The high-throughput acquisition and analysis of cellular image data, or high-content screening (HCS), can enable the analysis of complex cellular events not readily measurable with current high-throughput detection methodologies. HCS systems provide an integrated approach whereby image acquisition, processing, and data analysis may be performed in an automated and high-throughput fashion to create a dynamic cellular database (Price et al. 2002; Abraham et al. 2004). This database may then be mined according to specified metrics that represent the cellular phenotype of interest. Data may further be selectively extracted (or “gated”) from individual cells or subpopulations of cells, facilitating the analysis of heterogeneous cell populations. Although this platform presents several potential advantages over current detection methodologies for highly parallel assessments of cellular phenotypes, its utility in cell-based functional genomics applications has not been fully assessed.

To determine baseline transfection efficiencies for individual wells, all cells were marked with a DNA intercalating dye (Hoechst 33342) after 36 h, and image data for total cell counts and GFP-positive cells were collected (read 1) using a high-throughput fluorescence microscopy system (Q3DM EIDAQ100). To facilitate detection of proteins that would effect viability or growth, cells were then transferred to low-serum conditions. This resulted in a reduction of background cellular proliferation, and simultaneous sensitization to pro-apoptotic stimuli (data not shown). Finally, a subsequent imaging step (read 2) was performed to monitor alterations in cell number conferred by the overexpression of these individual cDNAs.

Image data were evaluated through the manipulation of a set of software modules (Cytoshop) that enabled the reduction of acquired image data to numerical descriptions of this cellular assay. Initially, nuclear images were corrected for shade distortion and background fluorescence through normalizing spatially segregated intensities and eliminating pixels that fell below the first (significant) mode of the signal distribution, respectively. These “corrected images” were then used to identify individual nuclei in a three-step process. First, to create a binary mask, a linear filter across all corrected nuclear images was applied (Price et al. 1996). This was followed by an automatic histogram-based thresholding step that allowed for the production of binary images (Price et al. 1996). Lastly, a connected components algorithm was then used on these binary mask images to establish separate identifiers for individual nuclei (Price et al. 1996). Approximately 80% of unthresholded objects were successfully classified as cells using this analysis [Supplemental Table 2, see “% identified cells (IC)”], while the remaining 20% represented spatially overlapping sets of nuclei, cell debris, and DNA precipitate. Visual analysis typically resulted in cell numbers within 15% of the results derived from the imaging algorithm (data not shown). Each of the nuclei, also referred to as areas of interest (AOI), and their associated masks, were then inserted into an image database called an “image table.” These tables also stored several properties of each AOI, including corrected pixel values, characteristics of the nuclear masks, spatial location, the coordinates of the AOI's bounding box, as well as others. Since the GFP signal was a cotransfection marker, the corrected GFP-channel (488-nm filter) AOI image pixels were selected relative to the nuclear AOI's centriod coordinates (x, y). These nuclear location values were used to define an expanded bounding box of pixels from the corresponding coordinates (x, y) in the GFP-channel images. The 488-nm AOIs were then formed by clipping the appropriate image pixels from the corrected GFP-channel image and inserting these pixels into the image table in association with their respective nuclear AOIs. Finally, positive transfection events (GFP+ cells) were established in AOIs where the summation of pixels within the GFP channel exceeded a calculated threshold.

Thus, cell-by-cell measurements were extracted from dual-emission images corresponding to cell nuclei and GFP expression, and used to identify transfected and untransfected populations in each well. Gene activities that perturbed cellular proliferation kinetics were quantified by recording changes observed in the GFP cotransfected cell population at read 2 relative to read 1. The results included a distribution of growth phenotypes as depicted by a significance factor, or z-score (Fig. 1C). z-scores are a function of standard deviations from the plate average (z-score = 0), and reflect the variance in cellular proliferation rates conferred by the introduced cDNA. Therefore, transfected cDNAs that enhance cellular growth rates relative to plate mean values have z-scores >0, while those that severely retard cell division or reduce cell viability have z-scores <0. For example, the pro-apoptotic proteins Bad and Bax induced z-scores of –5.47 and –4.35 (Fig. 1B; Table 1; Kaufmann and Hengartner 2001; Cory and Adams 2002). Conversely, cotransfection of PCNA, previously shown to disrupt growth arrest under serum-starved conditions (Fukami-Kobayashi and Mitsui 1999), yielded a z-score of 3.24 (Fig. 1B; Table 1). Additional known modulators of cell growth and cell death identified in our analyses include EZH2, CDK9, Mcl1, and Biklk, which produced z-scores of 5.16, 3.56, 3.09, and –7.93, respectively (Table 1; Hegde et al. 1998; Craig 2002; Varambally et al. 2002; Vrana et al. 2002; Bracken et al. 2003). In all, we identified 86 cDNAs (1.2%) that positively affected proliferative activities under reduced-serum conditions to a level more than three standard deviations from the trimmed experimental mean, and 84 cDNAs (1.15%) that restricted growth or induced cell death by a z-score of < –3.25 (Supplemental Table 2; data not shown). Analysis of Gene Ontology (GO) assignments and IntroPro domains contained within these identified proteins suggest that several of these molecules modulate cellular activities through assembly in dimeric or multimeric protein complexes and also at the level of transcriptional regulation (Supplemental Table 3).

To provide an approximation of false-negative rates in this assay, we analyzed the activities of control wells arrayed in each plate containing known regulators of apoptosis (Bad and Bax, Supplemental Table 2, wells O21–O24 and P21–P24). Of these controls, 71% (113/159) fell outside the cut-off range (z < –3.25) we used in the elucidation of our hits. However, if this standard is relaxed to a threshold of z < –2.5, >90% (144/159) of these control wells would be identified by our methodologies. These results suggest that, although our selection criteria were stringent, the overall performance of the assay and subsequent image analysis is robust and possesses a marginal false-negative rate.

Transfection efficiencies at read 1 were observed to be 30.2% on average, but varied widely from well to well, ranging from <1% to >80%. This distribution is a component of within screen variability, which may be derived from several factors in cell-based assay systems. These include heterogeneous cellular responses, plate variations, inconsistencies in library preparation, and other inherent methodological limitations (Lundholt et al. 2003). An important advantage of using image-based methods in functional genomics studies is the ability to differentiate affected and unaffected cells through in silico segregation of mixed cell populations. This enables the precise measurement of signal present only in individual or distinct subpopulations of cells (i.e., transfected cells), and simultaneously comparing their activities to nonresponders (i.e., nontransfected cells) within the same well. To better understand the potential advantages of this analysis methodology, we compared this approach to data generated through solely calculating whole-well fluorescent output, which is comparable to detection by a fluorescent plate reader. We find that image-based analysis of individual cells provides an overall increase in sensitivity, as assessed through a comparison of statistically significant outliers detected by each method (Fig. 1D). Single-cell analysis was dramatically more robust in extrapolating data from wells with low transfection efficiencies (Supplemental Fig. 4A), suggesting that this approach may be especially suitable for rare-event detection and high-variability assays. To further investigate the basis for this increased sensitivity of outlier detection, we compared the distribution of activities deduced from each analysis method (Supplemental Fig. 4B). Image-based analysis resulted in an ~10-fold compression of standard deviation across the assay when compared to whole-well fluorescent output (data not shown). When we assessed activities from this quantitative imaging assay to a parallel cDNA screen using a whole-well luminescence detection methodology, we observed a similar (13-fold) increase in the screen-wide standard deviation in the latter assay (Supplemental Fig. 4C). Since reduction in variability results in a higher confidence (smaller confidence interval) for probabilistic identification of outliers (“hits”), it is likely that the observed advantage in sensitivity for image-based analysis reflects an increase in the precision of acquisition and quantification of reporter events. Thus, we conclude the use of single-cell microscopic analysis results in superior data extraction and outlier detection in comparison to currently available high-throughput fluorescent detection protocols.

To ascertain the robustness of the screen and image analysis, and to determine false-positive rates, we next conducted a secondary analysis of proteins that were most proficient in inducing cell death or proliferation under serum-deprived conditions. In all, 92.5% (37/40) of sequence-verified clones demonstrated similar activities to those observed in the initial assay (Supplemental Table 4; data not shown). We postulate that the observed 7.5% false-positive rate results from inherent variabilities associated with high-throughput cellular genomics methodologies. We selected 15 cDNAs that, when overexpressed, exhibited robust proliferation-inductive phenotypes for further investigation (Supplemental Table 5). These included several gene products of unknown function including LRRK1, Ankrd25, Gatp1, and Gatp2, as well as prostaglandin E synthase (Ptges), an inducible enzyme that functions downstream of cyclooxygenase-2 (COX-2) in the prostaglandin E2 (PGE2) biosynthetic pathway (Levy 1997; Williams et al. 1999). To further confirm their proliferative potential under serum-depleted conditions, the activities of these encoded proteins were also examined in a human embryonic kidney cell line, HEK293. Ectopic expression of 11/15 clones reproducibly enhanced growth rates >1.5-fold, as measured by monitoring cellular metabolic activity (Fig. 2A; Supplemental Table 5). These findings validate the pro-proliferative activities of these molecules, and demonstrate that these gene products participate in regulating cell cycle activities in multiple cell types.

The abilities of these molecules to increase proliferation rates and induce oncogenic phenotypes were further assessed in primary cells. The retrovirus-mediated transformation of chicken embryo fibroblasts (CEFs) by oncogenes has been shown to result in cell immortalization, anchorage-independent growth, loss of contact inhibition, and increases in cell density at saturation (Bos et al. 1990; Vogt 2001). Avian retroviruses expressing these selected cDNAs were thus used to transduce these factors into primary fibroblasts. Analysis of CEFs stably expressing these genes demonstrated that several of these molecules can confer a growth advantage in this primary cell type (Fig. 2B; Supplemental Table 5), consistent with their putative role in cellular proliferation. Furthermore, ectopic expression of Gatp2 reproducibly potentiated CEF growth to saturation densities well beyond that observed with controls, resulting in large areas of multilayered growth (Fig. 2C). The introduction of Gatp1, LRRK1, and Ptges similarly led to the formation of localized focal lesions (data not shown; Supplemental Table 5).

Expression of LRRK1, Ankrd25, Gatp1, Gatp2, and Ptges was also demonstrated to promote anchorage-independent growth in CEF cells, underscoring their oncogenic potential. While these five clones induced this transformation phenotype to various degrees (Fig. 2D), Gatp1, Gatp2, and Ankrd25 directed substantially more aggressive colony formation in soft agar culture, as compared to infection with the empty RCASA parental virus. Taken together, these data confirm the proliferative activities of the putative growth-activators identified in our primary screen analysis, and establish their oncogenic potential in primary cells.

Since forced expression of Gatp1 resulted in increased proliferation rates in both immortalized and primary cells, we further investigated its necessity for cellular growth using RNA interference methodologies. Inhibition of Gatp1 expression was achieved in NIH3T3 cells using synthetic small-interfering RNA (siRNA) oligonucleotides, which reduced target mRNA and protein levels, respectively (Fig. 3A,B). Importantly, Gatp1 transcript and protein levels were unaffected by GL2 control siRNAs. Cells depleted for Gatp1 under standard growth conditions demonstrated normal growth kinetics comparable to mock- or control-siRNA-transfected cells (Fig. 3C, right panel). In contrast, the inspection of Gatp1 siRNA-transfected cultures under reduced-serum conditions (1% FCS) revealed a severe reduction of cellular proliferation (Fig. 3C, left panel). Cells transfected with the GL2 control siRNAs, as well as siRNAs directed against Gatp2, LRRK1, and Ptges, under these growth conditions displayed proliferation rates similar to mock-transfected cells (Fig. 3C; Supplemental Table 5; data not shown). To understand the basis for this diminished proliferative capacity, cell cycle profiles of NIH3T3 cells, transfected with Gatp1 or control siRNA, were measured by high-content image analysis at 24-h intervals post-transfection (Wan et al. 2004). Under reduced-serum conditions, treatment with Gatp1 siRNA led to the induction of a prominent sub-G₁ fraction (<2N DNA content) of cells, which appeared at later times post-transfection (Fig. 3D), indicative of apoptotic induction. This is consistent with the demonstration that siRNA-mediated abrogation of Gatp1 expression activates primary effector caspase-3 and caspase-7 (Fig. 3E; Kaufmann and Hengartner 2001; Cory and Adams 2002). Thus, we conclude that suppression of Gatp1 activity under growth-factor-limited conditions results in the induction of an apoptotic cascade.