Arf6 controls retromer traffic and intracellular cholesterol distribution via a phosphoinositide-based mechanism

Small GTPases play a critical role in membrane traffic. Among them, Arf6 mediates transport to and from the plasma membrane, as well as phosphoinositide signalling and cholesterol homeostasis. Here we delineate the molecular basis for the link between Arf6 and cholesterol homeostasis using an inducible knockout (KO) model of mouse embryonic fibroblasts (MEFs). We find that accumulation of free cholesterol in the late endosomes/lysosomes of Arf6 KO MEFs results from mistrafficking of Niemann–Pick type C protein NPC2, a cargo of the cation-independent mannose-6-phosphate receptor (CI-M6PR). This is caused by a selective increase in an endosomal pool of phosphatidylinositol-4-phosphate (PI4P) and a perturbation of retromer, which controls the retrograde transport of CI-M6PR via sorting nexins, including the PI4P effector SNX6. Finally, reducing PI4P levels in KO MEFs through independent mechanisms rescues aberrant retromer tubulation and cholesterol mistrafficking. Our study highlights a phosphoinositide-based mechanism for control of cholesterol distribution via retromer.

I ntracellular transport routes are under strict regulatory control in eukaryotic cells to ensure proper sorting of cargoes, maintain organelle identity and ultimately guarantee cell homeostasis. Among the key regulators of intracellular trafficking pathways, small GTPases, such as ADP ribosylation factor (Arf) family members, play a fundamental role in a compartmentspecific manner. Similar to other GTPases, Arf proteins cycle between an inactive GDP-bound form and an active GTP-bound form 1 . Unlike the other Arf family members (that is, Arf1-5), Arf6 is uniquely localized to the plasma membrane and to endosomes 2,3 , where it influences membrane trafficking. The role of Arf6 in various clathrin-dependent and -independent endocytic pathways as well as in recycling to the plasma membrane has been extensively studied 4,5 . A role for Arf6 in multivesicular body formation has also been recently described 6 . In addition, Arf6 regulates actin remodeling in such contexts as cell spreading, migration, cytokinesis, phagocytosis and neurite outgrowth 5,7 . In vivo, Arf6 ablation is embryonically lethal in the mouse 8 but a conditional knockout (KO) model revealed a non-cell autonomous role for neuronal Arf6 in oligodendrocyte precursor cell migration and myelination 9 .
One of the major mechanisms of action of Arf6 occurs through the control of lipid metabolism. Indeed, Arf6 binds and activates phosphatidylinositol-4-phosphate 5-kinases (PI4P5Ks), also known as type I PIPKs (PIPKIs), which phosphorylate PI4P into PI(4,5)P 2 (ref. 10,11). In addition, Arf6 can activate phospholipase D (PLD) 12 , whose product phosphatidic acid can in turn activate PIPKIs (ref. 13). Overexpressing a constitutively active mutant of Arf6 (Arf6 Q67L) also results in enlarged endosomes that contain high levels of PI(4,5)P 2 (ref. 14). More recently, Arf6 has been implicated in the regulation of cellular cholesterol distribution. In cultured cells, most cholesterol is derived from cholesteryl ester-rich LDL particles present in the media. LDL-particles are internalized by the LDL receptor (LDLR) and trafficked to the lumen of late endosomes/ lysosomes (LE/LYS). Cholesteryl esters are first hydrolysed by lysosomal acid lipase to free cholesterol, which is believed to be transferred by NPC2, a small soluble protein of the LE/LYS lumen, to the polytopic membrane protein NPC1. The latter mediates the egress of cholesterol from the endolysosomal system, allowing for its distribution to other cellular compartments and inhibition of cholesterol biosynthesis via the cholesterol-sensing machinery operating in the endoplasmic reticulum 15,16 . Perturbation of cholesterol traffic through mutations of NPC1 or NPC2 genes causes Niemann-Pick Type C (NPC) disease, a fatal neurodegenerative disorder associated with accumulation of free cholesterol and other lipids in the endolysosomal system 17 . Several studies provided hints that Arf6 is involved in the control of cholesterol homeostasis. First, Arf6 Q67L enlarged endosomes contain high levels of free cholesterol, labelled with filipin 18 . Second, silencing Arf6 was found to increase cellular cholesterol content in HeLa cells, reminiscent of an NPC disease phenotype 19 . In addition, Arf6 and components of its interactome were identified as 'hits' in a transcriptomic screen performed after cholesterol levels of cultured rat neurons were acutely increased 20 . More recently, Arf6 was a hit in a genomewide RNA-interference screen designed to identify genes required for intracellular transport of LDL-derived cholesterol, although this link was not explored further 21 . While these studies point to a link between Arf6 and cholesterol metabolism, the molecular basis underlying this relationship is unknown.
To address this fundamental question, we developed a novel model of Arf6 conditional KO in immortalized mouse embryonic fibroblasts (MEFs). We found that Arf6 depletion leads to cholesterol redistribution and accumulation in LE/LYS, a phenotype due to the mistargeting of NPC2 away from lysosomes. Because NPC2 is a cargo of the mannose-6-phosphate receptor (M6PR) 22 , we examined the function of retromer, which regulates the retrograde transport of the M6PR. We found that retromer function is impaired in Arf6 KO cells, thus leading to mislocalization of the cation-independent M6PR (CI)-M6PR away from the trans-Golgi network (TGN). We also found that PI4P levels were increased in Arf6 KO MEFs and that retromerpositive aberrant tubules emerged in part from PI4P-containing endosomes. Finally, we demonstrated that restoring PI4P levels rescues aberrant retromer tubules and cholesterol accumulation. This work therefore highlights a novel role for Arf6 in the regulation of retromer with critical implications for PI4P and cholesterol metabolism.

Results
Cholesterol redistributes to LE/LYS in Arf6 KO cells. To control the ablation of Arf6 acutely and inducibly, MEFs were generated from Arf6 Flox/Flox embryos also expressing a fusion of Cre recombinase with the ligand-binding domain of the oestrogen receptor (Cre-ER; Supplementary Fig. 1a). After immortalization, MEFs treated with tamoxifen showed gradually reduced levels of Arf6 protein by western blotting, leading to a complete absence of protein after 6 days ( Fig. 1a; Supplementary Fig. 1b). We then assessed the impact of Arf6 ablation on cholesterol. Cholesterol content was first analysed by liquid chromatography-mass spectrometry (LC-MS) and was similar in Arf6 WT and KO MEFs (Fig. 1b). MEFs were subsequently stained with filipin, a fluorescent probe for free cholesterol (Fig. 1c). The overall filipin intensity was comparable in the two genotypes (Fig. 1d), in agreement with the LC-MS results. However, while wild-type (WT) cells exhibited a diffuse staining at the plasma membrane and the cytoplasm with higher fluorescence intensity in the perinuclear region, the pattern of filipin staining was altered in KO MEFs, where it accumulated in cytoplasmic puncta (Fig. 1c). Quantification of filipin fluorescence revealed there were more cholesterol puncta in Arf6 KO cells compared with controls ( Fig. 1e). Filipin puncta in Arf6 KO MEFs also appeared significantly larger than in the controls (Fig. 1e). Aberrant filipin distribution could be rescued by expressing ARF6-HA (ref. 3) in Arf6 KO cells ( Supplementary Fig. 2), indicating that the cholesterol redistribution phenotype is a direct and specific consequence of Arf6 ablation in those cells. We then identified the specific compartment where free cholesterol accumulated in Arf6 KO cells. Filipin accumulated in Rab7-and LAMP1-positive organelles but not in EEA1-positive organelles (Fig. 1f), indicating that cholesterol is redistributed to LE/LYS but not to early endosomes in KO cells. LAMP1 protein levels ( Supplementary  Fig. 3a) as well as the number and size of LAMP1-positive organelles ( Supplementary Fig. 3b) were similar in both genotypes, indicating that Arf6 depletion did not cause any major structural disruption of LE/LYS. NPC2 mistrafficking causes cholesterol redistribution. Since the accumulation of cholesterol in LE/LYS is reminiscent of NPC cells 17 , we assessed whether the localization of NPC1 and/or NPC2 is affected by Arf6 depletion. As probed by Western blot, NPC1 protein levels were comparable (P40.05, Student's t-test) in Arf6 WT (100±9%,±indicates s.e.m., n ¼ 3) and Arf6 KO cells (75 ± 20%, n ¼ 3; Supplementary Fig. 3c). NPC1-GFP was found on the membrane of LE/LYS in both Arf6 WT and KO cells and displayed extensive co-localization with LAMP1, independently of Arf6 presence ( Supplementary Fig. 3d), suggesting that it is correctly trafficked in Arf6 KO cells. In contrast, co-localization of endogenous NPC2 and LAMP1 was greatly reduced in Arf6 KO cells compared with controls (Fig. 2a), ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11919 although NPC2 protein levels were comparable in both genotypes ( Fig. 2b; Arf6 WT 100 ± 6% and Arf6 KO cells 110±26%,±indicates s.e.m., n ¼ 3). The reduced presence of NPC2 in the LE/LYS compartment coincided with greater co-localization of this protein with Golgin97, a TGN marker ( Supplementary Fig. 4)     soluble lysosomal proteins, including hydrolases. The latter are then released in endosomes as a result of a lower pH and passively delivered to the LE/LYS where their function is needed. The retrograde transport of CI-M6PR from endosomes to the TGN is mediated primarily by the retromer complex 23,24 . We thus investigated whether NPC2 mistrafficking resulted from deficits in the traffic of CI-M6PR. In Arf6 WT cells, CI-M6PR was highly concentrated in the perinuclear region, in accordance with previous reports that it primarily accumulates in the TGN at steady state 25 (Fig. 3a). However, in Arf6 KO cells, CI-M6PR appeared more dispersed and was present not only in the TGN area but also on vesicular structures that were positive for Vps35, the core component of retromer (Fig. 3a).
The retromer complex function is impaired in Arf6 KO cells.
We then asked whether CI-M6PR mislocalization was due to a deficit in its retrograde transport caused by a malfunction of the retromer. The retromer complex is composed of two subcomplexes. The cargo-selective core subcomplex, composed of Vps35, Vps29 and Vps26, is involved in cargo recognition and sorting. The membrane-deforming subcomplex, composed of a sorting nexin (SNX) heterodimer, is involved in tubulo-vesicular carrier formation from endosomal membranes 26 . In the case of CI-M6PR transport, the specific SNX isoforms involved are heterodimers of SNX1 or SNX2 and SNX5 or SNX6 (refs 23,27-30).
We thus examined the relationship between Arf6 and retromer using confocal microscopy. Because various antibodies failed to specifically recognize endogenous Arf6 in immunofluorescence applications, two different ARF6 constructs, ARF6-HA (ref. 3) and CyPET-ARF6 (ref. 31), were exogenously expressed in HeLa cells, which were then stained for VPS35. A pool of Arf6 and retromer were found on the same endosomal compartments, suggesting a potential functional interaction ( Supplementary  Fig. 5). We then assessed the potential impact of Arf6 deletion on retromer levels by focusing on the core subunit of retromer and found that Vps35 protein levels were similar in Arf6 WT and KO cells (Fig. 3b). However, the co-localization between Vps35 and EEA1, an early endosome marker, was increased in Arf6 KO cells, compared with WT cells (15% versus 10%, respectively, Fig. 3c,d), showing that retromer was partially redistributed to early endosomes. While the number of tubulo-vesicular Vps35-positive retromer carriers was comparable, their size was increased in Arf6 KO cells (Fig. 3e). In contrast, the number and size of EEA1positive early endosomes were similar in both genotypes ( To test for potential changes in the dynamics of retromer tubules, we transfected MEFs with GFP-SNX6 and imaged them by spinning-disk microscopy (Supplementary Movies 5 and 6). As shown in Fig. 3g, there were more GFP-SNX6-positive tubules in Arf6 KO cells than in controls, and these tubules were both longer and more persistent. SNX6 tubules were still able to move towards the perinuclear region in Arf6 KO cells (Supplementary Movies 6), indicating no gross malfunction of molecular motors/microtubules transport. Altogether, our data suggest that Arf6 ablation causes alterations in retromer tubule dynamics, which are consistent with fission defects.

Retromer knockdown causes cholesterol increase in LE/LYS.
Our data suggest that Arf6 deletion causes a retromer defect which in turn leads to aberrant localization of CI-M6PR and its cargo NPC2 and finally to accumulation of cholesterol in LE/LYS. If this hypothesis is correct, then retromer deficiency should mimic the cholesterol phenotype observed in Arf6 KO cells. We thus treated HeLa cells with small interfering (siRNAs) to VPS35 and monitored the cholesterol distribution. After VPS35 siRNA transfection (VPS35 KD), VPS35 protein levels were decreased to 22 ± 11% ( ± indicates s.e.m., n ¼ 3) of those found in scramble siRNA-transfected cells, as assessed by western blot (Fig. 4a). Cells treated with siRNAs were fixed, immunostained for LAMP1 and VPS35 and labelled with filipin (Fig. 4b). Unlike Arf6 KO MEFs, VPS35 KD HeLa cells showed an overall increase in the levels of free cholesterol (Fig. 4c). However, similar to our observations in Arf6 KO cells, filipin accumulated primarily in large LAMP1-positive puncta (Fig. 4b). Quantification of filipin puncta showed that there was on average B77% more puncta/cell in VPS35 KD cells and that there was a significant shift in the relative distribution towards cells with a very large number of puncta (30 or more; Fig. 4d). Filipin puncta were also significantly larger in VPS35 KD cells than in controls (Fig. 4e). An increase in the number of filipin puncta was also observed when HeLa cells were treated with siRNAs targeting SNX1 and SNX2 or SNX5 and SNX6, the main sorting nexin isoforms involved in CI-M6PR trafficking 23,[27][28][29][30] (Supplementary Fig. 6). Taken together, these results show that depletion of retromer causes an increase in cholesterol levels and an accumulation of cholesterol in the LE/LYS compartment, consistent with our hypothesis. The retromer KD phenotype appears to be slightly more marked than the Arf6 KO phenotype vis-à-vis global cholesterol accumulation, perhaps reflecting the impact of a full versus a partial loss of function of retromer in VPS35 KD HeLa cells and Arf6 KO MEFs, respectively.
PI4P accumulates on retromer-positive early endosomes. To further delineate the molecular mechanisms underlying the retromer defect in Arf6 KO cells, we attempted to identify the Arf6 effector responsible for this phenotype. As stated earlier, some of the key effectors of Arf6 are PIPKIs, which catalyse the phosphorylation of PI4P into PI(4,5)P 2 (refs 10,11). We thus analysed the anionic phospholipid content of MEFs by HPLC combined with suppressed conductivity. Interestingly, we observed a B20% increase in PI4P in Arf6 KO cells, with no changes in phosphatidylserine, cardiolipin, PI or, more intriguingly, PI(4,5)P 2 (Fig. 5a). As it was reported that phosphatidic acid levels could also play a role in PIPKIs regulation 13 , we measured phosphatidic acid levels by HPLC and LC-MS. Phosphatidic acid levels measured by HPLC were similar (P40.05, Student's t-test) in Arf6 WT (100 ± 3 normalized molar % of measured lipids,±indicates s.e.m., n ¼ 3) and Arf6 KO cells (108 ± 16 normalized molar % of measured lipids, n ¼ 3). Similarly, phosphatidic acid levels measured by LC-MS were comparable (P40.05, Student's t-test) in Arf6 WT (0.077±0.010 molar % of measured lipids,±indicates s.e.m., n ¼ 11) and Arf6 KO cells (0.064 ± 0.007 molar % of measured lipids, n ¼ 12), making an indirect contribution of PLD unlikely.
Further investigation of PI4P levels by immunofluorescence also revealed a striking approximately twofold increase in PI4P fluorescence in Arf6 KO MEFs relative to controls (Fig. 5b). As expected 32 , a large pool of PI4P co-localized with the TGN marker Golgin97 in both genotypes (Fig. 5b). However, as recently reported by others 33 , PI4P was also present on more peripheral compartments likely corresponding to endosomes (Fig. 5b). We thus tested whether some of these PI4P-positive compartments are early endosomes by co-immunostaining MEFs for PI4P and EEA1 and found that the co-localization between PI4P and EEA1 was increased by twofold in Arf6 KO cells compared with controls (Fig. 5c). PI4P thus markedly accumulates on early endosomes. We also assessed whether PI4P accumulates on endosomal pools where retromer is present by immunostaining MEFs for endogenous PI4P and SNX6. We found not only that the overall PI4P/SNX6 co-localization was increased in Arf6 KO cells compared with WT cells (Fig. 5d, right panel) but also that the cellular region where this co-localization took place was different (Fig. 5d, left panel). While in Arf6 WT cells, PI4P and SNX6 mostly co-localized in the perinuclear region, they also co-localized in more peripheral compartments in Arf6 KO cells (Fig. 5d, insets). Comparable results were obtained when immunostaining MEFs for endogenous PI4P and retromer    Fig. 7). We thus concluded that in Arf6 KO cells, both PI4P and the retromer complex accumulate on endosomes.
Lowering PI4P levels rescues retromer function. These data led us to postulate that endosomal PI4P accumulation may be responsible for the defect in retromer tubules dynamics observed in Arf6 KO cells (Fig. 3g). To test this, we used phenylarsine oxide (PAO), a PI4K inhibitor 34 , to lower the levels of PI4P. After 30 min of 1 mM PAO treatment, levels of PI4P in Arf6 KO cells decreased by B20% and B35%, as assessed by immunostaining and HPLC, respectively (Fig. 6a,b). We quantified the number of GFP-SNX6 tubules in cells treated with dimethyl sulfoxide (DMSO) or PAO and visualized by spinning-disk microscopy ( Fig. 6c; Lowering PI4P levels rescues cholesterol redistribution. To test whether PI4P increase is also the cause of cholesterol accumulation in the LE/LYS of Arf6 KO cells, we assessed the impact of lowering PI4P levels on cholesterol localization using different but complementary approaches. First, we used PAO treatment and labelled cells for PI4P and filipin (Fig. 6d). After 30 min of treatment, the average number of filipin puncta per cell was slightly decreased, although it did not reach significance. However, the quantification revealed a shift towards cells with a lower number of puncta in the relative distribution (Fig. 6d,  center). The average size of filipin puncta was reduced with a specific decrease in the number of large (1.5-2.5 mm 2 ) puncta (Fig. 6d, right). This partial rescue of cholesterol distribution was accompanied by a normalization of CI-M6PR distribution and an increase in NPC2/LAMP1 co-localization after PAO treatment ( Supplementary Fig. 8b,c). Next, we overexpressed wild-type GFP-PIPKIgi1 (WT) or its kinase-dead K188A mutant in Arf6 KO MEFs (Fig. 7a). Quantification of filipin fluorescence in transfected cells after 48 h showed fewer filipin puncta of smaller size in WT-transfected cells (Fig. 7a). The kinase activity of PIPKIgi1 was thus required to rescue the cholesterol redistribution phenotype. To rule out that the changes in cholesterol distribution observed in the Arf6 KO were not due to the PI(4,5)P 2 levels increase we surprisingly observed after PAO treatment ( Supplementary Fig. 8a), we used a third strategy to selectively decrease PI4P levels in endosomes by exploiting the recently characterized endosomal PI4P-phosphatase Sac2 (refs 35,36) as a tool. We overexpressed wild-type GFP-Sac2 (WT) or its phosphatase-dead C458S mutant in Arf6 KO MEFs (Fig. 7b) and found that filipin puncta were fewer and of smaller size in WT GFP-Sac2 transfected cells (Fig. 7b). Cholesterol distribution rescue was thus dependent on the PI4P-phosphatase activity of Sac2. Altogether, these three approaches highlight the role of elevated PI4P levels in cholesterol accumulation in LE/LYS of Arf6 KO cells.

Discussion
Arf6 has been implicated in various aspects of cell physiology, primarily as a regulator of vesicular transport to and from the plasma membrane 4,5 . A key function of Arf6, however, appears to be the control of lipid signalling and, in particular, phosphoinositide signalling 11 . Parallel studies have also hinted that Arf6 may control cholesterol homeostasis, yet, through unknown mechanisms [18][19][20][21] . This study mechanistically links these disconnected bodies of literature, by identifying a phosphoinositide-based mechanism for the control of intracellular cholesterol distribution via the retromer pathway.
It is to our knowledge the first time that Arf6 is shown to regulate retromer function, although at least two previous studies presented circumstantial evidence consistent with this idea. First, even before retromer was identified, CI-M6PR was described as being present on Arf6-positive endosomes, but the functional relationship between these two proteins was not addressed 3 . Second, a more recent paper showed that EFA6A, a guanine ARTICLE nucleotide exchange factor for Arf6, interacts with SNX1 and regulates neurite outgrowth through this interaction, although neither the role of retromer nor the link with phosphoinositide signalling and cholesterol homeostasis was addressed in that context 37 . In this study, we propose a novel model whereby Arf6 controls the dynamics of retromer tubules in the endosome-to-TGN pathway responsible for the retrograde transport of CI-M6PR through negative regulation of an endosomal pool of PI4P (Fig. 8) 39,40 , an interesting possibility supported by our data is that the higher levels of PI4P found in the Arf6 KO cells may enhance the PI4P/SNX6 interaction on endosomal membranes at the expense of the SNX6/p150(Glued) interaction, resulting in retromer fission defects. Whether the lack of Arf6 and its NH 2terminal amphipathic helix directly alters membrane tubule dynamics as described for other Arf family members 41,42 , and contributes to retromer tubules defects, remains to be investigated. Similarly, a perturbation of the PI4P/PI(4,5)P 2 ratio on endosomal membranes may alter the fission efficiency of retromer tubules as we have previously reported for endocytic structures 43 .
Our study also further highlights the tight crosstalk between small GTPases and phosphoinositide metabolism 32,44 . In the context of retromer regulation, Rab5 has been shown to control class III PI3K (ref. 45), whose product PI3P mediates the recruitment of key components of the retromer assemblies, SNX1 and SNX2 (refs 27-29,46). In contrast, Rab7 directly interacts with the VPS core complex to facilitate the recruitment of retromer to endosomal membranes 46,47 . In this study, we propose a novel mechanism for GTPase-mediated control of retromer, involving the negative regulation of endosomal PI4P levels. As mentioned earlier, Arf6 controls PIPKIs, suggesting that PI(4,5)P 2 generation may be key to its biological function. However, the HPLC profile suggests that it is rather the accumulation of PI4P on endosomes than a reduction in PI(4,5)P 2 which accounts for the retromer phenotype. This is further supported by our rescue experiments showing that decreasing PI4P levels, rather than enhancing PI(4,5)P 2 levels, corrects the retromer phenotype (Figs 6 and 7). In addition, because PIPKIg overexpression rescues the cholesterol redistribution phenotype, we speculate that dysregulation of a PIPKI isoform on endosomal membranes is the basis of PI4P accumulation on these compartments. However, we cannot rule out contributions of class II PI3Ks, PI4Ks or PI4P phosphatases to this phenotype, particularly in light of the presence of some of their isoforms in the endosomal compartments, including the Sac2 inositol 4-phosphatase [34][35][36][48][49][50][51][52][53][54][55] .
While our study further establishes a link between Arf6 and NPC disease 19 , it also reports a role of retromer in endolysosomal cholesterol homeostasis. As retromer has been linked to several neurodegenerative diseases 56 and cholesterol accumulation in LE/LYS can lead to neurodegeneration 57 , this raises the question of the importance of the link between retromer and cholesterol in neurodegeneration. It is particularly intriguing that both decreased retromer levels 58 and increased cholesterol levels 59,60 have been linked to late-onset Alzheimer's disease. Further studies should give us insights into the role of the novel Arf6/retromer/cholesterol connection in the aetiology of neurodegenerative diseases.

Methods
Reagents. DMSO, tamoxifen, filipin, phenylarsine oxide (PAO) and bovine serum albumin (BSA) were purchased from Sigma-Aldrich. Saponin was from Acros. Generation of Arf6 Flox/Flox mice. To generate Arf6 conditional KO mice, the entire coding sequence (B4,000 bp) was flanked with loxP sites, with one (loxP) sub-cloned into a targeting vector upstream of the Arf6 coding region at the SspI site and a FRT-NEO-FRT-loxP cassette inserted downstream of the Arf6 sequence at the XhoI site. The targeting vector was electroporated into ES cells and successfully targeted ES cells were microinjected into blastocysts, which were then transferred into foster mothers for the generation of chimeric mice. Resultant Arf6 Flox-Neo/ þ mice were crossed with Flp Deleter mice (ACTB-Flpe from The Jackson Laboratory, Stock 005703) to generate Arf6 Flox/ þ mice. These mice were then crossed with the inducible Cre-ER mice (Rosa B6 Cre-ER/Cre-ER from the Jackson Laboratory, Stock 008463), to generate Arf6 Flox/ þ ;Cre-ER mice, which were then interbred to generate Arf6 Flox/Flox ;Cre-ER mice. Animals were used in full compliance with National Institutes of Health/ Institutional Animal Care and Use Committee guidelines. Specifically, adult mice were sacrificed via inhalation of carbon dioxide, followed by cervical dislocation, while embryos derived from the killed mothers were decapitated after inducing hypothermia on ice. The animal protocols were approved by the Committee on the Ethics of Animal Experiments of Columbia University. Arf6 Flox/ þ mice were made available to The Jackson Laboratory.
Cell cultures and transfection. Immortalized MEFs and HeLa cells were maintained at 37°C in a humidified 5% CO 2 atmosphere in DMEM with Glutamax supplemented with 10% fetal bovine serum and 1% penicillin/streptomycin (all from Life technologies). Cells were negative for mycoplasma contamination. Primary MEFs were generated from embryonic day 13.5 Arf6 Flox/Flox ;Cre-ER mice.  ARTICLE Torsos from Arf6 Flox/Flox ;Cre-ER embryos were minced and trypsinized for 15 min and homogenized by pipetting. MEFs were immortalized by multiple passaging and were used for experiments after 25 passages. Arf6 Flox/Flox ;Cre-ER MEFs contain a Cre-oestrogen receptor (Cre-ER) fusion protein, rendering Cre recombinase inducible by tamoxifen 65 . Arf6 Flox/Flox ;Cre-ER MEFs were plated at a high density and treated the next day with DMSO or 10 mM tamoxifen. After 72 h, cells were split in fresh media and left in culture for another 72 h. Arf6 deletion was complete by then. All experiments were performed between 3 and 7 days after complete Arf6 deletion. MEFs were transiently transfected using Lipofectamine LTX (Life Technologies) and protein expression was assessed after 24 or 48 h. HeLa cells were transfected using Lipofectamine 2000 (Life Technologies). Protein expression was assessed after 24 or 48 h and knockdowns were assessed at 72 h.
Anionic phospholipid analysis. Phosphoinositides and other anionic phospholipids were measured as deacylated lipids using anionic exchange HPLC with suppressed conductivity detection 66,67 . Samples were prepared from cells grown to half-confluency in 100-mm dishes, washed with ice-cold HBSS and processed for phosphoinositide enrichment, followed by deacylation 66,67 .
Free cholesterol analysis. Free cholesterol levels (and other lipids) were measured by separating complex lipid mixtures by reverse phase HPLC, followed by detection using multiple reactions monitoring 68 . Samples were prepared from cells grown to half-confluency in 60-mm dishes, washed with ice-cold HBSS and processed for lipid extraction using a modified Bligh and Dyer protocol 68 . Free cholesterol levels were quantified by referencing a D7-cholesterol (Avanti Polar Lipids) internal standard that was spiked into each sample before the analysis.
Immunofluorescence and filipin staining. For NPC2 uptake experiments, MEFs were exposed to fresh medium containing PBS (Boston BioProducts) or purified NPC2-Alexa488 (ref. 62) 60 nM for 24 h, rinsed twice with HBSS (Life Technologies) and fixed. For PAO experiments, MEFs were exposed to fresh medium containing DMSO or PAO 1 mM for 30 min at 37°C, rinsed twice with HBSS and fixed. For immunofluorescence experiments, cells grown on glass coverslips were washed once with HBSS and fixed with 4% paraformaldehyde (Electron Microscopy Sciences) for 20 min at room temperature or at 37°C (when preservation of the tubular structures was crucial). Cells were then washed twice in HBSS and incubated with NH4Cl (50 mM in PBS) for 10 min. For NPC2 staining, cells were fixed for 1 h in Bouin's solution (Sigma-Aldrich). Cells were then washed twice in PBS and permeabilized with solution A (saponin 0.05%, BSA 5% in PBS) for at least 45 min at 37°C. They were then incubated with primary antibodies diluted in solution A for 1 h at room temperature, washed three times in solution A, incubated with fluorescent secondary antibodies diluted in solution A for 1 h at room temperature and washed again three times in solution A. An extra step was performed for filipin staining. Cells were washed once in PBS, incubated in filipin (0.5 mg ml À 1 in PBS) for 1 h at room temperature and washed twice with PBS. Cells were finally washed once with PBS and coverslips were mounted in Vectashield mounting medium (Vectorlabs).
Confocal microscopy. Z-stacks were acquired by confocal laser scanning microscopy (Zeiss LSM 700). Fluorescence was collected with a Â 63 plan apochromat immersion oil objective (numerical aperture (NA) 1.4). Extraction of single z-frame and maximum intensity projections were performed with the ImageJ software. Integrated densities, that is, cell area Â [(mean cell intensities summed on the whole z-stack)-(mean background intensities summed on the whole z-stack)], were measured with Image J software. Z-stack co-localization percentage, that is, Â 100 (Pearson's coefficient) 2 , were obtained with the JACoP plugin of ImageJ 69 . The number and size of puncta were quantified with ICY software 70 . We checked that the cells we were comparing for fluorescence intensity, size and number of organelles or co-localization had comparable areas (P40.05 in Student's t-test). Each experiment was independently repeated at least three times, resulting in the analysis of 22 to 42 cells for each condition.
3D Structured illumination microscopy. Cells were grown on high performance #1.5 coverslips (Zeiss). Immunofluorescence was performed as described for confocal microscopy but secondary antibodies were used twice as concentrated (1/200) and coverslips were mounted in ProLong Gold (Life technologies). Z-stacks were acquired on a Nikon N-SIM microscope equipped with an iXon Ultra 897 camera (Andor) and a Â 100 plan apochromat immersion oil objective (NA 1.49). Image reconstruction and analysis were performed with the NIS-Elements software.
Spinning-disk confocal live imaging. Cells were grown in 35-mm glass-bottom dishes (MatTek) and imaged 48 h after transfection. Before imaging, medium was replaced to get rid of any debris and to ensure accuracy of volume for the PAO experiments. For kinetic PAO experiments, Arf6 KO cells expressing GFP-SNX6 were imaged for 2 min and their localization recorded. At time 0, DMSO or PAO (1 mM final) diluted in medium was added to the cells and three movies of 2 min were recorded (before 15 min, between 15 and 30 min and between 30 and 45 min). Only cells displaying at least two tubules at time 0 were kept for analysis. Live imaging was carried out using a Nikon Ti Eclipse microscope equipped with a Yokogawa CSU-X1 spinning-disk, a motorized stage and a Photometrics Evolve 512 camera (Roper Scientific) controlled with the NIS-Elements software. Movies were acquired at 37°C and 5% CO 2 with a Tokai Hit stage-top incubator and objective heater. Fluorescence was collected with a 60 Â plan apochromat immersion oil objective (NA 1.49). Images were collected at 1 Â 1 binning with an exposure time of 100 ms. The number, maximum length and persistence of GFP-SNX6 tubules were analysed with ImageJ software.
Protein biochemistry and immunoblotting. Cells were washed with ice-cold HBSS, scraped and centrifuged for 10 min at 16,000 g at 4°C. Pellets were resuspended in RIPA buffer (Pierce) complemented with cOmplete protease inhibitor cocktail (Roche) and proteins were extracted on a wheel at 4°C for at least 30 min. Samples were then centrifuged for 15 min at 16,000 g and proteins in the supernatant were processed for protein dosage (BCA, Pierce). SDS-PAGE was performed on samples (15 mg total proteins) loaded in NuPAGE 4-12% Bis-Tris gels (Life technologies). For quantification, wet transfer was performed at 40 V for 135 min at 4°C (Bio-Rad). For Arf6 deletion confirmation, transfer was performed with iBlot (Life Technologies). Primary and secondary antibodies were incubated overnight at 4°C and 45 min at room temperature, respectively. Revelation was performed with Immobilon Western Chemiluminescent HRP Substrate (EMD Millipore) and the chemiluminescent signal was imaged with ImageQuant LAS4000 mini (GE Healthcare). Quantification was performed with ImageJ. Uncropped blots are shown in Supplementary Fig. 9.
Statistical analysis. Statistical calculations were performed using GraphPad Prism software (version 5.02). All the data are given as mean ± s.e.m. In most cases, when comparing two samples, two-tailed Student's t-test was performed. If a clear hypothesis was postulated that made it valid, then we used a one-tailed t-test. When variances were not comparable, Welch's correction was applied. When the distribution could not be assumed to be Gaussian (for compartments size), we used a non parametric Mann-Whitney test. When more samples were compared and Bartlett's test showed that variances could be compared, we used one-way ANOVA or two-way ANOVA with Bonferroni post-tests. If variances could not be compared (P value in Bartlett's testo0.05), then we used t-tests. Chi-square test was used to compare distributions. Outliers, defined as values that were superior to (mean þ 3 Â s.d.) or inferior to (mean À 3 Â s.d.) were excluded.
Data availability. The authors declare that all the data supporting the findings of this study are available within the article and its Supplementary Information files or are available from the corresponding authors on request. Arf6 Flox/ þ mice generated in this study have been made available to the Jackson Laboratory.