A Role for the VPS Retromer in Brucella Intracellular Replication Revealed by Genomewide siRNA Screening

With >300,000 new cases of human brucellosis annually, Brucella is regarded as one of the most important zoonotic bacterial pathogens worldwide. The agent causing brucellosis resides inside host cells within vacuoles termed Brucella-containing vacuoles (BCVs). Although a few host components required to escape the degradative lysosomal pathway and to establish the ER-derived replicative BCV (rBCV) have already been identified, the global understanding of this highly coordinated process is still partial, and many factors remain unknown. To gain deeper insight into these fundamental questions, we performed a genomewide RNA interference (RNAi) screen aiming at discovering novel host factors involved in the Brucella intracellular cycle. We identified 425 host proteins that contribute to Brucella cellular entry, intracellular trafficking, and replication. Together, this study sheds light on previously unknown host pathways required for the Brucella infection cycle and highlights the VPS retromer components as critical factors for the establishment of the Brucella intracellular replicative niche.

performed a genomewide small interfering RNA (siRNA) perturbation screen on the human epithelial cell line HeLa (ATCC CCL-2) combined with bacterial infection at biosafety level 3. Infections of the siRNA-treated cells were performed with a green fluorescent protein (GFP)-expressing strain of Brucella abortus, and the outcome was analyzed at 48 h postinfection (hpi) using automated fluorescence microscopy (21,22). Infection scoring was determined with a tailored high-content analysis workflow ( Fig. 1 and Materials and Methods). In brief, a model of Brucella replication was fitted to the pathogen intensity distribution to gain an infection classification independent of absolute fluorescence intensity. Further, we implemented an image intensity normalization coupled with a novel pathogen-to-cell association approach, which enabled quantitative measurement of the pathogen intensity distribution (Materials and Methods). An overview of the results is presented in Fig. S1 in the supplemental material. To account for the well-known confounding off-target effects associated with siRNA technology (recently reviewed in reference 23), we applied a multiple orthologous RNAi reagent approach (MORR) (24) with n Ն 5 perturbations per host gene. Further, we applied the redundant siRNA analysis (RSA) algorithm (25) on the entire screening data to reduce the number of false-positive results caused by off-target effects of single siRNAs and to favor genes with a reproducible phenotype confirmed by independent siRNAs. Genes matching a Benjamini-corrected RSA P value of Յ0.01 with more than three hit wells were considered significant and selected for further analysis (see also Materials and Methods). As a result, we identified 425 significant hits affecting Brucella infection. These comprised 223 down-hits (Fig. S1A, red, and Table S1) and 202 up-hits (Fig. S1A, green, and Table S2). Single siRNA data points are presented in Fig. S2  (down-hits) and Fig. S3 (up-hits). A panel of representative images from the screen is presented in Fig. 2A. Hit genes were further stratified by gene annotation enrichment analysis and functional annotation clustering using DAVID (26), protein-protein interaction network using the STRING database (27), as well as manual data mining. The functional categories enriched in our hit lists are presented in Fig. 2B to D together with the high-confidence protein-protein interaction network for targets that reduced (Fig. 2E) or increased (Fig. 2F) Brucella infection upon knockdown. Gene ontology and functional clustering analysis indicated a rather small overlap in enriched pathways when considering up-or down-hits ( Fig. 2B to D). The most prominent clusters that positively affected infection upon knockdown comprised components involved in the control or modulation of central cellular processes, such as protein synthesis, transcription, and mRNA processing, and cell cycle progression, as well as clathrin-mediated endocytosis ( Fig. 2 and Table S2). The most prominent clusters that negatively affected infection upon knockdown comprised signaling pathways involved in actin remodeling and phagocytosis. These clusters included most core components of the actin-related protein-2/3 complex (ARP2/3) (ARPC2, ARPC3, ACTR2, and ACTR3), and one of its main modulator, the WASP regulatory complex (WRC) (CYFIP1, WASF3, NCKAP1, and ABI3). Down-hits also comprised multiple components involved in transforming growth factor beta (TGF-␤) and erythropoietin-producing human hepatocellular receptors (Eph) signaling as well as further vesicular/endocytic pathways ( Fig. 2 and Table S1). Among all these factors, we can highlight the Ras-related protein RAB7A, which is needed for Brucella trafficking to the replicative niche (10), the small GTPases RAC1 and CDC42, which are involved in Brucella internalization into nonphagocytic cells (13), as well as the transmembrane glycoprotein SLC3A2 (CD98hc), involved in both bacterial uptake and intracellular multiplication (28). Since the roles of these individual components have already been described in the context of Brucella infection, they can be considered benchmarks for our results and globally validate our systems-level perspective of the human infectome for Brucella infection.
Pathogen entry assay identifies a role for VPS35, VPS26A, and SEC61␥ in Brucella postentry trafficking. To further dissect the roles of the identified genes in the progression of Brucella infection, we took advantage of a pathogen entry assay previously developed in our laboratory (29). Briefly, at 4 hpi, host cell membraneimpermeable gentamicin was added to selectively kill extracellular Brucella, and concomitantly, cell membrane-permeable anhydrotetracycline was added to induce expression of a plasmid-encoded reporter in the viable intracellular bacteria. At 8 hpi, this approach allowed us to robustly identify individual intracellular bacteria and to quantify the bacterial load before intracellular replication is initiated ( Fig. 3A and Materials and Methods). For this assay, we selected a number of genes from the different pathways identified in the genomewide screen as well as additional genes supplementing them (Table S3). The results of this entry assay were plotted against a matching endpoint assay (intracellular replication at 48 hpi). Strikingly, most of the genes tested displayed a direct correlation between the results of the entry assay and the endpoint assay (r 2 ϭ 0.763). This was the case for components involved in the actin-remodeling pathway (RAC1, ACTR3, and CYFIP1) or those involved in the TGF␤ signaling (SMAD4, TGFBR1, and TGFBR2), which strongly reduced both entry and subsequent intracellular replication ( Fig. 3B and Table S3). Similarly, the components of the clathrin pathway GAK and AP2S1 both increased bacterial entry and replication (Fig. 3B and Tables S3  and S4). This supports their involvement in Brucella entry into nonphagocytic cells, without excluding an additional role at any further stage of the infection. To identify components with a divergent outcome between the entry and replication assays, we selected genes diverging by more than 1 standard deviation to the fitted data. Six genes matched this criterion (Fig. 3B). Three genes displayed an apparent higher effect on pathogen entry than subsequent replication (albeit at a rather modest level). These were the small GTPases CDC42 (13), the ␣1 subunit of the Na ϩ , K ϩ -ATPase ATP1A1 (30) and the subtilisin-like endoproteinases FURIN (31). Most strikingly, three genes displayed a stronger reduction in endpoint replication compared to entry. Knockdown of The VPS Retromer Is Required for Brucella Trafficking Sec61␥, a central element of the ER-protein translocation machinery (see, for instance, reference 32), which has been previously involved in Brucella infection (33), showed a strong decrease in intracellular replication albeit no effect on pathogen entry. Similarly, our assay identified the vacuolar protein sorting-associated proteins VPS35 and VPS26A-two essential components of the VPS retromer complex (recently reviewed in references 34 and 35). These genes and associated pathway(s) thus likely represent novel components controlling the postentry trafficking of Brucella toward its replicative niche and/or are themselves required for the establishment or maintenance of the   (Table S3). For the entry assay, cells containing single bacteria were considered infected, and the final readout is proportional to the median number of bacteria per infected cells. For the endpoint assay, only cells containing replicating bacteria were considered infected ( Fig. 1 and Materials and Methods). Each data point corresponds to the average of all siRNAs or esiRNAs targeted against the gene of interest (n ϭ 3). The straight fit (oblique line, r 2 ϭ 0.763) indicates a globally high correlation between both assays. The blue region shows all points within 1 standard deviation (SD) of the fitted data. The genes falling out of this range are marked in red. For ease of visualization, only the averaged values over all RNAi products targeting a given gene are displayed.
rBCVs. For the present study, we further focused on the roles of VPS35 and the VPS retromer in Brucella trafficking, as it was the most prominent hit in our entry assay.
The VPS retromer is important for Brucella intracellular replication. The retromer complex orchestrates the recycling of numerous transmembrane proteins from early and maturing endosomes either to the trans-Golgi network (TGN) or back to the plasma membrane. Formed by a heterotrimeric complex consisting of VPS26, VPS29, and VPS35, the VPS retromer is conserved from yeast to human. However, the individual retromer subcomplexes have functionally diverged to organize multiple distinct sorting pathways, depending on the association with different accessory factors (34,35). To further decipher the role of the retromer in Brucella trafficking and intracellular replication we specifically browsed our genomewide siRNA data for retromerassociated proteins ( Fig. 4A and D). In addition to VPS35 and VPS26A, already identified both in the genomewide screen and in the entry screen ( Fig. 2E and Fig. 3B), knockdown of VPS26B (the paralogue of VPS26A) resulted in a significant reduction in Brucella infection (Fig. 4D). Depletion of VPS29, the third core component of the VPS retromer, resulted only in a mild reduction of Brucella infection and did not reach significance due to the wide spread of data obtained for the cohort of nine individual siRNAs tested (suggestive of strong off-target effects). Next to the retromer components, knockdown of the small GTPase RAB7A showed the strongest reduction in intracellular Brucella (Fig. 4D). However, neither SNX3, that together with the VPS retromer forms the SNX3 retromer, nor SNX27, another retromer-associated component involved in endosome-to-plasma membrane trafficking (36,37), displayed significant effect (Fig. 4D). Depletion of SNX1 and SNX5, two of the four sorting nexins of the SNX-BAR retromer (38), even seems to enhance Brucella infection (although they did not pass our hit selection criterion), while the two others, SNX2 and SNX6, showed no effect (Fig. 4D). Noteworthy, the functional association of the SNX-BAR sorting nexins with the VPS retromer has been challenged by two recent publications, which rather support a VPS-independent action of SNX-BAR (39,40). Collectively, our data indicate that the observed postentry impairment in Brucella intracellular replication is specifically linked to the integrity of the heterotrimeric VPS retromer, although involvement of further components in this process cannot be excluded.
To validate the requirement of VPS35 on Brucella infection and to rule out any off-target effects, we performed a complementation experiment using a VPS35 cDNA insensitive to a coexpressed shRNA (41). While shRNA knockdown of endogenous VPS35 inhibited Brucella infection, as detected in our genomewide approach, ectopic expression of the shRNA-insensitive cDNA of VPS35 rescued the phenotype ( Fig. 4B and C), confirming that depletion of VPS35 indeed negatively affects Brucella infection. We further confirmed the observed requirement of the VPS retromer for Brucella intracellular replication by determining intracellular bacterial load at different infection times of siRNA-transfected cells using CFU determination (Materials and Methods). At 6 hpi, no significant difference to the control was detected, with the exception of a small increase in intracellular bacteria upon VPS35 knockdown (Fig. 4E). Importantly, at 20 h and 44 hpi, siRNA knockdown of either VPS35, VPS29, or VPS26 resulted in a significant decrease of viable intracellular Brucella compared to control-treated cells (Fig. 4E), confirming the data obtained by our microscopy-based entry screen (Fig. 3). Further, efficiency of siRNA knockdown was confirmed by Western blot analysis ( Fig. 4F and Table S4). Together, these results corroborate the importance of each constituent of the VPS retromer, including VPS29, for Brucella to reach and possibly to maintain its intracellular replicative niche.
VPS35 knockdown prevents Brucella escape from the lysosomal pathway. Transient association with the lysosomal marker LAMP-1 is a hallmark of BCV trafficking during the first hours of infection. This association is eventually lost for those bacteria that manage to escape the host degradative pathway. Thus, to investigate the role of VPS35 in Brucella trafficking and to assess at which stage it could be required for the establishment of the intracellular replicative niche, we quantified Brucella colocalization with LAMP-1 in siRNA-treated and control cells. To this end, we analyzed Brucella-  ) as determined by one-way analysis of variance (ANOVA) and Dunnett's multiple-comparison test (**, P value of Յ0.001; ns, not significant). (E) Normalized CFU recovered from siRNAtreated cells at 6, 20, or 44 hpi. The presented data correspond to CFU count normalized to control, siRNA-treated cells (n ϭ 3). Significance was determined using one-way ANOVA with Dunnett's multiple-comparison test (*, P Յ 0.01; **, P Յ 0.001; ns, not significant). (F) Western blot analysis of the indicated proteins in total lysate of HeLa cells treated with siRNA targeting the designated genes, 72 h posttransfection. Displayed is a representative example of an experiment performed in biological triplicate (n ϭ 3). See Table S4 for the matching averaged intensity quantification.
to D and F), indicating that VPS35 function is not required for the early trafficking of the BCVs. However, loss of LAMP-1 association at 18 hpi was mainly detected in control cells, whereas most Brucella bacteria remained in a LAMP-1-positive compartment upon VPS35 knockdown ( Fig. 5B and C). Accordingly, VPS35 depletion strongly reduced intracellular replication of Brucella 18 hpi (Fig. 5G) compared to control cells (Fig. 5E). Altogether, our single-cell colocalization analysis demonstrates the requirement of VPS35, and thus of a functional VPS retromer, for the diversion of BCVs from the lysosomal pathway and for the subsequent establishment of a successful replicative niche.

DISCUSSION
The different membrane-bound organelles that compose the secretory pathway and the endolysosomal system of eukaryotic cells constitute targets of choice for many intracellular pathogens, which have evolved highly diverse strategies to hijack and/or subvert these trafficking pathways to their benefit (2,3). In that context, the importance of retrograde trafficking for the infection cycle of a number of human pathogens (viruses and bacteria) has been recognized in the past years (e.g., references 42 and 43). This is, for instance, the case for Chlamydia trachomatis, which uses its effector LncE to subvert host restriction via direct interaction with SNX5, thereby disrupting retromer trafficking (44,45). Additional pathogens have been shown to specifically target the VPS retromer, using or subverting its function to their advantage. For instance, hepatitis C virus interacts with VPS35 through its protein NS5A. This viral protein is recognized as VPS retromer cargo, and its interaction with VPS35 supports viral replication (46). Among bacterial pathogens, the best-studied example thus far is the subversion of the VPS retromer function by the T4SS effector RidL of Legionella pneumophila. RidL was shown to interact with VPS29, inhibiting retromer activity by outcompeting the binding of the VPS retromer regulator TBC1d5 and thereby promoting Legionella intracellular replication (47)(48)(49). Most recently, integrity of the Salmonella-containing vacuole was shown to be maintained by the direct interaction of the SPI-2 T3SS effector SseC with the VPS retromer (50). In this study, we report the involvement of the VPS retromer in Brucella intracellular cycle. More specifically, we show that VPS retromer integrity is required for Brucella to escape the host degradative pathway, as supported by the inability of the eBCVs to mature upon VPS35 knockdown toward LAMP-1-negative rBCVs. We further show that this VPS retromer-dependent process takes place after internalization and early trafficking, between 6 h and 18 hpi, matching the estimated timing of the eBCV-to-rBCV transition. It is thus tempting to speculate that the VPS retromer, possibly together with RAB7A, plays a role in this yet elusive but essential branching point of Brucella intracellular trafficking, i.e., diversion from the lysosomal pathway toward its ER-associated replicative niche. An alternative, yet nonexclusive hypothesis is that the VPS retromer is involved in the establishment of the rBCV, possibly by providing host factors and/or membranes that follow retrograde trafficking. A further role for the VPS retromer in the maintenance of the rBCV cannot be excluded based on our results. With the accumulation of functional data and the increasing number of described interactors, the VPS retromer is nowadays largely appreciated as a recruiting hub that orchestrates the retrograde endosomal trafficking of numerous cargos to the TGN or the plasma membrane (34,35,51). This versatility however obscures the identification of the underlying mechanism(s) by which VPS35 and the VPS retromer may contribute to Brucella intracellular fate. Further browsing our data set for the effect of known VPS retromer interactors failed to resolve the VPS retromerdependent pathway(s)-if any-that is required for Brucella intracellular replication. The only VPS retromer interactor that we unambiguously identified is the small GTPase RAB7A, which is essential for the recruitment of the retromer to endosomal membranes (52,53). Importantly, association of this upstream interactor to the eBCV is a wellestablished hallmark of early Brucella intracellular trafficking (9,10,16). Moreover, overexpression of a RAB7 dominant-negative allele (RAB7 T22N ) impairs the establishment of the Brucella replicative niche (10). The importance of RAB7 was recapitulated by our siRNA knockdown approach, strengthening its role in controlling the fate of BCVs, albeit by an unknown mechanism. Considering that the recruitment of the retromer to endosomal membranes is strictly dependent on the presence of RAB7, it is conceivable to assume that depletion of RAB7 prevents the recruitment of VPS35 to the BCV, consequently explaining the RAB7 dependency observed for Brucella replication.
Alternatively, very recent findings have established that retromer depletion in HeLa cells actually results in the hyperactivation of RAB7, which causes an overall depletion of the RAB7 pool on endomembranes (54). That drastic consequence could also imply an indirect effect of the observed retromer requirement for Brucella trafficking, by acting at the level of RAB7 activity and its availability for the BCV maturation. However, our results indicate that the effects of RAB7A or VPS35 siRNA knockdown are not entirely congruent. Whereas depletion of either factor impairs Brucella intracellular replication, only RAB7A knockdown showed a marked effect on pathogen entry, whereas VPS35 appears to be required only at a later stage of the infection. The implications of the newly described feedback signaling on RAB7 triggered by the retromer depletion, as well as the relative contributions of the VPS retromer and RAB7 for the transition of the eBCVs to rBCVs should be addressed in future studies.
Besides the retromer complex, our study pinpointed the involvement of several host pathways in Brucella infection, which had not yet been associated with this process. The most prominent cluster negatively affecting infection upon knockdown comprises factors involved in actin remodeling and actin dynamics as well as associated signaling pathways. Apart from an early association of Arp2/3 with BCVs (55), surprisingly little is known about the exact role played by the Arp2/3 complex or the WASP regulatory complex and their regulators during and possibly after Brucella internalization. Here as well, further studies will be needed to decipher the precise nature of their involvement. Finally, we also found that members of the TGF-␤ and fibroblast growth factor (FGF) signaling pathways promote Brucella infection, as their depletion decreased the amount of intracellular Brucella. Interestingly, it has previously been reported that patients with brucellosis show higher TGF-␤1 serum levels, a finding that is correlated with depressed T cell function (56). Further, B cells were also shown to produce TGF-␤ at early stages of infection with Brucella in mice (57). A possible immunosuppressive role for this pathway during Brucella infection should be further investigated.
Summing up, we believe that the genes and pathways identified in this study constitute a rich resource toward the understanding of Brucella intracellular trafficking, which ultimately should allow development of new approaches to controlling Brucella infections in human.
siRNA reverse transfection. Reverse siRNA transfection was performed as described previously (22,29) with minor adjustments. In brief, genomewide screens were performed with Dharmacon ON-TARGETplus SMART pool (pool of four siRNAs per gene) and Qiagen Human Whole Genome siRNA Set HP GenomeWide (QU, four individual siRNAs for each target). Further validation screens included Ambion Silencer and Ambion Silencer Select custom libraries (with up to six additional siRNAs for about 1,000 genes) and Sigma MISSION esiRNA libraries for 1,900 genes. All screening experiments were conducted in a 384-well plate format. Each plate contained negative controls such as mock (transfection reagent only) and scrambled (nontargeting) siRNA. In addition, general siRNA controls for transfection efficiency and toxicity (e.g., Kif11 [see Fig. S1C in the supplemental material]) as well as positive controls (e.g., Cdc42, Rac1) that are known to have an effect on Brucella infection (13) were added to each plate. Based on Kif11, the average transfection efficiency reached 97.6% (91.5 to 99.99%). The following specifications apply to all siRNA screens except the QU siRNA library where specifications are given in brackets. RNAiMAX in Dulbecco modified Eagle medium (DMEM) without fetal calf serum (FCS) was added to each well containing 1.6 pmol siRNA (for QU, 1 pmol) or 15 ng esiRNA. Screening plates were then incubated at room temperature (RT) for 1 h. Following incubation, 500 HeLa cells were seeded per well in DMEM (final concentration of 10% FCS). Plates were incubated at 37°C and 5% CO 2 for 72 h prior to infection. For assays in 96-and 24-well formats, reverse transfections were performed in 6-well plates and subsequently reseeded in the respective plate format. On-target or control siRNAs were added to reach a final siRNA concentration of 20 nM together with RNAiMAx transfection reagent in DMEM without FCS. After 30 min of complex formation at room temperature, 110,000 HeLa cells in DMEM with 10% FCS The VPS Retromer Is Required for Brucella Trafficking (DMEM/10% FCS) were added to each well. After 48 h transfection, cells were harvested by trypsinization and reseeded in DMEM/10% FCS (for 96-well plates, 2,800 cells per well; for 24-well plates, 50,000 cells per well). The next day, cells were infected as described hereafter. The following siRNAs used for CFU determination, colocalization experiments, and/or immunoblotting validation were purchased from Qiagen (Hilden, Germany): VPS35 (SI00760690); VPS26A (SI00760543); VPS26B (SI00631267); VPS29 (SI00760613); all star negative (0001027281); all star death kif11 (0001027299).
Infection. For the genomewide and confirmation screens, infections were performed in 384-well plates as described previously (22,29). In short, B. abortus 2308 pJC43 (aphT::GFP) (58) was grown in TSB medium containing 50 g/ml kanamycin at 37°C to an optical density (OD) of 0.8 to 1.1. Bacteria were then diluted in DMEM/10% FCS and added at a final multiplicity of infection (MOI) of 10,000. Plates were centrifuged at 400 ϫ g for 20 min at 4°C to synchronize bacterial entry. After 4 h of incubation at 37°C and 5% CO 2 , extracellular bacteria were killed by exchanging the infection medium with DMEM/10% FCS supplemented with 100 g/ml gentamicin. After a total infection time of 44 h, cells were fixed with 3.7% paraformaldehyde (PFA) for 20 min at RT. For the entry assay, infections were performed as described previously (29). In brief, transfected cells were infected with B. abortus 2308 carrying pAC042.08 for 4 h after which GFP expression was induced for 4 h by the addition of anhydrotetracycline (100 ng/ml) during the gentamicin killing of extracellular bacteria. Follow-up experiments and colocalization assays were performed according to the above-described protocol in 96-well and 24-well plates, respectively. For the colocalization assay, cells were infected at an MOI of 2,000. At 2 hpi, cells were washed three times with DMEM/10% FCS containing gentamicin (100 g/ml). After the indicated incubation time, cells were washed three times with phosphate-buffered saline (PBS) and finally fixed for 20 min in 3.7% PFA (in PBS).
Imaging with high-throughput microscopy. Microscopy was performed with Molecular Devices ImageXpress microscopes. A MetaXpress plate acquisition wizard with no gain, 12-bit dynamic range, and 9 sites per well in a 3 ϫ 3 grid were used with no spacing and no overlap and laser-based focusing. A 4=,6=-diamidino-2-phenylindole (DAPI) channel was used for imaging nuclei, a green fluorescent protein (GFP) channel for bacteria, and a red fluorescent protein (RFP) channel for F-actin or dsRed of bacteria in the entry assay. Robotic plate handling was used to load and unload plates (Thermo Scientific). The objective was a 10ϫ S Fluor with 0.45 numerical aperture (NA). The Site Autofocus was set to "All Sites," and the initial well for finding the sample was set to "First well acquired." Z-Offset for Focus was selected manually, and manual correction of the exposure time was applied to ensure a wide dynamic range with low overexposure. Images from the different siRNA screens are available upon request.
Image analysis. Images were analyzed with the screeningBee analysis framework from BioData-Analysis GmbH. To correct for uneven illumination inherent in wide-field microscopic imaging, an illumination correction model was computed for every plate using Cidre (60). To ensure that the Cidre-corrected image intensities fall within the range [0.0, 1.0], a linear transformation for pixel intensities was computed that maps the 0.001 quantile to 0.01 and the 0.999 quantile to 0.99 postillumination correction. Illumination correction and intensity scaling were performed as preprocessing steps for every image prior to analysis.
To reduce the signal of Brucella DNA in the DAPI channel, a linear transform of the GFP channel was subtracted from the DAPI channel, with the linear transformation parameters f and o estimated in the following way: a mapping of GFP pixels to DAPI pixels was constructed so that for all intensities in the GFP images, the list of corresponding intensities in the DAPI images were recorded. For every list of DAPI intensities, only the mean intensities were retained. This creates a mapping of GFP intensities to their corresponding mean DAPI intensities. A linear regression was performed to obtain the linear parameters f and o that map the GFP channel image to the DAPI channel image. Cleaned DAPI images with a reduced Brucella signal were obtained by subtracting the linear transform of the GFP channel from the DAPI channel I= DAPI ϭ I DAPI Ϫ (f I GFP ϩ o) as preprocessing steps for every image prior to analysis. On a random subset of 128 images, CellProfiler (61) was executed to identify nucleus objects using "OTSU Global" segmentation in the DAPI channel, and the median, lower quartile, and upper quartile segmentation thresholds of the images were retained as T DAPI-m , T DAPI-lq , and T DAPI-uq , respectively. On the same images, the GFP background intensity B GFP was obtained as the position of the peak in the GFP intensity histogram, the dynamic range of the histogram D GFP was obtained as the difference between the 99% quantile and the 1% quantile of intensity values, and the bacterial segmentation threshold was computed as T GFP ϭ B GFP ϩ 2 100 D GFP . On all images, screeningBee CellProfiler was executed to perform object segmentation and measurements with the following steps. (i) Nuclei were detected as primary objects using the manual threshold setting. For each plate, it was manually chosen to use T DAPI-m , T DAPI-lq , or T DAPI-uq , depending on visual inspection of the segmentation results. Using the same threshold on all images improved site-to-site comparability. (ii) Cells were detected as secondary objects around the nuclei, with "OTSU Global" segmentation in the RFP channel. (iii) Bacteria were detected as primary objects using manual threshold setting with threshold T GFP . Using the fixed background intensity as a reference for T GFP allowed for segmenting even rather dim objects while avoiding site-to-site variability.
In order to accurately measure infection scoring, a reliable method to associate pathogen colonies to individual cells is necessary. A straightforward approach is to assume that pathogen colonies must be contained within the body of the host cell. However, high cell confluence can make actin channel-based cell body segmentation inaccurate. Single microcolonies are often split into pieces that are incorrectly assigned to neighboring cells by using this approach (Fig. 1B). To address this issue, we developed a novel algorithm to assign pathogen colonies to robust nucleus objects in an intelligent manner (Fig. 1C).
First, inexpensive "bridge" and "majority" morphological operations were applied to the pathogen objects to connect broken clumps. Next, a weighted distance metric was used to measure an attraction score a N,P between a pathogen P and individual nuclei N within a close proximity (d prox ). The attraction score is computed as the surface integral of the nucleus area in a continuous field emanating from the pathogen defined by an exponential function that is strongest within the microcolony itself and that decays exponentially as the distance from the microcolony increases: a N,P ϭ nN e Ϫ d n,p , where n is an element (pixel) belonging to nucleus object N, d n,p is the distance transform from the edge of microcolony P to n, and is a parameter controlling the strength of the decay. Attraction scores for all nuclei proximate to microcolony P are normalized such that the strongest nucleus attraction score is 1, Nuclei objects with normalized attraction scores above a threshold a min are associated with the pathogen microcolony. In the case that multiple nuclei are associated with the same microcolony, the microcolony is split so that each element is associated with the nearest nuclei. Large microcolonies are encouraged to split with greater ease than small microcolonies by weakening the minimum attraction score linearly according to the area of the microcolony , and a' min ϭ 0.5 a min otherwise (where A P is the area of the pathogen microcolony). Parameters settings d prox ϭ 45, ϭ 0.2, a min ϭ 0.5, and A large ϭ 8,000 were optimized by grid search on a data set of 7,566 hand-labeled segmentations resulting in a 95.58% correct association rate. Nucleus-to-pathogen microcolony associations were aggregated. The area and integrated intensity of the pathogen objects associated with each cell and the mean intensity of the nuclei in the GFP channel was computed as readout.
Infection scoring for endpoint assays. Wells that contain only 32 cells or less were excluded from infection scoring. In the remaining wells, bacteria were filtered in a decision tree (DT) classification to exclude objects of only one-pixel area. Based on the relation of bacteria to nuclei, for the remaining bacterial objects, the integrated GFP intensity was integrated over all bacteria relating to a cell. To reduce the impact of background intensity, an estimate for GFP background was computed using the 1% lower quantile of mean GFP intensity in the nuclei. For every cell, the estimated GFP background intensity was multiplied with the area of bacteria relating to this cell, and the result was subtracted from the integrated bacterial GFP intensity of the cell to arrive at a background-free estimate of "bacterial load" in each cell. The value range for this intensity was zero for cells with no segmented bacterial objects, and it was higher than zero for all other cells. This integrated GFP intensity was then log 2 scaled, to reflect the exponential growth of replicating Brucella. Before log 2 scaling, a small epsilon value of 2 Ϫ20 was added to every cell, so that the log 2 value of cells with no segmented bacteria will not be negative infinity. The arbitrary value 2 Ϫ20 is, by a large margin, smaller than the smallest actual intensity of our assays but large enough to be used in histogram binning. For every plate, the histogram of the log 2 -scaled integrated cellular GFP intensity was computed (Fig. 1D) with a bin size of 0.025. The histograms were normalized to an arbitrary "virtual plate cell count" of 10 10 . To extrapolate a continuous distribution from the possibly sparse histogram, kernel density estimation (KDE) was used with a manually optimized Gaussian kernel of standard deviation 16. The histogram shows a bimodal distribution. By correlating the plate histogram distributions to selected images from the plate, we could identify that the first mode of the distribution is composed of cells with a low number of infection events ranging from single Brucella to small clusters, whereas the second mode is composed of large colonies. The two peak positions of the bimodal distribution were identified. Normal distributions G S and G L were fitted to the peak positions for small and large colonies, respectively. For the fitting of G S and G L , the mean was given by the position of the peak, the height was given by the height of the peak, and the variance was optimized such that the distance between the KDE and the sum of G S and G L became minimal. To arrive at a binary infection scoring threshold, we identified a suitable value 3 standard deviations below the mean of G L . This threshold includes 99.8% of the events in G L . Cells with an integrated GFP intensity exceeding this threshold were considered true-positive infections and were labeled infected. The infection score was computed as the ratio of infected cells to the total number of cells in the well.
Redundant siRNA analysis and hit selection. Redundant siRNA analysis (RSA) (25) ranks genes by iteratively assigning hypergeometric P values to each of the multiple siRNAs targeting the same gene and picking the minimum value within a given group to represent this gene. The ranking score indicates whether the distribution of ranks corresponding to a gene is shifted toward the top, thereby aggregating the information provided by independent siRNA sequences with the same target in a robust manner. Individual siRNAs from the Qiagen library and the averages of independent replicates of the Dharmacon, Ambion, and Sigma libraries (repeated experiments with identical siRNA) were used as input. Prior to RSA analysis, siRNA targets were reidentified by searching against ENSEMBL cDNA and the REFSEQ mRNA nucleotide data in order to ensure comparability between libraries. Cases where matching failed were excluded from this analysis. Data were further filtered by removing all wells that do not pass quality control, control wells and wells where cell count was below the initially seeded cell number (500). As both up-hits and down-hits are of interest to this analysis, RSA was run twice, once with Z-scored infection scores ranked from low to high values and once ranked oppositely. The RSA parameters were set as follows: upper and lower bound (Ϫ0.5; Ϫ2) or (Ϫ0.5; Ϫ10) on averaged z-scored infection score for down-and up-hits, respectively. Bonferroni correction was applied to account for the different numbers of siRNAs per gene. Genes matching a Benjamini-corrected RSA P value of Յ0.01 with more than three hit wells were considered significant and selected for further analysis.

ACKNOWLEDGMENTS
We thank Julia Feldmann and Sonia Borrell from the Swiss TPH, Basel, Switzerland, for great assistance and support with the BSL-3 environment. We are thankful to the Imaging Core Facility of the Biozentrum, Basel, Switzerland, for excellent assistance and advice with confocal microscopy. We also thank Daniel Billadeau for sharing the VPS35 shRNA suppression/rescue constructs. Finally, we thank Simone Eicher for help with the RSA analysis and critical discussions. The genome sequence used in this research was derived from a HeLa cell line. Henrietta Lacks, and the HeLa cell line that was established from her tumor cells without her knowledge or consent in 1951, have enabled significant contributions to scientific progress and advances in human health. We are grateful to Henrietta Lacks, now deceased, and to her surviving family members for their contributions to biomedical research. This study was reviewed by the NIH HeLa Genome Data Access Working Group (http://acd.od.nih.gov/hlgda.htm). The genomic data sets used for analysis described in this manuscript were obtained from the database of Genotypes and Phenotypes (dbGaP) through dbGaP accession number phs000640.v1.p1. This work was supported by grants 31003A_173119 to C.D. from the Swiss National Science We declare that we have no conflicts of interests.