Image analysis of GFP-tagged protein localization bursts

Image analysis of GFP-tagged protein localization bursts

We are searching data for your request:

Forums and discussions:
Manuals and reference books:
Data from registers:
Wait the end of the search in all databases.
Upon completion, a link will appear to access the found materials.

I'm reading an article (full text here) that analyze the dynamics of localization of a GFP-tagged transcription factor (Crz1) over the time at the single-cell level, by taking movies in a fluorescent microscope.

In Methods section they say:

Fluorescence cell images were segmented using a Hough transformation algorithm in Matlab, provided by Sharad Ramanathan. Localization score was determined by the difference between the mean intensity of the 5 brightest pixels in the cell and mean intensity of the rest of the pixels in the cell.

The segmentation process here seems to be the identification of cells over the background. They then calculate a localization score, for each frame of the video, for every cell. Now there's the part that I can't understand:

Bursts were identified by thresholding traces at >1 standard deviations above background noise, estimated from the lowest 20% of values.

I searched some definitions of "background noise", but I can't figure out what does it mean in this particular context. Moreover, "lowest 20% of values" of what?

Is it plausible that they define it for the lowest 20% of values of localization scores over the time, at the cell each time considered?

Maybe can be useful a screenshot of a single cell in a photogram of the video:

Yep, the Hough Transform is a way to pick out shapes you're interested in, in this case they probably have it set to find circles, and they use that to segment the image.

I think that you have interpreted their methods correctly. For each cell they make a trace of localization score vs. time, localization score defined in arbitrary units as the difference between the mean of the five brightest pixels and the mean of the remaining pixels in the cell. I think the lowest 20% refers to the frames in the video that have the lowest 20% of localization scores. They take the lowest 20% of localization scores, calculate a standard deviation, and then for any frame that has a localization score that is more than 1 standard deviation above the mean of those 20%, you say that frame exhibits a burst of localization. If I'm understanding this correctly, this process would be repeated on every individual cell.

My interpretation is that the 20% doesn't have anything to do with the signal intensity from the background pixels, and it comes from analyzing the series over time, not a single image.

I don't know a whole lot about signal processing but I am a little familiar with artificial intelligence. Perhaps this wiki would be helpful I am familiar with k-means clustering discussed there and as a trivial segmentation method it would indeed identify the cell versus the background of the photo. The Hough transformation would be more sophisticated and probably more useful for this application but if you want to get a sense of the process the naive k-means algorithm may be helpful.

I interpret the "lowest 20% of values" to refer to the localization score values, the lowest being those who are darkest in your image. That is, the background the cell is imaged on. The background noise refers to the fact that a signal processing algorithm without any preprocessing may try to identify patterns from the background of the image rather than the cell subject matter, which is the part the researchers are after. This is why they perform the segmenting.

For instance, imagine a k-nearest neighbors algorithm for computing a "localization score" for image brightness. Your paper is interested in the intensity of GFP for a cell. In the image, pixels on the border of the cell with its background will have artificially low score values because of the background: the adjacent pixels that are in the background and not part of the cell are dark, but that doesn't mean anything related to the biology of the problem. This is the noise.

The lowest 20% of values, then, should refer to a single image, not to a series over time.

Dissecting DNA damage response pathways by analysing protein localization and abundance changes during DNA replication stress

Relocalization of proteins is a hallmark of the DNA damage response. We use high-throughput microscopic screening of the yeast GFP fusion collection to develop a systems-level view of protein reorganization following drug-induced DNA replication stress. Changes in protein localization and abundance reveal drug-specific patterns of functional enrichments. Classification of proteins by subcellular destination enables the identification of pathways that respond to replication stress. We analysed pairwise combinations of GFP fusions and gene deletion mutants to define and order two previously unknown DNA damage responses. In the first, Cmr1 forms subnuclear foci that are regulated by the histone deacetylase Hos2 and are distinct from the typical Rad52 repair foci. In a second example, we find that the checkpoint kinases Mec1/Tel1 and the translation regulator Asc1 regulate P-body formation. This method identifies response pathways that were not detected in genetic and protein interaction screens, and can be readily applied to any form of chemical or genetic stress to reveal cellular response pathways.


Epithelial cells cover the external and internal surface of the vertebrate body and are instrumental in maintaining homeostasis by separating distinct compartments of the body. Apical cell–cell junctions consist of tight junctions (TJs), adherens junctions (AJs), and desmosomes. AJs and desmosomes mechanically connect adjacent epithelial cells and contribute to maintenance of cell shape and tissue integrity (Hartsock and Nelson, 2008 Nekrasova and Green, 2013 Takeichi, 2014 Lecuit and Yap, 2015). TJs regulate the passage of fluids and solutes via the paracellular pathway and serve as a barrier (Hartsock and Nelson, 2008 Krug et al., 2014 Van Itallie and Anderson, 2014 Zihni et al., 2016).

Because epithelial tissues are continuously renewed, new cells must be generated by cell division, which is especially evident in the developing epithelium. Despite the drastic cell shape changes that occur during cytokinesis, cell–cell junctions must maintain cell–cell adhesion and barrier function during cell division. Although an understanding of how cell–cell junctions are maintained during cytokinesis is beginning to emerge (Higashi et al., 2016), how epithelial cells distinguish and coordinate the signaling mechanisms regulating contractile actomyosin arrays at cell–cell junctions and the cytokinetic contractile ring remains unclear.

Both cell–cell junctions and cytokinetic contractile rings are regulated by Rho GTPases (Kishi et al., 1993 Mabuchi et al., 1993 Nusrat et al., 1995 Braga et al., 1997 Miller, 2011 Arnold et al., 2017). RhoA switches between an active GTP-bound state and an inactive GDP-bound state. When RhoA is in the active GTP-bound state, it binds to and activates its effectors, including ROCKs/Rho kinases (Ishizaki et al., 1996 Matsui et al., 1996) and formins (Kohno et al., 1996 Watanabe et al., 1997 Alberts et al., 1998). It is not clear how RhoA effectors differentially regulate formation and maintenance of both RhoA-dependent junctional actomyosin bundles and cytokinetic actomyosin rings within the same cells. To address this question, we investigated the localization and functional roles of formins at epithelial cell–cell junctions and cytokinetic contractile rings in a developing vertebrate model system, the gastrula-stage Xenopus laevis embryo.

Formins constitute a family of actin regulators that is conserved among eukaryotes (Higgs and Peterson, 2005 Rivero et al., 2005 Chalkia et al., 2008). Formins mediate linear actin assembly through their formin homology (FH) 1 and FH2 domains. The FH1 domain recruits profilin-bound actin monomers and passes them to the FH2 domain. The FH2 domain directly binds to and caps the barbed end of actin filaments and simultaneously adds new actin monomers to the barbed end, which results in continuous elongation of F-actin at the barbed end (Pruyne, Evangelista, et al., 2002 Kovar et al., 2003). Vertebrate genomes have at least 15 formins (Higgs and Peterson, 2005 Rivero et al., 2005 Chalkia et al., 2008) (see Supplemental Figure S2). Among them, 10 formins (Dia1/2/3, Daam1/2, Fmnl1/2/3, Fhod1/3) are classified as Diaphanous-related formins (DRFs) (Alberts, 2002 Kuhn and Geyer, 2014). DRFs share several important regulatory domains in addition to the FH1 and FH2 domains. Binding of the Diaphanous inhibitory domain (DID) (Li and Higgs, 2005), which is located on the N-terminal (NT) side of the FH1/FH2 domains, to the Diaphanous autoinhibitory domain (DAD) (Alberts, 2001), which is located on the C-terminal (CT) side of the FH1/FH2 domains, keeps the actin-assembly activity of FH1-FH2 domains suppressed (Watanabe et al., 1999 Li and Higgs, 2003). At their NT end, DRFs have a GTPase-binding domain (GBD) (Watanabe et al., 1997 Otomo et al., 2005a Rose, Weyand, Lammers, et al., 2005). Binding of active Rho GTPases to the GBD (and part of the DID) releases the DID-DAD autoinhibitory interaction (Watanabe et al., 1999 Lammers et al., 2005 Nezami, Poy, and Eck, 2006). Additional factors can cooperate to release DID-DAD interactions, including Anillin binding to DID for mDia2 (Dia2, also known as DIAPH3, Diap3, or DRF3) (Watanabe et al., 2010), Flightless-I binding to DAD for mDia1 and Daam1 (Higashi, Ikeda, et al., 2010), and phosphorylation of DAD by ROCK for Fhod1 (Takeya et al., 2008) and mDia2 (Staus et al., 2011). Unleashing DID-DAD autoinhibition opens up DRF molecules, making the FH1-FH2 domains accessible to bind actin.

Formins have been implicated in the regulation of cell–cell junctions (for a review, see Grikscheit and Grosse, 2016). For example, mammalian Dia1 (mDia1, also known as DIAPH1 or DRF1) has been shown to localize at AJs and regulate the stability and contractility of AJs in many cell types (Sahai and Marshall, 2002 Carramusa et al., 2007 Ryu et al., 2009 Rao and Zaidel-Bar, 2016 Acharya et al., 2017). Fmnl3 is implicated in the regulation of AJs through F-actin polymerization and stabilization of E-cadherin in migrating mouse mammary epithelial EpH4 cells (Rao and Zaidel-Bar, 2016). A three-dimensional culture model of human breast epithelial MCF10A cells showed that Fmnl2 was involved in the formation of new cell–cell junctions between daughter cells downstream of Rac1 (Grikscheit et al., 2015). Finally, several studies indicate a role for formins in regulating cell–cell adhesion downstream of Rho during developmental processes. Drosophila Diaphanous regulates junctional Myosin II levels and activity and is required for properly regulated junctional stability and cell movements during morphogenesis (Homem and Peifer, 2008). Drosophila Diaphanous can also control E-cadherin endocytosis downstream of Rho, thus regulating the level of E-cadherin at the cell–cell junction (Levayer et al., 2011). Additionally, actin-based pushing controlled by Fmn1 acting downstream of RhoA drives apical emergence of new multiciliated epithelial cells in developing X. laevis embryos (Sedzinski, Hannezo, et al., 2016) however, how this specialized actin network is linked to junctions and whether Fmn1 regulates cell–cell junctions in this setting is not clear.

Formins are also known regulators of cytokinesis (for a review, see Bohnert et al., 2013). Fission yeast formin Cdc12 is concentrated at medial nodes and mediates formation and maintenance of contractile rings (Chang et al., 1997 Kovar et al., 2003 Wu et al., 2006). In budding yeast, two formins, Bni1p and Bnr1p, are required for successful cytokinesis (Imamura et al., 1997 Tolliday et al., 2002). Caenorhabditis elegans CYK-1 and Drosophila Diaphanous are required for early embryonic divisions (Castrillon and Wasserman, 1994 Severson et al., 2002). Although mDia1, 2, and 3 are all orthologues of CYK-1 and Diaphanous, only one vertebrate formin, mDia2, has been shown to control cytokinesis. mDia2 is localized at contractile rings in mouse NIH 3T3 fibroblasts, and knockdown of mDia2 caused cytokinesis failure in NIH 3T3 cells (Watanabe et al., 2008). Additionally, mDia2 knockout mice are embryonic lethal due to cytokinesis failure in fetal erythroblasts, which results in severe anemia (Watanabe et al., 2013). Because the nomenclature of Dia group formins is frequently confused between human and mouse genes (e.g., the human orthologue of mouse mDia2 [DIAPH3] is called hDIA3, DRF3, and DIAPH3), we consistently use Dia1 (mDia1 in mice, DIAPH1 in humans), Dia2 (mDia2 in mice, DIAPH3 in humans), and Dia3 (mDia3 in mice, DIAPH2 in humans) for X. laevis genes in this paper.

To date, there has been no comprehensive study of all 15 vertebrate formins in the same model system. Furthermore, it is unclear whether any formin(s) are involved in the regulation of both cell–cell junctions and cytokinetic contractile rings, or whether these two actomyosin-based structures actively influence each other through the regulation of formin proteins. Here, we cloned the 15 formins from X. laevis and characterized their localization in epithelial cells. We identified Dia1 and Dia2 as cell–cell junction localizing formins and found that perturbing the junctional localization of Dia1 and Dia2 resulted in a cytokinesis defect.


BacStalk is designed for label-free bacterial cell and stalk detection in phase-contrast images at pixel accuracy. As stalks usually have very low contrast and are hardly visible, conventional thresholding approaches fail to separate them reliably from the image background. BacStalk overcomes these difficulties by implementing a two-step approach: First, cells are identified by enhancing image features of a typical length scale by applying bandpass filtering, followed by automatic thresholding. In the second step, BacStalk detects connected stalks by performing local morphological operations. In detail, a shell is constructed around each cell, consisting of all pixels surrounding the cell that are positioned at a user-defined distance away from the cell. The intensity values of all pixels of this shell are then scanned for subtle intensity differences (either in the phase-contrast or bright-field image or in any desired fluorescence channel) that could indicate a potential stalk attachment point (Figure 1a). For each pixel of this shell, the intensity is compared to the mean intensity of all pixels of this shell, by computing the z-score (i.e., the number of standard deviations that an intensity value deviates from the mean shell intensity). If the lowest z-score in phase-contrast images (or highest z-score in fluorescence images) exceeds a user-defined threshold value, the pixel with this z-score is defined as the stalk attachment point. The stalk backbone is then generated through repeated dilation starting from the stalk attachment point, by looking for other pixels in a user-defined range that are again below (or above, for fluorescence) a defined z-score (Figure 1a). Thus, if present, the stalk is constructed from the stalk attachment point outwards in a directional manner. If another cell is encountered during stalk propagation, both the initial cell and the touched counterpart are defined as related and can be treated as a connected structure (mother cell and bud) during cell tracking and the downstream analysis. The larger cell is, hereby, defined as a mother cell. In cells and buds of stalked budding bacteria, the cell polarity is clearly defined by the stalk attachment point and indicated by a yellow dot at the end of the cell's medial axis (Figure 1a). In the case of stalk-free swarmer cells, the identity of the cell pole is guessed based on the cell morphology or the existence of intensity features, but it can also be interactively changed manually at any stage of the analysis. As unambiguous stalk assignment in clumps of multiple cells is typically impossible, cell clumps are automatically recognized and excluded from further analysis. The approach described above is relatively insensitive to uneven background illumination in the microscopy images and does not require any background correction.

As an alternative to BacStalk's inbuilt cell detection algorithm, BacStalk can also import the segmentation masks from other software packages, enabling the use of BacStalk's analysis features for images with custom segmentation requirements. BacStalk's cell and stalk detection process are relatively fast, requiring approximately 4.5 s per image (for 40 cells in a 2048 × 2048 pixel image using a computer with an Intel i7-6700K processor and 16 GB memory). On computers with multiple CPU cores, BacStalk can parallelize computing tasks to reduce processing time and to facilitate high-throughput analysis. Batch analyses of images can benefit from the possibility to differentiate between different strains or growth conditions using the custom metadata option during the import of images, thus allowing the analysis of multiple strains at the same time using the same settings.

To test the functionality of BacStalk, we first compared the results of its cell segmentation algorithm to that of MicrobeJ (Ducret et al., 2016 ) and Oufti (Paintdakhi et al., 2016 ) by quantifying the cellular dimensions of 105 C. crescentus cells as detected by each of these software packages (Supporting Information Table S1). This analysis showed some segmentation algorithm-specific differences in the cell length and width values but very similar standard deviations.

We then tested BacStalk on several species of stalked bacteria to verify the robustness of the stalk detection algorithm. The algorithm performed well for C. crescentus (Figure 2a), Brevundimonas aveniformis (Figure 2b), and H. neptunium (Figures 3 and 4). To examine the accuracy of the stalk detection algorithm, we used a C. crescentus strain in which the terminal segments of the stalk are fluorescently labeled by a GFP-tagged version of the stalk-specific protein StpX (Hughes et al., 2010 ). When correlating the stalks identified by BacStalk based on phase-contrast images in the presence of StpX-GFP signals (Supporting Information Figure S1), we observed that stalks were correctly identified for 91% of the cells (n = 616) without modifying the default image segmentation settings, indicating that the stalk detection algorithm works robustly. Most of the misidentifications occurred for cells that were slightly out of focus (only cells that were considerably out of focus had been excluded from the analysis manually) or in cases where the stalk was very small, clearly below the 5 pixels defined as the default minimum stalk length. To investigate whether stalks of different lengths are detected reliably, we imaged C. crescentus cells grown in the presence or absence of phosphate, as it has previously been shown that C. crescentus cells strongly elongate their stalk upon phosphate starvation (Schmidt and Stanier, 1966 ). The analysis of 500 cells grown in these two conditions verified that stalks of cells grown with phosphate were indeed considerably shorter and less variable in length (1.47 ± 0.75 µm) than stalks of cells deprived of phosphate (6.45 ± 2.90 µm) (Figure 2a). Interestingly, we found that in the latter condition the cell length correlated at least partially with the stalk length. Stalk detection also works in very low contrast conditions in which the stalk is hard to distinguish from the background by eye, as generally observed for B. aveniformis (Figure 2b). For this species, BacStalk identifies 60 ± 5% of all cells as stalked (n = 750, from five independent experiments), which is comparable to the stalked population fraction measurement of 46 ± 4% reported by a previous study (Curtis, 2017 ). Furthermore, our analysis showed that cells with stalks display on average slightly longer cell bodies (2.47 ± 0.56 µm) than cells without stalks (2.05 ± 0.45 µm).

In addition to stalks, the stalk detection algorithm can also identify other single polar cell appendages, such as flagella (Figure 2c). Since these structures are often not visible in phase-contrast or bright-field images, the algorithm can be applied to stained appendages visualized by fluorescence microscopy. BacStalk was able to reliably detect flagella of Shewanella putrefaciens cells that were labeled with a fluorescent dye (Figure 2c) (Kühn et al., 2017 ). This analysis showed that the flagellar lengths vary considerably between cells (from 0.33 to 6.88 µm, with an average of 3.35 ± 1.87 µm).

In case of the stalked budding bacterium H. neptunium, BacStalk can accurately distinguish between mother cells, stalks, and buds for the wild type as well as for mutants with altered morphology (Figure 3). Thus, BacStalk quantifies phenotypes in a more accurate, detailed, and faster manner than the current standard in the field (i.e., quantification via manual measurements). The power of such automated quantifications is illustrated by the fact that the morphological phenotype of an H. neptunium Δpbp1x mutant (i.e., an increased length of budding cells), which was described qualitatively before (Cserti et al., 2017 ), can now be precisely quantified (Figure 3): the combined length of mother cells and stalks was 3.2 ± 1.3 µm long for Δpbp1x cells and 2.7 ± 0.7 µm for wild-type cells (both n = 183). The quantitative analysis by BacStalk also provided additional new insights. For instance, BacStalk revealed that the distribution of stalk lengths is dramatically broadened in the Δpbp1x background and that the bud area correlates partially with the total length of the concatenated entity of mother cell, stalk, and bud.

The combination of features offered by BacStalk also facilitates detailed analyses of protein localization experiments based on fluorescence signals. To this end, BacStalk can detect the brightest focus inside cells after applying a 3-by-3 mean filter to the cell. The distance to the pole of this brightest fluorescence spot is then measured along the medial axis. By default, no fluorescence normalization or background subtraction is performed during the analysis, yet the user may choose whether to subtract the background or normalize fluorescence during the data visualization. To visualize protein localization, BacStalk can display combined intensity profiles of the mother cell, stalk, and bud, which can be aligned according to relevant cellular landmark locations and sorted by any measured property of the combined structure (Figure 1b). To generate intensity profiles along the medial axis of a cell, BacStalk fits a mesh of evenly spaced lines into each cell perpendicular to its medial axis, similar to Oufti (Paintdakhi et al., 2016 ). For each point on the medial axis, the mean or maximum of the intensity values along the corresponding line is calculated to obtain smooth intensity profiles along the cell medial axis. In addition, the mesh is used to perform a medial axis coordinate system transformation in order to reorient curved cells such that they can be arranged next to each other in two-dimensional (2D) demographs and kymographs to conserve the full spatial information of patterns inside the cells (Figures 2a, 4b,c, 5b). These 2D demo- and kymographs intuitively visualize the imaging data and provide important additional information in cases where the fluorescent protein of interest is located away from the medial axis. Furthermore, the fluorescence intensity profiles can be normalized and background-corrected per cellular entity.

BacStalk provides a very flexible data visualization environment for protein localization experiments. The 1D and 2D demo- and kymographs, the use of concatenated intensity profiles, the possibility to align these profiles based on the morphologically relevant criteria (e.g., alignment at the junction between the mother cell and stalk, or flagellum, or at a specific pole), and the option to display specific subsets of cells (i.e., swarmer cells, stalked cells without buds, and budding cells) are instrumental in understanding the localization behavior of proteins in different cell types in a mixed population. This is exemplified by an analysis of the localization dynamics of the histidine kinase CckA in H. neptunium using BacStalk, which verified the cell cycle-dependent localization previously observed in a qualitative manner (Leicht et al., 2020 ). BacStalk provides a means to automatically plot the intracellular localization of CckA-Venus in demographs separately for swarmer cells, stalked cells without buds, and budding cells (Figure 4a). All demographs were sorted according to the length of the cellular structures: the first demograph is sorted by the length of the cell, the second by the combined lengths of the mother cell and stalk, and the third by the combined lengths of the mother cell, stalk, and bud. Cells were aligned at the cell center, the cell-stalk junction, or the bud pole opposite to the stalk, respectively.

Apart from its powerful visualization tools, BacStalk includes special features to track individual cells in time-lapse experiments and to analyze dynamic protein localization during stalk-terminal budding. In time-lapse experiments, cells that display an overlap in position in consecutive frames are identified as the same cell and the lineage information of cells is saved. As an example, we reinvestigated the cell cycle-dependent localization of the guanylate cyclase PleD in H. neptunium cells (Figure 4b). The analysis by BacStalk provides a detailed quantification of the cell type-specific localization pattern that has previously only been described in a qualitative manner (Jung et al., 2015 ). In swarmer cells, PleD-Venus localizes at the flagellated pole of the mother cell. In stalked cells, the PleD-Venus focus also tends to be located at the flagellated pole of the mother cell, whereas in most budding cells it is detected in the bud at the pole opposite the stalk. Although the relocation of PleD from the old pole of the mother cell to the old pole of the daughter cell can be traced in a 1D kymograph, the 2D kymograph additionally facilitates the correlation of protein relocation with cell morphogenesis (e.g., bud formation). Quantification of the localization of PleD-Venus in different cell types is also possible by determining and plotting its distance from the old pole of the mother cell (Figure 4b).

To investigate the patterns of multiple fluorescence signals at different wavelengths simultaneously, BacStalk can create multi-channel kymographs (or demographs): the kymograph in Figure 4c shows H. neptunium cells undergoing replication, in which (a) the replisome component DnaN is tagged with the fluorescent protein Venus and (b) the origin of replication is followed with the help of a ParB-Cerulean fusion, which binds parS sites near the chromosomal origin of replication (Jung et al., 2019 ). This two-color approach corroborates the relative timing of replisome movement and origin segregation revealed previously by manual analysis (Jung et al., 2019 ), and verifies that the origin of replication already moves to the bud before the replication is completed (as visualized by the delocalization of DnaN-Venus). In addition, the 2D kymograph visualization clearly identifies both replication forks as separate entities (distinct DnaN foci inside the first cell in kymograph, Figure 4c). As in Figure 4b, we determined the distance of the ParB-Cerulean and DnaN-Venus foci to the old pole of the mother cell. Using this representation, Figure 4c confirms that the replication origin (tagged by ParB-Cerulean) is only transferred to the bud once a certain bud size has been reached and that the process of origin movement through the stalk must be fast, as the origin was captured inside the stalk in only

1% of all cells that were analyzed (4 out of 378 cells).

BacStalk provides several analysis tools that greatly simplify data exploration and visualization. Customizable cell measurements can be added by users and are listed, together with all other measured parameters, in an interface that allows for filtering and the identification of specific sub-populations of interest. The measured data for each cell can be exported to several standard formats. Similar to MicrobeJ (Ducret et al., 2016 ), all plots created with BacStalk are interactive: clicking on a data point in a scatter plot (Figures 2–4-2–4) or on a fluorescence profile in a demo- or kymograph (Figures 4 and 5) displays the underlying cell, so that the corresponding raw image data and phenotype can be assessed. Furthermore, the output images of BacStalk showing the results of analyses or images of cells are publication-ready: all images in Figures 2–5-2–5 have only been minimally edited after their export from BacStalk (e.g., by changing the background color, cropping or minor editing of the axes). BacStalk offers the possibility to obtain all underlying raw data, so that they can be used for downstream analyses with other software tools, as described in the BacStalk online documentation.

The features of BacStalk that are exemplified above for stalked and flagellated bacteria, most notably its interactivity, its ease of use (see Figure 6 for a description of the workflow), the one-click generation of 1D and 2D kymo- and demographs, and the ability to output publication-ready editable images, are also applicable to the investigation of classical, non-stalked model organisms, such as E. coli and M. xanthus (Figure 5). The generation of interactive 2D demographs is a feature that is currently not available in any other image analysis software package. Its usefulness is demonstrated by an analysis of the localization dynamics of YFP-tagged PadC, an adapter protein connecting the chromosome partitioning ATPase ParA to subpolarly located bactofilin polymers in M. xanthus (Lin et al., 2017 ). Here, the 2D representation provides important information about the spatial arrangement of the filaments within the cell that cannot be appreciated in standard 1D demographs (Figure 5B). It should be noted that for stalked and non-stalked species, images should only have a low to moderate cell density, as no cell-splitting functionality is implemented in the current version of BacStalk.

BacStalk was written in MatLab to make use of its built-in figure customization and editing capabilities for generating publication-ready vector graphics, and to provide advanced users with easy access to the underlying processed data. However, our main goal in the design of the software was to make it as user-friendly as possible and applicable on first-try without any programming knowledge. This ease of use is achieved by BacStalk's powerful and fast graphical user interface (Figure 6). In addition, the user is supported by a comprehensive documentation and detailed video tutorials, which are available online at together with the open source code and a stand-alone pre-compiled version that does not require a MatLab license.

Overall, BacStalk facilitates high-throughput, in-depth, single-cell image analysis of stalked and non-stalked bacteria. It thus enables the study of many interesting and environmentally relevant bacteria as novel model organisms, provides tools for more detailed analyses of established model organisms, and therefore, constitutes an indispensable tool for bacterial cell biology and physiology.


Training and validating a deep neural network (DeepLoc) for classifying protein subcellular localization in budding yeast

Toward our goal of building a transferable platform for automated analysis of high-content microscopy data, we constructed a deep convolutional neural network (DeepLoc) to re-analyze the yeast protein localization data generated by Chong et al ( 2015 ). We provide a brief overview of convolutional neural networks in Fig EV1 and refer readers to LeCun et al ( 2015 ) and Goodfellow et al ( 2016 ) for a more thorough introduction. To make a direct comparison of DeepLoc and ensLOC performance, we decided to train our network to identify and distinguish the same 15 subcellular compartments identified using the SVM classifiers (Fig 1A). We implemented and trained a deep convolutional network in TensorFlow (Abadi et al, 2015 ), Google's recently released open-source software for machine learning (Rampasek & Goldenberg, 2016 ). In DeepLoc, input images are processed through convolutional blocks in which trainable sets of filters are applied at different spatial locations, thereby having local connections between layers, and enabling discovery of invariant patterns associated with a particular class (e.g., nucleus or bud neck). Fully connected layers are then used for classification, in which elements in each layer are connected to all elements in the previous layer. Our network arranges 11 layers into eight convolutional blocks and three fully connected layers, consisting of over 10,000,000 trainable parameters in total (more detail in 4, network architecture shown in Fig 1B). To ensure the validity of our comparative analysis, we trained DeepLoc on a subset of the exact same manually labeled cells used to train ensLOC (Chong et al, 2015 ), totaling

22,000 images of single cells. However, instead of training a classifier on feature sets extracted from segmented cells, we trained DeepLoc directly on a defined region of the original microscopy image centered on a single cell, but often containing whole, or partial cells in the periphery of the bounding box. The use of these “bounding boxes” removes the sensitivity of the image analysis to the accuracy of segmentation that is typical of other machine learning classifiers. Despite using a substantially smaller training set than was used to train ensLOC (Chong et al, 2015 ) (

70% fewer cells), we found that training a single deep neural network using a multi-class classification setting substantially outperformed the binary SVM ensemble when assigning single cells to subcellular compartment classes (71.4% improvement in mean average precision, Fig 1C).

Figure EV1. Illustration of convolutional neural networks

  1. Illustration of how convolutional neural networks learn to identify location invariant patterns. The input shown is an illustration of a yeast cell with a nuclear periphery protein localization. The input is convolved with convolutional filters, each representing a unique pattern that is learned during training. When the pattern of a filter matches the input at some location, the corresponding feature map is activated at that location. Pooling layers smooth the activations in the feature maps by calculating an aggregation (such as the maximum) over adjacent elements, effectively down sampling the feature maps. Pooling layers reduce the number of parameters in the model and also contribute to the location invariance of network. The fully connected layers in the network are typically responsible for classifying the activations extracted by the convolutional layers into the desired output categories. Each element in the final feature map is connected to each element in the first fully connected layer. The final activations in the network are passed through an activation function, such as the softmax function, to produce a distribution over output classes.
  2. An example of computation carried out by a convolutional filter. The calculations below the figure illustrate that the activation in top left corner is calculated by the weighted sum of its receptive field weighted by the convolutional filter. In convolutional networks, the values in the convolutional filters are parameters that are updated during training to reduce the networks prediction error on labeled samples from the training set.
  3. An example of computation carried out by max pooling layers. The calculations below the figure illustrate that the activation in the top left corner is the maximum over the elements in its receptive field. These layers do not have parameters and subsample the feature maps to reduce the number of parameters in the network and introduce more spatial invariance.
  4. An example of computation carried out by the fully connected layers. The calculations below the figure illustrate that the activation is the weighted sum of the input elements. Once again the weights themselves are the parameters learned by the network during training. A non-linear activation function is typically applied to this activation (as well as activations in other layers in the network). The non-linear activation functions enable the network to learn non-linear mappings between layers, and ultimately enable the network to approximate complex non-linear mappings between the input data and output classes. In the final layer, the sigmoid (σ) or softmax functions are used to produce distributions over the output classes for binary and multi-class problems, respectively.

Figure 1. DeepLoc input data, architecture, and performance

  1. Example micrographs of yeast cells expressing GFP-tagged proteins that localize to the 15 subcellular compartments used to train DeepLoc.
  2. Architecture of DeepLoc illustrating the structure of typical convolutional blocks, max pooling, and fully connected layers. The flowchart focuses on a sample image with a GFP fusion protein that localizes to the nuclear periphery (input). The input is processed through a series of repeating convolutional blocks (orange) and max pooling layers (yellow). In the convolutional block, the activation images illustrate network representations of the sample image (input). The red box and dashed/solid lines illustrate the connections within convolutional layers. Max pooling (yellow blocks) down sample activations across spatial dimensions. After repeated processing through convolutional blocks and max pooling, three fully connected layers are used for classification (green). The last layer (output) represents the distribution over localization classes.
  3. Average precision of DeepLoc (red bars) and ensLOC (Chong et al, 2015 ) (blue bars) on classifying a single cell test set (n = 4,197 samples). The cell compartment is indicated on the x-axis and the average precision (area under the precision recall curve) on the y-axis. The dashed lines indicate the mean average precision across the localization classes (0.49 for ensLOC (Chong et al, 2015 ) and 0.84 for DeepLoc).
  4. Average precision of DeepLoc (red bars) and ensLOC (Chong et al, 2015 ) (blue bars) on assigning localizations to images of GFP fusion proteins with single or multiple localization classes according to manual annotations by Huh et al ( 2003 ) (n = 2,833 proteins). The cell compartment is indicated on the x-axis and the average precision (area under the precision recall curve) on the y-axis. The dashed lines indicate the mean average precision across the localization classes (0.70 for ensLOC (Chong et al, 2015 ) and 0.84 for DeepLoc).

The ensLOC method relied on aggregating across cell populations to achieve > 70% precision and recall in comparison with manually assigned protein localizations (Huh et al, 2003 ). To assess the performance of DeepLoc in a similar way, we aggregated cell populations by computing the mean for each localization category across single cells containing the same GFP fusion protein. Again, DeepLoc outperformed the binary classifier ensemble across all localization categories (Fig 1D), achieving a mean average precision score (area under precision recall curve) of 84%, improving on the classification accuracy of ensLOC by almost 15% with substantially less training input.

Visualizing network features

Having demonstrated the improved performance of DeepLoc over the analysis standard, we next investigated which components of our network were contributing to its success. One of the hallmark differences between deep networks and traditional machine learning is that the network's learned representations are better at distinguishing between output classes than extracted feature representations used by other classifiers. To address whether this difference was relevant in our experiments, we visualized the activations of the final convolutional layer in 2D using t-distributed stochastic neighbor embedding (t-SNE) (Maaten & Hinton, 2008 ) for a single cell test set (Fig 2A). t-SNE is a popular non-linear dimensionality reduction algorithm often used to visualize the structure within high dimensional data in 2D or 3D space. Similarly, we visualized the CellProfiler (Carpenter et al, 2006 )-based features used to train the ensLOC SVM ensemble (Chong et al, 2015 ) on the exact same test set of single cell images (Fig 2B). We observed that using the DeepLoc representations, cells appeared to be better arranged in accordance with their localization classes, suggesting that DeepLoc's convolutional layers learn to extract features that are meaningful in the distinction of protein subcellular localization. These results suggest that an important component of the improved performance of DeepLoc reflects the network's ability to learn feature representations optimized directly on pixel values for a specific classification task as opposed to training classifiers on static feature sets.

Figure 2. Visualizing DeepLoc features

  1. 2D t-SNE (Maaten & Hinton, 2008 ) visualization of activations in the last convolutional layer of DeepLoc for 2,103 single cells in the test set. We computed the maximum activation across the spatial coordinates for each of the 256 features prior to fitting t-SNE.
  2. t-SNE visualization of CellProfiler features extracted for the same cells. We normalized the 313 CellProfiler features to be in the range [0,1]. In these plots, each circle represents a single cell circles are colored by their localization as determined by manual annotation (Huh et al, 2003 ) (color code to the right).
  3. Filters and activations in the last convolutional layer of DeepLoc for sample input images containing GFP fusion proteins that localize to the bud neck (top), Golgi (middle), or nuclear periphery (bottom). The convolutional filter visualizations were generated by activation maximization (Yosinski et al, 2015 ). The maximally activated filter for each input is highlighted with a red box (bud neck at the top, Golgi in the middle, and nuclear periphery at the bottom). For the bud neck sample, the input patch, filter, and activation are presented together to visualize how features are activated in DeepLoc. Other input patches that also maximally activate the selected feature are displayed.
  4. Regularized activation maximization (Yosinski et al, 2015 ) of output layers based on inputs initialized to leftmost column (Initialization). Different localization classes (compartment labels at the top of the images) are grouped by their morphological similarity (labels at bottom of images).

Next, we wanted to display these features to assess how they differ between compartment classes. To do this, we visualized activations and patterns extracted in the last convolutional layer of the network (layer 8) for specific input examples (Golgi, bud neck, nuclear periphery, Fig 2C, 4). Different input patterns activated specific features in deeper convolutional layers (convolutional activations, Fig 2C), with representations being combined in the fully connected layers from the convolutional feature maps, ultimately producing unique signals for different input patterns. These signals differ by localization class in a biologically interpretable way. For example, images containing punctate subcellular structures like the Golgi (top panels, Fig 2C) activated similarly patchy, dispersed features, while images containing discrete compartments like the bud neck (middle panels, Fig 2C) activated features that appear localized and linear.

We extended our analysis by applying activation maximization (Yosinski et al, 2015 ) to visualize input patterns that maximally activate each output class (Fig 2D, see 4). This technique works by keeping the parameters of the network constant while updating input pixel values to maximize the activation of specific features. In our implementation, the network iteratively updates an input with a randomly initialized green channel to produce an example “input” that resembles a cell with a GFP fusion protein that localizes to the maximally activated output class. The visualizations produced by the network for different output categories were convincing in their similarity to real compartment architecture. For example, visualizations for compartments such as the actin cytoskeleton, peroxisomes, and the spindle pole body were all punctate and dispersed (Fig 2D). Although these general visualizations may place compartments in various locations in the cell due to variable compartment locations in different images (e.g., spindle pole), the general morphology remains biologically interpretable. These results further justify the use of deep learning for classifying protein subcellular localization.

Using DeepLoc to identify protein dynamics in response to mating pheromone

Next, we assessed the ability of DeepLoc to classify images of yeast cells generated in different microscopy screens from those that served as training input to the network. We opted to analyze images from a screen generated by our group at the same time and on the same HTP confocal microscope as our previously published wild-type screens (Chong et al, 2015 ), but that ensLOC had been unable to accurately classify. In this genome-wide screen, haploid MATa cells were exposed to the mating pheromone α-factor, causing cell cycle arrest in G1 phase and polarized growth of a mating projection (schmoo) (Merlini et al, 2013 ). We used DeepLoc to analyze 16,596 images of the ORF-GFP collection acquired after exposure to mating pheromone for 40, 80, and 120 min. Images and analysis are available on the Cyclops Database ( We reasoned that a pheromone response time course would be a challenging test case for DeepLoc, due to the dramatic changes in cell morphology associated with α-factor treatment. DeepLoc produced reasonable protein classifications for single cells within hours, without the need for additional, non-wild-type training, while re-implementing an SVM ensemble would have necessitated weeks of training and optimization.

We identified 297 proteins (Table EV1) whose localization changed significantly in response to α-factor using Welch's t-test to score localization changes and a mixture model to identify significance (see 4). The 100 proteins demonstrating the most substantial localization changes were significantly enriched for proteins with annotated roles in conjugation and sexual reproduction (Gene Ontology bioprocess P < 0.01). This subset was also enriched for proteins required for cell fusion (e.g., Fus1, Fus2, Fus3, P < 0.01), nuclear fusion during mating (e.g., Prm3, Fig2, Kar5, P < 0.01), and polarized growth of the mating projection (e.g., Bni1, Pea2, Cdc24, P < 0.05). DeepLoc's ability to identify the movement of proteins that are already implicated in the mating response program serves to validate our method for detecting biologically meaningful results.

To do this, in addition to the localization measurements calculated by DeepLoc, we also extracted pixel intensity measurements as a metric for protein abundance (Tkach et al, 2012 Breker et al, 2013 Chong et al, 2015 ) (Table EV2). In total, we detected 82 proteins whose abundance changed 2-fold or more in response to pheromone, with 75 proteins increasing in abundance and seven proteins decreasing in abundance. Although there are minimal data available for protein abundance changes in α-factor, we compared our abundance measurements to gene expression changes and found positive correlations that are largely driven by the strongest hits (Fig EV2). While unrelated to the localization analysis by DeepLoc, this evaluation of protein abundance further validates the effectiveness of our screening protocol it also provides a complementary overview of proteomic responses to those made by Chong et al ( 2015 ) in the Cyclops database.

Figure EV2. Correlation of protein abundance measurements with gene expression data in response to α-factor treatment

  • A–C. In these plots, only proteins with ∂PL > 1 in at least one time-point are compared to corresponding gene expression changes from three different data sources. In each instance, genes demonstrating a substantial increase in expression as well as protein abundance are indicated on the plots. Comparison to gene expression microarray data from (A) Pramila et al ( 2006 ), (B) Spellman et al ( 1998 ), and (C) Roberts et al ( 2000 ).

Next, we wanted to display a quantitative snapshot of these proteomic responses to α-factor treatment similar to those previously constructed to illustrate protein movement after treatment with rapamycin, hydroxyurea, or the deletion of RPD3 (Chong et al, 2015 ). We displayed proteins with the most substantial localization changes (t-test statistic with magnitude > 10) in a flux network, indicating if these proteins changed in abundance as well (Fig 3A). As previously reported (Chong et al, 2015 ), after exposure to an environmental perturbation, we observe that proteins change in abundance or localization but rarely in both. Representative micrographs illustrate interesting localization/abundance changes shown in the flux network (Fig 3B). Importantly, DeepLoc identified novel movements of proteins already implicated in the mating response, such as the movement of Kss1, a MAPK that functions primarily to regulate filamentous growth, from the nucleus to the cytoplasm. We also identified the appearance of cell fusion regulators Prm1, Prm2, and Fus1 at the vacuole, which presumably results from the endocytosis of these cell surface proteins. Importantly, DeepLoc also identified the known localization of Prm1 and Prm2 at the Schmoo/bud tip (Heiman & Walter, 2000 ), though this movement is not shown on the flux network as their localization at the vacuole is more substantial. Deeploc also identified changes in localization of a number of proteins that control bud site selection, including Bud2, Bud4, and Bud5, which presumably reflects the fact that pheromone signaling is controlling polarized growth and over-riding the bud site selection machinery.

Figure 3. Protein dynamics in response to mating pheromone

  1. Flux network (Chong et al, 2015 ) showing significant protein localization and abundance changes in response to the mating pheromone α-factor. Localization changes with t-scores above 10 are shown. Hubs represent cellular compartments, while nodes represent proteins. Nodes are colored to represent abundance changes for those proteins that are changing in both their localization as well as abundance. Edge thickness corresponds to the magnitude of the localization change score.
  2. Representative micrographs highlighting protein subcellular movements after treatment with α-factor. Group 1: proteins that move from the nucleus to the cytoplasm. Group 2: proteins that appear in the vacuole/vacuolar membrane. Group 3: proteins that are moving away from the spindle pole after treatment with α-factor.

In addition to these striking changes, DeepLoc also identified more subtle or partial localization changes. For example, Nvj1 localized primarily to the spindle pole in untreated cells, but was also present at the nuclear periphery, as previously reported, where it performs a role in the formation of nucleus-vacuole junctions (Pan et al, 2000 ). After treatment with α-factor, DeepLoc captured Nvj1's movement away from the spindle pole, and its enhanced localization at the nuclear periphery. A number of proteins with no or poorly annotated roles also show clear localization changes, implicating these proteins in the pheromone response. For example, an uncharacterized protein Yor342c moved from the nucleus to the cytoplasm after α-factor treatment, a relocalization that has been previously noted in response to DNA replication stress (Tkach et al, 2012 ).

Assessing the transferability of DeepLoc to new and different microscopy datasets

With the goal of generating an automated image analysis system that can be broadly implemented by the budding yeast community, we used transfer learning (Yosinski et al, 2014 ) to classify image sets that significantly diverge from the images used to train DeepLoc. First, we completed a new genome-wide screen in standard cell culture conditions, which we called wild-type (WT)-2017, using the budding yeast ORF-GFP fusion collection (Huh et al, 2003 ). To differentiate this image set from other datasets analyzed by DeepLoc, screens were performed using a new HTP confocal microscope, and strains contained different red fluorescent markers (See 4, cropped cell images available at:

okraus/ We incorporated five new localization classes, many of which are punctate (e.g., Cytoplasmic foci, eisosomes, and lipid particles) and likely difficult to differentiate using traditional machine learning approaches, explaining their absence from ensLOC (localization classes shown in Fig 4A). We transferred and fine-tuned DeepLoc to the WT-2017 dataset using an increasing amount of training input per class, and contrasted the performance of this network with one trained from scratch using the same amount of training input (See 4 Fig 4B). Remarkably, transfer learning using DeepLoc achieved an average accuracy of 62.7% when fine-tuned with only five additional supplemental training cells per class (Fig 4C, yellow highlight), with several localization categories achieving accuracies above 80% (Fig 4D) this is a 63.4% improvement in performance using transfer learning over training from scratch (Fig 4E). The classes with significant errors are mostly the new punctate localizations, including cytoplasmic foci, and lipid particles, which are difficult to differentiate with only a few samples, and are still identified with 63.8% accuracy when merged with peroxisomes into one class.

Figure 4. Performance of DeepLoc after transfer learning

  1. Example micrographs from a screen of wild-type yeast cells expressing ORF-GFP fusion proteins. The images are of single cells expressing fusion proteins that localize to 20 unique output classes (colored green). The cells also express a bright cytosolic marker (FarRed colored blue), as well as a nuclear RFP fusion protein (colored red).
  2. Illustration of transfer learning. All layers except for the last layer (in red) are initialized to the network trained on the Chong et al ( 2015 ) dataset.
  3. Comparison of classification accuracy (y-axis) for different training set sizes (x-axis) when transfer learning is implemented using DeepLoc (red line) versus training a network from scratch (blue line). Error bars indicate the standard deviation of the accuracy based on five different samplings of the training set for each training set size. A yellow box highlights network versions that are referred to in (D and E).
  4. Confusion matrix for transfer learning the DeepLoc network trained on the Chong et al ( 2015 ) dataset to the new dataset with five samples per class. The intensity of the yellow color in each block of the matrix indicates the fraction of cells classified from each class predicted to be in a given class (scale bar to the right). Prediction accuracy for each class is indicated in brackets on the y-axis.
  5. Confusion matrix for training DeepLoc from random initializations with five samples per class.

Next, we used our transfer learning protocol to classify images generated by the Schuldiner laboratory using a different microscope and fluorescent markers (Yofe et al, 2016 ). Because these images were never intended for automated analysis, they contain many cells that are often clustered and overlapping. Also, bright field imaging was used to identify outlines of the cells, which do not express a fluorescent cytosolic marker (Fig 5A). Despite these significant differences, we were able to use transfer learning with DeepLoc (Fig 5B) to classify protein localizations in this dataset with an average accuracy of 63.0% after training with only 100 samples per class (Fig 5C). Classification accuracy with transfer learning ranged from 79% for the mitochondrial and “punctate” compartments to 41% for the bud compartment (Fig 5D). The availability of unique cell images for training varied by localization class, which likely affected accuracy in some cases (see 4, Table EV3). In contrast, performance was reduced for all classes when DeepLoc was trained from scratch (Fig 5E). Despite these classification errors, the performance of DeepLoc is a significant achievement given that these images have previously only been classified by manual inspection, and that the imaging protocols were highly divergent from those that are optimized for automated analysis.

Figure 5. Performance of DeepLoc for classifying images of cells expressing ORF-RFP fusion proteins collected for manual assessment

  1. Example micrographs from a screen of wild-type yeast cells expressing ORF-RFP fusion proteins (Yofe et al, 2016 ). The images are of single cells expressing ORF-RFP fusion proteins that localize to 10 unique output classes. The cells express a single RFP fusion protein of interest cell outlines are visualized in brightfield.
  2. Illustration of transfer learning. All layers except for the last layer (in red) are initialized to the network trained on the Chong et al ( 2015 ) dataset.
  3. Comparison of classification accuracy (y-axis) for different training set sizes (x-axis) when transfer learning is implemented using DeepLoc (red line) versus training a network from scratch (blue line). Error bars indicate the standard deviation of the accuracy based on five different samplings of the training set for each training set size. A yellow box highlights network versions that are referred to in (D and E).
  4. Confusion matrix for transfer learning the DeepLoc network trained on the Chong et al ( 2015 ) dataset to the new dataset with 100 samples per class. The intensity of the yellow color in each block of the matrix indicates the fraction of cells classified from each class predicted to be in a given class (scale bar to the right). Prediction accuracy for each class is indicated in brackets on the y-axis.
  5. Confusion matrix for training DeepLoc from random initializations with 100 samples per class.


Overview of PLAST

PLAST can be generally applied to microscopy images of proteins labeled with fluorescent protein fusion tags, fluorophore-conjugated antibodies, or other labeling techniques. PLAST has five major steps: cell segmentation, feature extraction, protein localization profile (“P-profile”) construction, P-profile dissimilarity computation, and compartment mapping (Fig. 1A). First, we automatically segment cells from microscopy images. To avoid segmentation bias that may be introduced by protein-to-protein variations in expression levels [25], we do not use fluorescent signals from the labeled proteins. Instead, we have developed a segmentation algorithm based on differential interference contrast (DIC) illumination and fluorescent nuclear stains (Supplementary Fig. S1). Other segmentation algorithms based on fluorescent whole-cell stains [26] may also be used in this step.

Dynamics of the nuclear lamina as monitored by GFP-tagged A-type lamins

J.L. Broers, B.M. Machiels, G.J. van Eys, H.J. Kuijpers, E.M. Manders, R. van Driel, F.C. Ramaekers Dynamics of the nuclear lamina as monitored by GFP-tagged A-type lamins. J Cell Sci 15 October 1999 112 (20): 3463–3475. doi:

The behavior of chimeric proteins consisting of A-type lamins and green fluorescent protein (GFP) was studied to investigate the localization and dynamics of nuclear lamins in living cells. Cell line CHO-K1 was transfected with cDNA constructs encoding fusion proteins of lamin A-GFP, lamin Adelta10-GFP, or lamin C-GFP. In the interphase nucleus lamin-GFP fluorescence showed a perinuclear localization and incorporation into the lamina for all three constructs. Our findings show for the first time that the newly discovered lamin A 10 protein is localized to the nuclear membrane. The GFP-tagged lamins were processed and behaved similarly to the endogenous lamin molecules, at least in cells that expressed physiological levels of the GFP-lamins. In addition to the typical perinuclear localization, in the majority of transfected cells each individual A-type lamin-GFP revealed an extensive collection of branching intra- and trans-nuclear tubular structures, which showed a clear preference for a vertical orientation. Time-lapse studies of 3-D reconstructed interphase cells showed a remarkable stability in both number and location of these structures over time, while the lamina showed considerable dynamic movements, consisting of folding and indentation of large parts of the lamina. Fluorescence recovery after bleaching studies revealed a low protein turnover of both tubular and lamina-associated lamins. Repetitive bleaching of intranuclear areas revealed the presence of an insoluble intranuclear fraction of A-type lamins. Time-lapse studies of mitotic cells showed that reformation of the lamina and the tubular structures consisting of A-type lamins did not occur until after cytokinesis was completed.


We thank Bastian Oldenkott (Bonn), Esther Engelhardt (Cologne) and Florian Kotnik (Münster) for assistance with experimental optimisation, and Jörg Kudla (Münster) for access to the Leica SP5 confocal microscope. This work was supported by the Deutsche Forschungsgemeinschaft (DFG) through the Emmy-Noether programme (SCHW1719/1-1), the Research Training Group 2064 (‘Water use efficiency and drought stress responses: From Arabidopsis to Barley’), the priority program SPP1710 ‘Dynamics of thiol-based redox switches in cellular physiology’ (SCHW1719/7-1, ME1567/9-1) and a project grant (SCHW1719/5-1) as part of the package PAK918. The Boost Fund project ‘PlaMint’ of the Bioeconomy Science Center (BioSC) provided partial support. The scientific activities of the Bioeconomy Science Center were financially supported by the Ministry of Innovation, Science and Research within the framework of the NRW Strategieprojekt BioSC No. 313/323-400-002 13.


Polarizing cells extensively restructure cellular components in a spatially and temporally coupled manner along the major axis of cellular extension. Budding yeast are a useful model of polarized growth, helping to define many molecular components of this conserved process. Besides budding, yeast cells also differentiate upon treatment with pheromone from the opposite mating type, forming a mating projection (the ‘shmoo’) by directional restructuring of the cytoskeleton, localized vesicular transport and overall reorganization of the cytosol. To characterize the proteomic localization changes accompanying polarized growth, we developed and implemented a novel cell microarray-based imaging assay for measuring the spatial redistribution of a large fraction of the yeast proteome, and applied this assay to identify proteins localized along the mating projection following pheromone treatment. We further trained a machine learning algorithm to refine the cell imaging screen, identifying additional shmoo-localized proteins. In all, we identified 74 proteins that specifically localize to the mating projection, including previously uncharacterized proteins (Ycr043c, Ydr348c, Yer071c, Ymr295c, and Yor304c-a) and known polarization complexes such as the exocyst. Functional analysis of these proteins, coupled with quantitative analysis of individual organelle movements during shmoo formation, suggests a model in which the basic machinery for cell polarization is generally conserved between processes forming the bud and the shmoo, with a distinct subset of proteins used only for shmoo formation. The net effect is a defined ordering of major organelles along the polarization axis, with specific proteins implicated at the proximal growth tip.


This article is part of the Spatial and Temporal Proteomics special issue.


The authors acknowledge the facilities of the Australian Microscopy & Microanalysis Research Facility at the Centre for Microscopy and Microanalysis, The University of Queensland, and the Australian Cancer Research Foundation (ACRF)/Institute for Molecular Bioscience (IMB) Dynamic Imaging Facility for Cancer Biology, established with funding from the ACRF. Professor Michael P. Rout supplied the initial mCherry nanobody vector series. Dr Andy Badrock supplied the vector backbones for the split-mVenus expression vectors. Professor Fred Meunier provided intellectual input into experimental design. We are particularly grateful to Associate Professor Brett Collins for advice on the GBP/split-YFP interaction.

Watch the video: GFP tagging Green Fluorescent Protein fusion (November 2022).