automated quantitative live cell fluorescence microscopy

Automated Quantitative Live CellFluorescence Microscopy

Michael Fero1 and Kit Pogliano2

1Department of Developmental Biology, Stanford University School of Medicine, Stanford, California 943052Division of Biological Sciences, University of California San Diego, La Jolla, California 92093-0349

Correspondence: [email protected]

Advances in microscopy automation and image analysis have given biologists the tools toattempt large scale systems-level experiments on biological systems using microscopeimage readout. Fluorescence microscopy has become a standard tool for assaying gene func-tion in RNAi knockdown screens and protein localization studies in eukaryotic systems.Similar high throughput studies can be attempted in prokaryotes, though the difficulties sur-rounding work at the diffraction limit pose challenges, and targeting essential genes in a highthroughput way can be difficult. Here we will discuss efforts to make live-cell fluorescentmicroscopy based experiments using genetically encoded fluorescent reporters an auto-mated, high throughput, and quantitative endeavor amenable to systems-level experimentsin bacteria. We emphasize a quantitative data reduction approach, using simulation tohelp develop biologically relevant cell measurements that completely characterize thecell image. We give an example of how this type of data can be directlyexploited by statisticallearning algorithms to discover functional pathways.

Over the last decade, advances in micro-scope automation, fluorescent labeling,

sample preparation, image processing, patternrecognition, and statistical learning have ush-ered in a new era of cell biology experimentsbased on unbiased genome-wide cell-basedassays. In eukaryotic cell biology, these advanceshave been exploited to investigate the distribu-tion and patterning of labeled subcellular struc-tures and the “location proteome” defined bythe distribution and intensity of localizedprotein reporters (Chen et al. 2006). In the pro-karyotic world, the challenges surrounding flu-orescence microscopy at the diffraction limit

has meant a somewhat slower adoption ofhighly automated methods and analysis de-signed for large-scale genetic screens or assays.Nevertheless, large-scale approaches are nowbeing used more routinely during investigationsof bacterial systems, from the identification ofentire location proteomes (Werner et al. 2009)to the detection of modular transcription andsignaling pathways (Christen and Fero 2009).Because of the size and complexity of theeukaryotic cell, generalized pattern recognitionand classification approaches are often cru-cial. In these approaches, conditioned, param-eterized, preclassified image data are used by

Editors: Lucy Shapiro and Richard Losick

Additional Perspectives on Cell Biology of Bacteria available at www.cshperspectives.org

Copyright # 2010 Cold Spring Harbor Laboratory Press; all rights reserved; doi: 10.1101/cshperspect.a000455

Cite this article as Cold Spring Harb Perspect Biol 2010;2:a000455

1

on December 13, 2021 - Published by Cold Spring Harbor Laboratory Press http://cshperspectives.cshlp.org/Downloaded from

http://cshperspectives.cshlp.org/

statistical learning algorithms to build modelsthat capture the differences between classesof images. A properly constructed model canthen be used to automatically analyze a pre-viously unclassified experimental data set(Boland et al. 1998; Boland and Murphy 1999;Boland and Murphy 2001; Huang and Murphy2004; Chen and Murphy 2005; Jones et al.2009). In prokaryotic cell biology a similar sit-uation exists, with one major difference; thesimplicity of the bacterial cell allows us to enter-tain the idea of a cell parameterization based notonly on generalized image properties, but alsoon specific cell measurements of high biologicalrelevance. For example, instead of measuringhard to interpret properties based on general-ized image parameters and a global image trans-form, such as a Zernike or Fourier transform,one can conceive of measuring a more relevantset of parameters related directly to gene func-tion or to the biology of interest, such the po-sition and quantity of both localized anddelocalized fluorescent signals, the length andwidth of the cell, the degree of pinch at the divi-sion plane or sporulation septum, the preciselocation of the poles or other points of inflec-tion, as well as the location of membrane struc-tures such as pili or stalks and the membraneoutline itself. In this way, albeit with a certaindegree of additional work, the image can bereduced to a set of high information contentparameters. This type of data can be usedmore efficiently in any type of subsequent anal-ysis, whether direct or via a model-buildingstatistical learning exercise. In addition, reduc-ing the cell to a small set of highly informativeparameters allows the option of combiningthose single-cell measurements into quantita-tive ensemble-based phenotypes where theentire distribution of measured values is re-corded rather than just means and standarddeviations. With the addition of genotypeinformation, this data can then be used directlyin unsupervised statistical learning approachessuch as hierarchical clustering to do quantita-tive phenotypic profiling (Ohya et al. 2005).In this article we outline the process of doingsuch an automated analysis on a bacterial sys-tem. To keep our discussion firmly rooted in a

practical example, we will use a case-studyapproach based on work performed in thea-proteobacteria Caulobacter crescentus. How-ever, most of what we discuss can be appliedto other bacteria and single-cell eukaryotes. Itshould also be noted that the efficacy of a quan-titative analysis depends on the constructs usedto express fluorescently tagged reporter pro-teins. Even modest overproduction of someproteins, particularly those involved in signaltransduction and cell division, can have delete-rious effects on cell viability and on cell shape(Gregory et al. 2008). We will assume that forthe purposes of quantitative analyses, taggedprotein reporters are fully functional proteinsproduced from genes in their native chromoso-mal context.

AUTOMATED LIVE CELL FLUORESCENTMICROSCOPY

Live cell, 100�, oil immersion, fluorescencemicroscopy presents some problems for a high-throughput system not found in lower resolu-tion systems. First, the geometry is tightlyconstrained, with a �2 mm gap between thefocal plane and the condenser, and �0.24 mmspacing between the 100� objective and thefocal plane. The cells must be kept alive, whichusually means imaging on a 1% agarose þnutrient gel pad and limiting the amount oftime the samples are under observation. Fluo-rescence microscopy usually entails imaging inmultiple optical channels; the multiplexedmutant screen example described later usedone phase contrast plus three fluorescent chan-nels. This means the optical system has to bestable for several seconds while cubes or filterwheels are automatically switched. The needfor optical coupling oil also drives the designof the microscopy plate and choice of invertedversus upright scope geometry.

Data Acquisition and Autofocusing

Acquisition timescales range from tens of milli-seconds for optical channel exposure times toseconds required for some mechanical systemmotions, all the way to the hundreds of minutes

M. Fero and K. Pogliano

2 Cite this article as Cold Spring Harb Perspect Biol 2010;2:a000455



to needed to push through thousands of sam-ples. Several issues arise when trying to auto-matically image multiple live samples. Imagedrift, the availability of nutrients and oxygenover long time periods, the cells doublingtime, and drying out of agarose pads are amongthe issues that determine how long a sampleplate can sit on the microscope and thus howlong a data-taking run can last. Experienceshows that data-taking runs lasting approxi-mately 1 h work well. If imaging in a 96-wellmicroplate format this implies a throughputof 96 sample-wells per hour. Thus, we haveroughly 30 s to: move to the well, coarse focus,fine focus, image capture, and write the datain phase contrast plus several fluorescent chan-nels. Focusing is a large contributor to the timeit takes to image a sample well. Because of fluo-rophore bleaching, using any fluorescent chan-nel to optimize focus is usually unwise. Thisleads to using an alternative white-light channelsuch as phase contrast or DIC to set the focusfor subsequent fluorescent images. There areseveral schemes for calculating z-focus qualityscores (Groen et al. 1985; Firestone et al. 1991;Price and Gough 1994; Geusebroek et al.2000) but in our experience we have had goodresults with a simple autocorrelation function.Images are acquired in a fast loop while chang-ing the z-focus depth. The z-focus quality scoreis calculated for each z-focus depth. A fast para-bolic fit to the points surrounding the maxi-mum of the graph of all z-focus quality scoresyields the optimal z-focus depth. This is a veryfast way of finding the best focus and workswell for multiple closely spaced image fields.However, when the automated imaging systemhas to move over a macroscopic distance, forexample to the next well in a 96- or 384-wellmicroscopy plate, the fine focusing describedearlier will not be sufficient for finding a badlyout-of-focus image field. Even very small tiltsin the plane of the microplate can put an imagevery far out of focus after translating the objec-tive by several millimeters. To help the autofo-cusing heuristic an overall calibration or offsetis needed to account for any small misalign-ments between the perpendicular to the sampleplane and the optical axis. One method is to use

an offset calculated by manually taking focusedimage data at 3–4 points on the microscopyplate and then fitting a plane to the calibrationpoints. This is performed before automateddata taking begins. The z-coordinate of the ofthe fitted calibration plane at the location ofthe center of each microscopy-plate-well be-comes the zero offset for subsequent autofocus-ing. Alternatively, if available on the microscopesetup, one can use laser range finding to calcu-late the distance to the cover slip. This distanceplus the cover slip thickness gives the startingpoint for subsequent fine focusing. The finefocusing step is implemented with a precisionpiezo electric objective insert or peizo electricactuated stage insert. An approach using theimage data is used because the best focus isonly found using the full image as observedthrough the optical system and CCD camera.Even though the procedure outlined earlierwill yield satisfactory results in the fluorescentchannels, the correct focal plane in the white-light channel used for autofocusing will notnecessarily match the correct focal plane for dif-ferent fluorescent channels. However, this is arepeatable physical effect and can be compen-sated for by measuring the focal plane offsetsfor each fluorescent channel versus each otherand versus the optimal phase contrast focalplane. Once measured, this becomes a constantoffset when setting the z-focus for that particu-lar fluorescent channel.

Image Stability

Two important considerations for automatingdata taking and analysis are that one can takea well-focused fluorescent image without re-sorting to a z-stack, and that repeated frames,such as those found in a time series, are ingood alignment. During the analysis, it is possi-ble to correct for out-of-alignment frames, as itis possible to automatically search for the best-focused image out of a z-stack. However, cor-recting for issues resulting from poor data takestime and is likely to result in an increase inthe overall systematic error of a quantitativemeasurement. An environmentally controlledmicroscope setup will minimize drift in x–y

Automated Quantitative Live Cell Fluorescence Microscopy

Cite this article as Cold Spring Harb Perspect Biol 2010;2:a000455 3



or z-focus coordinates regardless of vibration,temperature, or humidity changes in the sur-rounding environment. Image drift over somerelevant time scale should be less than the maxi-mum resolution of the optical system. For 100�fluorescent microscopy this means stabilityagainst drift at less than a ten nano meters per sec-ond. Systems that achieve this level of stability canimage in phase contrast plus several fluorescentchannels while keeping the optical drift at thesubpixel level. Note that it can take severalminutes for the microscopy plate to come intothermal equilibrium after being introduced to acontrolled environment. Acquiring images tooearly can lead to significant drift and result inmisregistered and out-of-focus images.

Data Model and Storage

For flexibly analyzing large amounts of data theimages must be indexed, for example by theoptical channel imaged (c), the z-depth of focusin a multiple z-stack set of images (z), a timeindex for the purposes of time-lapse (t), theselected image field for a particular sample( f ), and finally an index for the unique sample(w). Once image data is acquired it needs tobe associated with its corresponding meta-data and stored. Getting complete metadataassociated with the correct image data can be

a vexing problem. Image information from themicroscope and camera that can be acquiredautomatically can be easily saved. Ensuringthat information provided by the experimenteris properly saved is more difficult. One schemerequires the upload of an experimental meta-data file at the start of a data-taking run. How-ever, user friendly tools for properly annotatingimage data with user-generated experimen-tal metadata after the fact are also important.Regarding computer power and data storage,the capability of software allowing individualinvestigators to organize, manage, and analyzea large amount of image data is improving sorapidly that any comment made here will bequickly out of date. In our experience, data anal-ysis ambitions tend to scale with increasingcomputer and data storage capabilities, to thepoint where dedicated image servers are usedfor image libraries and a multiple node com-pute clusters are used for image analysis.

AUTOMATED QUANTITATIVE IMAGEANALYSIS

Our example of an automated analysis workflow starts with a high-quality phase contrastor DIC image and one or more fluorescent-channel images for each image field as seen inFigure 1. In the first analysis stage, cell finding

Cell imagesCell detection

Feature extraction

Parameterization

Cell data

Quantifiedphenotype

Ensemble average

Reporter strain(multiply labeled)

High throughputautomated

live-cellimaging

Candidate generationmutagenesis,knockout, etc.

Candidateselection

Genotyping Genotype

Figure 1. Automated image analysis. Stored images are segmented by individual cell and parameterized using allavailable image information. Parameterized data is stored to disk for second-pass analysis where cell parametersare converted to biologically relevant metrics that can be used for quantitative phenotyping.





is followed by subcellular feature detection andparameterization at the single-cell level asshown in Figure 2. The data reduction factorat this stage is 10–20-fold depending on thecomplexity of the image. In the second stage ofthe analysis, the individual cell measurements

are combined into high-statistics ensemble aver-ages and standard deviations. A typical ensem-ble might consist of all of the cells from all ofthe images taken for a particular sample well.At this stage ensemble averages are translatedin to biologically relevant phenotypes and theresults can be tabulated, plotted, or subjectedto more advanced classification algorithms.Note that by making individual cell measure-ments, we have access to the distribution ofeach measured parameter on an ensemble basisand not just the average and standard deviation.There are observable phenotypes that do not sig-nificantly change the average or standard devia-tion, but do noticeably change the distributionof measured values. For example, high intensitymeasurements in the tail of a distribution oflocalized fluorescence signals may disappear ina particular mutant background, while thecore distribution remains intact (Christen andFero 2009). The reason for separating the celldetection and parameterization from the restof the analysis is that the processing time forthe two stages differs markedly. Because cellparameterization can require 100 times moreCPU time than ensemble characterization, it isnot wise to directly couple the two parts of theanalysis. Once the images have been processedand the cells reduced to their data summaryform, it is not necessary to repeat this processingstep. Instead, the investigator can concentrate onrefining the second stage of the analysis, which iswhere most of the biological content lies.

STAGE ONE: CELL LEVEL MEASUREMENTS

Cell Finding and Image Segmentation

Every automated analysis starts with findingcells. Under the controlled experimental condi-tions of a bacterial fluorescence microscopyexperiment, the cell density can be tuned to pro-vide a distribution of cells with a minimal num-ber of touching or overlapping cells. Unlikemany eukaryotic cell types, where cell findingand segmentation can be a challenging task(Kasson et al. 2005), cell finding in bacterialsystems is a fairly straight forward, usually in-volving either a thresholding or edge-detection,

Delocalized signal

Localized signal

A

B

C

Figure 2. Measurements made on an individualbacterial cell. (A) The membrane profile from asubpixel resolution spline fit to a phase-contrastimage is combined with, (B) subpixel resolutionmeasurements of fluorescent foci and delocalizedfluorescent signal to provide, (C) a parameterizedview of the cell including localized and delocalizedfluorescent contributions.





region-filling approach performed on the phasecontrast channel. To compensate for unevenillumination of the phase-contrast channel asimple block-filtering algorithm can be usedfor background flattening. This step followedby an automatically adjusted thresholding andregion-finding algorithm works well to identifycells. It is not crucial at this stage to perfectlyidentify the cell boundary. Rather, the aim isto properly tag the potential cell locations. Aftertagging the location of each cell, the originalimage can be segmented into a number of sub-images, one for each cell, in preparation fordetailed examination.

Membrane Profiling

When analyzing the smaller subimages the cell-finding algorithm is repeated, in a more refinedway, to get a good measure of the cell profile intwo dimensions. The goal, in our example, is amembrane profile representing a section half-way through the cell in the z-direction. Thiscan be a difficult step, if care is not taken delicatestructures that are marginally visible are lost. Insome cases it is possible to assume that themembrane boundary has no sharp discontinu-ities and a spline-fit (De Boor 1978) can be fit toa threshold boundary yielding a correct subpixelrepresentation of the membrane profile asshown in Figure 2A. Once a membrane profileis available, rod shaped cells can be defined byan internal coordinate scheme defined by thelongitudinal midcell line, the poles, and severalmeasures of cell width. Note that here we havetaken care to define an internal coordinate sys-tem defined by an midcell line that runs fromone cell pole to the other, with orthogonalwidth measures at right angles to the midcellline.

Cell Image Simulation

Vital to the process of developing algorithmsthat yield results with believable error estimatesis the implementation of a simple cell-imagesimulation, as shown in Figure 3A. With a sim-ulation we know, at the outset, what the “true”(expected) values are for the parameters that

0.1 0.2 0.3 0.4 0.5 0.6 0.7–3

–2

–1

0

1

2

3

Threshold (min = 0, max = 1)

Res

idua

l (pi

xels

, 1 p

ixel

= 8

0 nm

)

Cell length

Cell waist

Fluor location

B

0

10

20

30

40

50

60

70

80

90

100

–2 –1.5 –1 –0.5 0 0.5 1 1.5 2Residual (pixels), threshold = 0.4

Cou

nts

C

Simulated phase contrast image

A

Figure 3. Using cell image simulation when developingalgorithms. (A) An image simulation of a bacterial cellhas been used to test a membrane-fitting algorithm. (B)The difference between the predicted and measuredlength and width of the cell as a function of thresholdshows a systematic bias. (C) The effect on measuringa fluorescent spot relative to the measured pole of thecell. After correcting for the threshold effect asimulation run of 400 cells shows the resolution ofthe algorithms, given by the width of the residualhistograms as shown.





will eventually be measured by the analysis al-gorithm. Starting with an idealized cell profile,one can apply the effect of the optical system(in this case the optical point-spread-functionplus the finite pixel size of the imaging system),plus an additional dose of simulated noise thatemulates the noise seen in the actual images, tocreate a pseudo-cell image that for our purposescannot be distinguished from a real cell image.When an analysis algorithm is applied to thisimage, its output can be compared with theinput to the simulation to yield a residual plotof the expected value minus the measured valueas a function of the measured parameter, asshown in Figure 3B. As seen in the figure, thisanalysis has quantified the degree of the nega-tive correlation between a threshold parameterused in the membrane finding algorithm andtwo measures of cell size. This immediately tellsthe investigator that, depending on the thresh-old parameter used, the measurement mayhave a systematic error of up to four pixels.The measurement only falls into line with theinput value to the simulation for fairly low val-ues of the threshold. If these threshold values aretoo low for reliable cell finding, one option is touse a higher threshold, and then correct thelength and width measurements with a factordetermined from the simulation. The nextstep is to run a simple Monte Carlo simulationwhere numerous simulated cells are generatedwith varying sizes and amounts of noise. A his-togram of the residual values determined fromthe difference between expected and measuredvalues (with corrections) will produce aGaussian-like distribution, the width of whichcan be taken to be a lower bound on the system-atic error because of the optical properties of thesystem and the measurement algorithm itself.Figure 3C shows the result of such a MonteCarlo simulation for 400 randomly producedcells. The residual histograms show that sub-pixel resolution can be achieved if parameter-dependent algorithms are properly tuned.Without the adjustments determined from cell-image simulation, systematic errors larger thanthe inherent resolution of the optical systemcan be unwittingly introduced. Finally, the al-gorithm can be checked for accuracy with a

standard of known range in diameters such asthose given by commercially available polystyr-ene beads (Werner et al. 2009).

Localized and DelocalizedFlorescent Signals

Properly imaged fluorescent signals from GFP-protein fusions have excellent signal-to-noisecharacteristics. It is often a simple matter tofind the fluorescent maxima and, with criteriabased on proximity of the peaks to the struc-tures determined by the membrane profile,assign the fluorescence to the correct cell asshown in Figure 2B. If point-like in origin, thefluorescent signal will look like a well definedtwo-dimensional Gaussian peak with a widthdefined by the point spread function of the opti-cal system, as shown in Figure 2C. Knowledge ofthe underlying source of the fluorescent signalputs powerful constraints on the signal andhelps to eliminate spurious pixel noise (withwidths much smaller than the PSF) (Szelezniak2007) and out-of-focus channels or fluores-cent contamination or aggregates (with widthsmuch larger than the PSF). Slightly less simpleto determine is the level of the localized fluores-cent signal, and substantially more difficult todetermine is the level or gradient across thecell of nonlocalized signal. When one signalunderlies another it is often hard to disentanglethem because one may be a background contri-bution to the other. To make the analysis rigor-ous we recommend renormalizing the imagedata, scaling it between zero and one, where abaseline of zero is set to the value of the mini-mum-intensity pixel and a normalized intensityof one was assigned to the brightest pixel. Thescale factors and offsets should be preserved sothat after analysis the image can be translatedback into the original fluorescence units. Cellimage simulation is again useful here for devel-oping a well understood algorithm that willyield a measurement with reasonable systematicerrors. To characterize the amplitude of thelocalized signal one can fit a two-dimensionalGaussian to the data with starting parametersfor a nonlinear search algorithm given by(1) starting XY coordinates given by the peak





centroid, (2) a starting width determined bythe PSF, and (3) a starting amplitude given bythe amplitude of the brightest pixel in the fluo-rescent peak. The output from the fit can becompared with a simulated point like lightsource and compared as a method for tuningthe peak finding algorithm. In this way it is pos-sible to obtain subpixel resolution on the XYcoordinate of the center of the peak. The abso-lute amplitude of the localized signal is moredifficult and remains dependent on the amountof background attributed to delocalized, cyto-plasmic signal and whether it contributes tothe localized signal in an additive way. Wehave tested various techniques via cell imagesimulation to account for the delocalized signaland have found one that works reasonably well.The technique involves counting the number ofdiscrete “found” regions inside the cell boun-dary in the (unfiltered) fluorescent channel ofinterest, as a continually varying threshold isapplied. It is observed that the threshold yield-ing the maximum number of found regionscorresponds well with the median of the cyto-plasmic contribution to cell fluorescence.

STAGE TWO: ENSEMBLE LEVELMEASUREMENTS

Cell Cycle Example

We have divided image analysis into an imageprocessing stage where the individual cells areparameterized, and a second stage where en-semble averages are made and the cells are char-acterized biologically. We have mentioned theadvantages of doing this from a processingpoint of view, the second stage being muchless CPU intensive than the first. Anotheradvantage is procedural; we have left the diffi-cult technical task of reducing the cell imageto a set of numbers and can now concentrateon the biology. This stage of the analysis is oftenrepeated and tuned numerous times during thecourse of an analysis, with input from variouscollaborators, so it is useful not to have to rean-alyze all of the images every time someonehas a good idea about what to measure. Whatis measured at the ensemble level depends on

the experiment being performed and is thusoften unique. For example, if one were perform-ing a cell cycle time-course experiment on polarprotein abundance and localization withsynchronized cells, it would be reasonable toplot the peak polar signal averaged over anensemble of high quality cells for each timepoint in the cell cycle. An example of this typeof analysis is shown in Figure 4 for a Caulobactercrescentus cell cycle experiment. The valuesshown are from ensemble averages over manycells and the error bars reflect the error in thosemeasurements. Although it took several hun-dred seconds to parameterize all of the imagedcells during the first stage of this analysis, thesecond stage, involving selection of high qualitydata, determining the mean of the peak fluores-cent signals in the three channels and plottingthe data, takes only seconds.

Mutant Screening Example

A large-scale mutant screen provides a morechallenging example. In this case it is necessaryto process images from thousands of mutants,and then classify them on the basis of severalcriteria such as polar-localized signal amplitudeand location, cytoplasmic-delocalized signalamplitude, localized protein polarity, as wellas a host-cell shape parameters. As an examplewe show in Figure 5 the localized fluorescenceamplitude signals from a triply labeled strainof Caulobacter crescentus used in a highthroughput screen for mutants that mislocal-ized one or more of the reporter proteins.Each data point is an ensemble average over anumber of cells whose peak signal is in thehigh intensity tail of the peak signal distribu-tion. The data shows the distinct differencebetween reporter signals resulting from trans-poson disruptions in the PodJ and TacA genesand the unmutagenized control strains. InFigure 5 we can simultaneously observe three-dimensional phenotypic data, labeled by ge-notype. However, incorporating more fluores-cence based phenotypic information requiresanalysis tools capable of dealing with higher-dimensional data. Because we know the genotypeof every mutant analyzed for each fluorescent





phenotype, we can use hierarchical clustering asshown in Figure 6, taking advantage of all thefluorescent data available in the screen. Thecluster shown in Figure 6 represents a robustcluster of mutants that emerged from a muchlarger dendrogram of over 800 mutants. Thegenes identified in the cluster correspond tothe elements of a pathway controlling the relo-calization of the sensor histidine kinase DivJto the stalked pole at the swarmer to stalkedtransition (Biondi et al. 2006; Radhakrishnanet al. 2008; Christen and Fero 2009). Thus, it

is possible to completely automate a fluores-cence microscopy based screen that leads tofunctional gene module identification.

Shape Characterization

The shape of a mutant or non–wild-type cell isoften a critical part of a fluorescence microscopyexperiment. Although one of the easiest thingsto pick out by eye, it is surprisingly difficultto parameterize mutant shapes, because it ishard to anticipate all the shapes one might

DivJ-RFP

PleC-YFP

CFP-CpaE

120 15090300 60

120 15090300 60

Pea

k flu

ores

cenc

e (A

FU

)

Asymmetric cell division

PleC CpaE DivJ

C

A

B

Cell cycle progression (min)

0

200

400

600

800

Figure 4. Polar localized protein concentration as a function of cell cycle. (A) An example of cell cycle dependentprotein localization in Caulobacter. (B) An example single-cell image shows the correspondence with the visiblechanges in localized protein concentration (C) Automated analysis detects time variation in polar fluorescenceintensities of the CpaE-CFP, PleC-YFP, and DivJ-RFP reporter fusions.





encounter. The membrane parameterizationtechniques we have described do a reasonablygood job following a complex membrane profilebut the chances of missing an interesting phe-notype increase as the cells begin to differ mark-edly from what was expected by the algorithmdeveloper. Thus, mutant shape characterizationcan benefit from more global analysis techni-ques. In these approaches, the cell is notassumed to have a particular form that yieldsto parameterization in the way that we havedescribed above. Rather, the cells are interpretedas observations that can be compared either toeach other or to a control set and tested for sim-ilarity, or to a set of alternative models for clas-sification. This process lends itself to the use ofpattern recognition and supervised or unsuper-vised statistical learning techniques. Supervisedtechniques require a training data set and pro-duce a model that can then be applied to a testset for purposes of classification. The canonicalexamples of these approaches are classificationand regression trees and various types of neuralnets (Hastie et al. 2009). In an unsupervisedapproach, no assumption about what is ex-pected is made and the algorithm is expected

to bootstrap its way to finding unique classesbased on the data itself (Heidmann 2005).Image data is not often directly amenable tothese approaches and is usually first reducedin complexity by the computation of numericalrepresentations of the image. After a cell hasbeen thus re-interpreted it can be classified viaone of the supervised or unsupervised ap-proaches. Good examples of morphology anal-ysis are found in Pincus and Jones (Pincus andTheriot 2007; Jones et al. 2009).

CONCLUDING REMARKS

Well-defined prokaryotic fluorescence micro-scopy experiments can now be performed inan almost completely automated way. Once aprocess has been developed, image data acqui-sition, first and second stage analysis and thepresentation of final results can all be donewithout ever requiring the viewing, converting,or measuring of individual images by the biol-ogist. In making this statement we do not which

20004000

60008000

CpaE

0

20004000

60008000

PleC

0

2000

4000

6000

8000

DivJ PodJ

TacA

WTPodJTacA

Figure 5. Ensemble measurement of polar localizedprotein based on individual cell measurements. Thepolar fluorescence intensities of the CpaE-CFP, PleC-YFP, and DivJ-RFP reporter fusions are plotted forthe nonmutagenized control strain (black) as wellas gene disruptions in two separate open readingframes coding for the PodJ (blue) and TacA (red)proteins.

CpaE-CFP PleC-YFP DivJ-RFP

P D MB P D MB P D MB

z-score–5 –2.5 0 2.5 5

CC0138 shkACC3599 rpoNCC3599 rpoNCC0138 shkACC3315 tacACC0138 shkACC3315 tacACC1062 cckNCC3315 tacACC3315 tacACC3315 tacACC3315 tacACC2173 spmXCC2173 spmX

Figure 6. Example of Genotype-Phenotypemicroscopy based analysis. A collection of auto-matically determined cell metrics is used to identifya functional cluster indicating a gene module co-ordinating temporal and spatial protein localization.Heat map representation of the calculated z-scorevalues from localized protein abundance (L),delocalized protein (D), the fraction of cellsidentified as being biopolar (B), and the fractionsimilarly identified as monopolar (M) are shownfor a highly significant gene cluster designated bythe dendrogram.





to trivialize the effort involved in developingautomated acquisition and analysis tools. Abrief reflection on cost vs. benefit will make itclear to the experimenter whether the upfrontinvestment in automated data taking and anal-ysis is required. We have illustrated a seamlesspipeline approach, which requires a minimumamount of user intervention but a maximumof software developer effort. Unless the ex-periment is very high throughput, most ex-periments today can be performed quitesuccessfully using nonautomated image acquis-ition and semiautomated quantitative analysislike those available using stand-alone toolssuch as ImageJ (Girish and Vijayalakshmi2004; Collins 2007). However, as the interestin performing high-throughput assays andscreens at the genomic or meta-genomic level inprokaryotic systems grows, the value of inte-grated, automated systems for data acquisitionand analysis will become evident.

ACKNOWLEDGMENTS

This work was funded by National Institutes ofHealth Grant K25 GM070972-01A2 (M.F.).

REFERENCES

Biondi EG, Skerker JM, Arif M, Prasol MS, Perchuk BS,Laub MT. 2006. A phosphorelay system controls stalk bio-genesis during cell cycle progression in Caulobacter cres-centus. Mol Microbiol 59: 386–401.

Boland MV, Murphy RF. 1999. Automated analysis of pat-terns in fluorescence-microscope images. Trends CellBiol 9: 201–202.

Boland MV, Murphy RF. 2001. A neural network classifiercapable of recognizing the patterns of all major subcellu-lar structures in fluorescence microscope images of HeLacells. Bioinformatics 17: 1213–1223.

Boland MV, Markey MK, Murphy RF. 1998. Automated rec-ognition of patterns characteristic of subcellular struc-tures in fluorescence microscopy images. Cytometry 33:366–375.

Chen X, Murphy RF. 2005. Objective clustering of proteinsbased on subcellular location patterns. J Biomed Biotech-nol 2005: 87–95.

Chen X, Velliste M, Murphy RF. 2006. Automated interpre-tation of subcellular patterns in fluorescence microscopeimages for location proteomics. Cytometry 69: 631–640.

Christen B, Fero M, Hillson N, Bowman G, Hong S-H,Shapiro L, McAdams H. 2010. High-throughput identi-fication of protein localization dependency networks.Proc Natl Acad Sci 107: 4681–4686.

Collins TJ. 2007. ImageJ for microscopy. BioTechniques 43:25–30.

De Boor C. 1978. A practical guide to splines. Springer-Verlag, New York.

Firestone L, Cook K, Culp K, Talsania N, Preston K Jr. 1991.Comparison of autofocus methods for automated micro-scopy. Cytometry 12: 195–206.

Geusebroek JM, Cornelissen F, Smeulders AW, Geerts H.2000. Robust autofocusing in microscopy. Cytometry39: 1–9.

Girish V, Vijayalakshmi A. 2004. Affordable image analysisusing NIH Image/ImageJ. Ind J Cancer 41: 47.

Gregory JA, Becker EC, Pogliano K. 2008. Bacillus subtilisMinC destabilizes FtsZ-rings at new cell poles and con-tributes to the timing of cell division. Genes Develop 22:3475–3488.

Groen FC, Young IT, Ligthart G. 1985. A comparison of dif-ferent focus functions for use in autofocus algorithms.Cytometry 6: 81–91.

Hastie T, Tibshirani R, Friedman JH. 2009. The elements ofstatistical learning : Data mining, inference, and prediction.Springer, New York.

Heidmann G. 2005. Unsupervised image categorization.Image and Vision Computing 23: 861–876.

Huang K, Murphy RF. 2004. From quantitative microscopyto automated image understanding. J Biomed Optics 9:893–912.

Jones TR, Carpenter AE, Lamprecht MR, Moffat J, Silver SJ,Grenier JK, Castoreno AB, Eggert US, Root DE, GollandP, et al. 2009. Scoring diverse cellular morphologies inimage-based screens with iterative feedback and machinelearning. Proc Natl Acad Sci 106: 1826–1831.

Kasson PM, Huppa JB, Davis MM, Brunger AT. 2005. Ahybrid machine-learning approach for segmentation ofprotein localization data. Bioinformatics 21: 3778–3786.

Ohya Y, Sese J, Yukawa M, Sano F, Nakatani Y, Saito TL, SakaA, Fukuda T, Ishihara S, Oka S, et al. 2005. High-dimensional and large-scale phenotyping of yeastmutants. Proc Natl Acad Sci 102: 19015–19020.

Pincus Z, Theriot JA. 2007. Comparison of quantitativemethods for cell-shape analysis. J Micros 227: 140–156.

Price JH, Gough DA. 1994. Comparison of phase-contrastand fluorescence digital autofocus for scanning micro-scopy. Cytometry 16: 283–297.

Radhakrishnan SK, Thanbichler M, Viollier PH. 2008. Thedynamic interplay between a cell fate determinant and alysozyme homolog drives the asymmetric division cycleof Caulobacter crescentus. Genes Dev 22: 212–225.

Szelezniak M. 2007. Cluster finder in the HFT readout.

Werner JN, Chen EY, Guberman JM, Zippilli AR, Irgon JJ,Gitai Z. 2009. Quantitative genome-scale analysis of pro-tein localization in an asymmetric bacterium. Proc NatlAcad Sci 106: 7858–7863.





30, 20102010; doi: 10.1101/cshperspect.a000455 originally published online JuneCold Spring Harb Perspect Biol

Michael Fero and Kit Pogliano Automated Quantitative Live Cell Fluorescence Microscopy

Subject Collection Cell Biology of Bacteria

Electron CryotomographyElitza I. Tocheva, Zhuo Li and Grant J. Jensen

Cyanobacterial Heterocysts

James W. GoldenKrithika Kumar, Rodrigo A. Mella-Herrera and

Protein Subcellular Localization in BacteriaDavid Z. Rudner and Richard Losick Cell Division in Bacteria

Synchronization of Chromosome Dynamics and

Martin Thanbichler

Their Spatial RegulationPoles Apart: Prokaryotic Polar Organelles and

Clare L. Kirkpatrick and Patrick H. ViollierMicroscopyAutomated Quantitative Live Cell Fluorescence

Michael Fero and Kit Pogliano

MorphogenesisMyxobacteria, Polarity, and Multicellular

Dale Kaiser, Mark Robinson and Lee KroosHomologsThe Structure and Function of Bacterial Actin

Joshua W. Shaevitz and Zemer GitaiMembrane-associated DNA Transport Machines

Briana Burton and David DubnauBiofilms

Daniel López, Hera Vlamakis and Roberto KolterThe Bacterial Cell Envelope

WalkerThomas J. Silhavy, Daniel Kahne and Suzanne III Injectisome

Bacterial Nanomachines: The Flagellum and Type

Marc Erhardt, Keiichi Namba and Kelly T. HughesCell Biology of Prokaryotic Organelles

Dorothee Murat, Meghan Byrne and Arash Komeili Live Bacteria CellsSingle-Molecule and Superresolution Imaging in

Julie S. Biteen and W.E. Moerner

SegregationBacterial Chromosome Organization and

Esteban Toro and Lucy Shapiro

http://cshperspectives.cshlp.org/cgi/collection/ For additional articles in this collection, see

Copyright © 2010 Cold Spring Harbor Laboratory Press; all rights reserved


http://cshperspectives.cshlp.org/cgi/collection/

http://cshperspectives.cshlp.org/cgi/collection/

http://cshperspectives.cshlp.org/cgi/adclick/?ad=55917&adclick=true&url=http%3A%2F%2Fwww.genelink.com


automated quantitative live cell fluorescence microscopy

Documents