High-content genome-wide RNAi screens identify regulators of parkin upstream of mitophagy

An increasing body of evidence points to mitochondrial dysfunction as a contributor to the molecular pathogenesis of neurodegenerative diseases such as Parkinson’s disease. Recent studies of the Parkinson’s disease associated genes PINK1 (ref. 2) and parkin (PARK2, ref. 3) indicate that they may act in a quality control pathway preventing the accumulation of dysfunctional mitochondria. Here we elucidate regulators that have an impact on parkin translocation to damaged mitochondria with genome-wide small interfering RNA (siRNA) screens coupled to high-content microscopy. Screening yielded gene candidates involved in diverse cellular processes that were subsequently validated in low-throughput assays. This led to characterization of TOMM7 as essential for stabilizing PINK1 on the outer mitochondrial membrane following mitochondrial damage. We also discovered that HSPA1L (HSP70 family member) and BAG4 have mutually opposing roles in the regulation of parkin translocation. The screens revealed that SIAH3, found to localize to mitochondria, inhibits PINK1 accumulation after mitochondrial insult, reducing parkin translocation. Overall, our screens provide a rich resource to understand mitochondrial quality control.

An increasing body of evidence points to mitochondrial dysfunction as a contributor to the molecular pathogenesis of neurodegenerative diseases such as Parkinson's disease 1 . Recent studies of the Parkinson's disease associated genes PINK1 (ref. 2) and parkin (PARK2, ref. 3) indicate that they may act in a quality control pathway preventing the accumulation of dysfunctional mitochondria [4][5][6][7][8] . Here we elucidate regulators that have an impact on parkin translocation to damaged mitochondria with genome-wide small interfering RNA (siRNA) screens coupled to high-content microscopy. Screening yielded gene candidates involved in diverse cellular processes that were subsequently validated in low-throughput assays. This led to characterization of TOMM7 as essential for stabilizing PINK1 on the outer mitochondrial membrane following mitochondrial damage. We also discovered that HSPA1L (HSP70 family member) and BAG4 have mutually opposing roles in the regulation of parkin translocation. The screens revealed that SIAH3, found to localize to mitochondria, inhibits PINK1 accumulation after mitochondrial insult, reducing parkin translocation. Overall, our screens provide a rich resource to understand mitochondrial quality control.
Following the loss of mitochondrial membrane potential, PINK1 and parkin coordinate a ubiquitination 9 , proteasomal activation 10 and autophagic (mitophagy) 5 response that may attenuate cell death 11 . As mitophagy can ameliorate the deleterious consequences of mitochondrial dysfunction [12][13][14] , genes regulating parkin translocation could be useful drug targets for increasing mitochondrial quality control. To identify genes important for the PINK1-dependent recruitment of parkin to damaged mitochondria, we conducted genome-wide siRNA screens against two diverse libraries. Our approach used the well-characterized cellular phenotype of parkin accumulation on depolarized mitochondria 5 . HeLa cells stably expressing GFP-parkin (Extended Data Fig. 1a) and a mitochondrial-targeted red fluorescent protein (mito-dsRed) were transfected with siRNA duplexes in 384-well plates (Extended Data Fig. 1b). After siRNA treatment, mitochondria were chemically depolarized with carbonyl cyanide m-chlorophenyl hydrazine (CCCP) to induce parkin translocation. Chemical depletion of mitochondrial membrane potential (Dy) mimicked pathological conditions caused by genetic mutation in mitochondrial-or nuclear-encoded genes or other stresses that deplete Dy. The degree of Parkin translocation was then assessed by high-content microscopy and automated image analysis that extracted a multitude of phenotypic parameters (Extended Data Figs 1c, d and 2a, b). PINK1 siRNA-treated wells served as positive controls, abolishing parkin translocation following CCCP treatment (Fig. 1a). We observed a high degree of assay robustness (Z9 . 0.5, Extended Data Fig. 2c, d) and technical reproducibility (Extended Data Fig. 2e, f) in primary screens. Assay termination before saturation (Extended Data Fig. 2g) allowed us to detect both translocation inhibitors and accelerators (Extended Data Fig. 2h). For example, LMAN1 siRNAs accelerated parkin translocation, suggesting the gene negatively regulates this process (Fig. 1a). Our imaging-based assay was used to screen genome-wide arrays of Ambion single (3 unique siRNAs per gene) and Dharmacon pooled (pool of 4 unique siRNAs per gene) reagents to maximize discovery potential.
Candidate gene selection used the robust statistical measure of median absolute deviation (MAD) to standardize siRNA activities from the screens (Fig. 1b, c). Normalized data were analysed using both parkin translocation and cytotoxicity (cell count) measures (see Supplementary Methods and Extended Data Fig. 3). Mito-dsRed intensity was used to identify siRNAs that depleted mitochondria (Extended Data Fig. 3). As important modulators of parkin translocation may have been masked by low-potency gene knockdowns, selection of active reagents at lower thresholds was feasible with our non-pooled siRNA data set, where we observed coincident activity of unique siRNAs with the same gene target. After filtering for cytotoxic or mitochondria depleting siRNAs, a gene was chosen as a 'candidate' modulator if at least two siRNAs from the non-pooled data set were active based on phenotypic evaluation of parkin translocation (generally .1.5 MAD or ,21.5 MAD, Extended Data Fig. 4a, b). Pooled siRNA screen candidate selection criteria were set to yield similar numbers of genes as the non-pooled siRNA screen (, 62 MAD, see Supplementary Methods). The candidate list (Supplementary Table 1) had a gene selection rate (see Supplementary Methods) similar to other large-scale RNAi (RNA interference) screens 15 . Twenty-four candidates overlapped between the two screens ( Fig. 1c, yellow dots) including an ubiquitin-conjugating enzyme (UBE2J2) and a member of the Hedgehog pathway (HHAT). Additionally, knockdown of the predicted LOC401052 strongly accelerated parkin translocation with all four reagents tested. The limited number of shared candidates was consistent with previous studies 16,17 .
A large number of initial candidate selections were involved in gene expression (17% of non-pooled, 24% of pooled candidates). As PINK1 is very labile and essential for parkin translocation 6 , the high rate of activity from housekeeping gene siRNAs suggests they may influence PINK1 expression. PINK1 also may be a source of off-target effects that confound RNAi screens [18][19][20] . The non-pooled siRNA screening data enabled us to systematically profile the miRNA-like effects of the siRNA seed sequences. Complementarity between the 59 end or 'seed' region of the siRNA guide strand and the 39 untranslated region (UTR) of unintended messenger RNAs is a driver of off-target behaviour 20,21 . Seed sequences matching the 39 UTR of PINK1 exerted a highly biased inhibitory effect (P , 2.2 3 10 216 ) on parkin translocation compared to all other siRNA seeds in the library (Extended Data Fig. 4c). We also observed that strongly inhibitory siRNAs (.2 MAD) had ,10% more matches to the 39 UTR of PINK1 than the corresponding accelerator siRNAs (Extended Data Fig. 4d). In total, 10,935 siRNAs (,17%) in the non-pooled screen had at least one hexamer 22 seed match to the PINK1 39 UTR. To combat off-target effects we used common seed analysis (CSA) 23 . CSA exposes the activity of siRNA seed-based offtarget effects by weighting each reagent against the population of siRNAs in the screen sharing the same seed (Extended Data Fig. 4e-g). The gene-level scoring of the entire non-pooled data set was adjusted for seed bias (Supplementary Table 2). Analysis of C911 mismatch controls 24 supported the hypothesis that siRNAs with strong seed bias (low seedadjusted Z-score) modulate parkin translocation predominantly though seed-driven off-target activity (Extended Data Fig. 4h). In addition to CSA, we used gene pathway enrichment on the original non-pooled screen candidates (Supplementary Table 1) as a complimentary method to identify promising genes. While confirming the strong influence of gene expression modifiers in our non-pooled screen candidates, enrichment analysis (Fig. 2a) revealed the significant (P , 0.05) presence of pathways including muscle function, the ubiquitin-proteasome, and autophagy (Supplementary Table 3). Furthermore, queries of the original non-pooled candidates (Supplementary Table 1) against annotated databases (gene ontology (GO) and human MitoCarta) revealed ubiquitin and mitochondrial processes (Supplementary Tables 4 and 5). Within enrichment groups, STRING database analysis (see Supplementary Methods) indicated many putative gene interactions that may regulate parkin (Fig. 2b, c).
To query candidate genes for follow-up, we chose from four distinct categories (Supplementary Table 6). These categories were top performing genes in the seed-adjusted primary screening data (category 1); candidates within a pathway enrichment group (category 2); top candidates from the pooled screen (category 3), and rational selection of candidates guided by gene annotations (category 4). In total, 106 positive regulators of parkin translocation were selected for validation (Supplementary Table 6). Four additional siRNAs from a different vendor were used for high-throughput validation. We found 67 genes in follow-up studies recapitulated parkin translocation inhibition activity with at least two additional siRNA reagents (Fig. 2d), a common reconfirmation benchmark. This cutoff was further substantiated by quantitative PCR with reverse transcription (qRT-PCR) of randomly selected candidates with only 1 active follow-up siRNA (of 4). In all cases, equivalent knockdown of target mRNA was achieved by both parkin translocation inhibitory and non-active siRNAs in the set (Extended Data Fig. 5a-f). Therefore, on-target knockdown was unlikely to be source of the translocation phenotype in these low confidence reconfirmations. The 67 confirmed candidates were ranked by the number of active siRNA reagents out of the 8 assayed (4 from primary screens and 4 from follow-up analysis) and then by seed-adjusted Z-score (Top 10 in Fig. 2e and complete list in Supplementary Table 7). For 8 of the topranked genes, we examined mRNA levels in cells transfected with primary screen active siRNAs and observed .75% target knockdown in all but 1 of the 18 siRNAs tested (Extended Data Fig. 5g-n). Secondary screening of active follow-up siRNAs for PINK1 (expressed without its endogenous 39 UTR) immunofluorescence after CCCP treatment established if these gene knockdowns affected PINK1 protein accumulation (Fig. 2e, red, and Supplementary Tables 7 and 8).    Table 1). b, Candidate genes with ontology (GO) that included 'ubiquitin' and interactions by STRING association. Gene nodes are coloured green (positive regulator), red (negative regulator) and contain 'D' if they are members of the 'druggable' genome. c, Same as b but for the GO term 'mitochondria' or matches to the human MitoCarta database. d, Analysis of group specific (top panel) and aggregate (bottom panel) reconfirmation results (percentages are out of 105 tested). e, Top 10 positive regulator genes that passed reconfirmation ($2 active siRNAs) in order of total active siRNAs and ties settled by seedadjusted Z-score. Numbers in brackets indicate seed-adjusted Z-scores and red lettering indicates genes whose knockdown blocked PINK1 accumulation (see Supplementary Table 7). Asterisk indicates inclusion of siRNAs passing 'active' thresholds only in raw and non-transformed data sets because they exhibited excellent phenotypes upon inspection. Analyses in a-c were performed on original non-pooled candidate list (Supplementary Table 1) and d, e were performed on the candidates chosen for reconfirmation assays (Supplementary Table 6).

RESEARCH LETTER
TOMM7, a component of the protein translocase of the outer mitochondrial membrane (TOM) complex, was a reconfirmed candidate from the non-pooled screen and the sixty-third most potent positive regulator of parkin translocation in the seed-adjusted data set. Knockdown of this MitoCarta member in the pooled siRNA screen also showed a translocation deficit. Interestingly, a report on mammalian TOMM7 demonstrated that gene knockdown did not change the efficiency of mitochondrial protein import 25 . Our TOMM7 gene knockdowns in HeLa cells resulted in a reduction of GFP-parkin translocation (P , 0.001, Extended Data Fig. 6a, b) and a reduction of TOMM7 mRNA levels (P , 0.001, Extended Data Fig. 6c). To confirm TOMM7 siRNA activity was not an off-target effect, we generated a knockout HCT116 cell line using transcription activator-like effector nuclease (TALEN) mediated genome editing (see Supplementary Methods). Wild-type mRNA and protein was undetectable in the TOMM7 knockout cell line (Extended Data Fig. 6d, e), but PINK1 mRNA expression was unchanged (Extended Data Fig. 6f). Notably, the knockout of TOMM7 abolished the translocation of YFP-parkin after CCCP treatment (Fig. 3a, b and Extended Data Fig. 6g). Expression of HA-tagged TOMM7 in knockout cells restored the YFP-parkin translocation to near wild-type levels (P , 0.001 for partial and complete translocation). We examined if TOMM7 knockout affects mitophagy downstream of parkin translocation. After 24 h exposure to CCCP, ,5% of TOMM7 knockout HCT116 cells underwent detectable mitophagy compared to ,90% of the wild-type cells (P , 0.001) ( Fig. 3c and Extended Data Fig. 7a). We also observed a deficit in parkin-dependent degradation of MFN1 in knockout cells (Fig. 3d).
To explore TOMM7 function, we evaluated PINK1 levels in TOMM7 wild-type and knockout HCT116 cell lines. PINK1 undergoes rapid turnover in polarized mitochondria 6 and only trace amounts of PINK1 were detected in cell lysates from both lines (Fig. 3d). However, after CCCP treatment, normal accumulation of full-length PINK1 (ref. 6) failed to occur in TOMM7 knockout cells, suggesting that TOMM7 is important for PINK1 stabilization on the outer mitochondrial membrane of damaged mitochondria. To explore why, we performed in vitro import of radiolabelled PINK1 protein into isolated mitochondria from wild-type or TOMM7 knockout cells. Mitochondria from both cell lines were similar in their ability to import and process PINK1 precursor protein, as indicated by the amount of PARL-cleaved PINK1 (ref. 26) at each time point (Extended Data Fig. 7b, c). In vitro protein import and processing of the canonical Su9-DHFR precursor was also normal in TOMM7 knockout mitochondria (Extended Data Fig. 7d, e). When Dy is depleted, PINK1 associates with the TOM complex in the outer mitochondrial membrane 27 . However, PINK1 imported in vitro into TOMM7 knockout mitochondria failed to associate with the TOM complex (Fig. 3e, 2Dy lanes). Our results indicate that human TOMM7 functions in the TOM complex, not for generalized protein import, but to shunt and retain PINK1, and perhaps other proteins, to the surface of damaged mitochondria (Extended Data Fig. 7f).
To explore the wider importance of TOMM7, we depleted TOMM7 gene expression in human induced pluripotent stem cell (iPS cell) derived neurons (Extended Data Fig. 7g-l). Lentiviral short hairpin RNA (shRNA) achieved a modest knockdown of TOMM7 mRNA, without affecting PINK1 mRNA levels (55-75%, Extended Data Fig. 7m, n). As in other cell types, significantly less endogenous PINK1 accumulated following mitochondrial depolarization in TOMM7 knockdown neurons (Fig. 3f, g, P , 0.001 as compared to control). This suggests that TOMM7 functions to recruit parkin to damaged mitochondria by stabilizing PINK1 in neurons expressing tyrosine hydroxylase (TH 1 ).
In search of positive regulators of parkin function, we were intrigued by HSPA1L, a confirmed follow-up selection (category 4, annotation/ phenotype-based). As other HSP70 family members can interact with parkin 28,29 , HSPA1L may selectively promote parkin translocation activity. HSPA1L is a widely distributed, but low-abundance member of the HSP70 family 30 . The non-pooled siRNA screen yielded BAG4 as a negative regulator of parkin translocation. BAG-domain-containing proteins act as nucleotide exchange factors for HSP70 members and BAG5 can modulate the E3 ubiquitin ligase activity of parkin 29 . We proposed that HSPA1L and BAG4 co-regulate parkin localization following mitochondrial damage. HSPA1L knockdown in HeLa cells led to a significant (P , 0.01) decrease in parkin translocation (Extended Data Fig. 8a-c) and BAG4 knockdown enhanced parkin translocation (P , 0.001, Extended Data Fig. 8d-f). Neither of these knockdowns affected the level of PINK1 protein accumulation (Extended Data Fig. 8f, g). To ensure the phenotype was specific to HSPA1L, we used TALEN genome editing to knockout HSPA1L in HEK293 cells (see Supplementary Methods and Extended Data Fig. 8h). Parkin translocation was strongly inhibited in the HSPA1L knockout cells (P , 0.001, Fig. 4a, b) that have normal levels of HSPA1A (a homologous, more abundant HSP70), and equivalent levels of PINK1 (Extended Data Fig.  8i, j). This inhibition of parkin translocation was rescued by exogenous mCherry-HSPA1L (Fig. 4a, b and Extended Data Fig. 8k). Expression of exogenous HSPA1A did not rescue the HSPA1L knockout phenotype, indicating the translocation deficit is not from a loss of cytosolic chaperone capacity (Fig. 4b). To investigate the mechanism of HSPA1L  0.01, Fig. 4b). GFP-parkin immunoprecipitation from the HeLa cell line used in the screens was found to bind endogenous HSPA1L and other HSP70 isoforms by mass spectrometric analysis (Extended Data Fig. 8l and Supplementary Table 9). Immunoprecipitation of HA-BAG4 demonstrates that YFP-parkin also binds to BAG4 and this binding is diminished following mitochondrial depolarization (P , 0.01, Fig. 4c, d). The reciprocal immunoprecipitation of YFPparkin also shows HA-BAG4 binding to both full-length and to the DUBL (ubiquitin-like domain) form of parkin, with stronger binding to the DUBL form (Extended Data Fig. 9a). A different BAG domain containing protein, BAG5, had been previously shown to interact with parkin 29 . We confirmed this binding, but observed a greater binding to HA-BAG4 (Extended Data Fig. 9b). RNAi knockdown of BAG4 or HSPA1L alone promotes or inhibits parkin translocation, respectively. However, simultaneous knockdown of both genes in either HeLa (Fig. 4e and Extended Data Fig. 9c) or BE(2)-M17 neuroblastoma cell lines (Extended Data Fig. 9d, e) reduces their respective individual phenotypes. Therefore, neither protein is uniquely required for parkin translocation, but rather an imbalance between them causes robust phenotypes. BAG4 and HSPA1L appear to act together to regulate parkin translocation (Extended Data Fig. 9f).
We assessed the effect of HSPA1L on the translocation of diseaseassociated parkin mutants 6 . In HeLa cells HSPA1L enhanced the translocation of YFP-parkin(R275W) (P . 0.001, Fig. 4f, g and Extended Data Fig. 9g) to a greater extent than HSPA1A. Future studies on the molecular details of these interactions may provide attractive targets to enhance selective removal of damaged mitochondria.
As the siRNA retesting of negative parkin translocation regulators did not consistently recapitulate the generally more subtle acceleration phenotype, we used the seed-adjusted data set (Supplementary Table 2) to guide the pursuit of these genes. As the tenth most potent seedadjusted negative regulator, the common seed plot for SIAH3 siRNAs showed a consistent departure from their seed 'peers' (Extended Data Fig. 4g). SIAH3-Myc (SIAH/SINA superfamily member, see Extended Data Fig. 10a) co-localized with mitochondria (Extended Data Fig. 10b, c). We did not detect a stable interaction of SIAH3 and PINK1 in the presence of detergents required for immunoprecipitation (Extended Data Fig. 10d). SIAH3 knockdown in HeLa cells accelerated parkin translocation with an increase in partially translocated parkin at 30 (P , 0.001) and 60 (P , 0.05) minutes post CCCP treatment (Extended Data Fig. 10e, f) and complete translocation at 60 min (P , 0.001). RNAi rescue by SIAH3 cDNA expression (siRNA-resistant) significantly (P , 0.01) suppressed the complete parkin translocation phenotype (Extended Data Fig. 10e, f). We confirmed the SIAH3-mediated parkin translocation phenotype (P . 0.01) and .90% knockdown of SIAH3 mRNA in BE(2)-M17 neuroblastoma cells stably expressing parkin and mito-GFP (P . 0.001, Extended Data Fig. 10g-i). SIAH3 silencing increased PINK1 accumulation after CCCP treatment (P . 0.001, Extended Data Fig. 10j, k). SIAH3 probably acts as a negative regulator of PINK1 stabilization because knockdown of SIAH3 did not increase PINK1 mRNA or protein before mitochondrial damage (Extended Data Fig. 10j-l).
In conclusion, we used a diverse siRNA screening strategy to shed light on the genes regulating PINK1-dependent parkin translocation to damaged mitochondria. Using several analysis methods in parallel allowed us to identify and overcome the significant susceptibility of parkin translocation to off-target/pleiotropic PINK1 modulation and select strong candidates for further study. Our subsequent validation demonstrated unique molecular functions of HSPA1L, BAG4 and SIAH3, as well as defining a role for TOMM7 in outer mitochondrial membrane PINK1 stabilization. Our genome-wide analysis of PINK1 and parkin regulators illuminates mechanisms of mitochondrial maintenance.

METHODS SUMMARY
siRNA screening. siRNAs were arrayed in 384-well optical plates (2 ml per well, 0.8 pmol). For each well, 20 ml of DMEM medium (Life Technologies) containing 4 ml ml 21 RNAiMAX (Life Technologies) was added. After 30 min incubation, 20 ml of a 37,500 cells per ml suspension was added to each well for reverse transfection. Plates with cells were incubated at 37 uC/5% CO 2 for 48 h and the parkin translocation assay was initiated with 40 ml of DMEM containing CCCP (Sigma) at a final concentration of 10 mM. After 2.5 h of incubation with CCCP, cells in all wells were fixed, nuclear stained and analysed using an ImageXpress Micro (Molecular Devices). All screening was performed in environmentally controlled enclosures with automated liquid and plate handling. See Supplementary Methods for detailed methods.
Online Content Any additional Methods, Extended Data display items and Source Data are available in the online version of the paper; references unique to these sections appear only in the online paper.

RESEARCH LETTER
Extended Data Figure 1 | Automated screening strategy and image analysis for the parkin translocation assay. a, GFP-parkin expression in the HeLa screening cell line. Ubiquitinated GFP-parkin (*) indicates parkin activation in the presence of CCCP. b, The screening workflow used 'assay-ready' 384-well plates pre-printed with 0.8 pmol of library siRNAs. Non-targeting siRNAs (green wells) and PINK1 siRNAs (red wells) were in columns 23 and 24. Cells were reverse transfected with siRNAs 48 h before the parkin translocation assay (addition of 10 mM CCCP). Parallel plate scheduling ensured that the timing of each step in the protocol was consistent across all plates and minimized total length of the screening run. An 8-head peristaltic and 16-head syringepump microplate dispensers dispensed reagents during the assay and BioTek EL406 192-tube microplate washer/dispensers were used to remove media and wash. Screening was performing in an environmentally controlled robotic enclosure. High-content microscopy images encompassed the entire cell population of each well. RT, room temperature. All imaged wells were processed with two algorithms from the Molecular Devices MetaXpress PowerCore Server Suite. c, Assessment of parkin translocation exploited the loss of GFP-parkin signal in the nuclear region (in the z-plane) that occurred as parkin accumulated on mitochondria. Utilization of nuclear-to-cytosol correlation yielded a more robust measure of parkin translocation than measuring GFP-parkin co-localization with mito-dsRed since the latter technique was highly affected by cellular morphology. The automated image analysis algorithm first segmented each cell's nuclear regions by performing top hat and h-dome feature recognition on the Hoechst 33342 staining (DAPI channel image) intensity. Nuclear segmentation created defined 'windows' to observe the level of GFP-parkin. These windows were extended by a 3 pixel gap to correct for imperfect channel overlay. Each cell in the image (n, n 1 1, and so on) was then interrogated for pixel intensity overlap of the GFP signal (FITC channel image) in the window using a Pearson's correlation that also sampled pixel intensity in a ring (one-third the width of the nuclear window) that extended into the surrounding cytosol (as a signal reference). After calculating the Pearson's correlation of the FITC and DAPI image on the total region comprised of the nuclear window, gap and extended ring, cells scoring over a correlation threshold were scored as positive for parkin translocation inhibition. Well-level translocation data was reported as the percentage of cells in the well over the correlation threshold (exhibiting a lack of parkin translocation). d, Cell count and assessment of the mitochondrial signal from the cells in each well was accomplished with the same morphological filters as in c. Cell count was determined through the number of segmented nuclei in the DAPI channel image. Using the nuclear segmentation, the algorithm identified the mitochondrial mass associated with each nucleus in the TxRed channel image. The mito-dsRed signal was then integrated across the segmented mitochondrial region for each cell (n, n 1 1, etc.  Translocation overlay is shown for both the DAPI and FITC channels with mitochondrial segmentation shown for the TxRed channel. c, The average and standard deviation of the control siRNAs (PINK1 and NTC) on each plate were used to calculate Z9 scores across the Dharmacon pooled siRNA screen. d, Same as in c, but for the Ambion non-pooled siRNA screen. e, Randomly selected library plates from the pooled and non-pooled siRNA libraries were used to make duplicate assay plates and were run and imaged using the conditions of the original screens. After image analysis and quantification, raw parkin translocation data was plotted from the replicates and correlation was assessed for the non-pooled Ambion (right) and pooled Dharmacon (left) siRNAs. f, Additionally, triplicate copies of a Qiagen follow-up 384-well plate containing siRNAs sets for target reconfirmation were run in the automated parkin translocation assay. Translocation scores from each plate were normalized to a percentage of the mean NTC negative control score and plotted relative to one another. These experiments demonstrated a high degree of technical reproducibility for the parkin translocation assay present in both raw and normalized data. g, To understand the rate of parkin translocation in the automated assay, a time course format in a similar manner as described for the primary screen was assessed (see Supplementary Methods). Automated plate and liquid handling was performed on the Agilent robotic platform to execute the translocation assay (CCCP dispense, incubation, plate fixing). Successive plates were automatically fixed at intervals of 15 min, nuclear stained and imaged on the high-content microscope (as in the original screens). After automated image analysis, data from 45 wells (per siRNA treatment) per plate were plotted as mean 6 s.d. from each time point. Data points from PINK1 siRNA wells (squares) and data points from NTC siRNA wells (circles) are plotted. Parkin translocation rates presented are specific to large-scale automated screening in 384-well plates. Owing to differences in cell types, culture environment, temperatures and liquid handling factors, parkin translocation rates differed in low-throughput experiments and were therefore calibrated on an experiment-specific basis. h, The mean 6 s.d. of control reagents from each screen was plotted to illustrate the signal window for parkin translocation assessment (pooled screen: NTC n 5 1,056, PINK1 n 5 1056; non-pooled screen: NTC n 5 2,976, and PINK1 n 5 2,976). Accel, accelerators of parkin translocation.

RESEARCH LETTER
Extended Data Figure 3 | Overview of data analysis workflow. Raw numerical data from high-content data generated in pooled and non-pooled screens was first normalized to the same plate controls before being stored as screen-specific data sets. Aggregate data sets underwent MAD conversion and parkin translocation data was log transformed to achieve near-normal distributions between inhibitors and accelerators. Specific siRNA-level data points were excluded if they failed to pass cell count and mitochondrial intensity filters. Data frequency distributions of cell count (red dots) fit to Gaussian curve (black line) and mitochondrial signal (blue dots) fit to a Gaussian curve (black line) from each screen are presented. Finally, gene candidates that had been withdrawn or were absent from the human genome annotation were removed from respective candidate lists. After candidate lists had been generated based on defined thresholds, a fraction of the genes were selected for follow-up analysis using a diverse set of categories to maximize selection diversity. For the non-pooled candidate lists, gene function and GO annotation queries coupled to STRING database searches (category 2) and annotation analysis (category 4) were employed to select the most promising genes for follow-up studies. The most active gene targets from the pooled screen candidate list were also selected for follow-up analysis (category 3). Finally, a category was also developed for subset of genes having excellent seed-adjusted activity screens from common seed analysis of the non-pooled data set (category 1). See the Supplementary Methods for complete details. Green connecting arrows represent data processing or filtering operations. Red connecting lines indicate decisions made for candidate gene follow-up.