To identify genes involved in the regulation of FAs and cell shape, we conducted an siRNA screen using an automated, high-resolution, microscope-based assay (Fig. 1). For this purpose, an FA reporter cell line (HeLa cells expressing YFP-paxillin), specifically selected for its uniform FA morphology and distribution, was prepared (see Materials and methods). For the screen, cells were seeded on fibronectin-coated 384-well plates and transfected (24 h later) with the three human siRNA libraries chosen for this purpose: one targeting kinases (n = 576), a second targeting phosphatases (n = 192), and a custom library targeting MAR genes (n = 312). After fixation (72 h after transfection), wells were screened microscopically using a 60×/0.9 NA objective (Liron et al., 2006; Paran et al., 2006). A full screen of the three human siRNA libraries was repeated twice in duplicates, using SMARTpool concentrations of 100 nM and 50 nM.

Only plates with transfection efficiency better than 90%, based on the siTOX transfection control (which induces cell death; see Materials and methods), were taken for analysis. Moreover, we verified that the expression of fluorescent paxillin was suppressed by at least 99% using paxillin siRNA (Fig. S1).

The phenotypic features examined in this screen were subdivided into four main categories: cell coverage (percentage of total image area), which reflects cell number and spreading; FA morphology, which includes FA area (60th percentile, namely the FA area below which 60% of all FAs are found), FA mean intensity (60th percentile), FA length (90th percentile), and the percentage of small and round (dot-like) FAs, a rather prominent FA morphology feature; FA distribution, which includes FA abundance and peripheral FA; and cell shape, which includes spreading and elongation.

In this analysis, we define FAs as all distinct paxillin-containing structures larger than 0.25 µm². Such structures correspond to bona fide FAs, as well as to focal complexes and fibrillar adhesions. The distinction between these structures is beyond the capacity of a single molecule (paxillin)-based high-throughput assay.

After the collection of images from all wells (36 high-resolution images per well), each FA morphology parameter was quantified and statistically evaluated. To identify siRNA hits, z scores (subtraction of the controls mean from each individual raw score, followed by division of the difference by the controls standard deviation) were calculated for each individual feature of each siRNA, based on comparison with the corresponding values of the control wells in the same plate.

Those siRNA phenotypes that displayed an absolute z score value of >3.5 in both screens (using 100 nM or 50 nM siRNA) for at least one parameter were considered hits (see the “Data analysis” in the Materials and methods section for a detailed description of the data analysis process, and for the assessment of experimental variability).

There was usually good agreement between the 100 nM and 50 nM screens, except that at 100 nM, many of the siRNAs induced lower cell coverage, and the percentage of toxic wells was high (21% compared with 3.6% at 50 nM). We therefore repeated the screen at 50 nM in order to lower the toxicity and off-target effects. siRNAs that were hits at 50 nM and toxic at 100 nM were included in the hit list.

For hit siRNAs, FA distribution features and cell shape features were scored manually, and compared with controls. Fig. S2 shows the reproducibility of each of the measured features; moreover, to assess the reproducibility of our measurements, we calculated the mean coefficient of variance (standard deviation divided by the mean of all the z scores) for each siRNA in the screen (displayed in Table S2).

We also visually examined all the screen images in order to exclude wells with highly heterogeneous cell culture or other technical artifacts. In this manner, we identified hits that in the visual scoring were significantly different from controls, but had no effect on FA morphology.

Fig. 2 depicts single images of sample wells in which treatment with the specified siRNAs led to phenotypes with a low or high z score or manual score in the analysis. The images are grouped according to the features analyzed: control cells (Fig. 2 a), FA morphological features (Fig. 2 b), FA distribution features (Fig. 2 c), and cell shape features (Fig. 2 d).

Wells in which cell coverage was lower than 12% (altogether, 39 siRNAs in the 50-nM screen) were not scored for FA and cell shape features. Whether these siRNAs primarily target cell viability or cell adhesion remains to be explored. Nevertheless, based on the prominence, within this group, of siRNAs targeting cell cycle regulators and nuclear factor κB activation pathway components (Table S1), we assumed that most of the siRNAs affected cell number, target cell proliferation, and survival, rather than cell adhesion. The complete screening records, divided into hits, non-hits, and low cell coverage siRNA phenotypes, are available in Table S2.

The number of hits scoring positive for at least one parameter was very high, close to 45% of all siRNAs tested. The percentage of hits was particularly high among the MARs and phosphatases (53%), and lower among kinases (38%). The high frequency of hits may be attributed to the extensive involvement of these particular protein families in cell adhesion. Indeed, for individual features, the values were considerably smaller, ranging from 7.8% for cell elongation to 23.5% for FA length.

To determine which of the FA and cellular features were coregulated by the genes knocked down in this screen, correlations between pairs of measured features were calculated for all siRNA hits. Fig. 5 shows all the correlations found (significant correlations are marked with an asterisk), and representative images of the affected cells. As in the clustering analysis (Fig. 4 a), this analysis confirmed that FA morphological features were highly correlated in the screen hits, similar to their correlation in control cells. This finding indicates that only a small fraction of the siRNA hits (∼22%) interfered with these intrinsic correlations. These findings also demonstrated that cell coverage was positively correlated with FA mean intensity and abundance, and negatively correlated with the “small and round FAs” feature. Moreover, the higher the abundance of FAs, the greater the overall surface coverage by the cells.

However, FA abundance was negatively correlated with FA area, mean intensity, and length. Indeed, in many cases, siRNA perturbation led to a decrease in the amount of FAs, and to an increase in FA intensity and length (e.g., TLN1, P101-PI3K). The opposite phenotype; namely, formation of many small, faint, and short FAs, was induced by another set of siRNAs (e.g., AMFR, CDC25C). These results suggest that the establishment, growth, and maturation of FAs are differentially regulated by specific sets of genes, enabling cells to respond to matrix interactions either by enlarging existing adhesions or by creating new ones.

Particularly intriguing were those siRNAs that break the “common correlations,” such as ACTG1 and FMN1, which induce highly abundant, yet normal or even large FAs; or APC and DUSP5, which promote the development of small, faint, and sparse FAs. Another interesting effect of siRNA is the perturbation of the common correlation between cell shape and FA morphology features. Thus, elongated cells tend to form longer FAs, whereas highly spread-out cells are characterized by shorter adhesions. This apparent linkage is specifically perturbed by such siRNAs as TLN1 and CSK (see Fig. S3 for a summary of siRNAs that uncoupled the feature correlations).

Our findings of correlations between FA morphological features suggest that there is a common molecular pathway regulating FA size, geometry, and paxillin content as small focal complexes mature into large FAs (Fig. 6). The siRNAs of clusters MC1–4 perturb this common pathway. Nevertheless, the coregulation of these features is not absolute, and may be modulated by specific siRNAs, which apparently “uncouple” these features. For example, siRNAs of the MC7 cluster enhanced the formation of dot-like adhesions without blocking the formation of “mature FAs”; MC6 enhanced FA area without affecting its elongation and intensity; and MC5 interfered with paxillin recruitment and FA area without affecting FA elongation. These findings point to a possible hierarchy in the distinct stages of FA development, with sequential regulation of individual features by distinct gene families (Fig. 6).

To assess this interpretation, we selected three siRNA hits with robust effects and, using time-lapse video microscopy, checked their effects on FA dynamics. TLN1 siRNA from cluster MC6, which reduces the number of FAs and increases their size and intensity compared with the RNA-induced silencing complex (Risc)-free control (RF), was found to display a dramatic reduction in FA initiation and turnover (Fig. 7 and Videos 1 and 2 for RF; and Videos 3 and 4 for TLN1). However, siRNAs of EGF receptor (EGFR) and CAV2, both of which perturb the “main pathway” of FA maturation, exert their effect via two distinct mechanisms. EGFR siRNA induces fast turnover of FAs, which suggests that the EGFR is involved in FA stabilization (Fig. 7 and Videos 5 and 6). CAV2 siRNA, however, induces FAs that are as stable as those of control cells, but their growth is prematurely arrested, which suggests the involvement of CAV2 in FA maturation (Fig. 7 and Videos 7 and 8). The knockdown efficiency of TLN1, EGFR, and CAV2 siRNAs is shown in Fig. S4 b.

To validate the effects of the siRNAs, a group of 86 siRNAs, displaying the most prominent effects, were selected. These siRNAs were chosen based on the phenotypic clustering data, with priority given to siRNAs from different clusters, especially those displaying high z scores, phenotypes that uncouple the normal feature correlations, or those that induce rare phenotypes (Table S9). The validation itself involved screening for the cellular effects induced by the corresponding ON-TARGETplus (OTP) pools. OTPs have been shown to yield a reduced frequency of off-target effects (Jackson et al., 2006), but were shown to produce weaker phenotypes (Simpson et al., 2008); therefore, we used them at a concentration of 100 nM. We scored the effects as in the primary screen, and also compared the phenotypes visually. The siRNAs in which most of the features were similar to the SMARTpool phenotypes were designated as validated.

We then assessed the knockdown efficiency of eight siRNAs that were chosen for validation. We did not see a significant difference in the knockdown efficiency between the SMARTpool reagents and the OTP reagents, even in the nonvalidated siRNA phenotype (PHB; Fig. S4 b), which indicates that knockdown efficiency is not the only reason for the failure in validation. 44 siRNAs were validated (sample images of selected validated hits are shown in Fig. S4 a). Table I summarizes the validated siRNAs, together with their cluster assignment. The genes belonging to the adhesome, and whether they are intrinsic (FA component [FAC]) or associated (FA-associated component [FACA]) to FAs, are indicated. We also summarized their relevance to adhesion, cytoskeleton, FA structure, and cell migration, according to gene ontology annotations (supported by experimental evidence). 22 of our validated hits were not previously associated with adhesion and migration.

HeLa cells were retrovirally infected with YFP-tagged human paxillin in a pBabe vector. Single-cell cloning was used to obtain a morphologically uniform population.

Libraries were obtained from Thermo Fischer Scientific (http://www.dharmacon.com/HomePage.aspx) based on gene annotations from 2004. SMARTpools consisted of mixtures of four different siRNA sequences targeting the same gene. Three human siARRAY siRNA Libraries (Thermo Fisher Scientific) were screened: a kinase library targeting 576 kinases (protein kinases, selected based on (Manning et al. [2002], and lipid kinases), a phosphatase library targeting 192 phosphatases, and a custom library targeting 312 expanded family members of genes with known or predicted roles in cell adhesion and migration.

HeLa cells (250 cells per well) were plated in 50 µl DME (Invitrogen) + 10% FCS (Biological Industries) in 384-well plates (F-bottomed, μClear; Greiner Bio-One, GmbH), and cultured for 24 h at 37°C with 5% CO₂. The next day, 30 µl of the culture medium was removed from the cells, and cells were then transfected by direct addition of the 5-µl transfection mixture (1.25 µl siRNA [2 µM or 1 µM] in 2.5 µl DME with 0.25 µl Oligofectamine [Invitrogen] in 2.5 µl DME). Transfection was performed in duplicate. Cells were incubated for 24 h, and 40 µl of DME was then added so that the cells were grown for two more days in DME with 3% FCS. Control wells were left either untransfected or transfected with siCONTROL Risc-free siRNA (Thermo Fisher Scientific). Each plate contained two wells with siCONTROL TOX (Thermo Fisher Scientific) as an internal transfection control. The siCONTROL TOX is an RNA duplex that, when transfected into the cell, leads to cell death. Transfection efficiency was assessed by estimating the amounts of cells that remained viable in the well. Only plates with >90% transfection were taken for the screen.

After treatment, cells were fixed in 3% PFA for 20 min, and washed with PBS. Plates with fixed cells in PBS were screened.

Two screens were performed at final siRNA concentrations of 100 nM and 50 nM.

The automated microscopy systems (Liron et al., 2006; Paran et al., 2006) were based on IX71 and IX81 microscopes (Olympus). Automation was provided by ProScan (Prior Scientific), and included an xy stage, focus, shutters, and excitation and emission filter wheels. A fast laser AutoFocus attachment (Liron et al., 2006) was applied to focus the objective before image acquisition by means of a charge-coupled device camera (Quantix 57; Photometrics). The system was controlled by self-written Resolve6D software running in the RedHat Linux operating system. University of California at San Francisco Image Visualization Environment (http://www.msg.ucsf.edu/IVE/) and Priism libraries were used for image display and graphic user interfaces (GUIs). The acquired images in each well were tiled into montages in order to enable an overall visual impression of cell variability and for detection of consistent effects.

The general design of our computerized image analysis was described previously (Paran et al., 2007). In brief, images were first segmented into objects, using a binary threshold for the high-pass filtered images. For each object, morphological, fluorescence intensity, and estimated background parameters were calculated. In addition, the percentage of area covered by cells related to total image area (“cell coverage”) was calculated for each image, using the difference between the cell-free background fluorescence level, and cytoplasm diffuse fluorescence. Object-by-object, multiparameter data were saved for each image in separate files. User-controlled ranges for each parameter (minimum, maximum, and gates) defined the objects to be included, and excluded outliers.