Anal. Chem. 1995,67, 3253-3260
Single Molecule Fluorescence Burst Detection of DNA Fragments Separated by Capillary Electrophoresis Brian B. Haab and Richard A. Mathies*
Department of Chemistry, University of Califomia, Berkeley, Califomia 94720
A method has been developed for detecting DNA separated by capillary gel electrophoresis (CGE) using single molecule photon burst counting. A confocal fluorescence microscope was used to observe the fluorescence bursts from single molecules of DNA multiply labeled with the thiazole orange derivativeTO6 as they passed through the -2-pm diameter focused laser beam. Amplified photoelectron pulses from the photomultiplier are grouped into bins of 360-450 ps in duration, and the resulting histogram is stored in a computer for analysis. Solutions of M 1 3 DNA were first flowed through the capillary at various concentrations, and the resulting data were used to optimize the parameters for digital fUteringusing a lowpass Fourier filter, selecling a discriminator level for peak detection, and applying a peak-callingalgorithm. Statistical analyses showed that (i) the number of M 1 3 molecules counted versus concentration was linear with slope = 1, (ii) the average burst duration was consistent with the expected transit time of a single molecule through the laser beam, and (iii) the number of detected molecules was consistent with single molecule detection. The optimized single molecule counting method was then applied to an electrophoretic sepaiation of M 1 3 DNA and to a separation of pBR 322 DNA from pRL 277 DNA. Clusters of discreet fluorescence bursts were observed at the expected appearance time of each DNA band. The autocorrelation function of these data indicated transit times that were consistent with the observed electrophoretic velocity. These separations were easily detected when only 50- 100 molecules of DNA per band traveled through the detection region. This new detection technology should lead to the routine analysis of DNA in capillary columns with an on-column sensitivity of -100 DNA molecules/band or better. High-sensitivity detection methods are finding increasing importance and usefulness in many fields, such as bacterial and viral diagnostics, environmental studies and health care, and basic physical and biochemical research. The ultimate in high-sensitivity detection is the detection of individual molecules. This goal has been achieved in several different formats. In the solid phase, the spectroscopic isolation of individual fluorescent impurities in low-temperaturecrystalline hosts' has been used to study molec(1) Moerner, W. E.; Kador, L. Phys. Rev. Lett. 1989,62, 2535-2538.
0003-2700/95/0367-3253$9.00/0 0 1995 American Chemical Society
ular properties2 and changes in local environments and spectral characteristic^.^-^ Near-field optical scanning microscopy has been used to detect single fluorescent dyes on a ~ u r f a c e ,and ~,~ this method has also been used to study properties and dynamics of individual fluorophore~?*~ Individual dye molecules adsorbed on a surface have also been imaged using a position-sensitive photon counting apparatus.'O Many analytical and bioanalytical applications are anticipated for single molecule detection in the liquid phase, where single molecules of both fluorescent dyes and DNA labeled with multiple fluorescent dyes have been detected in confined drops," levitated microdroplets,12focused sheath f l o ~ s , l ~ and - ' ~ capillary flow^.'^-'^ Biological macromolecules have been observed at the single molecule level by microscopic imaging,Ig and this technique has been used to study conformational dynamics20-22 and physical properties of DNA23as well as the behavior of RNA p0lymerase.2~While many of these techniques have involved the analysis of DNA and other biological macromolecules, an electrophoreticor chromatographicanalytical (2) Omt, M.; Bernard, J.; Brown, R.; Fleury, L.; Wrachtrup. J.; von Borczyskowski, C. J. Lumin. 1994,60,991-996. (3) Basche, T.; Moemer, W. E. Nature 1992,355, 335-337. (4) Ambrose, W. P.; Basche, T.; Moemer, W. E. J. Chem. Phys. 1991,95,71507163. (5) Moerner, W. E. Science 1994,265, 46-53. (6) Betzig, E.; Chichester, R. J. Science 1993,262, 1422-1425. (7) Trautman, J. IC; Macklin, J. J.; Brus, L. E.; Betzig, E. Nature 1994,369, 40-42. (8) Ambrose, W. P.; Goodwin, P. M.; Martin, J. C.; Keller, R A. Science 1994, 265, 364-367. (9) Xie. X. S.; Dunn, R. C. Science 1994,265, 361-364. (10) Ishikawa, M.; Hirano, K; Hayakawa, T.; Hosoi, S.; Brenner, S. Jpn. J. Appl. Phys. 1994,33, 1571-1576. (11) Nie, S.; Chiu, D. T.; Zare, R N. Science 1994,266, 1018-1021. (12) Whitten, W. B.; Ramsey, J. M.; h o l d , S.; Bronk, B. V. Anal. Chem. 1991, 63, 1027-1031. (13) Nguyen, D. C.; Keller, R. A;Jett, J. H.; Martin, J. C. Anal. Chem. 1987,59, 2158-2161. (14) Shera, E.B.; Seitzinger, N. K; Davis, L. M.; Keller, R A; Soper, S. A. Chem. Phys. Lett. 1990,174, 553-557. (15) Soper, S. A; Shera, E. B.; Martin, J. C.; Jett, J. J.; Hahn, J. H.; Nutter, H. L.; Keller, R. A. Anal. Chem. 1991,63, 432-437. (16) Peck, IC;Stryer, L.; Glazer, A. N.; Mathies, R A Proc. Natl. Acad. Sci. U S A . 1989,86, 4087-4091. (17) Lee, Y. H.; Maus, R G.; Smith, B. W.; Winefordner,J. D.Anal. Chem. 1994, 66, 4142-4149. (18)Wilkerson, C. W., Jr.; Goodwin, P. M.; Ambrose, W. P.; Martin, J. C.; Keller, R. A. ApPl. PhyS. Lett. 1993,62, 2030-2033. (19) Hirschfeld, T. Appl. Opt. 1976,25, 2965-2966. (20) Schwartz, D. C.; Koval, M. Nature 1989,338, 520-522. (21) Smith, S. B.; Aldridge, P. K; Callis, J. B. Science 1989,243, 203-206. (22) Perkins, T. T.; Smith, D. E.; Chu, S. Science 1994,264, 819-822. (23) Smith, S. B.; Finzi, L.; Bustamante, C. Science 1992,258, 1122-1126. (24) Schafer, D. A; Gelles, J.; Sheetz, M. P.; Landick, R. Nature 1991,352, 444-448.
Analytical Chemistfy, Vol. 67, No. 18, September 15, 1995 3253
separation technique employing single molecule detection has not yet been demonstrated. We present here the application of single molecule fluorescence burst detection to the separation of DNA in capillary gel electrophoresis (CGE). CGE is an attractive method to couple with single molecule detection since it is a widely used, reliable technique which offers high-resolution separation of DNA and other biological m a c r o m ~ l e c u l e s Applications . ~ ~ ~ ~ ~ employing CGE would be further enhanced by improvements in signal-to-noise ratio and sensitivity. Single molecule detection would provide the ultimate improvement. Optimization of the signal-to-noise ratio (S/N) is the main challenge in developing a single molecule detection system for CGE. The limiting factor for high-sensitivity detection in electrophoresis has been the background count rate caused by scattering and impurity fluorescence from the separation matrix as well as limited signal from the sample molecules. To address these issues, this work builds upon that of Zhu et al.,27who recently evaluated a variety of intercalation dyes for the detection of CGE separations of double-stranded DNA. The dye [(N,N'tetramethylpropanediamino)propyl]thiazole orange (T06) 28 was determined to have the best detection limits of the monointercalators, and for that reason it was chosen for use in this work. TO6 is advantageous because it has a high quantum efficiency and because its extremely high affinity for DNA permits DNA detection with very low concentrations of dye in the electrophoresis buffer. The S/N was further enhanced through the use of confocal optics, efficient spectral filtering, and a high numericalaperture oil immersion objective to increase the light gathering efficiency and to reduce scattering caused by the air-glass interface. The development of a digital photon counting interface for single molecule detection was also critical to this effort because it offers the potential for improved time resolution and S/N over analog detection.29 Because a single molecule in a typical electrophoresis experiment will travel through the laser beam in only 2-10 ms, amplified photoelectron pulses must be captured in bins of 300-500 p s in duration to achieve a desired minimum of 4 or 5 bins per DNA transit time. At this high rate of counting, commercially available counting circuits typically do not have enough on-board memory to allow extended, unintermpted data collection. We therefore constructed a computer-based counting circuit to accomplish this task efficiently. With these technical improvements, we have successfully developed a single molecule fluorescence burst detector and method for CGE. To optimize the spectroscopic apparatus and data analysis methods, solutions of M13 DNA were first flowed through the detector and the fluorescence bursts recorded. Statistical analyses were performed off-line to confirm single molecule detection and to better understand and optimize the system. In particular, we examined whether (i) the number of single molecule events is linear with concentration, (ii) the (25) Ewing. A. G . ;Wallingford, R. A.; Olefirowicz, T. M. Anal. Chem. 1989,61, 292A-303A. (26)Landers, J. P.; Oda, R. P.; Spelsberg, T. C.; Nolan, J. A,; Ulfelder, K. J. Bioteckniques 1993,14, 98-111. (27) Zhu, H.; Clark, S. M.; Benson, S. C.; Rye, H. S.; Glazer, A. N.; Mathies, R. A. Anal. Ckem. 1994,66, 1941-1948. (28) Benson, S. C.; Mathies. R. A.; Glazer, A. N Nucleic Acids Res. 1993,21, 5720-5726. (29) Ingle, J. D.;Crouch, S. R Spectrochemical A n a l i s ; Prentice-Hall: Englewood Cliffs, KJ, 1988; Chapter 5.
3254
Analytical Chemistry, Vol. 67,No. 78,September 15, 1995
duration of fluorescence bursts corresponds to the expected transit time of one molecule through the laser beam, and (iii) the number of molecular events corresponds to that expected for single molecule counting. In addition, these M13 data were used to optimize the noise filtration and peak-calling algorithms. Finally, M13, pBR322, and pRL 277 DNA were injected into the capillary and separated by CGE. Clusters of fluorescence bursts were observed at the expected appearance time of each DNA band. The statistical tests and single molecule counting system were applied to these data and used to confirm that the individual bursts within the clusters arose from single molecules of DNA This work establishes the feasibility of single molecule photon burst detection of DNA in all types of capillary separations. EXPERIMENTAL SECTION
Samples. M13 (7250 bp) and pBR 322 DNA (4631 bp) were purchased from New England Biolabs (Beverly, MA). The M13 DNA was linearized with BamHI, and the pBR 322 DNA was linearized with EcoRI. pRL 277 DNA (6798 bp)30was a gift from Dr. Yuping Cai of the UC Berkeley Molecular and Cell Biology Department and was linearized with EcoRI. DNA samples were diluted with a 0.4 mM Tris-acetate, 0.1 mM NazEDTA buffer (PH 8.0) and stored at -20 "C. Capillary Electrophoresis. Fused silica capillaries (100 pm i.d., 200 pm 0.d.) were purchased from Polymicro Technologies (Phoenix, Az) and the interior surfaces derivatized according to the protocol of Hjerten.3l The capillary length was 50 cm, and a 2-mm observation window was burned 25 cm from the injection end. The capillary was filled with a solution of 40 mM Trisacetate, 1mM EDTA, pH 8.0 ( l x TAE), and 0.10%-0.50% (w/v) hydroxyethyl cellulose (HEC, M,,= 438 OOO;Polysciences Inc., Warrington, PA). This HEC-containing buffer was first degassed under vacuum for 20-25 min and then centrifuged at 17 OOO rpm for 40 min (Sorvall Superspeed RC2-B). The monomeric intercalating dye [ (N,N'-tetramethylpropanediamino)propyllthiazole orange (T06)28 was added to the solution (10 nM) before centrifuging. At this concentration, the DNA should be nearly saturated with dye, giving a labeling density of 2-4 bp/dye. The HEC buffer solution was manually injected into the capillary using a 100-pL syringe. Samples were electrokinetically injected (100150 V/cm for 5-8 s) from a 5pL volume in an Eppendorf tube, which was then replaced by a buffer reservoir. Electrophoresis was performed at 150-200 V/cm. At the conclusion of each experiment, the capillary was flushed with water and acetone and purged with nitrogen gas. For the experiments using flowing solutions of DNA, a syringe pump (Harvard Apparatus, Millis, MA) and a 1OC-pL syringe were coupled to the capillary. Instrumentation. The experimental apparatus described in Figure 1is similar to that presented earlier.27 Excitation (488 nm) from a Coherent Innova 70 kryptodargon ion laser was passed into a confocal microscope (Axioskop; Carl Zeiss), reflected off a 510-nm long-pass dichroic beam splitter, and focused into the capillary by a 40x, 1.3 NA oil immersion objective (Plan Neoflor; Carl Zeiss) . The capillary was placed on a microscope slide, and a drop of immersion oil (Carl Zeiss; n = 1.518) was placed between the objective and the observation window of the capillary. The laser power was optimized according to the theory described (30) Black, T. A.; Cai, Y.; Wolk, C. P. Mol. Bid. 1993,9, 77-84. (31) Hjerten, S. J. Chromatogr. 1985,347,191-198.
"4
0.2w
.D
r--u computer
Figure 1. Schematic of the laser-excited, confocal fluorescence detection system and photon counting circuit. The 488-nm excitation light is reflected by the long-pass beam splitter (6s) and focused by the objective (0) into the capillary. Emission from the capillary is collected by the objective. passed through the beam splitter and green bandpass filter (F),and then focused by the lens (L) through pinhole (P)to the photomultiplier (PMT). The photoelectron pulses from the PMT pass through an amplifierldiscriminator and are counted by the counting circuit. The circuit toggled between two 8-bit counters so that as one counter was counting, the other was being read and cleared. The capillary (C) was mounted on a microscope slide. and a drop of immersion oil (n = 1.548) was placed between the objective and the detection window of the capillary.
earlie132 and was set between 0.3 and 0.5 mW at the sample. Fluorescence was collected by the objective, passed through the beam splitter, filtered by a bandpass filter (530DF30, Omega, Braffleboro,vr),focused on a 50pm pinhole, and detected with a photomultiplier (Hamamatsu F94502) in a cooled housing CE21ORF; Products for Research, Danvers, MA). After being passed through an amplifier and discriminator (SSR 1120), the photoelectron pulses were fed to a counting circuit that was built on a "breadboard" which plugged into one of the expansion slots of a 386based AT personal computer. This circuit toggled between two &bit counters Uameco, 74F193) at a rate of between 360 and 450 &bin, set by a function generator. While one counter was counting, the other was being read to the computer's RAM disk and cleared, so that there was no dead-time in collecting the data. Data Analysis. Data were first filtered to remove any anomalously large onebin noise spikes. The spikes occurred approximately once per 100 OOO points and were replaced by the average of the two adjacent points. Data were then Fourier transformed using an algorithm from ref 33, low-pass filtered in frequency space, and reverse Fourier transformed. The peakcalling algorithm searched for maxima in the data set which were above a defined discriminator level and also had immediately flanking points at a lower value. (321 Clalhicr. R A, Peek. K, Swyn.
(33) ~
P
W S
L Anol. Ckm. 1990,62,1786-1791 H..Venerling. W T;Teukolrkj. S. A. Flannery R. P N ~ m c m o l
Tlm.
(rsonds)
MI3 DNA dilution series. Solutions of MI3 DNA in 1 x TAE stained with 10 nM TO6 were flowed through the 100pm-i.d. capillary using a syringe pump. The trials were 108 s long (300 000 data points with a binwidth of 362 ps), and a 6-s poriion of each trial is shown. The law power at 488 nm was 0.3 mW at the sample, and the mean flow velocity was 0.044 c d s . Figure 2.
RESULTS AND DI!3CUSSION M13 DNA Dilution Series. To test the single molecule detection apparatus and to develop data analysis methods, solutions containing linear M13 DNA (7.25 kb) were flowed through the capillary with a syringe pump. Figure 2 presents 6 s portions of each 108sbial in a M13 DNA dilution series. The c o n c e n b tions ranged from 0 to 20 pM. When the blank solution was flowed through the capillary, no large fluorescence bursts were observed, and the average count rate was 0.38 counts/bin. When DNA solutions were introduced, discreet fluorescence bursts were seen, and the number of these bursts increased with concentration, suggesting that they were due to the passage of individual DNA molecules through the f m s e d laser beam. If the probability of more than one DNA molecule simultaneously occupying the probe volume is extremely low, then the observed fluorescence bursts can be atbibuted to single molecules of DNA An occupancy. or average number of molecules per probe volume, of -0.01 has been shown to be an effective criterion for single molecule detection.16 The e-2 probe volume in our case was approximated to be a cylinder 4.0 pm in height and 2.4 pm in diameter, giving a volume of 18 fL. 'The 2.4-pm diameter is our best estimate of the effedive beam diameter based on the capillary electrophoresis measurements described below, and the 4.0pm height is an estimate of the convolution of the depth of focus of the laser beam with the confocal detection efficiency." For an lML. probe volume, the occupancies range from 0.002 to 0.2 for
R~~pn11C2nd~.:CambridgrUnivrruryReEs NerYork. i992,Chapln 12
(34) Wilson. T.:Carlini, A R Opt. Lm. 1987. 12. B7-229.
Analyiical Chemisiy, Vol. 67, No. 18, September 15, 1995 3255
60
20
RaWD8''
10I C
.o 1.0 0.8
a
0.4
0 0
-
r j
= 0.2 5
0
I
I 0
10
20 30 40 Offset (bins)
50
60
Figure 3. Autocorrelation functions calculated from the M13 DNA dilution series data. The concentrations were 20 ( A ) , 10 (O), 2 (O), and 0 pM (A).(top) Unnormalized autocorrelation functions. (bottom) Normalized autocorrelation functions. The noise spike at t = 0 was removed from all data, and data were normalized to have the same value at T = 1. Each offset bin is 362 ps in duration, and the formula used to calculated the autocorrelation function was G ( t ) = (1/ N)Zn(t)n(f r ) , where G is the autocorrelation, N is the size of the data set, n(f) is the value at time t, and t is the offset.
+
the 0.2-20 pM concentrations. For the 2 pM concentration, assuming a Poisson distribution, the probability of no molecules in the probe volume is 0.980, the probability of 1 molecule in the probe volume is 0.0196, and the probability of two or more molecules simultaneously in the probe volume is 1.97 x For the 20 pM solution, the probability of two or more molecules simultaneously in the probe volume is 0.0175, which means that even at this concentration, the majority of the observed fluorescence bursts must arise from single molecules. The autocorrelation function has been shown to be effective in demonstrating the presence of non-Poissonian bursts due to single m01ecules~~J~ and in characterizing transit times and particle number^.^^,^^ Figure 3 presents the calculated autocorrelation functions from the M13 DNA dilution series data. The autocorrelations are approximately Gaussian in shape. The unnormalized autocorrelations in Figure 3 (top) show that the magnitude of the autocorrelations increases with concentration, and the normalized autocorrelations in Figure 3 (bottom) show that the shape and width of the autocotrelation functions change very little with concentration, as expected for single molecule detection. The values of the autocorrelation function at t = 0 and t = m should be linear with c~ncentration,'~ and this is observed within the uncertainties of the concentrations for these very dilute DNA samples. The transit time, defined by the e-2 width, can be determined by fitting a Gaussian to the autocorrelation function. The autocorrelation of a Gaussian yields a half-Gaussian with a standard deviation that is increased by a factor of (2)'12, so the transit time is given by 40/(2)~/~,where u is the standard deviation of the Gaussian fit. The data fit approximately to a Gaussian with (35) Petersen, N. 0. Biophys. J. 1986,49, 809-815.
3256 Analytical Chemisfry, Vol. 67, No. 18, September 15, 1995
0.40
0.42
0.44
0.46
0.48
Time (seconds)
Figure 4. Data illustrating the digital filtering procedure. Representative M13 fluorescence burst data from the 20 pM run are shown, followed by the same data after applying a 670-, 330-, and 130-Hz low-pass filter. The 330-Hz filter was chosen to analyze the full data set because it provided the greatest high-frequency filtration without significantly broadening the bursts.
an average standard deviation of 7.1, which yields a mean transit time of 7.3 ms. The beam was focused 10-15 pm beneath the top wall of the capillary, where, based on the average linear velocity from the syringe pump of 440 pm/s and assuming a parabolic flow profile, the flow velocity was calculated to be 300 pm/s. This velocity estimate allows us to calculate an effective e-2beam diameter of 0.0073 s x 300 pm/s = 2 pm. This observed beam diameter was larger than that calculated from diffractionlimited optics but agreed well with that determined in subsequent electrophoresis experiments (see below). The increased size of the beam is presumably due to defocusing of the laser by refractive index mismatches at the oil-quartz-water interfaces36and by the cylindrical lens formed by the capillary. Single Molecule Counting. To count single molecules, it is first necessary to digitally filter the data to remove high-frequency noise, making the single molecule bursts easily identifiable by a peak-calling algorithm. Figure 4 demonstrates this digital filtration. As described in the Experimental Section, the data were Fourier transformed, low-pass filtered, and reverse Fourier transformed. As the frequency cutoff is lowered, the data become smoother, but after a certain point, the underlying peak shape is distorted. From the data in Figure 4, a 330-Hz cutoff was determined to be best since it removed most of the high-frequency noise without changing the underlying peak shape. Previous workers have enhanced the S/N in single molecule burst detection by applying a weighted quadratic summing filter14,37 or a weighted sum algorithm.38 The weighted quadratic summing filter squares the data and then convolutes the data with a weighting function, which smoothes the data and enhances (36) Hell, S.; Reiner, G.; Cremer, C.; Stelzer, E. H. IC]. Microsc. (Oxford) 1993, 169, 391-405. (37) Soper, S. A,; Davis, L. M.; Shera, E. B. J. Opt. SOC.Am. B 1992,9, 17611769. (38) Castro. A,: Fairfield, F. R.; Shera, E. B. Anal. Chem. 1993,65,849-852.
0
1103 m
f
lo2
(c
;
10
1 1 0
20
40
60 80 Peak Height
100
\20
Figure 5. Distributions of fluorescence burst heights from the M i 3 DNA dilution series after applying a 330-Hz low-pass filter. A histogram of the number of peaks versus peak height for the blank (A),2 ( A ) , 10 (O),and 20 pM (0)data is presented.
peaks but also has the effect of greatly heightening and broadening the data plus biasing peaks toward the shape of the weighting function. We have chosen to use a Fourier filter in our applications since it maintains the height, width, and underlying shape of the peak. Information on the number of photons per peak is retained, and peaks which are close together are still resolvable after filtration, which often is not the case in our experience with the weighted quadratic summing filter. For the use of the Fourier filter, one does need to know the approximate duration of the bursts so that a proper low-pass cutoff can be set; this information can be obtained from the autocorrelation function. It is observed empirically that the best low-pass cutoff is at a frequency 1.5-3 times that of the reciprocal burst duration. For example, if a burst lasts 8.0 ms, its reciprocal duration is 1/0.0080 s = 120 Hz, and the optimum low-pass cutoff is from 180 to 360 Hz. The application of the filter to model data showed that cutoff frequencies more than 1.5 times that of the reciprocal burst duration broaden’the burst by no more than 2%, while cutoff frequencies more than 3 times that of the reciprocal burst duration leave high-frequency noise on top of a burst. We have applied this criterion in the analysis of all subsequent data sets presented here. After filtration, peaks were identified by an algorithm which searched for maxima in the data set having immediately flanking points at a lower value. The requirement for two flanking points at a lower value was used to reject small fluctuations which may be left over after filtration, but this criterion is not so restrictive that it rejects actual peaks that are close together. The final parameter to be set before performing single molecule counting is the discriminator level (D), which is the level above which peaks are accepted and below which they are rejected. D should be set to maximize detection of true single molecule events while minimizing false positives from random fluctuations and noise in the background. To determine the proper value for D,it is useful to plot a histogram of the number of peaks counted versus peak height. For the M13 DNA dilution series data, such a plot is presented in Figure 5 . This plot shows that the blank falls as a perfect Poisson distribution about its average (0.38) until about a height of 5, where a shoulder occurs, indicating a small amount of fluorescent impurities in the blank. The distributions of the higher concentrations of M13 DNA show increasing numbers of peaks at each peak height. The wide
0
20
40 60 80 100 D l r c r i m l n a t o r Level (D)
120
140
Figure 6. Plots of the number of
peaks counted versus discriminator level (0) for the blank (A),2 (a),10 (@), and 20 pM (0)M13 DNA dilution series data. The dashed line indicates D = 6. distribution of burst heights is probably due to the fact that the spatial intensity distribution of the laser beam is nonuniform, and the burst height is then dependent on where a molecule travels through the beam.39 From this plot, the optimum D will presumably lie between 5 and 10, the lowest discriminator setting that distinguishes a single molecule burst from fluctuations in the blank. The effect of a change in D on the count rate can be visualized more clearly by plotting the number of peaks detected versus discriminator level. Figure 6 presents such a plot for the M13 dilution series data. In the blank, 32 molecules were detected when D = 5, and 15 were detected when D = 7, well above the numbers expected by the Poisson distribution, once again indicating the presence of some chemical noise.40 As D is raised from 0 to 5, the counts from the DNA solutions fall rapidly but reach a semiplateau extending from level -6 on out to about -80. D should be set in this “single molecule counting plateau” region because the number of peaks counted is then a less sensitive function of D, just as one would select a single photon counting plateau. As a consistency check, we calculated the number of molecules expected and compared this result to the data in Figure 6. For a 2.4 x 4.0;um cross section probe area, a flow rate of 0.030 cm/s, a collection time of 108 s, and a concentration of 2.0 pM, 370 molecules are expected if we neglect uncertainties such as wall adsorption and false positives. For the 2 pM solution, this count was found for a value of D between 5 and 6, coniirming that this region at the edge of the single molecule counting plateau gives us the proper choice for D. Thus, a D value of 6 was adopted for subsequent analysis of the dilution series data. The peak-calling algorithm was then applied to the data set using the defined discriminator level of 6. Figure 7 presents a representative portion of the filtered 10 pM dilution series data, where the circles indicate peaks counted at D = 6. The background fluctuations are well below this level, and nearly all of the obvious peaks have been identified, demonstrating the effectiveness of this peak selection method. The relationship between molecules counted and concentration using various discriminator levels is examined in Figure 8. In (39) Mets, U.; Rigler, R. J, Fluoresc. 1994,4, 259. (40) If the average background count rate is 0.38 and the fluctuation in counts is purely random, the Poisson probability of a count of 5 or greater is 4.8 x or 14 times per 300 W p o i n t data set, and the probability of a count of 7 or greater is 1.6 x lo-’, or less than once per data set of 300 OOO points.
Analytical Chemistry, Vol. 67, No. 18, September 15, 1995
3257
80
I
iI!
-
N In 0
I
I I
I
I
,
k40
20
Q 0
11170
11.85
11.75
11.’85 (seconds)
11.80
Tlme
11.90
11.95
Figure 7. Representative filtered M13 fluorescence burst data from the 10 pM flow experiment. The circles indicate peaks selected by the peak counting algorithm. The algorithm searched for maxima above a defined discriminator level, set at 6 for this analysis (dashed line). Proper peaks were also required to have immediately flanking points at a lower value.
I o4 P
1
E
I
Single Molecule Counting
1 5 1
+
300 Time (semnds)
350
~~~~
t.........
..................
1 Concent rat ion
I
1
10
I
100
(pM)
Figure 8. Examination of the linearity of the single molecule detection system at various discriminator levels. The plot displays the numbers of molecules counted versus concentration at discriminator levels of 3 (0),6 (O), and 15 (A ). The number of “molecules” counted in the blank at each discriminator level has been subtracted. The slope of the D = 3 least-squares fit is 0.994 with R = 0.971; the slope of the D = 6 fit is 1.01 with R = 0.995; and the slope of the D = 15 fit is 1.12 with R = 0.999.
single molecule counting, the number of molecules counted versus concentration should be linear with a log/log slope of 1. If, for example, the simultaneous presence of two or more molecules in the beam were the source of the observed bursts, the plots would have a slope of 2. At each discriminator level, the slope of the least-squares fit to the data is very nearly 1, with good linearity, which confirms single molecule detection and proper peak identification. The D = 3 data have the smallest correlation coefficient due to the large contribution of background fluctuations at this low discriminator level. In addition to background fluctuations, uncertainty in the concentration is probably the largest source of error in all the data sets. There is no systematic change in slope with discriminator level, indicating that the distribution of burst heights is nearly the same for all the concentrations. 3258
d
4
250
I 0.1
PI
+-.
:::::::::::::
............................... ..................
g-
Analytical Chemistry, Vol. 67, No. 18, September 15, 1995
Figure 9. Capillary gel electrophoresis separation of M13 DNA detected by single molecule counting. D N A was electrophoresed at 200 V/cm in 0.10% HEC with 10 nM T06. The injected D N A concentration was 10 pg/pL, and the bin width was 380 ps. (A) A 0.6-s region of raw data from the center of the M13 band, showing individual, well-resolved fluorescence bursts. (B) A 12-s region from the center of the same band. (C) A 200-s region of the raw data, showing the entire M13 band. (D) Single molecule counting applied to the electrophoresis data. A 500-Hz filter was applied to the data, and a discriminator level of 10 was chosen on the basis of the peak height distribution function. A value of 1 indicates a molecule counted. (E) Frequency of molecules counted versus time. Counts from the above trace were added in a sliding window of 13 000 bins, or 4.9 s. (F) Analog detection of M13 DNA (100 pg/pL) separated under the same conditions as above.
Single Molecule Detection of Capillary Gel Electrophoresis. With these methods developed and optimized, we were ready to apply the single molecule counting system to capillary gel electrophoresis. Figure 9 presents a study of the electrophoretic separation of M13 DNA, A large number of bursts appeared about 300 s after injection, as indicated in Figure 9C. Figure 9B presents an expansion of an g s portion of the band, and Figure 9A presents a 0 6 s region, showing that the fluorescence bursts from the middle of the band are clear and discreet. At this injected DNA concentration (10 pg/pL, or 2 pM), the calculated occupancy is only -0.006, a strong indication that bursts arise primarily from single molecules. The autocorrelation function from this data set indicated a transit time of 2.7 ms, in close agreement with the 2.9 ms value expected for molecules having a velocity of 830 pm/s (25 cm/300 s) through a 2.4-pm beam waist. Since the velocity profile in the capillary is flat and the velocity of the DNA is precisely known, this number for the beam waist was deemed to be more accurate than that determined from the flow pump experiments, where the velocity profile is parabolic. The single molecule counting method was then applied to these data. A 500Hz filter was used, and since the background was slightly higher
than that in the flow experiments due to the presence of HEC,D was set at 10. A digital plot indicating the counted molecules versus time is presented in Figure 9D. A total of 489 molecules were counted from 285 to 340 s. To compare this detection method with standard analog detection, the frequency of the counted molecules was calculated. The molecular counts were summed over a 4.9s (13 000 bin) sliding window to produce a plot of count rate versus time, presented in Figure 9E. The resulting DNA band has approximately the same shape as the analog band, presented in Figure 9F, confirming that this method has successfully detected the component M13 DNA molecules after an electrophoretic separation. Interestingly, even some of the fine structure of the analog band is preserved in the single molecule frequency plot. For example, the small shoulder on the right portion of the band and the broad tail at the end appear in both. This molecular level resolution of small features in a DNA separation suggests that this technique could be used to examine the chemical origins of the differences in the migration rates of these molecules. The true power of electrophoresis is that it can separate molecules on the basis of differing size. To demonstrate the combination of this capability with single molecule detection, the single molecule counting system was applied to the separation of pBR 322 DNA (4631 bp) from pRL 277 (6798 bp) DNA, Figure 10 presents the results of this experiment. Two distinct clusters of fluorescent bursts were observed about 790 and 810 s after injection. The top panels in Figure 10 show that the fluorescence bursts from the middle of each of these bands are discreet and well resolved. The calculated occupancy, assuming that the band has the same concentration as the injected solution, is only about 0.01. The transit times calculated from the autocorrelation function were 7.4 ms for the first band and 7.8 ms for the second band. The velocities of 320 pm/s (25 cm/790 s) for the first band and 310 pm/s (25 cm/810 s) for the second band give a 2.4pm beam waist, in agreement with the width determined in the M13 electrophoresis experiment. Based on the relative mobilities, we assign the first band to pBR 322 and the second band to pRL 277. The single molecule counting method was applied to the data using a 4WHz filter and a discriminator level of 12, and two separate clusters of molecular counts were observed. Fifty-two molecules were counted in the pBR 322 band, and 106 molecules were counted in the pRL 277 band. When the digital count rate is converted to a frequency, the frequency profiles in Figure 10D are very similar in shape to the analog bands, as expected. Also, the small features of the analog bands are again visible in the frequency plot, such as the small bump at the end of the pRL 277 band. The variation in appearance times of the bands between the analog run and the single molecule counting run is due to natural mobility differences seen after the capillary is refilled with buffer.41 We have demonstrated that single molecule fluorescence burst counting can be used as an ultrasensitive detection method for electrophoretic separations. This accomplishment has been based on two fundamental improvements in the detection technology. First, it was important to optimize the S/N by binding large numbers of highquantum-yield dye molecules to the DNA and to minimize the fluorescence and scatter from the gel matrix. We have exploited the optimization of intercalation dye technology developed by Glazer and co-workers to select a thiazole orange (41) Clark, S. M.; Mathies, R. A. And. Biochem. 1993,215, 163-170.
40
I
A
"I
I
B
lOO-1
I
7io
7bo
aio
ado
760
aio
si0
4
pRL 2 7 7
800
820
Time
840
ad0
(rcondr)
Figure I O . Capillary gel electrophoresis separation of two different kinds of DNA detected by single molecule counting. pBR 322 DNA (4.6 kbp) was separated from pRL 277 DNA (6.7 kbp) in 0.50% HEC with 10 nM TO6 at a field strength of 150 V/cm. The injected sample concentration was 50 pglpL, and the bin width was 450 ps. (A) Expansion of a 0.6-s region of raw data from the center of each DNA band, showing individual, well-resolved fluorescence bursts. (B) An 80-sregion of the raw data, showing the fluorescence bursts in both DNA bands. (C) Single molecule counting applied to the electrophoresis data, using the methods described above, with a 400-Hz filter and a discriminator level of 12. A value of 1 indicates a molecule counted. (D) Frequency of molecules counted versus time. Molecules counted were added in a sliding window of 5000 bins, or 2.25 s. (E) Analog detection of the pBR 322 and pRL 277 separation using the same conditions as above.
derivative ("06) as the most sensitive of the mono- and bisintercalator~.~~ Second, we have constructed a 0 dead-time photon counting detector with the capability of recording the number of counts received in arbitrary length microsecond or longer bins. This has permitted full temporal resolution of the bursts and optimization of both the experiment and data analysis. Despite these excellent results, a number of experimental and technical improvements are desirable. The first issue is the efficiency of detection, or the fraction of molecules traveling through the capillary which get detected. For this discussion, the detection efficiency will be considered to be given by the fraction of the flowing stream or electrophoretic band that actually passes through the illumination beam defined by the e+ intensity profile. This definition makes the implicit assumption that the molecules that pass through the beam are all counted. Our present configuration has been useful to demonstrate the feasibility of single molecule detection in capillary electrophoresis, but the detection efficiency is very low (-0.1%). This low efficiency reduces the overall on-column mass sensitivity. For a detection limit of S/N = 3, our current configuration would require a total of -10 OOO M13 molecules in a band. which is about the same as Analytical Chemistry, Vol. 67, No. 18,September 15, 1995
3259
our analog detection limits.26Lee et al.17 have addressed the issue of detection efficiency by confining the sample to a 1@x 10-pm square capillary and focusing the laser over the entire inner bore of the capillary. Using this method, they achieved near unity detection efficiency of an infrared dye in flowing methanol. A similar approach could be taken in capillary electrophoresis by performing separations in micromachined channels. DNA separations have recently been demonstrated in micromachined channels by Woolley and mat hie^.^^ Using our probe beam cross section of 2.4 pm x 4.0 pm and detection channels of 20 pm x 6 pm, the detection efficiency would be -8%, and as few as 125 molecules on-column would constitute a detectable band (S/N = 3). Increased efficiency could be achieved by increasing the beam cross section and further decreasing the channel dimensions. Further improvement in the S/N is also desired in order to allow detection of smaller DNA fragments. Many applications of interest would require detection of DNA that is 100 bp or less in length. Since the number of fluorescent dyes intercalated into the DNA is proportional to the length, the fluorescence signal strength is also proportional to the length of the DNA43 Keller and co-workers have used the fluorescence burst intensity as a measure of DNA size and have reached a lower size limit of about 1500 bp.44 Based on preliminary experiments with shorter DNA samples and extrapolation of the burst height distributions reported here, the detection size limit in our current configuration is around 1000 bp, which means that an order of magnitude improvement in S/N is desirable. This improvement may be achieved through the use of heterodimeric energy-transfer intercalating dyes. Dyes such as TOTAJ3 and its derivatives45 have the advantage of extremely high binding afiinities and excellent spectral characteristics. A high binding afiinity means that less dye is required in the running buffer, thus lowering background, ~~~
~~
~
(42) Woolley,A. T.; Mathies, R A. Proc. Nutl. Acud. Sci. U S A . 1994,91,1134811352. (43) Glazer, A. N.; Peck, K; Mathies, R A. Proc. Nutl. Acad. Sci. U.S.A. 1990, 87, 3851-3855. (44) 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. (45) Benson, S. C.; Singh, P.; Glazer, A. N. Nucleic Acids Res. 1993,21,57275735.
3260 Analytical Chemistry, Vol. 67, No. 18, September 15, 1995
and a high extinction coefficient and fluorescence quantum yield mean that the signal strength from these dyes will be at least as strong as that from T06. Another advantage of energy-transfer intercalators is that the fluorescence of these dyes can be Stokesshifted into the red region of the spectrum, where background is much lower. We anticipate that these and other improvements will yield a substantial increase in the S/N of single DNA molecule counting in the future and a corresponding reduction in the minimum size of detectable single molecules of DNA, thereby enhancing the applicability of this new detection technique. PROSPECTS The successful application of single molecule fluorescence burst detection to capillary electrophoresis opens up new horizons in high-sensitivity DNA detection. One possible application of this technology would be the identification of an incipient cancer or trace viral or bacterial contamination when the amount of target nucleic acid is minuscule. A second application of this technique would be the quantitation of low-abundance species of mRNA Single molecule detection and CGE would enable direct detection of a desired mRNA without the introduction of large uncertainty due to amplification. Standard assays would also benefit from this detection method. For example, the polymerase chain reaction could be performed with fewer cycle numbers of amplification, which would save time, reduce noise, and improve sensitivity. ACKNOWLEDGMENT
We thank the members of the UC Berkeley High Sensitivity DNA Analysis Project for many valuable interactions. This research was supported by the Director, Office of Energy Research, Office of Health and Environmental Research of the US. Department of Energy under Contract DEFG91ER61125. This work is dedicated to the memory of our friend and colleague Dr. Huiping Zhu (1959-1995). Received for review June 5, 1995. 1995.a
Accepted July 7,
AC950542C Abstract published in Advance ACS Abstracts, August 15, 1995.
