Automated Subcellular Location Determination and High-Throughput MicroscopyIntroductionApproaches to Systematic Analysis of Subcellular LocationKnowledge CapturePredictionDeterminationImage Acquisition Considerations for Automated AnalysisLabelingMagnification and ResolutionDimensionChannelsNumber of ImagesSubcellular Location Feature ExtractionSubcellular Object FeaturesSingle-Cell FeaturesField FeaturesTemporal FeaturesMajor Computational Questions in Subcellular Pattern AnalysisStatistical Tests: ComparisonSupervised Learning: ClassificationUnsupervised Learning: ClusteringExtending Single-Cell Methods to TissuesDatabase versus File System ApproachesConclusionsReferencesDevelopmental CellReviewAutomated Subcellular Location Determinationand High-Throughput MicroscopyEstelle Glory1and Robert F. Murphy1,*1Center for Bioimage Informatics, Molecular Biosensor and Imaging Center, and Departments of Biological Sciences, BiomedicalEngineering and Machine Learning, Carnegie Mellon University, 4400 Fifth Avenue, Pittsburgh, PA 15213, USA*Correspondence: [email protected] 10.1016/j.devcel.2006.12.007Dramatic advances in methods for protein tagging and the development of fully automated micro-scopes enable collection of unprecedented volumes of image data on the subcellular location of pro-teins in live cells. Combining these approaches with machine learning methods promises to providesystematic, high-resolution pattern information on a proteome-wide basis.IntroductionTo understand the intricate pathways that regulate biolog-ical processes at the cellular level, we need to be able tocapture data about the subcellular distributions of proteinsand how these vary within cell populations. Automatedanalysis of fluorescence microscope images providesa powerful way of acquiring such information. The highspecificity of fluorescent probes for labeling componentsof interest and the availability of advanced light micro-scopes permit high spatial and temporal resolution imag-ing of living cells. The determination of accur ate protein lo-cation provides valuable information for understanding themolecular mechanisms that underlie the functions of cells(Ehrlich et al., 2002; Suzuki et al., 2002; Wu et al., 2003).Knowledge of the localization of proteins within cellularcompartments is critical to understanding their functionfor many reasons. Each compartment is defined by its ownchemical and physical characteristics, such as the acidicpH in the lysosome, the viscoelasticity of the cytoskeleton,or the hydrophobicity of membrane. Thus, location canprovide useful information for improving predictions of pro-tein conformation. Besides, since organelles are the loca-tion of specialized functions in the cell, such as oxidativemetabolism in mitochondria, transcription of ribosomalRNA in nucleoli, and maturation of newly synthesized pro-teins in the endoplasmic reticulum, the determination ofsubcellular location for a protein can yield hypothesesabout the metabolism in which it is involved and the pro-teins with which it interacts. Changes in location overtime are also critical to cell behavior. For example, in a sig-nal transduction pathway, the transportation from the cy-toplasm to the nucleus induced by the activation of theprotein is characterized by the location of the protein be-fore and after its activation, the activation moment, andits duration. Lastly, once the subcellular distribution of aprotein is defined for healthy adult cells, comparison withdiseased or developing cells can yield important insightsthat can lead to improved diagnostics and therapeutics.Information on the subcellular location of proteins is in-creasingly being collected in parallel for large numbers ofproteins (Hoja et al., 2000; Jarvik et al., 2002; Korolevaet al., 2005; Rolls et al., 1999; Simpson et al., 2000) or evenfor entire proteomes (Huh et al., 2003). As for many previ-ous studies of individual proteins, the primary means ofanalyzing and annotating images depicting subcellular lo-cation in these large-scale studies has been visual exam-ination. Over the past decade, however, the feasibility ofusing machine learning methods to automate the determi-nation of subcellular location from fluorescence micro-scope images has been demonstrated convincingly (Bo-land et al., 1997, 1998; Boland and Murphy, 2001; Huangand Murphy, 2004b). In fact, these methods can performbetter than visual examination (Murphy et al., 2003). Overthe same time period, automated systems for performingcell-based assays were developed and used by pharma-ceutical companies to screen for drugs with desired effects(Taylor et al., 2001; Zhou and Wong, 2006). These systems,variously referred to as high-content screening or high-throughput microscopy systems, are increasingly beingused for basic research on biological pathways (Pepperkokand Ellenberg, 2006; Perlman et al., 2004; Price et al., 2002;Sigal et al., 2006; Starkuviene et al., 2004; Yarrow et al.,2005). This article reviews the methods currently availablefor automated, large-scale determination of the intracellu-lar location of fluorescent-labeled molecules within cells.Approaches to Systematic Analysis of SubcellularLocationProteins can display highly specialized locations withincells, such as being present in just the mitochondrial innermembrane, just the rims of a particular Golgi cisterna, orjust specific regions of a chromosome. However, the num-ber and resolution of locations considered in subcellularlocation classification varies greatly between studies.The simplest location studies are interested in three com-partments (nucleus, cytoplasm, and extracellular environ-ment), while more accurate studies have considered 7 to22 different subcellular structures (often depending onthe organism).Knowledge CaptureSystematic efforts to catalog the subcellular locationsof proteins typically use either knowledge capture ordata-driven approaches. The first seeks to collect and or-ganize information on location that has been collectedDevelopmental Cell 12, January 2007 ª2007 Elsevier Inc. 7over many years and published in the archival literature. Acritical starting point for these efforts is the developmentof a standard vocabulary to describe location. Whilemany vocabularies have been used over the years, thecreation of the Cellular Component ontology by theGene Ontology (GO) Consortium (Harris et al., 2004 ) hashad a major impact. This ontology describes locations atthe levels of subcellular structures