Design Considerations for a Conventional Microscope based Microarray-Reader to Improve Sensitivity

Transcription

1 Design Considerations for a Conventional Microscope based Microarray-Reader to Improve Sensitivity L.R. van den Doel, L.J. van Vliet, and I.T. Young Delft Interfaculty Research Center - Intelligent Molecular Diagnostic Systems, Pattern Recognition Group, Faculty of Applied Sciences, Delft University of Technology, Lorentzweg 1, NL-2628 CJ, Delft, The Netherlands, phone: , fax: {L.R.vandenDoel,L.J.vanVliet,I.T.Young}@ph.tn.tudelft.nl Abstract. An unmodified wide-field fluorescence microscope system equipped with a scientific grade CCD camera can be used to measure fluorescent signals in subnanoliter vials. The lowest detectable fluorescence signal in a 0.8nl vial of µm 3 corresponds to a concentration of fluorescent molecules on the order of a nanomolar (10 9 M) for a dye such as Rhodamine. This paper will explain that the lowest detectable concentration is limited by one of the following instrumental factors: stray light, originating in the excitation path of the microscope system or readout noise in the photon detection part of the system. This paper will present a number of modifications to our wide-field microscope system to improve its performance in terms of the detection limit, under assumption that readout noise is the limiting factor. 1 Introduction In high-throughput screening applications many samples are tested against a large variety of targets in microarrays. In our TU Delft interfaculty research center we are developing microarrays fabricated in silicon. These microarrays consist of a number of vials with typical dimensions of µm 3, which translates to a volume of 0.8nl. The biochemical reactions in the vials generate fluorescent signals, which are a measure for the concentration of the substrate in the well, from which enzyme activity levels can be derived. We have been using a conventional microscope system as a microarray reader. From past experiments we have learned that the detection limit of our unmodified wide-field fluorescence microscope system equipped with a scientific grade CCD camera is on the order of nanomolar (10 9 M) concentrations for a dye such as Rhodamine. The three dimensional volume of a vial with dimensions of µm 3 is projected onto a two-dimensional sensor. For a vial with an area of µm 2 this projection translates to 12 detectable fluorescing Rhodamine molecules per square micrometer of the vial. In this paper we will present some modifications to our microscope system to improve its performance in terms of sensitivity. The detection limit is determined as follows: a microarray is filled with solutions with varying concentrations of Rhodamine. The fluorescence levels of the Rhodamine are measured with a conventional microscope system equipped with a scientific grade CCD camera. The detection limit of the complete detection system is defined as the lowest concentration of Rhodamine that can be discriminated from a blanco solution. The fabrication process of the microarrays, the filling procedure of the vials, and the details of the microscope and detection system are described in [1]. 1.1 Limiting Factors of the Instrument Reading the microarray, i.e. measuring fluorescence levels from all the vials, is hampered by many unwanted factors from: (a) the sample (biological / biochemical variations), (b) the optical system, and (c) the electronics. The latter two characterize the instrument and are independent of the sample. These variations should be smaller than the biological / biochemical variations. There are two possible limiting factors (noise sources) that determine the detection limit of the system. The first limiting factor of the detection limit is stray light. Stray light originates in the optical (excitation) system by reflections (flare) at air-lens interfaces and scattering at diffuse surfaces within the optical system (glare). Other contributions to stray light are (Raman) scattering in the solvent and autofluorescence in the objective. Furthermore, there are contributions to the fluorescent signal from non-specific (auto-)fluorescence of the solvent, or ingredients. The second limiting factor is at the detection side of the microscope system: readout noise. Readout noise originates in the preamplifier of the CCD camera. The readout noise is a lower bound of the camera noise. Proceedings of the Second Euroconference on Quantitative Molecular Cytogenetics, Salamanca, Spain, April 2001, pp

2 1.2 Detection Limit of the Microarray Reader We will look in closer at fluorescence measurements to determine the detection limit. Figure 1(a) shows the response of a detection system as a function of the concentration of the fluorescent particles. The dashed lines indicate the distribution of the blanco solution. The solid lines indicate the distributions of the different fluorescent solutions. These distributions imply that the blanco solution as well as the fluorescent solutions are measured with a certain precision and accuracy. Both precision and accuracy are defined by the system. The accuracy of the fluorescent solutions equals the accuracy of the blanco solution. This means that a possible bias can be removed by subtracting the average blanco signal from the fluorescent signals. We assume in that the probability density functions of the blanco solution and the fluorescent solutions can be estimated extrinsically during a calibration procedure, i.e. each concentration can be independently measured many times on the same microarray (built-in redundancy). The results of this calibration procedure are a probability density function of the instrument s response to a blanco solution and probability density functions of the instrument s responses to various fluorescent solutions. Figure 1(b) shows the definition of the detection limit. When we measure the fluorescent signal of a solution in a vial, we need to estimate the underlying concentration of fluorescent particles. We define the minimum detectable signal such that 97.5% of all blanco solutions will be classified as blanco and only 2.5% will be erroneously treated as originating from the fluorescence process to be measured. As a consequence, the probability that a fluorescent solution corresponding to the detection limit will produce a signal larger than the minimum detectable signal is 50%. In other words, which concentration (and corresponding expected flux λ) has the highest probability of producing this measurement. For a single measurements this amounts to assigning the concentration according to the estimated calibration curve. In statistical terms, for one measurement as well as for many measurements, we select the molecular concentration that maximizes the a posteriori probability that a fluorescent solution λ generates this response: argmax λ P (λ X) = N i=1 p λ(x i ), assuming equal a priori probabilities. Note that averaging measurements results in a probability density function with a variance (of the sample mean) that is inversely proportional to the number of measurements.) For an infinite number of measurements, the sample mean equals the expectation value of the probability density function. As a conclusion, repeating the number of measurements will result in a lower detection limit. In theory, it is possible to distinguish between a blanco solution and a solution with a single fluorescent particle under the assumption that a very large number ( ) of measurements can be performed. Note that an exact measurement does not contribute to an exact interpretation of the underlying biology / biochemistry, since they will still be limited by the biological variation. 1.3 Improving the Detection Limit (for 1 and N measurements!) From Figure 1(a) the following possibilities to improve the detection limit can be derived. 1. Reduce the variation of measuring the blanco solution. The instrumental variation of the blanco has contributions of stray light (i.e. scatter, autofluorescence, reflections, and non-specific fluorescence) and detector noise (i.e. dark current, and readout noise). Contributions of stray-light can be minimized by closing the fieldstop as far as possible. 2. Reduce the variation of measuring the fluorescent solutions. The instrumental variation of the fluorescent solutions has contributions of stray light and detector noise. A third contribution to the variation of the fluorescent solutions is photon shot noise. This is a Poisson distribution due to the quantum nature of light [3]. A fourth contribution is caused by differences in injected sample volume. Repeating identical biochemical processes in different vials will generate the same Poisson process. Due to differences in volume, and therefore different amounts of ingredients, will result in a different observed emission rate λ. The error in the injected liquid volume in the vials has a coefficient of variation CV = 5.9% [1]. A new method to fill the vials in a more accurate, massive parallel manner, avoiding fast evaporation of the liquid sample is currently being developed and tested. If the differences of liquid volumes could be taken into account, e.g. with a dedicated electronic liquid volume sensor [4], or could be neglected, then at high fluorescent concentrations the limiting factor of the variation is photon shot noise. For low concentrations, the same argument holds for the fluorescence solutions as for the blanco solution. 3. Improve the sensitivity of the detection system. This is achieved by using high power objectives. The area of the field of view is proportional to 1/M 2, where M is the magnification of the objective. As a result, the excitation light intensity per unit area in the field-of-view is proportional to M 2. The detection (emission) efficiency is mostly influenced by the Numerical Aperture of the objective. The light gathering power of an objective in Köhler illumination is proportional to NA 2 /M 2. Under the assumption that there is a linear 109

3 relation between excitation and emission, we can say that the intensity of the emission light is proportional to M 2. Combining these results, it follows that the light gathering power of an objective (with epi-illumination adjusted to Köhler illumination) is proportional to NA 2. In the case of readout noise limited detection, the Quantum Efficiency of the detector becomes of equal importance w.r.t. the readout noise. In the case of stray light limited detection, a better detector does not help, since it will increase the signal level as well as the noise level, leaving the Signal-to-Noise Ratio unchanged. This means that the graph in Figure 1(a) will only be scaled in the vertical direction. In the following sections we will present possible modifications to overcome the limiting factors of our conventional microscope system and improve the performance of the detection system in terms of the detection limit. 2 Readout Noise Limited Detection In Figure 1 it is assumed that the response of the detection system is linearly proportional to the emission power, and that fluorescence is a linear process, i.e. the emission power depends linearly on the excitation power. This assumption is true for low excitation power and low concentrations of fluorescing particles [2]. Furthermore, this figure suggests that the variation of the fluorescence measurements grows linearly with increasing concentrations. This linear behavior is not true, and this figure is only drawn this way for clarity. In the following subsection we will estimate the strength of the minimum detectable signal in the case that readout noise is the limiting factor. In the previous section a definition for the minimum detectable signal was derived. Another statistic of interest is the coefficient-of-variation CV = σ/µ. We will show that the CV can be improved either by repeating the measurement in the same vial (n T ), or by increasing the measurement time (1 nt ). The result of this improvement in CV, however, has no biological / biochemical meaning, since it only implies a better estimate of the biochemical process in that single vial, characterized by the emission rate λ vial in that vial. It does not say anything about the biological variation of the process involved, characterized by the probability density function of λ. 2.1 Ideal Detection System First, we will consider the case that we have an ideal imaging and detection system: there are no external noise sources, the system consists of perfect components, and each emitted photon that enters the detection system is detected. This means true photon counting. The quantum nature of light shows that the number of photons emitted in a fixed time interval T varies. This number obeys a Poisson distribution. Therefore, the exact concentration of fluorescent molecules cannot be measured: the coefficient-of-variation CV = σ/µ = λt /λt = 1/ λt, where λ is the emission rate of the fluorescent solution. If we want to discriminate a fluorescent solution from a blanco solution (λ = 0) with this ideal system within a measurement time of T = 1s, then only a single photon must be detected. If we want to detect this single photon with a probability of 99%, then the probability of not detecting a single photon is 1%. In terms of Poisson statistics this translates to a emission rate λ = 4.61photons/s. This single measurement has a CV = 57%, which is much larger than the biological variation. Repeating this measurement n times (n T ) or increasing the measurement time T to 1 nt, changes the CV : CV = 1/ nλt. With n = 87 the CV of these measurements becomes 5%. Note that either repeating the measurement n times or increasing the measurement time n times amounts to the same improvement of CV due to the absence of detector noise. 2.2 Single Element Detection System with Readout Noise Secondly, we will replace the ideal detector with a more realistic detector: a detector with readout noise. Readout noise originates in the preamplifier of the sensor (photodiode or CCD camera). The readout noise is independent of the signal and the integration time, but it depends on the readout rate. Readout noise is additive Gaussian noise with zero mean. The readout noise of the sensor is expressed in electrons in terms of the RMS error at a certain readout frequency, e.g. the Photometrics Series 200 CCD camera has a readout noise level σ r = 11.7e RMS@500kHz [5]. This translates to the same number of photon equivalents, assuming a Quantum Efficiency of 100%. To avoid that reading out the camera results in a negative number of photons, a constant bias n bias is added to all signals: reading a blanco solution will not result in zero photons, but in a number of photons related to the normal distribution of the readout noise with a constant bias added. The probability density function related to the blanco solution is a gaussian distribution with mean n bias and standard deviation σ r : N(µ = n bias, σ = σ r ). In the case of reading a fluorescent solution, there is first the probability that the fluorescent 110

4 solution generates n fl photons, and secondly the probability that n d photons are detected, given that n fl photons were emitted. Repeating this experiment many times results in a probability density function, that is the convolution of a Poisson distribution (defined by the emission rate λ and the measurement time T ) and the readout noise (defined by the Gaussian distribution of the readout noise). For λt large enough, this convolution results in a normal distribution with mean n bias +λt and standard deviation σ f = σ 2 r + λt : N(µ = n bias +λt, σ = σ 2 r + λt ). In this case, the minimum detectable signal equals kσ r, which translates to an emission rate λ = 35.1photons/s, given k = 3 and T = 1s. This measurements has a CV = σ 2 r + λt /λt = 37.4%. Repeating this experiment n times (n T ) yields CV = 1/ n σ 2 r + λt /λt. With n = 56 measurements follows that CV = 5%. Increasing the measurement time T to 1 nt yields CV = σ 2 r + nλt /nλt. With n = 14.5 follows that CV = 5%. A longer measurement time is superior to repeated measurements in terms of CV. 2.3 Many Pixels Detection System with Readout Noise Thirdly, the point sensor is replaced by a CCD camera, which is built as a 2-D array of photo-sensitive elements. Each element can be characterized as point sensor. In this case, the photon detector is CCD camera built-up of a large number M of pixels. Each and every pixel suffers from readout noise. We assume here that the readout noise is stationary, and ergodic, i.e. the noise of all pixels is independent, identical and constant in time, such that the variances of the readout noise per pixel can be added: σr,m 2 = Mσ2 r. Reading a blanco solution will result in a normal distribution with mean M n bias and standard deviation M σ r : N(µ = M n bias, σ = M σ r ). The fluorescent signal is spread over all pixels. This results in a spatio-temporal Poisson distribution. The variance of the fluorescent signal for all pixels σf,m 2 is given by M M ( σf,m 2 = σf 2 = σr 2 + λt ) = Mσr 2 + M i=1 i=1 M i=1 λt M = Mσ2 r + λt, (1) where λt/m means that the total emission photon flux is spread over M pixels. Reading a fluorescent solution many times results in a normal distribution N(µ = M n bias + λt, σ = Mσr 2 + λt ). For this detector the detection limit (λt ) equals k Mσ r = photons with k = 3, and M = The CV of this measurement yields CV = Mσr 2 + λt /λt = 33%. Repeating this experiment n times gives a CV = 1/ n Mσ 2 r + λt /λt. With n = 45, CV = 5%. Increasing T to 1 nt, yields CV = Mσr 2 + nλt /nλt. With n = 6.7 follows CV = 5%. As a conclusion, measuring low levels of fluorescence requires multiple measurements in order to achieve a sufficient CV. Furthermore, the minimum detectable signal can be lowered by imaging the fluorescent signal from the vial onto as few pixels as possible. Thus, measuring very low levels of fluorescence is at the expense of spatial resolution. Further on in this paper we will show some modifications how this can be achieved. Furthermore, we will estimate the minimum detectable signal in photons for different configurations based on the model presented in this section. 3 Stray Light Limited Detection In the case of stray light limited detection a better detector (higher QE, lower readout noise) does not result in a better detection limit: the detector does not only have a better response to the signal, but also to the noise, i.e. the Signal-to-Noise Ratio will not improve. This means that Fig. 1(a) will only be scaled in the vertical direction and the detection limit will not shift to lower concentrations. The amount of stray light is al ready minimized by using high power objectives, dedicated filter blocks and closing the field stop as far as possible. Another option to gain signal with respect to stray light is to use dyes with a much higher quantum yield. This can also be achieved by using other solvents. To avoid fast evaporation we have used a mixture of glycerol / water in the past. Practically all dyes, however, have a higher quantum yield in an aqueous solution. The new filling procedure uses water as solvent. This will result in a better quantum yield. Stray light can be avoided at all by detecting luminescence instead of fluorescence. With luminescence light is emitted in the absence of excitation light. Furthermore, the limiting factor of stray light can be overcome by exploiting other techniques, such as Fluorescence (Cross-)Correlation Spectroscopy. With this technique the fluorescence signal from a confocal volume is detected as a function of time and the correlation spectrum is computed. This techniques is limited by Raman scattering of the solvent. The dynamic range of these techniques is from 10 9 M down to M. In principle, lower concentrations can be detected with this technique, but the measurement time is inversely proportional to the concentration of the fluorescent dye. 111

5 4 Minimum Detectable Signal for Different Configurations We will now compute the strength of the minimum detectable signal for different configurations in the case of readout noise limited detection. The excitation part of all configurations is the same. This implies that the generated emission photon flux is the same for all configurations. The area of the vials on the microarray is µm 2. From these examples we will predict the detection limit for our new configuration, if readout noise limits the detection. Other important parameters are listed in Table The configuration used for the first experiments consisted of a Zeiss 20 /0.75 FLUAR objective, a 1 camera mount, and the Photometrics camera (no binning). The vial of a microarray is imaged onto pixels. The readout noise per pixel is 11.7e. This translates to a minimum detectable signal of e. Rhodamine emits around 600 nm. The Quantum Efficiency of this camera is approximately 45% at this wavelength. The minimum detectable signal converts then to photons. The CV of this measurement is 33%. 2. The new configuration consists of a 20 /0.75 objective. The tube lens is replaced by an inverted 5 objective. The overall magnification is 4. The Photometrics (no binning) is replaced by the Princeton camera (no binning). The vial will be imaged onto pixels. The readout noise per pixel is 6e. This translates to a minimum detectable signal of e. That is equivalent to photons (QE 70%@600nm) without binning. In theory this will translate to a detection limit which is a factor of 25 lower. The CV of this measurement is 34% 3. Another interesting configuration is the following: the optical configuration remains identical, but instead of the Princeton CCD camera, a photo multiplier tube is mounted on the microscope. The limiting factor for a PMT in photon counting mode is the dark current. The dark current is expressed in Equivalent Noise Input (ENI). A typical range for the ENI is W. This translates to 0.3 to 3 thousand photons per second [6]. For short integration times the minimum detectable signal for a PMT is on the same order of magnitude as for the new configuration. Future experiments must prove whether or not our modified microscope system is operating under the limitation of readout noise, and what the true detection limit is. 5 Modifications to the Microscope System These modifications to our microscope system consist of the following: 1. The amount of stray light is al ready minimized by closing the field stop as far as possible, such that only (a fraction of) a single vial is illuminated. Note, that this has no effect on the excitation power: only that excitation light is blocked, that would not lead to excitation anyway. We have chosen a (relatively) high magnification objective (20 /0.75), such that the excitation power per unit area is large. 2. The high NA objective implies a high collection efficiency. Furthermore, the trinocular tube with the eyepieces and the relay optics (M = 1 ) has been replaced by an inverted 5 /0.25 objective. This enlarges the emission photon flux in photons per unit area by a factor of 25: we are not interested in spatial resolution, only in sensitivity. In principal, instead of a 5 objective, an identical 20 objective could be used. But a very large working distance is necessary to make a sharp image. The 5 objective has a working distance WD = 9.3mm. Placing the infinity-corrected objectives back-to-back creates a proper imaging system. 3. The scientific grade CCD camera with small pixels is replaced by a CCD camera with large pixels. This implies that only one tenth of the photon flux is necessary to get the same increase in signal level per pixel, given that both cameras have the same overall conversion factor from photons to gray level. This gain can be further increased by using on-chip binning. This gain in signal at reduced readout noise level implies a better SNR. 6 Conclusions and Discussion This paper discusses the measurement of absolute fluorescence levels in a conventional microscope system equipped with a scientific grade CCD camera. Two limiting factors for the detection limit can be distinguished: stray light 112

6 originating in the excitation part of the microscope system or readout noise in the photon detection part of the system. This paper showed possible modifications to a conventional microscope system to improve the sensitivity in terms of the detection limit. This paper showed that measuring very low levels of fluorescence is at the expense of spatial resolution. In the case readout noise is the limiting factor, the modifications consisted of a lower overall magnification by replacing the tube lens in the microscope with an inverted 5 objective and replacing the detector with a new CCD camera with large pixels, a lower readout noise level and a higher Quantum Efficiency. With these modifications we aim at a detection limit of a single fluorophore per µm 2. The simple model we have presented in this paper predicts that this limit can be reached, but future experiments must prove whether or not this prediction is true. Acknowledgements This research is supported by the Delft Interfaculty Research Center Intelligent Molecular Diagnostic Systems. References 1. L. van den Doel, M. Vellekoop, P. Sarro et al. Fluorescence detection in (sub-)nanoliter microarrays. In M. Ferrari (editor), Micro- and Nanofabricated Structures and Devices for Biomedical Environmental Applications II, volume 3606 of Proc. SPIE, Progress in Biomedical Optics, pp F. Rost. Quantitative Fluorescence Microscopy. Cambridge University Press, Cambridge, Great Britain, J. C. Mullikin, L. J. van Vliet, H. Netten et al. Methods for ccd camera characterization. In Proceedings of the SPIE Image Acquisition and Scientific Imaging Systems San Jose, volume 2173, pp K. Hjelt, L. van den Doel, W. Lubking et al. Measuring liquid evaporation from micromachined wells. Sensors and Actuators A Physical 85, pp , H. Netten. Automated Image Analysis of FISH-stained Cell Nuclei. Ph.D. thesis, Delft University of Technology, H. Kume (editor). Photomultiplier Tubes, Basics and Applications. Hamamatsu Photonics K.K., second edition, Table 1. Specifications of CCD cameras: the Princeton Versarray 512B Backilluminated CCD camera with Tektronix chip and Photometrics Series 200 CCD camera with Kodak KAF1400 chip. Princeton Versarray 512B Photometrics Series 200 readout noise: 3 pixel size 24 24µm µm 2 electronic gain 4.6e /#ADU 8.009e /#ADU Quantum Efficiency 70%@600nm 45%@600nm Detector response [#ADUs] mean of fluorescent solution kσ blanco mean of blanco (C=0) solution distribution of fluorescent solution distribution of blanco solution Detector response [#ADUs] expected response to fluorescent solution kσ blanco minimum detectable signal mean of blanco (C=0) solution distribution of blanco solution ConcentrationC of dye molecule [M] ConcentrationC of dye molecule [M] detection limit (a) Due to instrumental variations the response to a concentration of fluorescent particles has a certain distribution. (b) The minimum detectable signal equals the response below which 97.5% of the responses the blanco lie. Figure 1. Fluorescence response model. 113