Parallel Imaging Microfluidic Cytometer

Abstract

By adding an additional degree of freedom from multichannel flow, the parallel microfluidic cytometer (PMC) combines some of the best features of flow cytometry (FACS) and microscope-based high-content screening (HCS). The PMC (i) lends itself to fast processing of large numbers of samples, (ii) adds a 1-D imaging capability for intracellular localization assays (HCS), (iii) has a high rare-cell sensitivity and, (iv) has an unusual capability for time-synchronized sampling. An inability to practically handle large sample numbers has restricted applications of conventional flow cytometers and microscopes in combinatorial cell assays, network biology, and drug discovery. The PMC promises to relieve a bottleneck in these previously constrained applications. The PMC may also be a powerful tool for finding rare primary cells in the clinic.

The multichannel architecture of current PMC prototypes allows 384 unique samples for a cell-based screen to be read out in approximately 6–10 minutes, about 30-times the speed of most current FACS systems. In 1-D intracellular imaging, the PMC can obtain protein localization using HCS marker strategies at many times the sample throughput of CCD-based microscopes or CCD-based single-channel flow cytometers. The PMC also permits the signal integration time to be varied over a larger range than is practical in conventional flow cytometers. The signal-to-noise advantages are useful, for example, in counting rare positive cells in the most difficult early stages of genome-wide screening. We review the status of parallel microfluidic cytometry and discuss some of the directions the new technology may take.

1. Introduction

Relatively narrow sets of methods define eras like genomics and proteomics. The instruments used to practice these methods are often badly mismatched to the biological agenda. We argue that such a bottleneck now exists in cell-resolved measurement. The various “omics” have increased the encyclopedia of molecules and interactions to the point where we can practice broad combinatorial experiments in cells. The primary tools for read-out of these experiments remain microscopy, cytometry, arrays, fluorimeters and a handful of biochemical assays.

Because it can quickly produce a statistically significant reading, one of the most important of these tools is the fluorescence-activated flow cytometer (FACS) (Shapiro, 2003; Carey, 2007; Givan, 2001). However in several dimensions FACS is inadequate to the agenda. It is only practical to make measurements on a few variables at a time and at a compromised sample throughput. In contrast, high-content screening (HCS, i.e., automated microscopy) (Taylor et al., 2007; Bullen et al., 2008; Haney et al., 2008; Gough et al., 2007; Lee et al., 2006; Pepperkok et al., 2006; Eggert et al., 2006) is an attempt to add more information content to FACS. Throughput of both FACS and HCS is an issue for readout of combinatorial biology in general, but particularly with live cells. For example, nuclear transcription kinetics, often have a half-time response of 5–10 minutes (Ding et al., 1998). In a live-cell kinetic study it is usually not possible to read a single 96-well HCS plate in this time. Furthermore for either flow cytometry or HCS, fixing cells causes protein reorganization, and many cytokine modifiers can show alternatively agonism or antagonism in a dose-dependent fashion. Therefore, the biology of combinatorial biology such as large RNAi screens or small molecule studies calls out for dose-response curves taken over many concentrations, on live cells, and with time response on the order of several minutes. The current methods remain orders of magnitude mismatched in speed for the real needs of network biology. As an entirely separate axis of comparison the 1-D imaging ability of the PMC is new to high-speed flow cytometry. The addition of 1-D imaging can be thought of continuing the trend of FACS towards higher content, as has been expressed in recent years by adding lasers and color channels to FACS. Adding 1-D imaging is equivalent to adding many color channels however.

Limitations implicit in the architecture of single-channel flow cytometers restrict applications for studying rare cell types and for massively parallel screening. These are, principally, (i) serial sample processing, which is bounded by sample changeover and, (ii) a short (usually microsecond) data acquisition time, which in turn limits signal averaging. Commercial flow cytometers have been demonstrated with positive abundances as low as parts per million. However, depending on available sample and background noise, single-channel machines are generally not seen as practical for screening when the abundance of “positives” is lower than about 1:10,000 or when the total sample is less than 10–50 thousand cells (Shapiro, 2003). In many cases, auto-fluorescence and non-specific markers limit minimum abundances to higher ratios (1:1000, or 1:100). Recent developments in cytometers have explored automated sample loaders to minimize the disadvantage of serial analysis (Edwards et al., 2004), however sample changeover times still remain on the order of a minute for most commercial FACS machines that are in the field.

This chapter will review considerations in adding a high degree of microfluidic parallelism to flow cytometry. Specifically we review results from a prototype parallel flow cytometer (PMC), which was designed with particular attention to the needs of rare-cell counting (McKenna et al., 2009). Rare cell capability (detection of rare positives within a high background of negatives) is the priority for detection in cancer and also, quite generally, at the early stages of genome-wide screening.

2 Background

Flow Cytometry

Flow cytometry is an impressive technology that has been optimized to extraordinary refinement (Shapiro 2003, Carey 2007, Givan 2001). There is also a large body of work on elegant microfluidic manipulations of cells, including sorting and switching of biological cells in single channels and in dispensing of cells into arrayed well devices. Some examples are cited here (Dittrich et al., 2003; Wang et al., 2005; McClain et al., 2001; Fu et al., 2002; Wolff et al., 2003; Gawd, et al., 2004; Emmelkamp et al., 2004; Yi et al., 2006) but the full microfluidics literature is far too extensive to review in this chapter. In part because of the higher systems-level engineering that is required, less work has been undertaken in practical microfluidic cytometry systems for use in the biology laboratory.

High-Content Screening

High content screening (HCS) is frequently done with CCD-based microscopes in open wells (Taylor et al., 2007, Bullen et al., 2008, Haney et al., 2008, Gough, et al., 2007, Lee et al., 2006, Gonzales, R.C and Woods, R.E., 2002. Ding et al. 1998), on spotted slides (Carpenter et al., 2006, Wheeler et al., 2005), or in flow (George et al., 2006, http://www.amnis.com). Even on high-density slides, the state of the art is largely determined by the performance of low-signal scientific CCD cameras. At 1024x1024-pixel image size, the frame rate due to buffering restrictions is either 15 or (conditionally) 30 frames a second. However even much slower rates are often mandated by low signal. Analysis of a single high-density spotted slide may take many hours (Carpenter et al., 2006, Wheeler et al., 2005). Autofocusing and mechanical motions, further limit throughput (accounting for the majority of the time budget on wide-field imaging systems (Taylor et al., 2007). CCD-based imaging flow cytometers are more limited in throughput. Users typically report raw data acquisition (unclassified cells) from such a machine at 100–1000 objects/min.(http://www.amnis.com/applications.asp#link2). The bottom line is high content microscope-based systems for HCS are frequently too slow for scaled-up applications. A second drawback can be that, with full 2-D imaging, data storage rapidly requires terabytes and overflows even large data-storage resources.

Several of the most common high-content assays implemented on microscopes (in 2-D) are (Taylor et al., 2007, Bullen et al., 2008, Haney et al., 2008, Gough, et al., 2007, Lee et al., 2006, Gonzales, R.C and Woods, R.E., 2002. Ding et al. 1998): (a) nuclear translocation (NT). The most common NT assay is NF-kB translocation. NF-kB is a transcription factor that is critical to cellular stress response. The p65 subunit is a sensitive to several known stimulants, for example, by altered interleukin ILa1 or tumor necrosis factor. The translocation to the nucleus is required to induce gene expression. (b) apoptosis. Image-based assays for apoptosis can provide more information than FACS. For example, by determining nucleus size, it is possible to ascertain necrotic or late apoptotic cells. The nucleus is stained and the image algorithm determines shape and size relative to the cell dimensions. (c) target activation. A very wide class of assays measure localization and total intensity from GFP fusions or other fluorescent markers. Cell cycle, receptor internalization, or drug resistance are commonly measured. (d) co-localization of markers. Co-localization is highly informative about biological mechanism. This is enormous area of active research particularly in the field of biological development. Imaging information is highly useful. (e) intracellular trafficking. Several microscope-based assays track the intracellular migration of molecules by programmed endocytosis. Amnis Inc. has introduced an assay where the antibody CD20 is monitored and correlated with markers for endosomes and lysosomes. (f) morphology. The most obvious markers for phenotype are cell shape and area, however more subtle rearrangements of the cytoskeleton and location of organelles are also often used in microscope assays. (g) cell cycle. The progression of cell cycle is widely used in screening cancer therapies. The phase of individual cells is correlated with markers for specific proteins. Measurements are often also made on the dimensions or total DNA of the nucleus.

High-Content Screening Instruments

Several commercial 2-D HCS instruments are: (i)-CCD/automated microscopes (Thermo Scientific – Cellomics ArrayScan^TM, GE Healthcare- inCell^TM, PerkinElmer- EvoTech Opera^TM, Molecular Devices IsoCyte^TM); (ii)-TDI CCD/flow cytometer (Amnis ImageStream^TM); (iii)--low-res laser scanners (CompuCyte iColor^TM, Acumen- Explorer^TM and Cyntellect, LEAP). These systems generally achieve assay rates of about 2–6 wells/min. for real HCS assays. The Amnis ImageStream is unique as a CCD-based flow imaging system. However it is a single-channel instrument. The laser scanning instruments (CompuCyte iColor^TM, Acumen- Explorer^TM, and Cyntellect, LEAP) are not flow-based.

3 Instrument Design

The design of a PMC differs from a FACS in (1) its need for a wide field of view detector (rather than focused point detector), (2) its need for automation to support parallel sample transfer, (3) its differing needs for data processing and (4) the design of the microfluidic itself. The microfluidic, when all fabrication and flow considerations are taken into account, becomes a big opportunity for widely novel design. One specific consideration is how to rethink flow focusing in order to make best use of the small-sample capability of microfabricated devices. The detector becomes more complex than a FACS since the wide-field requirement more or less mandates a scanner (arguments below). However once the additional mechanical complexity of scanning is accepted, there is a large and important freedom in signal-collection strategies. This is also what permits high-speed imaging. We discuss these design aspects below.

3.1 PMC System Architecture

A prototype automated PMC is shown in Figs. 1 and 2. The microfluidic flow devices are mounted on a top plate and are serviced with a gantry robot combined with a sample elevator that handles 384-well microtiter plates. The fluid handling is via an integrated 96-tip pipettor that permits automated maintenance of 384-well plates on the temperature-controlled base. The sample deck includes positions for nutrient/wash trays that can also be accessed by the pipettor. As a result, live-cell cultures can be sustained for several days on the system or can be loaded from off-system culture apparatus. All 384 channels can be loaded from a microtiter plate in <30 seconds. Flow is actuated by suction using syringe pumps. The optical detector is a photomultiplier-based rotary scanner located under the microfluidics (see below). The system is operated from a graphical user interface that displays data during real time. However data reduction is done on exported files off line (see below).

Figure 1.

(a) Parallel microfluidic cytometer (PMC) for cell-based assays. The system is designed for automated fluorescence measurements on 384-channel microfluidic plates and comprises up to two temperature-controlled microfluidic “chips” (16 to 384 channels each), a scanning detector, and automated pipettor/sample elevator for automated maintenance of cell suspensions/cultures. Cell suspensions are pulled by vacuum suction from injection wells using a positive-displacement syringe pump. Multicolor detection is via a scanned confocal detector that oscillates below the microfluidics. ((Modified from (El-Difrawy 2005) with permission. Copyright 2005, American Institute of Physics.))

Figure 2.

(a) Plan-view detail of the PMC showing microfluidics, sample and wash plates and the optical scanner located beneath the fluidics; (b) A 384-channel microfluidic plate and (c) a segment of data collected from several channels. A short time sequence from one of four photomultipliers is shown with each pixel representing 35 μm in the (horizontal) scan direction. Data is collected at a rate of 3 scans/s, (0.33 s vertical displacement of each row of pixels in the data image); 240 s of data shown. Signal amplitude is shown in RGB color scale with blue representing low signal and red high signal. [From (McKenna et al. 2009). Reproduced with permission of the Royal Society of Chemistry.]

3.2 Robotics

In the early prototype of a PMC system it was thought to be important to include competent robotics. The stages of the X-Y-Z axis (GL16S with 1250 mm of travel for the x-axis, a KR46 with 540 mm of travel for the Y axis, and KR26 with 65 mm of travel for the Z-axis T.H.K. Ltd, Tokyo, Japan) are driven by a set of three servo drivers (SGDH-01AE, Yaskawa Electric, Japan) and are coordinated by a six-axis motion controller (Model 6k6 Compumotor, Rohnert Park, CA). The pipettor head has programmable suction and injection capability for volumes between 2–20 μl, and is driven by a DC brushless servo amplifier (Model 503, Copley Controls Corp, Westwood, MA). The 96-tip robot head accesses water and buffer reservoirs, an ultra sound washing station, a microtiter plate elevator with up to 32 sample plates for continuous operations (Packard Instrument Co, Meriden, CT).

3.3 PMC Detector

The requirements for a PMC detector are meso-scale sensor field, high time response, and variable integration time. We also favored out-of-plane (confocal) background-light rejection. The combined factors led us to strongly favor photomultiplier or avalanche photodiode detection (rather than a CCD or CMOS imager). We therefore modified a PMT fluorescent scanner that we had developed for DNA sequencing (El-Difrawy et al., 2005) and applied it to the new application.

A 100-mW multi-line argon-ion laser beam (Melles Griot #532-MA-A04, Melles Griot, Carlsbad, CA) is passed through a rotating head that moves a 0.5-numerical-aperture (NA) aspheric lens (# 350240, Thor Labs, Corp., Newton NJ) and is driven by a DC brushless motor (BEI# DIP20-17-0027, BEI Kimco, Vista, CA). The laser focus is adjusted to NA ~0.01 and excites fluorescence as the rotating head moves under the detection window. The fluorescence is collected at NA 0.5 through the rotating head (Fig. 2), is separated into four wavelength bands using dichroics and bandpass filters and distributed onto four PMTs (H957-8 Hamamatsu, Bridgewater,NJ).

Non-uniform velocity and cycle-to-cycle fluctuations cause variations in the level of the collected signal and results in added noise. Therefore we have built a PID controller that maintains the speed profile of the rotating head. The parameters of the PID controller are tuned to minimize speed fluctuations during the collection arc while maximizing accelerations outside the arc. The position of the rotating head is measured using an optical encoder (HT30P156 D14 N4096, Dynamic Research Corporation. Wilmington, MA.) A proportional-integral-derivative (PID) controller maintains the speed profile and is tuned to minimize scan-to-scan speed fluctuations while maximizing accelerations outside the collection arc.

The rotating head is programmed to a saw-tooth velocity profile at 12 or 3 Hz recording 300- 2,800 data points across a window of ~3.5–80 mm, or in a modulated circular motion. For each encoder point the values of the 4 PMTs are recorded and the 16-bit digital value is saved to the PC hard drive 12(3) times a second.

The constancy and reproducibility of the speed profile have been measured and show a standard deviation of less than 1% from the target velocity (10,000 scans, all flow channels). The sensitivity of the system was evaluated with fluorescence standards and shown to have a 10pM (fluorescein) detection limit in the 60-μm-deep channel, which is near the state of the art for on-column LIF detectors.

3.4 Data Processing

The data acquisition and PID control are synchronized using a digital signal processor (DSP, ADSP-2181, Analog Devices, Norwood, MA) running at a clock rate of 33 MHz. The PID controller runs at a servo rate of 1 kHz while the data acquisition is performed at a faster rate of 200 kHz. The collected sequencing data is uploaded from the DSP memory to the PC through the PC parallel port, which runs in the enhanced parallel port (EPP) mode. A simple control circuit is used to provide the PC with direct memory access to the program and data memory of the DSP processor.

Raw PMT data is saved in 4 files representing 16-bit data at each location separated by 10μm in the scan window. Each file contains the data for one PMT. The data is first processed to eliminate scan areas without cells. The 2800-bit (80-mm scan) by n (scan) data is then segmented into channel-specific sections of ~20 bits by n scans. This data is then reformatted into 512 x 512 x 16 bit grayscale images and saved as TIFF images for each PMT channel. A calculated image is then created by subtracting the red channel from green (negative values are set to zero). For the initial levels of the screen the TIFF images for the virtual channel are then reviewed and all “bright” cells (those having at least one pixel greater than 10,000 RFUs) counted.

3.5 Microfluidics

Microdevices with 16, 32 and 384 channels (Figs. 5–8) were fabricated in aluminasilicate glass (Corning, Eagle^TM). Unaligned single-mask contact lithography was followed by high-temperature fusion bonding of 0.7-mm-thick plates of 25x50 and 25x25-cm size. (Aborn et al., 2005, Goedecke et al., 2004). The microchannels had a hemispherical cross section with a radius of 60 μm and converged to a density of 5-channels/mm in the scan zone. In the network layouts, channels (20–40 mm length, 1–3μl internal volume) were matched to a few percent in flow resistance. Access for introduction of the cell suspension was through laser-drilled ports, which were conical in shape and terminated at the flow channel with an exit of ~80–100μm diameter. Composite G-10 fiberglass boards were mechanically machined with 2-mm-diameter sample wells distributed on 4.5-mm or 9.0-mm centers, and were glued with thermally curing epoxy on top of the bonded glass devices.

Figure 5.

A plate of 32-channel PMC microdevices at the lithography stage of fabrication. Five devices are fabricated simultaneously on a 250x250mm alumina silicate glass plate. There are economies of scale from batch fabrication -particularly yield improvements at bonding stage. As a last step individual devices are separated by diamond sawing.

Figure 8.

A 384-channel PMC microdevice plate at mask stage, finished device shown in Fig. 2(b). The flow channels fan out on the “loading” (top) end to allow room for the sample-well array that must match the 4.5-mm spacing of the robotic pipettor. At the “scan” end the flow channels converge to a maximum density allowed by the bonding process, 5 channels per mm. The channel cross-section is hemispherical, 60-μm radius. This channel structure is etched into the glass plate (flat-panel display glass), the access holes are laser drilled, conical shape terminating with a 80-μm diameter at the etched channel, then the plate is sealed by high-temperature fusion bonding.

3.6 Flow and Flow Focusing

Microfluidic systems, created by lithographic methods, are generally constrained as two-dimensional (X,Y) flow networks. One-dimensional squeezing, in the plane of the flow network, is relatively easy to accomplish simply by using T-junctions. However “vertical” hydrodynamic focusing (in the plane perpendicular to the network) is more germane for narrow-depth-of-field optical detection of the PMC. In order to focus microfluidic flows vertically, it is necessary to utilize a torque (out of the plane of the network) or to merge flows as vertically distinct layers. From a fabrication standpoint, the geometry in which layers are introduced by intersecting two vertically displaced channels is easiest; this approach requires only a simple unaligned (or weakly aligned) two-level network structure, with no significant microfabrication changes from our normal unaligned procedure. From a modeling point of view the geometry is slightly more complicated since the normal isotropic wet etching procedure produces a nearly hemispherical channel cross-section and flow profiles are highly sensitive to relatively small changes in channel cross-section.

To understand how to design focusing devices we explored low-Reynolds-number, fully reversible, pressure-driven Stokes flow, in the geometry of Fig. 9 through two CFD simulation packages (Lin et al., 2009). Based on the resulting models, we constructed simple three-level and four-level vertical focusing devices and tested their performance via 3-D optical imaging in a confocal microscope (Lin et al., 2009). The models show that the profile created by combining two flows in isotropically etched channels progresses nonlinearly as a function of the flow-rate ratio of the several fluid streams. That is, an addition of 50% fluid B to A does not give the same result as two sequential 25% additions of fluid B to A. Through comparison with experimental data, we found that the models are highly accurate in predicting flow profiles.

Figure 9.

(left side) A simple crossing junction used as a design element in software and imaging calibrations; two inlet flows from PA and PB. single outlet flow from PD. No flow allowed through PC (wall boundary condition). The analysis channel is on top. The sheath channel is on bottom. Percentages of flow from PA and PB are in reference to PD, the total flow after the junction; (right side) illustrates the four-level compensated vertical focusing device modeled in Figs. 10 and 11. Additive sheath (symmetric sheath inputs S1 and S3) and additive analysis (symmetric S2) are combined upstream of a correction flow (symmetric S4). The device is driven by suction from a port at the right end. Adjustable flow resistances on the channels S1–S4 are used to tune the device. [(Reprinted from (Lin et al. 2009) with permission. Copyright 2009, American Institute of Physics.)]

3.7 Sorting

A number of innovative microfluidic cell sorting devices have been designed and implemented on single-channel microfluidic cytometers (Dittrich et al., 2003; Wang et al., 2005; McClain et al., 2001, Fu et al., 2002, Wolff et al., 2003; Gawd, et al., 2004, Emmelkamp et al., 2004; Yi et al., 2006). However, many of these single-channel switches are difficult to multiplex, or lack the switching speed needed for a PMC. A truly impressive parallel switch has been designed and implemented on a PMC by Bohm, Gilbert, and Deshpande (Bohm et al. 2007). This system uses 144 parallel channels and a flow switch capable of a 0.5-ms activation cycle. These researchers have announced ambitious applications in the purification of therapeutic quantities of human blood (http://www.cytonome.com).

4 Operating Methods

For the most part the operating methods and the sample preparation for PMC applications is identical to the well-established protocols of flow cytometry and microscopy. The very few differences are summarized in the paragraphs below. We also provide specific protocols used to prepare the samples used in the demonstrations described in Section 5 (Results) below.

4.1 Microdevice maintenance

All flow cytometers require certain routine operating procedures and maintenance. The PMC is no exception. A 1%-concentration of bovine serum albumin in PBS buffer is periodically pumped through the microdevice to reduce protein adhesion (not more than once a week even with heavy use). As with single-channel cytometers, cell suspensions are treated with established cytometry prefiltration methods (Shapiro, 2003). An iodixanol (OptiPrep®, Sigma Aldrich) gradient-medium buoyancy-agent is typically added to the samples to assist buoyancy of the suspended cells. After about 100 hours of use, the microdevices are usually cleaned with chlorine bleach, however there are no extraordinary difficulties with channel fouling or clogging. With careful handling, devices appear to be reusable for an indefinite number of cycles.

Sample loading onto the microdevice is with the automated pipettor out of 96-well or 384-well plates. To counteract settling, the pipettor is also used to periodically mix the sample suspensions by returning at an interval of ~10 minutes to each well, aspirating, then re-loading a portion of each well volume on the microfluidic device.

4.2 Sample preparation

Cytometry samples were prepared by standard protocols. Several details relevant to Section 5 (below) are given in the next several paragraphs.

4.2.1 Samples for the primary-cell (lymphoma model) dilution studies

For the sensitivity trials (Sect. 5.1), EBm-BRD2-/GFP large B-cell lymphoma cells were obtained from the spleens of female 20-week-old FVB mice (Greenwald et al., , 2004). Unstained splenocytes (negatives) were obtained from female 16-week-old FVB mice. Fresh cells were frozen in freezing media (50% complete-10 RPMI-based medium, 40% FBS, 10% DMSO), then thawed in small batches as needed, diluted to calibrated ratios in PBS buffer and scanned on the PMC.

4.2.2 Cell Line for CPTHR Screen

For this large-scale screen (Sect 5.3), clonal osteocytic cells, expressing a high level of the carboxy-terminal region of parathyroid hormone receptor (CPTHR), were derived from fetuses in which the majority of exons encoding PTH1R had been ablated by gene targeting. These clonal osteocytic (OC) cell lines expressed 1,900,000 to 3,400,000 CPTHR binding sites per cell, a level 6- to 10-fold higher than observed on osteoblastic cells obtained from the same fetal calvarial bones and at least 5-fold higher than in ROS 17/2.8 cells. Biotinylated [Tyr 34 ] human PTH (24–84) was synthesized at the Massachusetts General Hospital Peptide and Oligonucleotide Core Laboratory (Boston, MA).

4.2.3 The cDNA Library for CPTHR Screen

The cDNA library (Sect. 5.3) was constructed using both random and oligo dT primers to synthesize the first strand DNA. This approach enriches the library with the 5’ portions of large cDNAs compared to cDNA libraries prepared using oligo dT primers only. Inserts were cloned in Lambda Zap pCMV-script expression vector (Stratagene). Since insert size represented in the library is crucial for the successful expression cloning, we examined the insert size in single colonies from different pools of the library. For this purpose, we used PCR analysis approach using T3 and T7 primers and cDNA preps from the single colonies. An average size of 2 kb was obtained. The library was divided into 100 pools of 10,000 PFUs/each and single pools were transiently transfected into COS-7 cells using Fugene 6 (Roche) according to the manufacture’s protocol.

The cDNA library, average insert size 2 kb, was divided into 100 pools of 10,000 PFUs/each and single pools were transiently transfected into COS-7. We calculated that a 200-μL sample (1,000 cells/μL) would produce 20–40 positive events in a positive pool. To reduce the false-positive count we used a simplified labeling protocol and evaluated the results using the image-processing algorithm above. Osteocyte cells without fluorescently labeled ligand were used as a negative control.

5 Results

5.1 Sensitivity Trials on Primary B-cell Lymphoma cells

From work to date we know that two of the strengths of the PMC are (1) rare-cell measurements and, (2) measurements on primary cells or on cultures were available sample is limited. Below we show results for a simple dilution study using murine B-cell lymphoma cells. The study was undertaken to prepare for larger studies that will use, in one case, human clinical samples and, in a second, murine blood samples for active monitoring of cancer treatment and regression in mouse models. We used blood samples from an existing transgenic mouse model that constitutively expresses a double bromodomain-containing 2 (Brd2) GFP fusion (Greenwald, 2004).

Samples were prepared by quantitative dilution from cell stocks, then presented to the PMC at a flow rate of ~200μm/s using the detection arrangement of Fig. 4(b). Frozen extracts were used, hence the preliminary study represents a more difficult case in terms of signal-to-noise (weaker GFP marker) relative to fresh clinical samples. However, we expect additional sources of variability in the clinical samples.

Figure 4.

Two configurations of the optical detector to match cell assays used for the PTHR screen (Fig. 4(a)), and for the dilution study on primary leukemia cells (Fig. 4(b)). The scattered forward light sensor (Fig. 4(b)) is a fiber optic (910-μm dia., 0.22 NA) on a rotatable mount that can be adjusted in the range from 20 degrees to 70 degrees off the forward direction.

5.2 Dilution Study on Clonal Osteocytes

A second dilution-curve study is shown in Fig. 18. In preparation for a large-scale screen (Sec 5.3), clonal osteocytic cells. Positive dsRed-expressing cells were serially diluted in a background of GFP-expressing cells.

Figure 18. Results for a 384-channel run (clonal osteocytes).

(a) histogram of dsRed-cell counts for a cell dilution curve (dsRed-expressing cells diluted serially with GFP-expressing cells). The histogram is organized by well placement on the PMC fluidics. Counts for all 384 microfluidic channels are shown. Sample dilutions are run redundantly in 2-ea. columns of 8-well rows (layout on the microfluidic device), i.e., 24 channels for each dilution. (b) Total counts are summed for each sample and used to generate the serial dilution curve (log vertical scale) which shows slight saturation at the highest concentration of positives (100% positives, right side of the figure). [From (McKenna et al. 2009). Reproduced with permission of the Royal Society of Chemistry.]

Figure 18 plots all microfluidic channels for a 384-lane microdevice, but uses eight channels redundantly to collect data for each dilution. This procedure makes use of one of the inherent attributes of the PMC, namely high channel count, to average out flow non-uniformities. The results are much as for the primary cell study above, but with different scan settings. The high-concentration saturation artifact of Fig. 16 is not seen up to concentrations well over 50% positives.

Figure 16.

Calibration of dilution study on primary leukocytes. For all the objects identified by the scatter detector we plot the maximum GFP channel value vs. the yellow channel value. Note that most objects in the negative sample have lower fluorescence then the positive sample, a more sensitive measure is made by comparing the ratio of the two PMTs.

5.3 Genome-wide cDNA Screen

The longer integration times of a PMC should increase rare-cell selectivity and thereby allow increased pool sizes for early-stages of large screens. This has major implications for a genome-wide screen where the target must be found in an initial pool of a many negatives and where the number of positive cells may number in the single digits per microliter. As a test (McKenna et al., 2009), we chose an on-going genome-wide cDNA screen for the carboxy-terminal parathyroid hormone receptor (CPTHR). The classical way to approach a screen of this kind is to, (a) separate the several million potential target sequences into a manageable number of initial pools (usually about 10–100 pools), (b) to identify the pool containing the positive sequence then to (c) subdivide this pool. This process is repeated until the positive pool is enriched to the level of a single candidate. The most demanding part of the screen occurs in the initial stage -since it requires finding as few as several-dozen positives (antibody-stained clonal oesteocytes) in a background of a million negatives.

Clonal oesteocytes were incubated with 0.5 mM EGTA for 20 minutes at 4 °C. Cells were then centrifuged for 3 minutes at 3000 rpm at room temperature and re-suspended in binding buffer. Cell suspensions were incubated with 10-6 M biotinylated hPTH (24–84) and streptavidin Texas red for one hour at 4 °C. Cells were then washed by centrifugation for 3 minutes at 3000 rpm at 4 °C, then were re-suspended in binding buffer.

Cells in a 200-μl buffer volume were loaded into multiple sample wells and pulled through the detection zone of the PMC at a flow rate of 10–20uL/Hr per channel. This corresponds to a flow velocity of several hundred μm/s. The laser spot (nominal diameter ~30μm) was adjusted to traverse the biological cell at a much faster scan rate of 10–40 mm/s (0.8–3 ms nominal dwell). Each sample was sampled in 4 to 10 duplicate channels in our experiments.

To partially automate data reduction, we developed a post scan data process using Matlab. First, the raw data signal of the red PMT (4) is subtracted from the green PMT (2) (Fig. 3(a)) to compensate for auto-fluorescence. The channel locations are then overlaid to segment the data into individual-channel time sequences – about 15 pixels wide by the total number of scans (~50,000 pixels) long. Each channel segment is searched for scans that contain signal above a noise threshold. These scans are then automatically “cut-and-pasted” to a new image that represents the objects in one channel (accumulated for the run), and the number of events are determined by a software counting algorithm.

Figure 3.

(a/b) optical diagram of the laser-induced fluorescence (LIF) detector including detail of rotary scanner that introduces the 488-nm laser beam and returns fluorescence onto four photomultipliers (PMTs, other configurations shown below in Fig. 4). The scanner uses a 3-inch radius of rotation, a DC rotary motor, and optical encoder; (c) detail of a typical condensed-time data scans for a single microfluidic channel [(c), left and center] and reduced-difference scan to identify positives [(c)), right]. See text. [From (McKenna et al. 2009). Reproduced with permission of the Royal Society of Chemistry.]

Final bright cell counts were entered into a spreadsheet and compared across samples in order to determine run-group statistics median, average and standard deviation. These values were used to determine the probability that a given pool was negative. Those pools that were above the median plus two standard deviation were retested, and if still contained outliers, were designated for further expansion.

The workflow of the screen is shown in Fig. 19. The initial stage included nine sample pools and one control, all of which were run in redundant microfluidic channels. All samples showed a few positive cells with a median count of 4 and a standard deviation of 12.58. We calculated the boundary for outliers, median plus two sigma, to be 28. One pool was an outlier with 39 positives, and when tested again produced 35 positives. The outlier was subdivided into twenty sub-pools and each was tested twice. A count of positives produced a median of 5.5, a sample deviation of 12.47 and an outlier boundary of 47. One sub-pool showed 95, then on recount 98, positive cells. This process was repeated for two more subdivisions until a sample was produced that was overwhelmingly positive (> 10,000 on the PMC). Levels 3–5, which had much higher abundances of positives, were conducted in parallel on the PMC and on a conventional single-channel cytometer (FACS-Caliber^TM, BD Biosciences). Finally we isolated a candidate cDNA, which was sequenced by capillary electrophoresis and found to include a seven-transmembrane domain belonging to a family of G-protein coupled receptors (GPCRs). The sequence was run against the BLAST database and found to be a novel candidate. The end result is that the PMC was able to rapidly perform a full genome-wide cDNA-screening assay with statistically significant results on positive counts of only several dozen cells in background of several million negatives and with sample pools of 200μl.

Figure 19.

Schematic representation of a cDNA expression cloning study that identified a new target for the CPTHR receptor. The most difficult first two stages were completed on the PMC using the rare-cell detection advantages of the variable integration detector.

6 Adding 1-D Imaging to the PMC

The PMC offers a way to increase the throughput of image-based HCS into the domain of FACS through a flow architecture rather than static imaging. Our approach circumvents the rate limitations of the CCD (microscopes and CCD-based flow cytometers) by using a 1-D scanner and photomultiplier detection. The principal PMC instrument adjustment relative to the results reviewed above is to increase the spatial resolution of the scanner in Fig. 3, and thereby collect multiple intracellular pixels on each cell that is detected in the flow. The scanner then collects a multicolor “line-scan”, from each microfluidic channel.

6. 1 Classification of phenotypes by 1-D images

Our objective was to develop 1-D imaging for high-throughput, high-content image-based screening. The economy of 1-D images (when compared to CCD images) is a computational advantage [Gonzales et al. 2002]. However, less image information means more ambiguity. The question becomes: “Can a 1-D image provide sufficient information for a high-content screen?” A key aspect for fluorescence localization assays will be a fast analysis algorithm for binning of image events.

The classification ambiguity typical from 1-D imaging as it relates to a protein-localization assay is illustrated for 2-colors in Fig. 20. With 1-D images, the feature set is greatly reduced (relative to 2-D), with new classes of indeterminate phenotypes; distinguishing features become asymmetries, profile shape factors, and relative curve heights. However, there is a great deal of information available in 1-D; furthermore even in our initial system, there are 4 different 1-D color images for each cell. Fig. 20 also illustrates an interesting problem that was unknown at the onset; it is not clear if the problem of phenotype classification in 1-D will become easier or harder with higher resolution in the scan (smaller laser spot). More detail is not necessarily better for easy phenotype classification.

Fig. 20. Cartoon of typical 1-D images that are encountered in a protein localization assay.

The left column shows 2-D (microscope) images with the marker (green) and cytoplasm (pink). Three positives are shown top; three negatives on bottom. The confocal slit in our detector discriminates strongly against out-of-focus images. The right column shows the several principal 1-D image types that are generated depending on how the laser scanner traverses the cell. The dashed arrow shows the location of the single line scan that is taken per cell. Some of the most diagnostic signatures are surprising.

Our task is to acquire/model various forms of multicolor 1-D images from typical cells and to then partition them into the “positives” and “negatives” typical of a cell assay. The problem is complicated by the several trade-off choices in the optical system and the illumination. In addition, a real assay sample will contain both positives and negatives. The cell types are on a continuum of size/shape/cell-cycle factors, which causes a heterogeneous distribution of 1-D images. The exact position of cells in the Z-focus is a complicating factor for all imaging methods (although it is minimized for our confocal detector). The traditional way to approach these problems, all of which are also encountered in CCD imaging, is to set up data filters and thresholds that eliminate ambiguous data. We used the same approach, however the algorithms and filters are unique to 1-D imaging. The metrics of success are partitioning confidence factor (for example, the Students T test) and the sampling efficiency (as measured in time per assay). For a simple binary (yes/no) assay the number of discriminating (qualified) objects is as few as 50–100 objects [Taylor et al. 2007]. Therefore since many thousand events per second can often be processed, it is possible to “throw away” a large number of the events and still end up with a fast high-confidence assay.

The problem was addressed with a combination of empirical modeling and data reduction from our data libraries and from specifically acquired 2-D and 1-D images. Obviously the actual data as acquired on real live and fixed-cell samples must ultimately be used to confirm the models and, conversely, the models depend on optics choices.

6. 2 Confirmation of 1-D Imaging on a PMC

For a feasibility study we began by modifying one of our prototype PMC systems to reduce the spot size of the scanner to the extent possible (from 30-μm to 3.5-μm). Next we programmed the signal processing hardware to collect 100 points at 1-μm spacing across the channel (Fig. 21). However, our current hardware had the limitation of processing pixels at a maximum 8,000 per second. To get around this restraint we limited the range of the scanner to 320 data points at 1- μm resolution and 20 Hz. (However the optical resolution remained at 3.5 μm.)

Figure 21.

1-D images of 6-μm beads (one of four PMTs). On left 350 pixels in line scan (X) vs. scan number (Y) with fluorescence intensity values shown as growing from blue to green to red. On right two sections are magnified, showing image of 6-μm bead moving at a higher speed through the detector (top) and slower at bottom. Slower speed yields a more 2-D “picture-like” image.

We utilized S. cerevisiae mutants engineered to over-express the amyloid protein a-Synuclein (aSyn-GFP) (Shorter and Lindquist, 2005). In the native state, cells show a uniform distribution of the fusion protein along the membrane and in the cytoplasm. Under induction, the protein condenses to one or several focal conjugates per cell of 1–2μm diameter. Cells were fixed and fluorescently labeled with a red whole cell dye. Our samples contained a negative control with normally expressed a-Syn and positive sample with ~50% of cells over expressing. Cells were fixed and suspended in PBS at a density of 1000cells/μl. A post-scan algorithm identified cells, created a Gaussian-smoothed image for each color channel then used various comparative color-channel algorithms to categorize images and identify cell metadata. Fig. 22 displays the raw traces for the red (580nm) and the green (540nm) color channels. These data were filtered to select a target diameter (red FWHM) of 4 - 6 μm, then an algorithm modelled after (two-dimensional) “roundness” was applied to the red and green channels. As shown in Fig. 22, the two populations are clearly distinguished. This was repeated for filters set to various signal X –widths. When we analysed some subgroups we were surprised to find that we could separate the positive and negative samples using some less obvious signatures. For example, for small-width thresholding (red FWHM ~ 2 or 3 μm after deconvolution of the laser spot), we found that green signal would occur over threshold in 5–20% of negative samples, but less then 1% for positive samples. Our explanation is that this group represents scans that skirt the center of the cell, and that such scans often entirely miss the aSyn-GFP focal conjugates. This is a novel indirect way to infer the condensed-state positive.

Fig. 22. Results showing 1-D HCS data using a 3.5-μm laser spot to scan αSyn-GFP expression patterns in S. cerevisiae.

The first version of the detector is (just) able to distinguish the localization patterns. Top: raw scans for whole-cell (red) and αSyn-GFP (green). Left (a) negative cells, right (b) positives showing αSyn aggregates. Below: Filtered data using a modified “roundness” parameter distinguishes positive (induced) sample from a negative with baseline αSyn expression.

6.3 Proof of Principle for Nuclear Translocation Assay by 1-D Imaging

Next we simulated the nuclear translocation assay. We used mouse fibroblast cells (swiss-3T3) that were treated with Trypsin EDTA (Cellecgro) to make them non-adherent, and then fixed (3.7% formaldehyde) and labeled these cells with Sytox Orange nuclear stain (2.5mM, Invitrogen). Half of this sample was dyed with a second nuclear stain, (0.5mM Sytox Green, Invitrogen) and the other half with CDFE whole cell stain (5mM, Invitrogen). Three singly stained, control, samples were also scanned in order to obtain PMT color correction information. As above, the cells were scanned with our first-generation PMC with a laser spot size of ~3.5 micron and an image digital capture resolution of 1-μm per time point under the lane. A post-scan algorithm identified the cells, smoothed, digitally zoomed the images, then color corrected and normalized the fluorescence levels. We found two methods to separate the samples. The first was by comparing the object width (FWHM) of the orange and green channels (Fig. 23). A more accurate method appears to be to use the orange channel as a mask of the nucleus, and quantifying the green signal outside that mask (Figs. 24,25). Therefore even with 3.5-μm spot resolution, 1-D line scans can resolve nuclear vs. cytoplasmic location of the green marker. A next-generation optical scanner with 1-μm resolution (rather than 3.5-μm) and updated digitizing electronics, will greatly increase the number of color channels, allow 1-D and 2-D line scanning, and enable data collection at increased speed.

Figure 23. Proof of principle for NT assay.

The FWHM for the green and orange channels are compared in a scatter plot. For the “un-stimulated” CDFE sample cells, the wider green line scan skews the sample above the center diagonal line (proving marker in the cytoplasm). For the “stimulated Sytox Green sample cells the distribution centers along the diagonal (marker confined to the nucleus).

Figure 24. Proof of principle for nuclear translation (NT) assay.

(a) Representative line scans from each sample show green fluorescent signal difference between whole cell marker and nuclear stain when compared to orange nuclear stain (FWHM point for normalized scans marked by blue line). In (b) an algorithm classifies objects by first eliminating all green signal outside of green FWHM and inside of orange FWHM, then measuring the remaining green signal. This value is significantly larger for “positives” (stimulated cells) in the NT assay. 38

Figure 25. Box plots of the two samples that simulate nuclear translocation.

The green signal outside the nuclear area (nuclear area defined by the Orange PMT channel) is plotted (vertical axis in the plot). As a group, the “un-stimulated” cell type (left column) shows more green signal outside the nucleus then the “stimulated” cell sample (right column).

7. Conclusions

Although parallel microfluidic cytometry is at early stages of development, nonetheless some of the predicted features have been proven. Four key aspects of the architecture are (i) parallelism from the microfluidics, (ii) high-sensitivity from an optical scanner with variable integration time, (iii) Parallel flow imaging with a high-speed analog detector (rather than CCD), (iv) a small-sample capability from the microfluidics.

The 384-channel parallelism, most importantly improves sample-throughput, but also sidesteps the time biasing between samples due to sample changeover in a single-channel. The elimination of time biasing addresses issues with unstable samples or degrading markers and, importantly, allows rigorously time-synchronized comparative assays, e.g., for biological process with fast kinetics. The scanner permits practical adjustment of integration time, including lengthened signal averaging, which greatly improves performance in rare-cell analyses. The microfluidic flow allows efficient handling of very small and rare cell samples, e.g. a few microliters of primary cells. Single-channel cytometers continue to be improved in some of these features (e.g., Goddard et al. 2007, Haynes et al., 2009), however none combines these features.

An increased detection sensitivity relative to conventional flow cytometers, as seen in our dilution studies, is reasonable given simple signal-to-noise (S/N) arguments. The PMC and single-channel cytometers utilize nearly the same spectral separation and PMT-based photodetection, both operate in the high-signal (rather than photon-counting) regime, both have a dominant noise contribution from the shot noise, and both systems can be operated near photobleaching. This implies a comparable number of signal photons for the two detectors. In the experiments above we have varied the integration time between 0.8 and 60 ms, up to 3–4 orders of magnitude longer than is typically used in a single-channel cytometer. This permits a 3–4 order-of-magnitude smaller amplification bandwidth and, for a Poisson distribution of noise, an advantage of 1.5 to 2 orders of magnitude in the S/N for our detector.

Moreover the integration time of the PMC is an elective setting in the system; it is set by the scanner velocity and is independent of minimum flow requirements. On the high-count-rate end both PMC and single-channel cytometers (FACS) are ultimately limited by essentially the same digitizing electronics; therefore the PMC, when it is run at high flow velocity, can achieve approximately the same total count rates as a high-end FACS. In some initial trials, we have adjusted the PMC for rare cell capability and high sample-number throughput. This is the optimization for early stages of a genome-wide screen. We have been able to confirm the improved performance in this domain of optimization.

For a binary assay, closer to a classical flow-cytometer assay, i.e., abundance of “positives” ~0.1% or higher, we can operate the PMC at an integration time closer to that used in FACS. A realistic sample throughput for a binary assay on the PMC with this tuning is 384 unique samples in six minutes (384-ea. 1 μL samples, 10³ cells/μL). This might compare with a maximum of approximately ten unique samples in six minutes for a typical commercial single-channel FACS. However the 384-well-plate automation that permits the PMC to be integrated with existing high-throughput cell culture is important in order to realize these advantages.

In the imaging application, the PMC has demonstrated an unexpected proficiency in separating samples with highly economical 1-D images. Even with a 3.5-μm resolution on a relatively small (5–6μm) yeast cell, we are able to see condensation of a GFP marker, and on mammalian cells, the classical nuclear translocation assay has also been simulated. We definitely expect that 1-D imaging on the PMC will be further developed as a means to add “high-content” to FACS.

There are several intriguing directions that will be developed in the near future. First, there remains a space to be explored at ultra-high count rate on a PMC. This has only recently become possible with improvements in digitizing electronics. Digitization rates now exceed the maximum rates that can be used on as single-channel flow cytometer. By expanding the flow stream over parallel channels, the throughput of the latest digitization electronics can be used to full effect. The S/N trade-offs need to be explored and low integration time can introduce a trade-off in single-cathode PMTs. However multicathode PMTs can be used, so there is little real question that substantial count-rate improvements can be achieved over the first PMC prototypes with a further large advance over single-channel cytometers. The sample throughput of the PMC already exceeds FACS, in the future the PMC will also exceed FACS in absolute (single-sample) count rate. We can anticipate an improvement of ten or more over the current state. This will push flow cytometry into the domain near two 384-plates per minute for a binary assay, i.e., well into a space useful for drug discovery.

A second area that needs expanded engineering is the integration of cell sorting onto the PMC (e.g., Bohm et al. 2007), with good independent logic controllers on each channel and with isolated-well fraction collection. The added value of fraction sorting on a PMC is enormous. It will allow downstream analysis, e.g., qPCR or mass-spectrometry, on sorted fractions in a massively combinatorial way.

A third unexplored direction for the PMC is into high-time-response kinetics. This is a domain were the comments in the introduction about “tools limiting science” apply. Since it has not been practically possible, without heroic measures, to do cell-resolved studies of kinetics in a massively parallel way, it has not been possible to do statistically significant studies of many aspects of biological kinetics with high time response. We know that the majority of signaling pathways are dynamic on time scales of minutes. But since there has been no efficient way to measure them, this fundamental aspect of systems biology has remained outside the realm of practical investigation.

Figure 6.

A finished PMC microdevice similar to those in Fig. 4 (slightly different design) but after attachment of G-10 fiberglass pumping block and fluid reservoirs. The suction port and wash port are threaded to receive standard 10–32 HPLC fittings. The 32 open sample ports are 2-mm diameter and 10-mm deep, on 9 mm centers (other designs use 4.5mm centers), and are compatible with a standard multi-tip pipettors.

Figure 7.

A 16-channel PMC microdevice with 3-sided hydrodynamic focusing. This design can be fabricated with one microlithographically-defined fluidic level and captures the three (out of four) directions for flow focusing, The glue-on fiberglass block (see e.g., Fig. 5 above) is machined to combine the three “blue” buffer flows into a single manifold and reservoir. Other flow focusing designs are provided in Section 3.4.

Figure 10.

Plan view layout of the device designed to test vertical flow focusing and subtractive compensation. Eight variations of the 4-sheath configuration shown in Fig. 9(b), labeled “A”– “H”(right side of die), are included on the single test die. The two layers of etched channels are indicated as red (top plate) and black (bottom plate) respectively. A single laser-drilled hole is provided for each input or output (S1–S4, Fig. 9(b)) and for a common suction port (common to configurations A–H, right side of die). The full die size is 3x7 cm. ((Reprinted with permission from (Lin et al. 2009). Copyright 2009, American Institute of Physics.))

Figure 11.

Simulations of four-layer focused flow, (a) before and, (b) after, the channel S4 junction and subtractive correction flow (plane V4). As the traces pass beyond the channel S4 junction (Fig. 9) they are preferentially pulled downward and outward. The flow interface indicated by the arrows is most strongly altered by the subtractive flow. ((Reprinted with permission from (Lin et al. 2009). Copyright 2009, American Institute of Physics.))

Figure 12.

Raw data. Two sample types showing raw signal on all four photomultipliers (PMT’s ) using the detector layout of Fig. 4(b). The x-axis is time, designated as scan number, 12 scans per second. The labeled and unlabeled cells show up as events on the scatter detector, while GFP cells appear as fluorescent spikes on the P1–P3 color PMT’s. Weak autofluorescence is occasionally observable on P3 (unlabeled cells). For the GFP sample, the signal ratios vary significantly, e.g., P2 (GFP channel) compared to P3 (Yellow channel).

Figure 13.

Unreduced image data: Raw data (e.g., Fig. 12) is plotted as an image of the microfluidic channel cross-section (vertical axis labeled spinner position) versus time, raw data is collected at 300 pixels per 16-channel (or 384-channel) scan and 12 scans per second.

Figure 14.

Distinguishing positives from autofluorescence. A scatter-plot comparison of the ratio of the GFP channel to yellow channel for objects with a sufficient maximum GFP value (threshold) in positive and negative samples. The low ratio in the negative sample shows how auto-fluorescence cells can be rejected as negatives. By determining the mean and standard deviation for cells in the negative sample it is possible to calculate an outlier threshold (> mean + 4 sd).

Figure 15.

Calibration for dilution study on primary splenocytes. Ratio of GFP/yellow channels as a plot of objects and as histogram for a positive sample (right) and a (mostly) negative sample (left). From the histogram we conclude that cells with a PMT ratio greater than 0.8 would (a) definitely be GFP cells and (b) these cells would represent about half of the cell number that was contained in the GFP source sample.

Figure 17.

Results for dilution study on primary splenocytes. Measured percentage and expected percentage of GFP labeled cells for all samples (ordered by expected percentage) shows a clear distinction between negative samples and positive samples down to dilutions of 0.01%. There is observable saturation of the count at high abundance (likely due to multiple-cell counts).