A whole genome RNAi screen of Drosophila S2 cell spreading performed using automated computational image analysis

Cells can adopt a wide range of morphologies, from round lymphocytes to highly branched Purkinjee neurons. Major challenges of modern cell biology are to identify the proteins that are involved in cell shape determination and understand how these proteins are regulated by external and internal signals. Many of the best-studied proteins involved in cell shape determination are components or direct regulators of the actin cytoskeleton (Faix and Rottner, 2006), although many other proteins appear to play roles as well (Randazzo et al., 2007). Not surprisingly, different sets of proteins contribute to the determination of morphology in different cell types (Liu et al., 2009).

RNAi screens provide powerful approaches for identifying proteins involved in cell biological activities. However, only a limited number of screens have been performed on cell shape. Initial screens by Kiger et al. (2003) and Rogers et al. (2003) examined the effect of knocking down a limited set of genes (994 and 96 genes, respectively) on cell shape in Drosophila melanogaster tissue culture cells. At that time, the acquisition of image data was performed manually on the microscope, and the image data were analyzed by laborious visual inspection. Since that time, technological advances in robotic microscopy have allowed fully automated image acquisition, making it possible to obtain images from a whole genome RNAi screen in only a few weeks. As a result, the image analysis has become the rate-limiting step, and in most cases, the amount of image data for a whole genome screen exceeds what can be reasonably and reliably analyzed through visual inspection. Complex visual data, such as cell shape, are also best compared quantitatively rather than by qualitative assessment. Thus, developing new approaches for automated quantitative image analysis has now become a greater challenge than collecting the primary image data. An important step in automated, computational analysis was taken by Bakal et al. (2007), who used computational methods and neural networks to classify phenotypes in a screen with 249 genes predicted to play a role in cell morphology. More recently, Sepp et al. (2008) performed an automated whole genome screen of Drosophila primary neuron cells for genes that regulate neurite outgrowth using algorithms designed to analyze neurite formation, and Liu et al. (2009) screened the kinome and looked for kinases that influenced cell morphology.

In this study, we performed a whole genome RNAi screens for the uniform spreading of Drosophila S2 cells on Con A–coated surfaces (Rogers et al., 2002) and used different computational approaches to analyze the image data. In addition to the expected actin nucleating proteins, the screen also uncovered a set of genes involved in membrane secretion, supporting a link between membrane dynamics and cell spreading.

The goal of a whole genome RNAi screen is to generate a “hit list” (e.g., 20–200) of genes that can be subjected to further, more labor-intensive characterization. In this study, we sought to perform a whole genome RNAi screen and identify a hit list of genes involved in the spreading of round Drosophila S2 cells into a pancake shape after they are plated on a Con A surface (Rogers et al., 2002). Con A likely cross-links together several cell surface receptors that signal to activate the actin cytoskeleton and form a large and symmetric lamella. Because S2 cells are thought to be similar to hemocytes, this Con A response may be similar to a phagocytic response in which the cell is attempting to engulf a large particle.

To perform the RNAi screen, S2 cells were treated with double-stranded RNA (dsRNA) corresponding to a unique gene sequence for 5 d using a previously described full-genome Drosophila RNAi library (Fig. 1; Goshima et al., 2007). To test protein knockdown under our screening conditions, we performed immunoblot analysis of seven different proteins; the reductions of these proteins ranged from 75–99% (mean of 92%; Fig. S1). After dsRNA treatment, cells were placed on Con A–coated glass-bottom 96-well plates for 3 h, fixed and stained for DNA, α-tubulin, and actin, and imaged by automated microscopy (Fig. 1).

Because of the very large amount of image data from this screen, the major challenge was analyzing cell morphology in an automated manner. The first computational challenge is to reliably outline the cell boundaries (known as segmentation; Fig. 1). We found that the microtubule images provided the most accurate means of segmenting cells because of a gap in signal intensity between adjacent cells compared with actin staining (a more obvious cell boundary marker but one that gave rise to more errors because the fluorescence was often continuous between adjacent cells). From the segmented cells, we extracted ∼70 quantitative features (such as cell area, actin staining intensity, etc.), which were used for the classification of phenotypes via multifeature machine learning or single-feature sorting (see Materials and methods). For machine learning, we chose to use decision trees instead of neural networks because they are relatively easy to construct, provide reliable results, and allow for easy visualization of the most salient parameters. In this method, a training set consisting of normal and phenotype of interest is selected, and the associated set of extracted image features is compiled to create the training set. A decision tree classifier is constructed to separate the normal and phenotype of interest cells in the training set and is applied across the genome.

We also optimized or developed new single-parameter methods to identify phenotypes (see Materials and methods). Based on morphological RNAi phenotypes characterized by Rogers et al. (2003), we sought to identify (a) stellate morphology (star burst projections instead of the normally round, symmetrical shape, which is seen after RNAi to subunits of Arp2/3 and Scar and more mildly in knockdowns of Rac1 and Rac2), (b) failure to spread, as is seen with RNAi to cofilin (an actin-depolymerizing protein), profilin (involved in actin monomer binding and nucleotide exchange), and capulet (cyclase-associated protein involved in actin monomer binding), and (c) increased membrane ruffling and peripheral actin accumulation, which is seen with RNAi to capping protein β (a barbed-end actin-capping protein).

For stellate cells (e.g., Arp2/3 or Scar knockdown), we tested two methods. First, we used adaptive corner detection (He, X.C., and N.H.C. Yung. 2004. Proceedings of The Pattern Recognition 17th International Conference), which is well suited to analyze features of curvature. We found that this algorithm detects few corners (i.e., sharp curvature of the plasma membrane) in wild-type S2 cells (Fig. 2, C and E) but many points of curvature in a Scar-depleted cell (Fig. 2, D and E). Second, we developed a new method for analyzing the topology of the cell perimeter using power spectral analysis (see Materials and methods; Fig. 2, F–H; and Fig. S1). We also developed additional single-feature sorting algorithms to identify cells with very bright actin staining along the cell boundary (which is observed after capping protein α depletion as a result of uninhibited actin filament elongation; Fig. S2) or with very low peripheral boundary staining (e.g., after profilin RNAi, which inhibits actin polymerization; Fig. S2). Measuring cell area also allowed us to identify cells that fail to spread on Con A–treated surfaces, which is an actin- and membrane-dependent phenomenon (Rogers et al., 2003). This parameter in conjunction with actin staining intensity was a good discriminator of the RNAi phenotype of actin-depolymerizing proteins (e.g., cofilin and capulet), which yielded bright actin filament staining combined with a small cell footprint (Fig. S2). We validated that the cells displaying all of the aforementioned phenotypes, as recognized by visual inspection, also scored highly by the automated computational metrics (Table S3).

The automated screen successfully assigned high ranks to actin-associated proteins that were previously shown to produce phenotypes in S2 cells. Notably, many subunits of Arp2/3, the Wave–Scar complex, two Rac GTPases, and capping protein β were identified (Table I; Rogers et al., 2003). As exceptions, the automated screen failed to identify Cdc42, Aip1 (actin-interacting protein 1), and Slingshot (a phosphatase that regulates cofilin), which were reported to yield morphology phenotypes by Rogers et al. (2003) in S2I cells. However, when we tested these genes in multiple single-RNAi trials and examined the cells visually as well as by automated analysis, we did not see phenotypes for these gene knockdowns in S2U cells. Thus, the failure to identify phenotypes for these genes did not reside in the automated screen but, rather, most likely reflects differences between the S2U cells used in this study and the S2I cells used by Rogers et al. (2003). Although our screen was very effective in identifying subunits in the Rac–Scar–Arp2/3-mediated pathway, it did not find signaling molecules further upstream. This may be because of the fact that Con A activates many redundant signaling pathways that feed into Rac GTPase. However, the screen did identify casein kinase 1α as producing a phenotype similar to Scar and Arp2/3, as is subsequently described in more detail.

In addition to the Arp2/3 and Scar complexes, all three computational approaches identified the casein kinase 1α gene (CG2028; CK1α) as producing a stellate phenotype after RNAi knockdown (Table I). Representative fields of CK1α-depeleted, Scar-depleted, and normal cells are shown in Fig. 3 (A–C). Casein kinase 1 has not been directly linked to an actin phenotype in previous literature, although it has been identified in a previous screen as influencing cell morphology (Liu et al., 2009). dsRNAs directed to nonoverlapping regions of the CK1α gene that were not present in the dsRNA used in the genomic screen yielded the same phenotypes (see Materials and methods).

To better understand the CK1α knockdown phenotype, we performed live cell imaging in a GFP-actin cell line. Wild-type cells displayed robust retrograde flow of actin from the periphery toward the interior, as monitored by speckle analysis (Iwasa and Mullins, 2007). Wave/Scar depletion dramatically reduced actin retrograde flow. Similarly, CK1α knockdown substantially reduced actin retrograde flow (P = 2.4 × 10⁻¹³; Fig. 3 D). The reduction in actin retrograde flow may explain the stellate phenotype of these knockdown cells. Actin retrograde flow has been shown to push back microtubules, impeding their invasion toward the leading edge (Waterman-Storer and Salmon, 1997). In an extreme case of total actin depolymerization with latrunculin, microtubules push out long and thin processes in Drosophila S2 cells (Kural et al., 2005). The stellate phenotype may represent an intermediate phenotype where reduced actin retrograde flow weakens the barrier to microtubule growth.

In addition to the expected actin-related genes, another major group of genes that emerged from the screen were ones related to membrane function. These genes were identified computationally as having a high filamentous actin intensity divided by area (actin/area; Table II), which reflect their normal actin levels, but a small footprint from a failure to spread on the Con A surface. RNAi knockdown of several genes involved in ER function or vesicle trafficking scored as highly defective for cell spreading, including several COP proteins, SNARE and SNARE-interacting proteins (Syx5 and Sly1), and the GTPase Sar1 (Table II and Fig. S2).

Depletions of two proteins involved in the regulation of lipid biosynthesis unexpectedly gave nonspreading phenotypes: sterol regulatory element–binding protein (SREBP), termed HLH106 in Drosophila, and acetyl CoA carboxylase (ACC; Table II, Table S2, and Fig. S2). SREBP regulates several genes involved in fatty acid production (Seegmiller et al., 2002), including the rate-limiting enzyme ACC. Secondary rescreening confirmed that SREBP and ACC were the only two genes in the SREBP pathway that gave the nonspreading phenotype (Table S2). As Drosophila is a cholesterol auxotroph and has little cholesterol in its plasma membrane (Karlson, 1970), the role of SREBP in cell spreading is unlikely to involve cholesterol, and it is thought that the primary role of SREBP is to enhance the transcription of genes involved in fatty acid synthesis (Seegmiller et al., 2002).

A cell-spreading defect from RNAi depletion of membrane proteins might result from general toxicity (many result in lower cell numbers ranging from 15 to 100% of normal; Table II). However, the RNAi phenotype for this gene set (nonspreading yet showing normal to high levels of actin filament formation) appears to be quite specific and was not observed with RNAi of other genes that produced a low cell number (e.g., knockdown of ribosomal, proteosome, RNA polymerase II, and ATP synthetase subunits (Fig. S2). The actin/area ranks for these essential genes are generally >14,000 in contrast to many of the trafficking genes that ranked in the top 100 (Table II). Additionally, knockdown of vesicle trafficking genes did not substantially activate caspase 3, indicating that this nonspreading phenotype is not linked to apoptosis (Fig. S3). One exception is the ER-associated protein CG8465, which showed moderate activation of caspase 3 after its knockdown. A specific link between membrane trafficking and cell spreading has not been previously shown in previous literature to our knowledge, although it is not surprising that reduction in the flux of membrane or critical proteins to the cell surface might interfere with ability of the actin network to extend the lamella.

We found three uncharacterized genes (CG8465, CG9175, and CG15231) that exhibited nonspreading RNAi phenotypes (Table II). CG8465, a 128-kD protein with two ankyrin repeats, is potentially homologous to human ANKLE2 and ankyrin repeat–containing protein, although it seemingly lacks an LEM domain. CG9175, a 49-kD protein that contains two WD40 repeats, has been previously hypothesized to be the Sec12 homologue in insects, although sequence homology is low, and no functional data have been presented (Weissman et al., 2001). CG15231 is a small, unconserved peptide with a predicted signal sequence, which was also named DIM4 (Drosophila immune-induced molecule 4) based on its presence in the hemolymph of bacterially challenged flies (Uttenweiler-Joseph et al., 1998). CG15231 does not have an obvious orthologue outside of insects.

To learn more about these genes, we tagged the proteins with GFP at both the N and C termini. Both N- and C-terminally GFP-tagged CG8465 and CG9175 localized to a tubular network that colocalized with the ER marker Sar1 (Fig. 4, C and D); at higher expression levels, CG8465 localizes throughout the cytoplasm (not depicted). CG15231-GFP localized to small puncta, possibly membrane vesicles, that moved rapidly in the cytoplasm (Fig. 4 E). Interestingly, these GFP-containing puncta occasionally fused with the plasma membrane, producing a burst of CG15231-GFP into the solution (Video 1). Fusion events were not observed when GFP was fused to the N terminus, where the putative signal sequence resides. Thus, the three uncharacterized genes associated with a nonspreading RNAi phenotype are all associated with membrane compartments.

We further characterized the dynamics of GFP-tagged CG8465 and CG9175 using FRAP (see Materials and methods; Fig. 4 G). CG9175-GFP showed a slightly slower rate of photobleaching recovery (estimated diffusion coefficient [D] of 0.27 µm²/s) than Sar1-mCh (D of 0.56 µm²/s) and the KDEL receptor–GFP (D of 0.38 µm²/s), two well-characterized ER proteins. Our estimated diffusion coefficient for the KDEL receptor is in good agreement with a published value of 0.41 µm²/s (White and Stelzer, 1999). In contrast, GFP-CG8465 showed a significantly faster recovery (D = 0.82 µm²/s), which might be explained by the presence of a quickly recovering cytoplasmic population.

Next, we tested whether CG8465 and CG9175 play a role in protein secretion. Because we visualized GFP-tagged CG15231 being secreted from the cell, we measured the amount of CG15231-GFP that accumulated in the medium as a secretion assay (see Materials and methods). Depletions of known secretory proteins Sar1, Syx5, and COP (CG5484, a COPII protein) resulted in an expected 5–10-fold decrease in secretion compared with the control. RNAi of CG9175 also resulted in a strong secretion defect, which is on par with other known genes in the secretory pathway. However, RNAi of CG8465 exhibited only a moderate decrease, suggesting that it may influence but is not an essential component of the secretory pathway (Fig. 4 F). We also examined ER and Golgi morphologies for these depletions but found no obvious defect (Fig. S3).

CG9175 was previously identified in an RNAi screen for secretion (Bard et al., 2006) but was thought, based on its sequence, to be involved in translation and discounted. Although not annotated as such in FlyGene, CG9175 was previously postulated to be the Drosophila version of yeast Sec12, although the sequence identity is weak (∼7%; Chardin and Callebaut, 2002). Our finding of ER localization and strong secretion defect after RNAi all support the notion that CG9175 is the Drosophila homologue of Sec12. Sec12 in Saccharomyces cerevisiae is a guanine nucleotide exchange factor that activates Sar1, a key GTPase involved in assembling COPII coats (Weissman et al., 2001). Although metazoan Sec12 has not been studied in vivo, biochemical experiments suggest that mammalian Sec12 is important for recruiting Sar1 to the ER (Weissman et al., 2001). However, our results show that Sar1-mCh still localized to ER tubules after CG9175/Sec12 depletion (Fig. S3). Thus, our results are most consistent with CG9175/Sec12 being necessary for Sar1 activation but not essential for its recruitment to the ER.

In summary, we performed the first completely automated whole genome RNAi screen for cell morphology of a nonneuronal cell type. The methods of analyses that we used in this study can be used for other RNAi-based screens of shape and in other quantitative analyses of cell morphology, such as the formation of filopodia, neuronal dendrites, or other cellular extensions. Perhaps somewhat surprisingly, the screen identified more uncharacterized genes associated with membrane systems than the actin cytoskeleton and signaling. The reason may be that the most important actin-associated proteins have been already identified and that Con A stimulates several redundant upstream signaling systems that feed into actin nucleation. The mechanism by which knockdown of proteins involved in membrane synthesis and secretion produces a specific cell-spreading defect is probably complex but is likely to involve a depletion of key membrane proteins and insufficient lipids for cell expansion. The screen also led to the discovery of three uncharacterized proteins, two of which are part of the ER and a novel secreted protein.

Drosophila S2U cells were grown, and dsRNA treatment was performed as described previously in Goshima et al. (2007). For RNAi, we used the V2 RNAi library described previously in Goshima et al. (2007), which is now available commercially at Thermo Fisher Scientific. After 5 d, cells were resuspended and transferred to Con A–coated glass-bottom plates (Matrical). Con A treatment was performed by drying 70 µl of a 0.05 mg/ml Con A solution onto the bottom of each well. Cells were fixed in 6.4% formaldehyde and incubated overnight with an anti-tubulin antibody (1:1,000; YL; AbD Serotec) in the presence of phosphate-buffered saline containing 0.1% (vol/vol) Triton X-100 and 3% (wt/vol) BSA. The cells were incubated with 1 µg/ml DAPI (Sigma-Aldrich), donkey anti–rat antibody, FITC-conjugated antibody (1:250; F2658; Sigma-Aldrich), and 66 nM TRITC-phalloidin (Sigma-Aldrich) for 1 h. High throughput images were acquired using an ImageXpress Micro (MDS Analytical Technologies) with a 40× 0.95 NA lens (Nikon). 15–36 images per well were obtained so that 100–200 cells were imaged per well. Linear contrast adjustments were performed on all images.

Images were taken on an ImageXpress Micro–automated microscope with a 40× 0.9 NA objective using MetaMorph software (MDS Analytical Technologies). Plates were images in mounting media (Dako) at 20°C. Linear contrast adjustments were performed on all images.

Phenotypes for CG2028, CG8465, CG9175, and the nonspreading genes were verified by visually confirming the depletion phenotype using nonoverlapping RNA directed toward a different part of the gene as described previously in Goshima et al. (2007). To be considered confirmed, a gene had to have a score that would place in the top 2% of its identifying category in the original screen in at least half of all rescreens. Rescreen number varied from 1 to 12 per gene. CG2028 repeated 2/4 times visually, 2/4 times by fast Fourier transform (FFT) analysis, but only 1/4 times by corner detection. CG15231 is too small to design a nonoverlapping dsRNA. There are two computationally predicted off-target genes, CG13159 and CG9771, which show one 16-bp segment of overlap with the dsRNA for CG15231. However, RNAi of these two genes did not produce a phenotype. This makes it unlikely that the observed phenotype for CG15231 is because of an off-target effect, although the possibility cannot be excluded. All other genes indicated in Table S1 were subjected to visual confirmation.

cDNAs for cloning were obtained by PCR from Drosophila S2 cell cDNA. CG8465 was cloned from a cDNA obtained from Thermo Fisher Scientific. CG9175 cloned from our S2U cells had several mutations (all but one noncoding) compared with the FlyBase sequence. Vectors were cloned into pMT-GFP (Invitrogen) or pMT-Cherry vectors as described previously (Goshima et al., 2007) using the Gateway system (Invitrogen). GFP was imaged after induction of gene expression with 50 µM CuSO₄ overnight. Time-lapse imaging was performed using a microscope (TE2000; Nikon) with a 100× 1.49 NA objective and µManager microscopy software (Stuurman et al., 2007). Cells were imaged in growth media at 20°C with a camera (iXon^EM+; Andor). Linear contrast adjustments were performed on all images. For measuring actin retrograde flow in actin-GFP–expressing cells, kymographs were constructed by selecting lines of pixels perpendicular to the edge of the lamella in an image stack. Kymographs were analyzed for at least 10 cells with at least four measurements taken per cell using at least two different areas of the lamella. For CK1α- and Scar-depleted cells, actin velocities in the protrusions were measured, as flow does not occur in the regions between protrusions.

Secretion assays were performed using our CG15231-GFP cell line expressed from a metallothionein promoter. On the beginning of day 4 of RNAi treatment, the cells were induced with 500 µM copper sulfate. After 12 h, the media from the cells were removed, resuspended in 2× sample buffer, and immunoblotted for GFP, and the cells were lysed in 4× sample buffer and immunoblotted for tubulin. Immunoblots were developed using ECL and scanned. The ratio of intensity of GFP band to the tubulin band (to normalize for cellular protein) for each condition was quantitated by measuring total pixel intensity in the band of interest corrected for background intensity.

Caspase assays were performed as previously described (Wei et al., 2002). In brief, cells were lysed, and total protein concentrations were normalized to 0.74 mg/ml. DEVD.pna (California Peptide Research) cleavage was measured initially and at 6 h as OD 410 on a plate reader. Experiments were performed in quadruplicate.

Experiments were performed on a confocal microscope (510; Carl Zeiss, Inc.) with a 300 mW 488 laser. A 2.6-µm slice was bleached through the entire cell width and much of the cell depth (100% laser power and 100% transmission), and recovery of fluorescence was observed every 1.5 s after photobleaching (100% laser power and 0.75% transmission). No significant photobleaching was observed in control experiments. To obtain diffusion coefficients, the first 65% of the recovery toward the asymptote was fit as previously described (Ellenberg et al., 1997).

We sought to develop image analysis routines that are capable of analyzing the phenotypes described previously in the Rogers et al. (2003) screen. The first step in such analysis is to distinguish individual cells and outline the boundary between cells, referred to as segmentation (Fig. 1). A common approach for performing this is through the watershed method, which we performed using the images of DNA and tubulin staining. Actin staining proved to be difficult to use for segmentation because the actin structures of adjacent cells often touch, making groups of more than two cells very difficult to segment accurately. However, with the tubulin signal, there was a gap in signal intensity between adjacent cells, and watershed lines closely followed the true cell edges when segmenting through groups of cells (Fig. 1). We found that the most robust means of detecting cell edges was by using a simple threshold after the watershed operation, which accurately divided the image into regions containing single cells. From these individual cells, we extracted upwards of 70 features, such as cell area, actin staining intensity, etc., which were used for the classification of phenotypes via multiple-feature machine learning or single-feature sorting.

All image analysis was performed in MATLAB (MathWorks) using the image processing toolbox. After loading images into MATLAB, appropriate contrast-stretching operations were performed on all images. Groups of pixels <300 and 200 pixels were excluded from the tubulin image and DAPI image, respectively. A composite image of the tubulin and DAPI channels was created by adding the DAPI image to the tubulin image and dividing by 1.5. This image was inverted and slightly blurred using a convolution filtering technique with a 2-pixel-diameter disc filter. The extended minima was calculated (using the imextendedmin function), and these minima were imposed onto the image (using the imimposemin function). The purpose of the blurring and minima imposition was to prevent the common problem of oversegmentation via the watershed transform. A watershed segmentation was then performed on this image. The goal of this operation was not to identify cell borders but to isolate one cell in each watershed segment. Any cell edges touching the border of this segment were blurred with a convolution operation with a disc-shaped filter with a diameter of 10. This blurring ensures that measurements were not skewed by an artificially created jagged edge of the watershed line. The actin, tubulin, and DAPI images contained within this segment were thresholded. If this new image only contained one group of DAPI pixels, and the DAPI, tubulin, and actin signals overlapped after all objects <400 pixels were excluded, it was fed through the rest of the analysis.

Corner detection is described in Fig. 2. The code for corner detection was previously described by MathWorks (http://www.mathworks.com/matlabcentral/fileexchange/7652-a-corner-detector-based-on-global-and-local-curvature-properties). FFT analysis is described in Figs. 2 (F–H) and S1. A total of 70 extracted features were used for machine learning. The following is a description of each of these parameters. (1–6) Actin, distribution analysis: the cell is divided into six concentric rings by repeated erosion with a circular structural element. The amount of actin in each ring is summed and reported independently (six measurements). (7) Actin, total intensity: the summed intensity of the actin staining in the cell. (8) Actin, perimeter analysis: the intensity of the actin signal up to 6-pixels deep into the lamella is calculated and divided by the intensity in the rest of the cell. (9) Area: the area of the thresholded cell. (10) Area, convex hull: the area of the convex hull of the thresholded cell. (11) Cells per image: the mean number of analyzable cells per image. (12) Corner detection: see Fig. 2. (13) DAPI fragments: the number of DAPI fragments in the cell. (14) DAPI fragment distance: the mean distance separating DAPI fragments. (15) Difference of Gaussian filter: the area of resulting image after the difference between the original image filtered with Gaussians of two different kernels (Yi and Coppolino, 2006). (16) Elongation: the length of the major axis of the cell divided by the minor axis. (17–45) FFT perimeter analysis: see Figs. 2 (F–H) and S1. (46) Perimeter: the number of pixels that compose the perimeter of the thresholded cell. (47) Perimeter, smoothed: the number of pixels that compose the perimeter of the thresholded cell after a blurring operation. (48–58) Polynomial perimeter analysis: the distance from the centroid to each perimeter pixel was calculated as in FFT analysis. The resulting plot is then fit to an 11th-order polynomial, and the coefficients are recorded (11 measurements). (59) Texture, contrast: MATLAB built-in function for assessing contrast of images. (60) Texture, correlation: MATLAB built-in function for assessing correlation of images. (61) Texture, energy: MATLAB built-in function for assessing energy of images. (62) Texture, homogeneity: MATLAB built-in function for assessing homogeneity of images. (63) Texture, range: MATLAB built-in function that calculates the maximum difference in every 3 × 3 neighborhood. (64–69) Tubulin extension analysis: the cell was divided into six concentric rings by repeated erosion with a circular structural element. The amount of tubulin in each ring was summed and reported independently (six measurements). (70) Tubulin, total intensity: the summed intensity of the tubulin staining in cells.

For machine learning, training sets of at least 100 cells were selecting based on the images of the cells. Alternating decision trees were constructed using Weka (Witten and Frank, 2005) on all of the features extracted for each selected cell. Because of the small training set size, twofold cross validation was performed to roughly assess the performance of the tree, but no cross validation was performed when training to evaluate the entire dataset. To calculate a rankable score, the number of cells classified as having the phenotype of interest was divided by the number of cells classified as normal for each knockdown. The analysis code and primary image data are available at http://www.cellmigration.org/resource/discovery/vale/dambrosio2010_rnai.cgi.

Fig. S1 shows FFT power spectral analysis, the protein depletion after RNAi treatment to the relevant gene, and sample Western blot data used to generate Fig. 4 F. Fig. S2 shows RNAi phenotypes. Fig. S3 shows Golgi and ER morphology in the indicated knockdowns and DEVD cleavage activity. Table S1 shows the top and bottom 40 ranked genes for each indicated sorting method. Table S2 shows an analysis of toxicity and the nonspreading phenotype in the depletion of SREBP pathway genes. Table S3 shows validation of quantitative parameters. Video 1 shows time-lapse total internal reflection fluorescence microscopy of CG15231-GFP expressed in S2U cells.

This project was supported by the Cell Migration Consortium (grant U54GM064346) from the National Institute of General Medical Sciences. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institute of General Medical Sciences or the National Institutes of Health.