An ESCRT-dependent step in fatty acid transfer from lipid droplets to mitochondria through VPS13D−TSG101 interactions

Upon starvation, cells rewire their metabolism, switching from glucose-based metabolism to mitochondrial oxidation of fatty acids, which require the transfer of FAs from lipid droplets (LDs) to mitochondria at mitochondria−LD membrane contact sites (MCSs). However, factors responsible for FA transfer at these MCSs remain uncharacterized. Here, we demonstrate that vacuolar protein sorting-associated protein 13D (VPS13D), loss-of-function mutations of which cause spastic ataxia, coordinates FA trafficking in conjunction with the endosomal sorting complex required for transport (ESCRT) protein tumor susceptibility 101 (TSG101). The VPS13 adaptor-binding domain of VPS13D and TSG101 directly remodels LD membranes in a cooperative manner. The lipid transfer domain of human VPS13D binds glycerophospholipids and FAs in vitro. Depletion of VPS13D, TSG101, or ESCRT-III proteins inhibits FA trafficking from LDs to mitochondria. Our findings suggest that VPS13D mediates the ESCRT-dependent remodeling of LD membranes to facilitate FA transfer at mitochondria-LD contacts.

T he fitness of an organism depends on its ability to adapt to changes in nutrient availability in the environment. To do so, cells inside organisms store energy as neutral lipid esters in lipid droplets (LDs), a type of lipid storage organelle conserved from yeast to humans, and then reprogram their metabolism from glucose catabolism to the mitochondrial oxidation of fatty acids (FAs) when nutrients become limited. Starvation triggers the release of free FAs from triacylglycerols stored in LDs and the transfer of FAs to the mitochondrial matrix for oxidation, which produces ATP to sustain cellular activities 1,2 .
Although the biochemical basis of FA activation and import into mitochondria has been extensively studied for decades, our understanding of how FAs are transferred from LDs to mitochondria to drive FA oxidation under starvation is still limited. Current models propose that FA transfer relies on mitochondria-LD membrane contact sites (MCSs), where the membranes of the two opposing organelles are in close proximity (typically 10-30 nm) [3][4][5][6] . Consistent with this notion, the association of LDs and mitochondria increases in response to nutrient starvation in cultured cells 3,[7][8][9] and exercise in skeletal muscle 10 . This spatial proximity between mitochondria and LDs can decrease the prevalence of cytosolic FAs, preventing lipotoxicity 3,8 . Another model suggests a role for mitochondria−LD MCSs in adipocytes, where mitochondria closely associated with large LDs exhibit reduced β-oxidation and display increased ATP synthesis, suggesting that the major function of these LD-associated mitochondria is to support the growth and expansion of LDs 11 . Therefore, it is plausible that mitochondria−LD MCSs serve as sites for both lipogenesis and lipolysis in different cell types or under different metabolic conditions [12][13][14][15] . Nevertheless, the molecular machinery that promotes FA transfer at mitochondria −LD MCSs remains poorly understood.
Vacuolar protein sorting-associated protein 13D (VPS13D) belongs to the metazoan VPS13 family. The human VPS13 family contains four members, namely, VPS13A, B, C, and D, whereas a single VPS13 exists in yeast 16 . A pioneering study demonstrated that VPS13A and VPS13C are lipid transporters at endoplasmic reticulum-mitochondria/LD and ER-late endosome/LD MCSs, respectively 17 , while VPS13B is suggested to be a Golgi apparatusassociated protein required for Golgi integrity and neurite outgrowth 18,19 . Human VPS13 proteins are of great biomedical interest, as loss-of-function mutations in each of them are associated with genetic diseases [20][21][22][23][24] . Loss-of-function mutations of VPS13D cause a type of autosomal recessive ataxia with abnormal mitochondrial morphology, reduced energy generation, and lipidosis 23,24 . Human VPS13D is a giant protein consisting of 4363 amino acids (transcript variant 2, NM_018156), with a molecular weight of 493 kDa. VPS13D gene is the only essential gene in the VPS13 family, suggesting that VPS13D participates in cellular processes fundamental for organism survival 25 . A pioneering study identified VPS13D as a ubiquitin binding protein and a key regulator of mitochondrial size and clearance in Drosophila 26 .
Here we demonstrated that VPS13D coordinated FA trafficking in conjunction with the endosomal sorting complex required for transport (ESCRT) protein tumor susceptibility 101 (TSG101). The VPS13 adaptor-binding (VAB) domain of VPS13D and TSG101 directly remodels LD membranes. The lipid transfer domain of human VPS13D binds glycerophospholipids and FAs in vitro. Depletion of VPS13D, TSG101, or ESCRT-III proteins inhibits FA trafficking from LDs to mitochondria.

Results
The localization of mammalian VPS13D. To date, the localization and cellular functions of mammalian VPS13D remain largely unknown. To explore the cellular functions of VPS13D, we initiated with an examination of the localization of endogenous VPS13D by immunofluorescence staining (IF). Although a commercial antibody was not initially effective in IF, its specificity for VPS13D could be improved by pre-clearing the antibody against fixed and permeabilized VPS13D-suppressed cells according to a protocol modified from a previous study 27 (Supplementary Fig. S1a). The efficiency and specificity of smallinterfering RNAs (siRNAs)-mediated VPS13D suppression in the cells used for pre-clearing the antibody was confirmed by quantitative PCR (Fig. 1a) and western blotting (Fig. 1b). IF experiments using this pre-cleared antibody revealed that VPS13D was preferentially localized to the junctions between mitochondria and LDs in response to oleic acid (OA) stimulation in complete medium (CM) followed by 5 h of starvation in Earle's Balanced Salt Solution (OA/EBSS) (Fig. 1c, top panel); however, its localization at these junctions was almost completely lost upon VPS13D suppression (Fig. 1c, bottom panel). We found that 49% of endogenous VPS13D puncta were present at mitochondria −LD junctions (Fig. 1d), and 79% of these junctions were marked by VPS13D in wild-type cells (Fig. 1e).
To confirm the localization of VPS13D, we ectopically expressed a superfolder green fluorescent protein (sfGFP)-tagged VPS13D (transcript variant 2, NM_018156) (VPS13D^sfGFP), in which sfGFP was internally fused to VPS13D at a site shown to preserve yeast VPS13 function 28 , in HEK293 cells. We used HEK293 cells for this purpose as they were well suited for transfection and expression of the huge VPS13D^sfGFP construct (~20 K base pairs in size). Consistent with the results of endogenous VPS13D IF experiments, a significant portion of VPS13D^sfGFP puncta was present at mitochondria-LD junctions in response to OA/EBSS by high−resolution confocal imaging ( Supplementary Fig. 1b, c).
Two amphipathic helices in the VPS13_C domain of VPS13D target LDs. The localization of VPS13D at mitochondria−LD junctions requires the presence of binding sites responsible for recognizing these two membranes. Based on bioinformatics analysis, VPS13D was found to consist of an N-terminal putative LTD, a UBA domain followed by a VAB domain, and a VPS13_C domain at its C terminus (Fig. 2a, top panel). To identify the regions responsible for recognizing LDs and mitochondrial membranes, respectively, we performed molecular dissection of VPS13D. Among the five truncated versions of VPS13D tested (tVPS13D1-5, Fig. 2a, bottom panel), a truncated version of the C-terminal region (tVPS13D-5) was associated with the LD membrane, as revealed by the co-localization between tVPS13D-5 and the class I LD membrane protein ACSL3 (Fig. 2b). Video analysis confirmed that the association of tVPS13D-5 with the LD membrane was stable and specific ( Supplementary Fig. 2a). OA stimulation strongly enhanced the recruitment of tVPS13D-5 to the LDs (Fig. 2c). We further found that the VPS13_C domain alone (3981-4339 residues), which is present in tVPS13D-5, was exclusively localized to LDs in response to OA treatment (Fig. 2d), whereas under normal conditions, the VPS13_C domain could be recruited to both LDs and mitochondria (Supplementary Fig. 2b).
Previous studies have shown that the VPS13_C domain of VPS13A/C binds LDs through an amphipathic helix 17 . To examine whether the VPS13_C domain of VPS13D binds LDs by a similar mechanism, we searched for amphipathic helices in the VPS13_C domain using the HeliQuest tool 29 . Two amphipathic helices were identified, with the first helix (Helix-1) ranging from residues 3981 to 3998, while the other (Helix-2) ranging from residues 4050 to 4071 (Fig. 2e). To determine whether these two amphipathic helices were required for VPS13_C domain binding to LDs, point mutations were introduced into the hydrophobic interfaces of each helix (Helix-1: L3991Q; Helix-2: L4052Q and L4053Q) to break their structures (Fig. 2e). Disruption of both Helix-1 and Helix-2 not only blocked the binding of VPS13_C to LDs under normal conditions ( Supplementary Fig. 2c) but also in response to OA stimulation (Fig. 2f), indicating that the binding of VPS13D to LDs is mediated through the two amphipathic helices in the VPS13_C region.
We further tested whether the two helices were required for LD binding of full-length VPS13D. VPS13D with three point mutations in the two helices (L3991Q L4052Q L4053Q) failed to associate with LDs. Instead, this mutant associated with mitochondria which appeared to be fragmented (Fig. 2g).
The N-terminal region of VPS13D targets mitochondria. To determine which region of VPS13D is required to target mitochondria, we generated a set of N-terminal VPS13D truncations (Fig. 3a). A region at the N-terminal portion of VPS13D (residues 404-913) showed mitochondrial localization in response to OA treatment (Fig. 3b, bottom panel), whereas mitochondrial localization was barely detectable by microscopy without OA treatment (Fig. 3b, top panel). We found that a region of VPS13D containing residues 404-612 was sufficient to target the outer mitochondrial membrane (OMM) in an OA-dependent manner (Fig. 3c), whereas the other region from residues 613-913 failed to localize to the OMM ( Supplementary Fig. 2d Fig. 1 The localization of mammalian VPS13D. a qPCR assays confirmed the efficiency and specificity of siRNAs-mediated VPS13D suppression in HEK293 cells. Two-tailed unpaired student's t-test. Mean ± SD. b Western blots demonstrated the efficiency of siRNAs-mediated VPS13D suppression in HEK293 cells transfected with VPS13D^sfGFP with anti-GFP and anti-tubulin antibody. c Confocal image of fixed HEK293 cells stained with pre-cleared anti-VPS13D antibody (red), BODIPY493/503 (LD; green), and anti-CoxIV antibody (mitochondria; blue) upon treatments with either scrambled (top panel) or VPS13D siRNAs (bottom panel). Left: merged image of whole cells; right: three insets from boxed regions in whole cell image. Contrast range of images was set to the same level for both scrambled and VPS13D siRNAs treated cells. Yellow arrows denoted VPS13D puncta at potential mitochondria-LD junctions. Four independent imaging sessions were performed with similar results. d Percentage of VPS13D puncta (n = 644 from 20 cells in four independent assays) present at mitochondria-LD junctions upon OA/EBSS treatment. Mean ± SD. e Percentage of mitochondria-LD junctions (n = 286 from 20 cells in four independent assays) marked with VPS13D puncta upon OA/EBSS treatment. Mean ± SD. Scale bar, 10 μm in whole cell image and 2 μm in insets in c.   stimulation in COS7 cells stably expressing VPS13D_N (Fig. 3d). Mitochondrial localization of VPS13D_N was further confirmed in HEK293 cells by confocal microscopy. A fraction of VPS13D_N was localized to mitochondria even in the absence of OA (Fig. 3e, top panel), and its recruitment to the OMM was enhanced upon OA stimulation (Fig. 3e, bottom panel).
The VAB domain of VPS13D interacts with TSG101. Next, we investigated the cellular localization and functions of the VAB domain of VPS13D. The VAB domain with a GFP tagged at its Cterminus (VAB-GFP) formed bright puncta associating with LDs and mitochondria upon OA/EBSS stimulation in HEK293 cells ( Supplementary Fig. 3a, b). We noted that the VAB domain was preferentially localized to one side of each LD instead of being distributed over the entire LD surface ( Supplementary Fig. 3a, c). Time-lapse images confirmed that the association of VAB puncta with LDs ( Supplementary Fig. 3d) and mitochondria (Supplementary Fig. 3e) was stable. The VAB domain contains six repeats (R1−R6), which are conserved in four human VPS13 proteins 30 . Expression of either R1-R3 or R4-R6 alone resulted in diffused localization all over the cell without the formation of puncta ( Supplementary Fig. 3f, g), suggesting that all six repeats (R1-R6) were required for VAB-GFP localization. Cell fractionation analysis demonstrated that the majority of VAB-GFP was found in the cytosolic fractions, whereas its presence in the crude mitochondrial fraction (P2/P3) could be detected upon Dithiobis (succinimidyl propionate)-mediated protein−protein crosslinking ( Supplementary Fig. 3h), suggesting that the association of VAB-GFP with LDs or mitochondria is transient and might be mediated by protein−protein interactions. We found a Pro-Ser-Ala-Pro (PSAP, residues 3044-3047) motif in the VAB domain of VPS13D (Fig. 4a), which was highly conserved in mammals and specific to VPS13D rather than other mammalian VPS13 paralogs ( Supplementary Fig. 4a). The PS/ TAP motif was previously identified in HIV-1, the Ebola virus, and other enveloped viruses that can recruit TSG101, an ESCRT protein 31,32 . The ESCRT complex plays essential roles in multiple core cellular events involving membrane remodeling 33,34 , such as cytokinesis 32 , endosome maturation 35,36 , autophagy 37 , membrane repair [38][39][40][41] , and FA trafficking from LDs to peroxisomes 42 . The presence of a PSAP motif in the VAB domain of VPS13D suggested a potential interaction between VPS13D and TSG101. Indeed, GFP-Trap assays confirmed the interaction between VPS13D and TSG101 in HEK293 cells ( Supplementary Fig. 4b), in agreement with a previous study on Drosophila 26 . The VAB domain was sufficient for VPS13D interactions with both endogenous TSG101 and Halo-TSG101 (Fig. 4b). The interaction between VAB domain and endogenous TSG101 was also confirmed in COS7 cells stably expressing VAB-GFP (Supplementary Fig. 4c). Deletion of the PSAP sequence in the VAB domain reduced its interaction with TSG101 by~60% (Fig. 4c), as assessed by GFP-Trap assays, but did not entirely eliminate the interaction, indicating that this motif has an important role in the VAB-TSG101 interaction but also suggesting that additional regions of the VAB domain and TSG101 interact. TSG101 protein contains several unique domains, such as an inactive ubiquitinconjugating domain (UEV), a proline-rich (PR) region, a coiledcoil (CC) domain, and a steadiness box (SB) (Fig. 4d) 43 . The UEV domain and PR-CC-SB domains independently interacted with the VAB domain, as revealed by GFP-Trap assays (Fig. 4e, Supplementary Fig. 4d), indicating a bimodal interaction between the VAB domain and TSG101 in a manner reminiscent of TSG101 interaction with the ESCRT-0 protein, namely, hepatocyte growth factor-regulated tyrosine kinase substrate (HGS) 44 or a TSG101-associated ligase Tal 45 . For instance, the UEV domain of TSG101 was found to independently interact with the PR domain and C-terminal clathrin-binding domain of HGS 44 . Although TSG101 was previously found to interact with HGS on endosomal membranes 44 , we found that VAB-GFP puncta did not associate with HGS-decorated endosomal membranes under OA/EBSS treatment ( Supplementary Fig. 4e, f), suggesting that the VAB-TSG101 interaction is independent of HGS.
Next, we investigated the cellular localization of TSG101. Previous studies have shown that TSG101 preferentially localizes to the late endosomal structures, and that its cellular localization is under cell cycle regulation 46,47 . To avoid the artifacts due to high levels of Halo-TSG101 expression driven by the cytomegalovirus promoter (CMV), we generated a stable HEK293 cell line that was infected with a lentiviral vector expressing the transgene under the control of a weak promoter EF1α; this cell line showed a much lower and more homogeneous level of Halo-TSG101 among cells (hereafter referred to as Halo-TSG101), as revealed by flow cytometry analysis ( Supplementary Fig. 4g). TSG101 was not localized to LDs under normal conditions ( Supplementary  Fig. 4h, i), whereas a significant portion (~25%) of Halo-TSG101 puncta was recruited to LDs after OA/EBSS stimulation (Fig. 4f, g). In addition, the recruitment of endogenous TSG101 to LDs was confirmed in OA/EBSS-treated HEK293 cells by IF staining with TSG101 antibody, the specificity of which was validated by siRNAs-mediated suppression in IF experiments (Supplementary Fig. 5a-d).
Next, we sought to determine the role of VPS13D in TSG101 localization to LDs. Depletion of VPS13D abolished the recruitment of TSG101 to LDs under OA/EBSS stimulation ( Fig. 4f, g). In contrast, overexpression of VAB-GFP increased TSG101 recruitment to LDs upon OA/EBSS treatment by~3.2fold. Deletion of PSAP significantly reduced the TSG101 recruitment resulting from VAB-GFP overexpression (Fig. 4f, g and Supplementary Fig. 5e, f). The UEV domain alone and PR-CC-SB domain alone failed to associate with LDs in response to VAB-GFP overexpression under OA/EBSS stimulation (Supplementary Fig. 5g, h), suggesting that the bivalent interaction of VAB and TSG101 might be critical for TSG101 recruitment to LDs.
One of the most noticeable features of VAB/TSG101-decorated LDs was that the morphology of LDs was dramatically modified; we observed both constriction of LDs (Fig. 4h, insets 1, 3, 4, 6) and budding/tubulation from LDs (Fig. 4h, insets 2, 5, 6) with specific co-localization of VAB-GFP and Halo-TSG101 at constricted/budding sites, as shown by high-resolution confocal live cell microscopy. This suggests that TSG101 and the VAB domain of VPS13D have direct roles in modulating LD morphology. To further confirm the roles of the VAB domain of VPS13D and TSG101 in the remodeling of LD membranes, we directly examined the LD morphologies upon OA/EBSS stimulations by performing transmission electron microscopy (TEM) of HEK293 cells co-expressing VAB-GFP and Halo-TSG101. Electron micrographs revealed that a substantial portion of LDs was dramatically deformed upon overexpression of Halo-TSG101 and VAB-GFP ( Fig. 4i, j): LDs underwent either budding (Fig. 4i, right panel; inset 1) or constriction (Fig. 4i, right panel; inset 2). However, the LDs in OA/EBSS-treated cells expressing the empty vectors were typically spherical with mild deformations (Fig. 4i, left panel). In addition, we noted that the mitochondria were highly condensed in the VAB/TSG101-overexpressing cells, suggesting that the mitochondrial functions might be impaired in these cells.
Given the dynamic nature of mitochondria and LDs, we tracked the VAB/TSG101-decorated LDs relative to mitochondria over time and found that the VAB/TSG101 could be specifically localized at mitochondria−LD contacts, and that the size of BODIPY-C12-labeled LDs appeared to gradually decrease over time as they contacted mitochondria ( Fig. 4k and Supplementary Movie 1). Collectively, our findings suggest that VPS13D and TSG101 directly remodel LD membranes in a cooperative manner.
VPS13D is required for maintenance of mitochondria-LDs interactions. The presence of VPS13D at mitochondria−LD junctions upon OA/EBSS treatment led us to hypothesize that VPS13D might play a role at such MCSs. Therefore, we initially explored the role of VPS13D in the maintenance of mitochondria −LD MCSs using Doxycycline (Dox)-inducible mitochondria −LD MCSs reporter HEK293 cell lines, in which the proximitymediated GFP reconstitution of Tom20-GFP 1-10 and Plin2-GFP 11 specifically occurs at mitochondria−LD MCSs. Because the reconstitution of GFP was irreversible, a low concentration (0.1 μM) of Dox and short incubation time (12 or 18 h) were applied to avoid artificial tethering between mitochondria and LDs. Under this condition, the morphology of both mitochondria and LDs appeared to be normal, and GFP signals were specifically confined to mitochondria−LD MCSs (Fig. 5a). Suppression of VPS13D substantially decreased the formation of mitochondria −LD MCSs in response to OA/EBSS treatment by confocal microscopy (Fig. 5b) and by high-throughput flow cytometry ( Fig. 5c and Supplementary Fig. 6a-c). Immunoblot analysis showed that the expression level of GFP 1-10 was not significantly   We further confirmed the role of VPS13D in the maintenance of mitochondria-LD junctions by using a reversible dimerizationdependent fluorescent proteins (ddFPs) system 48 . We observed that the ddGFP signals were specifically enriched at mitochondria-LDs junctions but these signals were much weaker compared with the splitGFP signals, possibly because ddGFP dimerization is reversible ( Supplementary Fig. 6i). Suppression of VPS13D significantly reduced the mitochondria-LD interaction but TSG101 depletion did not substantially alter such MCSs, as revealed by flow cytometry analysis of ddGFP fluorescence ( Supplementary Fig. 6j).
In addition, we examined the effects of VPS13D suppression on mitochondria-LD interactions by co-localization analysis. Suppression of VPS13D substantially reduced the extent of mitochondria−LD interactions in HEK293 cells (Fig. 5d, e), as revealed by the decreased number of co-localized pixels between mitoBFP and ACSL3-Halo. In addition, VPS13D suppression resulted in more LDs at the cell periphery and fewer LDs at perinuclear regions (Fig. 5f). VPS13D depletion led to a significant increase in both LD size (Fig. 5g) and density (number of LDs / cell area) (Fig. 5h). Moreover, the duration of mitochondria−LD contacts decreased upon VPS13D suppression, as determined by live-cell imaging (Fig. 5i, j and Supplementary Movie 2).
VPS13D and TSG101 are required for efficient transfer of FAs from LDs to mitochondria. Given that the transfer of FAs from LDs to mitochondria may rely on mitochondria-LD MCSs, we next explored whether VPS13D plays a role in the transfer of FA using fluorescent FA analog BODIPY-C12 (Red C12)-based pulse-chase assays as previously described 3 ( Supplementary  Fig. 7a). Red C12 was initially incorporated into LDs (t = 0) and was gradually transferred from LDs to mitochondria after starvation (t = 5 or 24 h), as revealed by the presence of Red C12 fluorescence in mitochondria (Fig. 6b). In addition, we found that the transfer rate of Red C12 from LDs to mitochondria contacting LDs was significantly higher than that to mitochondria not adjacent to LDs by live cell microscopy after 5 h of starvation ( Supplementary Fig. 7b, c).
We next use RNA interference to determine the role of VPS13D in FA transfer from LDs to mitochondria. VPS13D suppression substantially inhibited the transfer of Red C12 to mitochondria upon starvation, as revealed by the large amount of Red C12 that was still incorporated in LDs, while less was trafficked to mitochondria (Fig. 6a, b, k). Both the size and number of Red C12-labeled LDs markedly increased after 5 h of starvation in VPS13D-suppressed cells compared with control cells (Fig. 6a, b, i and j).
We further investigated whether TSG101 and other components of the ESCRT complex were required for FA trafficking. Distinct from a canonical role of ESCRTs that function in inverse budding reactions on negatively curved membranes, four ESCRT proteins, namely, CHMP6, CHMP4B, CHMP1B, and IST1, were previously suggested to mediate classical-topology membrane shaping reactions on positively curved membranes 34,42,49 , which more likely fits into the scenario of remodeling of LDs membranes during FAs trafficking. Therefore, we performed siRNA-based analysis of the four ESCRT proteins as well as TSG101 and ALIX, an ESCRT-associated protein required for cytokinesis 32 , using Red C12 pulse-chase assays in cells starved for 5 h. Suppression of TSG101 or CHMP4B resulted in a~3-fold increase in the size of red C12-labeled LDs after 5 h of starvation (Fig. 6c, f, i). Moreover, depletion of VPS13D, TSG101, CHMP4B, CHMP6, or IST1 led to a significant increase in the number of Red C12-labeled LDs compared with control cells (Fig. 6a-i, j). In addition to VPS13D, depletion of TSG101, CHMP6, or the ESCRT-III proteins resulted in a significant reduction of Red C12 transfer to mitochondria ( Fig. 6c-g, k), suggesting that these ESCRT proteins play important roles in FA trafficking from LDs to mitochondria. In contrast, depletion of ALIX did not significantly affect either the number or the size of red C12-labeled LDs, or the transfer of FA upon 5 h of starvation (Fig. 6h, i, j & k), suggesting that ALIX might be dispensable for this process. In addition, we found no substantial changes in the total amounts of BODIPY-C12 taken up by either control, VPS13D, or ESCRT protein-depleted cells at t = 0 (right before starvation) (Fig. 6l), suggesting that the increase in the size and number of LDs was not due to more FAs taken up by VPS13D, or ESCRT proteins-suppressed cells. Taken together, our findings suggested that VPS13D and the ESCRT complex were required for efficient FA trafficking from LDs to mitochondria. However, it should be noted that the fluorophore-conjugated FAs likely have different biophysical properties from the endogenous FAs, which could affect on their trafficking and metabolism 50 , the in vivo FA trafficking results using BODIPY-conjugated FAs should be further verified with the use of isotopically labeled FAs.
As an alternative approach, we confirmed the roles of VPS13D and TGS101 in the transport and oxidation of endogenous FAs in  mitochondria via seahorse FA oxidation (FAO) assays. siRNAmediated suppression of either VPS13D or TSG101 in HepG2 cells cultured in substrate-limited medium led to a slight reduction in basal respiration but a higher reduction in the maximal respiration of mitochondria. As a control, cells were pretreated with Etomoxir (Eto, a carnitine palmitoyl transferase-1 inhibitor), which substantially reduced both the basal and maximal respiration of mitochondria (Fig. 6m). The reduced FA oxidation observed in VPS13D-suppressed cells was consistent with the findings of two clinical studies, in which cells from patients of ataxia with spasticity (caused by loss-of-function mutations of VPS13D) showed lipidosis and reduced energy productions 23,51 . In addition, a previous study showed that the cristae of mitochondria appeared to be intact in VPS13D-depleted cells, as observed by TEM, indicating that the capacity of mitochondria for energy generation might be largely unaffected upon VPS13D-suppression 26 . Therefore, these results suggest that the low energy production and lipodosis observed in the patient cells may be attributed to the impaired transfer of FAs to mitochondria. The lipid transfer domain of VPS13D directly binds lipid fatty acid moieties in vitro. VPS13D harbors a putative LTD at its N terminus, which is highly conserved among VPS13 paralogs and ATG2 27,52 . Two recent studies demonstrated that the LTD of fungi VPS13 proteins binds and transfers glycerophospholipids 17,53 . It is still unknown whether human VPS13D-LTD could directly binds and transfers lipid fatty acid moieties. To test whether VPS13D-LTD could directly bind lipids in vitro, we performed protein-lipid co-migration assays. Purified VPS13D-LTD co-migrated with nitrobenzoxadiazole (NBD)-labeled glycerophospholipids including phosphatidic acid (PA), phosphatidylserine (PS), phosphatidylethanolamine (PE), phosphatidylcholine (PC), and to a lesser extent, ceramide (Cer) (Fig. 7a), which was consistent with the findings of a previous study in which yeast VPS13α bound glycerophospholipids 17,53 . Besides phospholipids, purified VPS13D-LTD could bind NBD-labeled fatty acids C12 (NBD-C12) and BODIPY-labeled fatty acids C16 (BODIPY-C16) (Fig. 7a). The binding of purified VPS13D-LTD to FAs was stronger compared with that of VPS13D-LTD to phospholipids, as indicated by the ratio of FA fluorescence intensity to the amount of protein (Fig. 7b).
To further confirm that VPS13D binds FAs, we sought to estimate the stoichiometry of the binding of VPS13D-LTD to FAs via in vitro protein-lipids co-migration assays. Through titration, we found that one VPS13D-LTD molecule (320 residues) could bind~5 molecules of BODIPY-C16 (Fig. 7c, d), and we further estimated that a full-length VPS13D molecule with a long groove serving as a lipid transfer channel at its Nterminus (~1500 residues) could bind~25 molecules of BODIPY-C16, which agreed with the estimated stoichiometry of the binding of yeast VPS13 and ATG2 to phospholipids 17,27 , and this result supported the notion that VPS13 is a bridge lipid transporter that is capable of binding tens of lipids at once 53 . In addition, to validate the finding that VPS13D binds endogenous, unlabeled FAs, we examined whether unlabeled FAs could outcompete BODIPY-C16 for binding to VPS13D-LTD. Competition assays indicated that palmitic acid, but not OA, could sufficiently outcompete BODIPY-C16 for binding to VPS13D-LTD in a dose-dependent manner (Fig. 7e, f). Collectively, these results supported the hypothesis that VPS13D-LTD directly binds FAs in vitro.
We further validated the role of the lipid transfer activities of VPS13D in FA transport in cells. Inspired by a recent study 27 , in which a fragment of the ATG2A lipid transfer domain (mini-ATG2A, residues 1-345) could almost completely rescue the phenotype of ATG2 knockout lines upon overexpression, we tested whether VPS13D-LTD was indispensable for FA transfer from LD to mitochondria by rescuing the FA transfer defects induced by VPS13D suppression. Based on the high conservation of the LTDs between VPS13 and ATG2, we made a hydrophobic surface mutant of VPS13D-LTD (L42R, L66R, L85R, I148R, I152E, I174R, M181R, F203E, L278R, L293R, W319R) (Supplementary Fig. 7d), which could still bind but not transport lipid fatty acids moieties 27 .
VPS13D suppression substantially inhibited the Red C12 transfer from LDs to mitochondria, and this defect was significantly rescued by expression of siRNA-resistant full-length VPS13D^sfGFP (Fig. 7g-j). The wild-type VPS13D-LTD (VPS13D-LTD-WT) protein significantly rescued the VPS13D suppression-induced FAs trafficking defects after 5 h or 24 h of EBSS starvation (Fig. 7g, k). In contrast, the VPS13D-LTD mutant (VPS13D-LTD-mut) did not significantly rescue the FA trafficking defects (Fig. 7g, l). This line of evidence supported the hypothesis that VPS13D-LTD plays an important role in FAs trafficking at mitochondria-LD junctions. In addition, we found that VPS13D-LTP was substantially less effective in rescuing the FA trafficking defects than full-length VPS13D, suggesting that both the lipid binding/transfer activities and the LD remodeling activities of VPS13D-VAB and TSG101 are required for efficient FAs transfer from LDs to mitochondria.

Discussion
Our results revealed a function of the disease-causing protein VPS13D at mitochondria-LD MCSs under starvation, and we proposed that VPS13D facilitated FA transfer at such MCSs through remodeling the morphology of the LD membranes via interaction with TSG101; and through binding FAs (Fig. 8).
Our findings suggest that VPS13D is a mitochondrial/LDs protein that is present at mitochondria−LD MCSs under starvation. However, it is possible that the localization of VPS13D may be subject to regulation by metabolic states or the cell cycle, as yeast VPS13 was found to localize to different organelle MCSs via different protein adaptors under different metabolic conditions 28,30 . In addition, adaptors that are responsible for recruiting VPS13D to the OMM are still unknown. This fundamental question awaits further investigation.
ESCRTs are considered to function in inverse budding reactions, in which ESCRT-III proteins form filaments that assemble on membranes with a negative curvature 33 . However, recent studies revealed that ESCRTs could also mediate classical-topology membrane remodeling, in which the ESCRTs form a double-stranded helical polymer that coats positively curved membranes 34 . For instance, IST1 promotes fission of endosomal tubules via recruitment of the microtubule-severing ATPase Spastin to mediate endocytic recycling 36 , and both IST1 and CHMP1B mediate the transfer of FAs from lipid droplets to peroxisomes, possibly by modifying the LD morphology 42 . Our results further supported a direct role of ESCRT complexes in FAs trafficking from LDs to mitochondria The p values were labeled in front of the corresponding data points for Eto-pretreated (red), VPS13D siRNA (green), and TSG101 siRNAs (magenta) treated groups, compared with the scrambled siRNA treated group (blue). Oligo: oligomycin; FCCP: Carbonyl cyanide-p-trifluoromethoxyphenylhydrazone; RTN: rotenone; AA: Antimycin. Four independent assays were performed with similar results in (i-l), and at least three biological replicates were performed with similar results in m. The statistic analysis was performed with twotailed unpaired student's t-test, and data are means ± SD in i-m. Scale bar, 10 μm in whole cell image and 2 μm in insets in a-h. through the remodeling LD morphology. Because of the unique phospholipid monolayer of the LD membrane and the spatial segregation between triglycerides (inside LDs) and lipase (associated with the cytoplasmic face of the LD membrane), it is plausible that such remodeling of the LD membrane by VPS13D and TSG101 or other ESCRT components promotes the access of triglycerides to the membrane-associated lipase, and eventually facilitates triglyceride mobilization and subsequent FA trafficking.
A previous study showed that yeast Vps13α binds to glycerophospholipids but not to other lipid species in Expi 293 F cells by liquid chromatography-tandem mass spectrometry (LC/MS/MS) 17 . We postulated that two reasons to explain why free FAs were not found to co-purify with VPS13α in this previous work. First, the ionization efficiency of free FAs under electrospray ionization MS conditions without derivatization is poor, rendering the construction of a robust platform for profiling of various FAs in biological samples using LC-MS-MS very challenging 54  identification of free FAs is also easily affected by matrix effects and the sensitivity of detection is low 55 . Second, the release of FAs from triglycerides by lipases exclusively occurrs upon stimulation by, for instance, starvation or exercise. Consistent with this, our results showed that barely any Red C12 was mobilized and transferred from LDs to mitochondria in complete medium (Fig. 7h, t = 0). Therefore, we postulated that the low abundance of FAs in cells cultured in complete medium might be another reason.
In addition, it should be noted that the results from the seahorse assays only indicated that VPS13D depletion modestly reduced the oxidation of FAs (Fig. 6m), but did not indicate a direct role of VPS13D in the transfer of FAs. Although our results demonstrated that the LTD of VPS13D bound FAs in vitro and was required for the efficient transfer of Red-C12 to mitochondria in vivo, more direct evidence is required to confirm that VPS13D directly transports endogenous FAs at MCSs.  Table 1. DNA primers used for generating constructs and for quantitative PCR assays reported in this work are listed in Supplementary Table 2. Oligonucleotides for siRNA reported in this work are listed in Supplementary  Table 3.
GFP-trap assay. GFP-Trap (GFP-trap agorose beads; GTA-100; ChromoTek) was used for detection of protein-protein interactions and the GFP-Trap assays were performed according to the manufacturer's protocol. 5% input was used in GFP traps unless otherwise indicated.
In vitro lipid-binding assay. 19 μl purified VPS13D LTD (1.5 mg/ml) was mixed with 1 μl of either NBD-labeled PA, PC, PE, PS, ceramide, C12 (1 mg/ml in methanol) or BODIPY-C16 (1 mg/ml in DMSO) in 20-μl total reaction volumes and incubated at 4°C for 3 h. Samples were loaded onto 12% Precast Native gels with Hepes-Tris buffer and run for 4 h at 90 V on ice. NBD/BODIPY fluorescence was visualized using a Bio-rad ChemiDoc XRS + (170-8265). Then gels were stained with Coomassie blue G250 (20279; ThermoFisher Scientific) to visualize total proteins. Image analysis. All image analysis and processing was performed using imageJ automatically identified by colocalization plugin with white pixels representing potential MCSs. For the measurement of mitochondrial Red C12 fluorescence intensity, an area encompassing the whole mitochondrial network was manually selected, and was measured via imageJ. The Red C12 fluorescence intensity of the whole cell areas was also calculated via imageJ.
Statistics and reproducibility. All statistical analyses and p-value determinations were performed with GraphPad Prism6 and Microsoft Excel (2011, 2013). All the error bars represent mean ± SD. To determine p-values, ordinary one-way ANOVA with Tukey's multiple comparisons test were performed among multiple groups; and a two-tailed unpaired student t-test was performed between two groups. A p-value of <0.05 was considered statistically significant. All western blots, immunoprecipitation, seahorse assays and in vitro protein-lipid binding assays were from at least three biological replicates and representative results are shown. Images from indirect IF of endogenous VPS13D using anti-VPS13D antibodies in HEK293 cells are representative from four independent experiments. Images from indirect IF of endogenous TSG101 using anti-TSG101 antibodies in HEK293 cells are representative from three independent experiments. Live imaging of transiently transfected cells expressing full length VPS13D^sfGFP was performed in HEK293 cells as VPS13D^sfGFP was too large to efficiently transfect other cell lines. Live imaging of transiently transfected cells expressing VPS13D mutants was performed in two cell lines: COS7 and HEK293. Cells expressing VPS13D^sfGFP or VPS13D mutants were imaged between 5 and 40 independent live-cell imaging sessions; cells expressing Halo-TSG101 and TSG101 mutants were imaged between 3 and 25 different sessions. Cells expressing VAPB-GFP and Halo-TSG101 were imaged at least 30 independent live-cell imaging sessions and repeated in three biological replicates in TEM with similar results. The splitGFP-based mito-LD MCSs reporter cells were imaged in at least four independent live-cell imaging sessions, and were applied to flow cytometry three times under two different conditions of Dox inductions, respectively, with similar results. The ddGFP-based mito-LD MCSs reporter cells were imaged in three independent live-cell imaging sessions, and were applied to flow cytometry three times with similar results. HEK293 cells treated either with scrambled or VPS13D siRNAs were imaged between 3 and 10 independent live-cell imaging sessions for the quantifications of mitochondria-LD interactions. HEK293 cells treated either with scrambled, VPS13D, TSG101, CHMP4B, CHMP1B, CHMP6, IST1 or ALIX siRNAs were imaged in at least 4 independent live-cell imaging sessions. VPS13D-suppressed HEK293 cells rescued either with siRNA-resistant full length VPS13D^sfGFP, siRNA-resistant GFP-LTD or siRNA-resistant GFP-LTD Mutant were imaged in three independent live-cell imaging sessions.