van Orden, A.; Keller, R. A. Anal. Chem. 1998, 70, 4463-4471.

Analytical Chemistry, Vol. 74, No. 6, March 15, 2002

1419

statistics. Ultimately, measurements at zero external field would be expected to yield the free diffusion coefficient of DNA molecules trapped in the channel. However, the diffusive passage of a DNA molecule through the channel is a very rare event. YOYO-stained double-stranded M13 DNA inside the channels under the influence of an electric field is shown in Figure 3B (Corresponding time series can be downloaded at http:\www.hgc. cornell.edu\Microfluidics\Movies.html). The first movie shows the movement of DNA molecules when a small voltage (15 V) is applied across the fluidic reservoirs. The motion of the DNA molecules changes immediately from Brownian motion to a directed flow. The molecules get accelerated upon entering the narrow channel region of 1-µm width. DNA fragment sizing experiments were performed at much higher voltages, typically around 500 V, where the speed of DNA molecules through the channel is much faster. This is illustrated in the second movie, where the very rapid movement of DNA through even the large channel regions is visible. The actual analysis of fragments is carried out in a mode typical for FCS, using a diffraction-limited excitation volume rather than a wide-field imaging mode. Here, the focused laser beam illuminates only a small region of the device. This is shown in the final movie, at a constant applied voltage of 100 V. At this voltage, the passage of DNA molecules is visible as flashes of fluorescence light. Photon burst analysis was performed on this sample under conditions described in the figure legend, with a sample output of the computer program shown in Figure 4. The middle left window shows the histogram analysis calculating the burst size distribution according to the algorithm described in the Experimental Section. One peak centered around 1123 photons/burst is obtained, with a standard deviation of 7%, corresponding to a DNA size of 7.3 ( 0.5 kb. During the acquisition time of 131 s, 704 molecules, or 5.6 fg of DNA were recorded. This corresponds to 86 pl of solution consumed during the measurement time. On average, it took 1.6 ms to analyze each molecule, as determined from the photon burst width histogram (lower left window). The lower right window shows the histogram of time durations between bursts, from which the concentration of the DNA sample can be obtained. The exponential tail of the histogram was fit to the n ) 1 term of a Poisson distribution, using the value of the flow speed deducted from FCS measurements (see the Experimental Section). The measured value of 1.06 × 1011 molecules/ mL (17.1 nM) is fairly close of the expected value of 0.82 × 1011 molecules/mL (13.6 nM). Photon burst histograms of a mixture of DNA molecules containing several fragment sizes are shown in Figure 5. Arrays of Gaussian peaks were used to fit the histograms. The smallest fragment of 0.13 kb could not be resolved from the background in these experiments. All other fragments, except for the 2- and 2.3-kb doublet, are clearly resolved. Table 1 shows the fit parameters for a typical run. Relative standard deviations were around 10%, decreasing with increasing fragment length, as previously found.17 The relative molar fractions of the fragments were correctly obtained from the histograms (Table 1, relative burst frequency). Because the 4.4-kb fragment had partially annealed to the 23-kb fragment, the fractions of these bands are much lower than 14%. From the fraction of the annealed 27.5-kb 1420 Analytical Chemistry, Vol. 74, No. 6, March 15, 2002

Table 1. Fit Parameters from the Photon Histogram Analysis of the DNA Mixture (Data Shown in Figure 5) fragment size (bp)

average burst size (photons)

relative standard deviation (%)

relative burst frequency (%)

564 2027 and 2322 4361 6557 9416 23130 27490

1188 5075 10670 15894 22741 58016 69326

25 14 15 7 10 6 6

18 30 8 19 24 6 12

annealed product (Table 1), it can be deduced that ∼60% of these two fragments were annealed to each other in this sample. A total of 10 000 bursts were required for generating this histogram, corresponding to 76 fg of DNA sample. The burst size was linearly proportional to the DNA fragment size over the entire range of the molecular weight marker. Figure 5C shows the combined data from the two histograms, with the data from the first histogram multiplied by the transmission factor of the neutral density filter used to attenuate the laser. Figure 5D shows the same DNA sample, as obtained by agarose gel electrophoresis. In comparison to the histogram, the doublet of 2 and 2.3 kb is well separated; however, the two largest bands of 23.2 and 27.5 kb are not resolved, typical of the nonlinear separation properties of gel electrophoresis. In addition, the brightness of each band is related not only to the number of molecules present in each band but also to the length of the DNA fragment, due to the different amount of fluorescent dye bound to the fragments. For example, the largest band results in a very bright, overexposed band, whereas about the same number of molecules in the 0.6-kb band are barely visible above the background (Figure 5D). This makes quantitation of the relative proportions of multiple fragments of significantly different sizes difficult by gel electrophoresis. DISCUSSION The striving for ever decreasing dimensions of microdevices is rooted in the understanding that observations on single molecules with high signal-to-noise properties enhance the attainment of previously inaccessible knowledge. Analogous to the computer industry, smaller device dimensions also allow for higher densities of features on a single chip, thereby increasing their performance. In this article, the use of channels with such small dimensions is described for the application of single molecule detection-based DNA fragment sizing. All fabrication steps are fully compatible with semiconductor industry CMOS processing. As in previous reports involving larger channels or capillaries, every molecule passing the detector is analyzed and contributes to the information obtained about the sample, which is important in applications involving costly samples, the detection of very rare molecular species in a complex sample mixture, or both. With the ultrasmall channels used here, high signal-to-noise detection allows very fast flow speeds, up to 5 mm/s, thereby shortening the analysis time per molecule to a few milliseconds, and thus the overall analysis time of the entire sample. DNA fragments over a wide dynamic range can rapidly be detected using this system. The amount of sample required is minute, several orders of magnitude less than for the case of gel

Figure 5. (A) Photon burst histogram obtained for a mixture of several DNA fragments. The red curves correspond to the fitting of a set of six Gaussian peaks by a least-squares method. The positions of the peaks depend on the size of DNA fragments; the peak area is proportional to the relative concentration of each fragement. For fitting parameters, see Table 1. (B) Section of photon burst histogram showing the 0.6-kb fragment resolved over the background. The power of the sample for this trace was 1 mW. (C) Plot of the burst size as a function of the (known) fragment size. Error bars correspond to the standard deviation of the peak sizes (see Table I). The dashed line is a linear least-squares fit. The intercept was -256 photons, and the slope was 2.498 photons/bp. The correlation coefficient is 0.9998. (D) Image and line scan through the center of the image of the same sample, as obtained by agarose gel electrophoresis. The original photon burst data were corrected for nonlinearities of the detector, as described in the Experimental Section.

electrophoresis, and essentially limited by abilities to handle the fluid exterior of the channel. Single molecule techniques furthermore offer the advantage of intrinsic quantitationsthe molecules simply have to be counted to obtain this information. This was demonstrated by extracting the overall as well as fractional concentrations of each DNA fragment from a single experiment. The resolution of the technique is influenced by several factors. Related to the staining of the DNA with fluorescent dyes are fluctuations in the number of dye molecules attached on fragments of equal length, fluctuations in the number of photons emitted by each dye molecules during observation, fluctuations in the number of photons detected by the collection optics, and variations in binding as a function of DNA sequence. Unlike gel electrophoresis, the resolution improves with increasing fragment size since

all factors (except for the last mentioned) obey Poisson statistics.17 The first parameter sets an intrinsic limit on the lower resolution of the method. For the smallest fragment of 0.13 kb in the DNA mixture studied, the variance on the number of dye molecules bound to a given fragment is quite large. Additionally, in our experiments, a low background fluorescence originating from the device was observed which set the lower limit of resolution for fragment sizing. Thus, it was difficult to resolve this fragment from the background. We are investigating alternative nanofabrication methods to avoid this source of background noise, thus extending the lower resolution limit to its intrinsic value. It is also likely that, with the current protocol, the number of dye molecules bound to a fragment did not reach equilibrium because of the low off rate of YOYO-1 from DNA, thus causing a further broadening of the histogram peaks. The use of other dyes, such Analytical Chemistry, Vol. 74, No. 6, March 15, 2002

1421

as PicoGreen,24,25 and different staining protocols may result in better resolution. Effects of laser polarization on sizing resolution have also been reported.26 Further experiments aimed at addressing these questions, establishing the upper DNA fragment size limit, and improving the overall performance are currently in progress. The possibility of multidimensional analysis permits the distinction of species that otherwise would be hidden by measuring one parameter alone. This principle has proven powerful in solution analysis27 and flow cytometry;28 here it can be exploited to discriminate differences in electrophoretic mobilities in addition to molecular brightness. This could be applied, for example, to binding studies of proteins to DNA molecules. Other applications, such as the detection of rare single-nucleotide polymorphisms (SNPs) in a complex mixture of DNA, or the occurrence of several SNPs on a single DNA molecule, are currently being studied in our laboratory. The extreme confinement of fluid volume by these nanofabricated devices also implies that the concentration required for single molecule detection can be increased proportionally. This is important not only for DNA fragment sizing but for many other applications involving the study of single fluorophores.29,30 As many biochemical processes occur efficiently only at concentrations (24) Yan, X.; Grace, K. W.; Yoshida, T. M.; Habbersett, R. C.; Velappan, N.; Jett, J. H.; Keller, R. A.; Marrone, B. L. Anal. Chem. 1999, 71, 5470-5480. (25) Yan, X.; Habbersett, R. C.; Cordek, J. M.; Nolan, J. P.; Yoshida, T. M.; Jett, J. H.; Marrone, B. L. Anal. Biochem. 2000, 286, 138-140. (26) Agronskaia, A.; Schins, J. M.; de Grooth, B. G.; Greve, J. Appl. Opt. 1999, 38, 714-719. (27) Eggeling, C.; Berger, S.; Brand, L.; Fries, J. R.; Schaffer, J.; Volkmer, A.; Seidel, C. A. M. J. Biotechnol. 2000, 86, 163-180. (28) Steinkamp, J. A.; Habbersett, R. C.; Hiebert, R. D. Rev. Sci. Intrum. 1991, 62, 2751-2764. (29) Do ¨rre, K.; Stephan, J.; Lapczyna, M.; Stuke, M.; Dunkel, H.; Eigen, M. J. Biotechnol. 2001, 86, 225-236.

1422

Analytical Chemistry, Vol. 74, No. 6, March 15, 2002

much higher than the pico- or nanomolar concentration regime typically required for single molecule analysis in solution, nanofabricated devices offer the potential to study these processes at physiologically more appropriate concentrations. We have shown that single fluorophores can be detected with high signal-to-noise ratios in these nanofabricated channels.31 Fluorescence correlation spectroscopy and single-fluorophore characterization can be performed at higher concentrations, ranging into the micromolar regime, thus extending the range of biochemical reactions that can successfully be studied on a single molecule level.19 Nanochannels hold potential interesting innovations in other areas. As the surface-to-volume ratio of microfluidic systems is large, surface modifications could lead to experiments involving binding of molecules. Incorporation of waveguides for efficient light delivery not involving complex microscope setups has already been demonstrated.20 ACKNOWLEDGMENT We thank Dr. Stephen Turner and Dr. Mike Levene for insightful discussions. Dr. Levene’s assistance with the optical setup is gratefully acknowledged. This publication was made possible by Grant DE-FG02-99ER62809 from the U.S. Department of Energy, and by Grant P41-2RR04224 from the National Center for Research Resources, National Institutes of Health.

Received for review October 9, 2001. Accepted December 20, 2001. AC011076W (30) Levene, M.; Larson, D.; Korlach, J.; Foquet, M.; Turner, S. W.; Craighead, H. G.; Webb, W. W. Biophys. J. 2000, 78, 2368. (31) Korlach, J.; Levene, M.; Turner, S. W.; Larson, D. R.; Foquet, M.; Craighead, H. G.; Webb, W. W. Biophys. J. 2000, 80, 147a.