AP-4 vesicles unmasked by organellar proteomics to reveal their cargo and machinery

Adaptor protein 4 (AP-4) is an ancient membrane trafficking complex, whose function has largely remained elusive. In humans, AP-4 deficiency causes a severe neurological disorder of unknown aetiology. We apply multiple unbiased proteomic methods, including ‘Dynamic Organellar Maps’, to find proteins whose subcellular localisation depends on AP-4. We identify three highly conserved transmembrane cargo proteins, ATG9A, SERINC1 and SERINC3, and two AP-4 accessory proteins, RUSC1 and RUSC2. We demonstrate that AP-4 deficiency causes missorting of ATG9A in diverse cell types, including neuroblastoma and AP-4 patient-derived cells, as well as dysregulation of autophagy. Furthermore, we show that RUSC2 facilitates the microtubule plus-end-directed transport of AP-4-derived, ATG9A-positive vesicles from the TGN to the cell periphery. Since ATG9A has essential functions in neuronal homeostasis, our data not only uncover the ubiquitous function of the AP-4 pathway, but also begin to explain the molecular pathomechanism of AP-4 deficiency.


Introduction
Eukaryotic cells use a highly regulated system of vesicular and tubular transport intermediates to exchange molecules between organelles. Adaptor protein complex 4 (AP-4) is one of five related heterotetrameric AP complexes, which selectively incorporate transmembrane cargo proteins into nascent vesicles, and recruit machinery for vesicle budding and transport 1 . AP-4 consists of four subunits (β4, ε, μ4, and σ4) forming an obligate complex [2][3][4] (Fig. 1a). Loss-of-function mutations in any of the genes (AP4B1, AP4E1, AP4M1, and AP4S1) cause a severe recessive neurological disorder with early-onset progressive spastic paraplegia and intellectual disability [5][6][7] . AP-4-deficient patients have brain abnormalities including thinning of the corpus callosum 6,[8][9][10] , indicating an important role for AP-4 in neuronal development and homeostasis. Axons of Purkinje and hippocampal neurons from Ap4b1 knockout mice contain aberrantly accumulating autophagosomes that are immuno-positive for AMPA receptors 11 . However, the link between AP-4 deficiency and dysregulation of autophagy remains unclear.
While the clathrin adaptors AP-1 and AP-2 are well characterised, the function of AP-4, which does not associate with clathrin, has remained elusive. At steady state AP-4 localises to the trans-Golgi network (TGN) 2,3 and so is presumed to mediate cargo sorting at the TGN. The destination of the AP-4 trafficking pathway remains controversial, with conflicting reports suggesting transport to early endosomes 12,13 , to late endosomes/lysosomes 14 and, in polarised epithelial cells, to the basolateral membrane 15 .
Likewise, AP-4 has been suggested to influence the sorting of various cargoes, including amyloid precursor protein 12,13,16 , low-density lipoprotein receptor 11,15 , AMPA receptors 11 , and δ2 glutamate receptors 11,17 . However, several of these studies relied on exogenously 4 Davies et al., 2017 expressed proteins, while the potential endogenous cargoes (e.g. δ2 glutamate receptor) have cell-type-limited expression, unlike AP-4, which is ubiquitously expressed 2,3 . There is currently no consensus which proteins are genuine cargoes of the AP-4 pathway, and hence no consensus as to its function. Similarly, AP-4 vesicle machinery is largely uncharacterised.
The only identified AP-4 accessory protein is a cytosolic protein of unknown function called TEPSIN 18 .
Due to the very low abundance of AP-4 (ca. 40-fold lower than AP-1 or AP-2 in HeLa cells 19 ), its functional characterisation has proved challenging. Nonetheless, given its ubiquitous expression in human tissues, AP-4 is likely to play a ubiquitous and important role in protein sorting, whose identification will be paramount to understanding AP-4 deficiency. Here, we combine orthogonal, unbiased and sensitive proteomic approaches to define the composition of AP-4 vesicles. From the intersection of our analyses we identify physiological cargo proteins of the AP-4 pathway, and novel AP-4 accessory proteins. We demonstrate that AP-4 is required for the correct sorting of three transmembrane cargo proteins, including ATG9A, which has essential functions in neuronal homeostasis. Thus, our data suggest a potential mechanistic explanation for the pathology caused by AP-4 deficiency.

Dynamic Organellar Maps identify ATG9A, SERINC1 and SERINC3 as AP-4 cargo proteins
In AP-4-deficient cells, transmembrane cargo proteins that are usually transported in AP-4 vesicles, and cytosolic accessory proteins that are normally recruited by AP-4, are likely to be mislocalised. As an unbiased screen for such proteins, we used the 'Dynamic Organellar Maps' approach recently developed in our laboratory 19,20 . This mass spectrometry (MS)based method provides protein subcellular localisation information at the proteome level ( Fig. 1b). A comparison of maps made from cells with genetic differences allows the detection of proteins with altered subcellular localisation.
We prepared maps from wild-type, AP4B1 knockout and AP4E1 knockout HeLa cells ( Fig. 1c and Supplementary Fig. 1a, b), in biological duplicate ( Fig. 1d and Supplementary Table 1). For every protein, we calculated the magnitude of localisation shifts between the wild-type and each knockout, and the reproducibility of the shift direction (Fig. 1e). Three proteins underwent significant and reproducible shifts in both knockout cell lines: SERINC1 and SERINC3 (Serine incorporator 1 and 3), multi-pass membrane proteins of unknown function, and ATG9A (Autophagy-related protein 9A; Fig. 1f). ATG9A is the only transmembrane core autophagy protein and is thought to play a key (though poorly defined) role in the initiation of autophagosome formation 21 . The altered subcellular distribution of these proteins in AP-4-deficient cells identified them as candidate cargo proteins for the AP-4 pathway.
To begin to interpret the nature of the detected shifts, we used subcellular localisation information inferred from the maps. In both wild-type and AP-4 knockout cells,  ATG9A and SERINCs mapped to the endosomal cluster ( Fig. 1g-i). However, this cluster comprises different types of endosomes, as well as the TGN 19 . Scrutiny of the map visualisations ( Fig. 1g-i) and marker protein neighbourhood analysis (Supplementary Table   2) suggested that in the knockouts both SERINCs shifted intra-endosomally, while ATG9A localisation shifted from endosomes towards the TGN.

Comparative vesicle proteomics identifies RUSC1 and RUSC2 as AP-4 accessory proteins
Cytosolic proteins that only transiently associate with membrane may be missed by the Dynamic Organellar Maps approach, especially if they have low expression levels. We hence applied another proteomic approach developed in our lab, comparative vesicle profiling 18 , to identify proteins lost from a vesicle-enriched fraction in the absence of AP-4 ( Fig. 2a). This is particularly suited for identifying vesicle coat proteins. Cargo proteins are sometimes less strongly affected, as they may exist in several vesicle populations 18 .
We used four different methods to ablate AP-4 function: (i) knockdown; (ii) knocksideways (a protein is rerouted to mitochondria, acutely depleting its cytosolic pool 22 Table 3) and principal component analysis was used to combine the information from all datasets (Fig. 2b). Proteins that were reproducibly lost from vesicle fractions from AP-4depleted cells included the AP-4 subunits and TEPSIN, while SERINC1 and SERINC3 were the most depleted membrane proteins. ATG9A was affected to a lesser extent. Two related cytosolic proteins, RUSC1 and RUSC2 (RUN and SH3 domain-containing protein 1 and 2; Fig.   2c), were also reproducibly lost (Fig. 2b). They were not included in the maps since they are  barely detectable in subcellular membrane fractions. However, both proteins are highly enriched in the vesicle fraction, suggesting they are vesicle-associated proteins.
To investigate the RUSC-AP-4 relationship, we analysed total membrane fractions and whole cell lysates from control and AP-4 knockout cells by deep sequencing MS. RUSC2 (which has a lower whole cell copy number than RUSC1 19 ) was not consistently detected, but RUSC1 was dramatically lost from the total membrane fraction (>4-fold; Fig. 2d) and the whole cell lysate (>3-fold; Supplementary Fig. 1c) in knockout cells (Supplementary Table 4).
This suggests that without AP-4, RUSC proteins are no longer recruited to the membrane, leading to their destabilisation. Mutations in RUSC2 have been reported in patients with phenotypes reminiscent of AP-4 deficiency 23 . Collectively, these data suggest that RUSC1 and RUSC2 are novel AP-4 accessory proteins.

Proximity labelling and sensitive co-immunoprecipitation support that AP-4 interacts with ATG9A, SERINCs, and RUSC2
AP complex-cargo interactions are transient and so have largely proved refractory to standard co-immunoprecipitation approaches. As an alternative we applied BioID, which uses a promiscuous biotin ligase, BirA*, to biotinylate proteins proximal to a protein of interest 24 . We stably expressed AP4E1-BirA* in HeLa cells and quantified biotinylated proteins by MS relative to controls ( Fig. 2e and Supplementary  Fig. 2b-d).
We next performed co-immunoprecipitation of the AP-4 complex via overexpressed TEPSIN-GFP, under sensitive low-detergent conditions ( Fig. 2f and Supplementary Table 4).
ATG9A, SERINC1 and SERINC3 were co-precipitated, confirming that they are cargo proteins of the AP-4 pathway. As expected, these interactions were not observable with a conventional immunoprecipitation protocol ( Supplementary Fig. 2e).
Most AP complex accessory proteins interact with one or both C-terminal "ear" domains of the large subunits, including TEPSIN, which binds to the AP-4 β and ε appendage domains 26,27 . As RUSC2 was identified by BioID, we tested whether GFP-RUSC2 would interact with either of the appendage domains in a GST pull-down experiment. GFP-RUSC2, but not GFP alone, was pulled down with both appendage domains, notably at higher levels with the ε ear (Fig. 2g), confirming that RUSC2 is a bona fide AP-4 ear interaction partner.
In sum, using orthogonal proteomic approaches, we have identified three novel AP-4 cargo proteins, ATG9A, SERINC1 and SERINC3, and two novel AP-4 accessory proteins, RUSC1 and RUSC2. These are all low abundance proteins, expressed at comparable levels to AP-4 in HeLa cells 19 (Fig. 2h) and primary mouse neurons 20 .

AP-4 is required for the correct subcellular localisation of ATG9A and SERINCs
To further characterise the ATG9A and SERINC missorting phenotype we used immunofluorescence microscopy. In wild-type cells, ATG9A was detected as fine puncta with increased density in the juxtanuclear region (Fig. 3a), consistent with previous data 28 .
In contrast, there was a striking accumulation of ATG9A in the TGN region in both AP-4 knockout lines ( Fig. 3a and Supplementary Fig. 3). Importantly, the mislocalisation of ATG9A 9 Davies et al., 2017 in the AP4B1 knockout was completely rescued by stable expression of the AP4B1 subunit ( Fig. 3a). This was confirmed by quantitative automated imaging (Fig. 3b). To determine whether ATG9A is similarly affected by loss of AP-4 in a cell line more relevant to the neuronal phenotypes of AP-4 deficiency, we knocked out AP4B1 or AP4E1 in SH-SY5Y neuroblastoma cells (Fig. 3c). As before, loss of AP-4 caused a striking accumulation of ATG9A at the TGN (Fig. 3d).
Since there are no commercial antibodies that allow detection of endogenous SERINC1 or SERINC3, we used CRISPR to knock in fluorescent Clover tags at the endogenous loci ( Supplementary Fig. 4a). Confocal microscopy with Airyscan enhanced resolution revealed the tagged SERINCs to localise to the perinuclear region and fine puncta throughout the cell (Fig. 4a, c). Strikingly, the peripheral SERINC-positive puncta showed considerable overlap with ATG9A, and AP-4 knockdown resulted in loss of these puncta, suggesting they were AP-4-derived vesicles (Fig. 4a-d and Supplementary Fig. 4b). In the AP-4-depleted cells, the SERINCs accumulated in the perinuclear area, although unlike ATG9A, they displayed little colocalisation with TGN46 ( Supplementary Fig. 4c, d).

AP-4-derived vesicles accumulate at the cell periphery in RUSC2-overexpressing cells
We created HeLa cell lines that stably overexpress GFP-tagged RUSC2 and found it to localise to fine puncta throughout the cell, with a concentration in clusters at the periphery ( Fig. 5a and Supplementary Fig. 5a). Overexpression of RUSC2 resulted in a dramatic relocation of ATG9A and SERINCs to the cell periphery, where they colocalised with the GFPtagged RUSC2 (Fig. 5a, b and Supplementary Fig. 5a, b). This effect was specific for AP-4 Davies et al., 2017 cargo proteins; other membrane proteins and organellar markers did not respond to overexpression of RUSC2 ( Supplementary Fig. 5c).
The localisation of AP-4 itself was not affected by RUSC2 overexpression ( Supplementary Fig. 6a), but we hypothesised that the peripheral structures in RUSC2overexpressing cells might be AP-4-derived vesicles. Consistent with this, GFP-RUSC2 expressed in AP-4 knockout and knockdown cells neither accumulated at the cell periphery, nor colocalised with ATG9A ( Fig. 6a and Supplementary Fig. 6b, c). Importantly, both the peripheral distribution of GFP-RUSC2 and its colocalisation with ATG9A were restored in the AP4B1 knockout by transient expression of the missing AP-4 subunit (Fig. 6a). This demonstrates that the peripheral RUSC2/ATG9A/SERINC-positive structures are AP-4dependent compartments. The absence of AP-4 from the accumulating structures suggests that it is released soon after budding, as is typical of most known vesicle coats 1 .
Given the colocalisation between RUSC2 and ATG9A, we tested whether ATG9A and GFP-RUSC2 interact physically. ATG9A co-precipitated with GFP-RUSC2 from wild-type but not AP4B1 knockout cells (Fig. 6b). This demonstrates that the interaction between RUSC2 and ATG9A (which could be indirect) requires AP-4, suggesting that the three proteins come together transiently during the formation of AP-4 vesicles.
Ultrastructural analysis by correlative light and electron microscopy (CLEM) revealed that the peripheral clusters of GFP-RUSC2-positive puncta corresponded to an accumulation of uncoated vesicular and tubular structures (Fig. 6c). Their peripheral localisation suggested targeting to the plus-ends of microtubules, and RUSCs have been implicated in microtubulebased transport 29 . Treatment with nocodazole prevented the peripheral localisation of GFP-RUSC2 but, unlike in AP-4-deficient cells, GFP-RUSC2 still co-localised with ATG9A (Fig. 6d).  These data suggest that the distribution of AP-4 vesicles requires microtubule-based transport, whereas their formation does not.

AP-4 deficiency causes dysregulation of autophagy in HeLa cells
Ap4b1 knockout mice show aberrant accumulation of autophagosomes in neuronal axons, and increased levels of the autophagic marker protein LC3B 11 . We investigated if there were similar effects on autophagy in our AP-4 knockout HeLa cells. Wild-type and AP4E1 knockout cells were grown in complete or starvation medium (to induce autophagy), with or without bafilomycin A1 (which blocks autophagosome degradation). The level of LC3B was then assessed by Western blotting (Fig. 7a). In untreated cells, there were increased levels of unlipidated LC3B-I and lipidated LC3B-II in the AP4E1 knockout. Under starvation conditions (when LC3B-I is mostly converted into LC3B-II), there was also increased LC3B-II in the AP4E1 knockout, which increased further in the presence of bafilomycin A1. This suggests that the elevated level of LC3B-II was not due to a block in degradation. Comparison of the relative amounts of LC3B-II with and without bafilomycin A1 in the wild-type versus the AP-4 knockout cells also suggested comparable levels of autophagic flux. The same trends were seen when comparing AP4B1 knockout cells with the AP4B1 rescued cell line (Fig. 7b). Quantitative MS confirmed the increased level of total  Table 4).
Elevated LC3B-II may reflect an increase in autophagosome size 30 . To test this, we visualised LC3B-positive structures by immunofluorescence microscopy in wild-type, AP4B1 knockout, and AP4B1 rescued cells, following two hours starvation (Fig. 7c). Quantification through automated imaging showed an increase in the size of LC3B puncta in the knockout Davies et al., 2017 cells (Fig. 7d), but no significant change in the number of puncta ( Supplementary Fig. 7a).
Collectively, these data suggest that lack of AP-4 in HeLa cells causes dysregulation of autophagy.

ATG9A mislocalisation is a ubiquitous phenotype in cells from AP-4 deficient patients
We analysed fibroblasts from patients with homozygous mutations in one of the AP-4 genes 5,6,31,32 by immunofluorescence microscopy (Fig. 8a). Mutations in any of the four subunits caused a striking accumulation of ATG9A at the TGN. In addition, fibroblasts from an individual with a heterozygous loss-of-function mutation in AP4E1 (the phenotypically normal mother of the homozygous AP4E1 patient) displayed normal ATG9A localisation, so mislocalisation of ATG9A is a cellular phenotype that correlates with disease in AP-4 deficiency.
The microscopy indicated that there was also an increase in the overall ATG9A signal in the patient cells. Western blotting confirmed a large increase in the amount of ATG9A in whole cell lysates from all four patient cell lines and an intermediate level in the unaffected AP4E1 heterozygote (Fig. 8b). These data suggest that AP-4-deficient cells compensate for the missorting of ATG9A by increasing its expression, further supporting the importance of AP-4 in the regulation of ATG9A trafficking. 13 Davies et al., 2017

Discussion
The nature of AP-4 vesicles and their role in membrane trafficking has remained elusive for the two decades since their discovery. Here we have used orthogonal global proteomic tools to delineate the function of the AP-4 pathway. We identified three transmembrane cargo proteins, ATG9A, SERINC1 and SERINC3, and two novel AP-4 accessory proteins, RUSC1 and RUSC2. Our approach was hypothesis-free and analysed the subcellular distribution of endogenous proteins. The latter is critical for assessing the role of trafficking pathways, especially for AP-4, which is of comparatively low abundance. When we overexpressed our AP-4 cargo proteins, they were no longer trafficked in an AP-4dependent manner (data not shown), suggesting that investigations based on overexpressed candidate cargo proteins are likely to lead to spurious results. The AP-4associated proteins identified in this study all have low expression levels similar to those of AP-4 itself 19 , highlighting the sensitivity of our approach. They are also, like AP-4, expressed ubiquitously. ATG9A localisation depends on AP-4 not only in HeLa cells, but also in neuroblastoma-derived SH-SY5Y cells and in fibroblasts from AP-4-deficient patients ( Fig. 3 and Fig. 8a), suggesting that trafficking of ATG9A from the TGN is a ubiquitous function of AP-4.
Our data strongly support a model whereby AP-4 packages ATG9A, SERINC1 and SERINC3 into vesicles at the TGN, which associate via the RUSCs with machinery for microtubule plus-end-directed transport to the cell periphery (Fig. 8c). Neuronal deficiency of Atg9a in mice leads to progressive axonal degeneration, ataxia and convulsions 33 . Thus, provided that AP-4 has equivalent functions in neurons and HeLa cells, which our neuroblastoma data support (Fig. 3d), a hypothesis for neuronal AP-4 pathology emerges. 14

Davies et al., 2017
Neurons require efficient long-range transport, especially towards the distal axon, rendering them susceptible to disturbances in membrane trafficking 34 . Furthermore, microtubules are unipolar in axons, with distally localised plus-ends 35

, and it has been shown in
Caenorhabditis elegans that Atg9-containing vesicles are transported towards the distal axonal microtubule plus-ends 36 . The distal axon is also an important site of autophagosome biogenesis 37,38 . Our model suggests that in neurons lacking AP-4, ATG9A will not be packaged correctly at the TGN, and will not efficiently reach the distal axon. This may interfere with the spatial regulation of autophagy and/or with other functions of ATG9A, disrupting neuronal homeostasis.
The identification of ATG9A as an AP-4 cargo may also explain the aberrant accumulation of autophagosomes in neuronal axons of Ap4b1 knockout mice 11 . There is a similar dysregulation of autophagy in AP-4 knockout HeLa cells, which contain enlarged autophagosomes and increased levels of LC3B (Fig. 7). Basal elevation of LC3B-I and LC3B-II has been observed in ATG9A knockout HeLa cell lines [39][40][41] , so our data are consistent with ATG9A mistrafficking leading to impaired ATG9A function. The role of ATG9A in autophagy is poorly defined, but it is thought to act early in autophagosome biogenesis without becoming incorporated into autophagosome membrane itself 21 . In yeast the phagophore assembly site originates from Atg9-positive clusters of vesicles and tubules (the "Atg9 reservoir") 42 . Likewise, in mammalian cells Atg9 localises to a tubulovesicular compartment, distinct from other organellar markers 21 . The AP-4-derived ATG9A-positive vesicles and tubules we have observed at the periphery of RUSC2-overexpressing cells fit the description of this compartment (Fig. 6c). In AP-4-deficient cells, ATG9A trafficking may be stalled at the TGN and the peripheral "ATG9A reservoir" depleted. Further investigation is necessary to understand how this may lead to the observed effects on autophagosomes in AP-4-deficient cells. ATG9A mistrafficking has been linked previously to an accumulation of enlarged immature autophagosomes in Niemann-Pick type A patient fibroblasts 43 . However, ATG9A has functions independent of autophagy 33,44,45 , which may also be relevant to potential pathogenic effects caused by its missorting.
Loss-of-function mutations in RUSC2 cause a neurological disorder with considerable overlap with the AP-4 deficiency phenotype 23 . This is in keeping with our identification of RUSCs as AP-4 accessory proteins. It will be interesting to look for similar ATG9A and SERINC sorting defects in cells from RUSC2-deficient patients. The RUSCs are poorly characterised but implicated in vesicular transport 29,46 . RUSC1 has been proposed to act as a vesicletransport adaptor by linking syntaxin-1 to kinesin-1 motors 29 . Our data suggest that RUSC2 may similarly link AP-4 vesicles to microtubule transport machinery (Fig. 6).
The presence of SERINC1 and SERINC3 in the ATG9A tubulovesicular compartment warrants further investigation. Little is known about the function of SERINCs, despite their recent identification as HIV restriction factors [47][48][49] and their high degree of conservation in all branches of eukaryotes. They were originally proposed to mediate the incorporation of serine into membrane lipids 50 , but recent functional studies have shown no effect on membrane composition 51,52 . Our discovery that ATG9A and SERINCs are trafficked together suggests a functional relationship.
In conclusion, this study has greatly expanded our understanding of the AP-4 pathway. The identification of novel AP-4 cargo and accessory proteins provides tools for further investigation of AP-4 function, generates hypotheses about the pathomechanism of AP-4 deficiency, and marks a significant step towards the development of possible treatments. 16 Davies et al., 2017

Constructs
A modified retroviral pLXIN vector (pLXINmod) was a gift from Andrew Peden (University of Sheffield). Myc-BirA* cDNA was amplified from pcDNA3.1_mycBioID (a gift from Kyle Roux; Addgene plasmid #35700 24 ) by PCR. AP4B1 cDNA was amplified from a full-length IMAGE clone (2906087), AP4E1 cDNA from a full-length IMAGE clone (40146497), and AP4M1 and AP4S1 cDNAs from sequence verified EST clones. The control BioID construct, pEGFP-myc-BirA*, was made by cloning myc-BirA* cDNA into a pEGFP-N2 vector (Clontech) using BsrGI and XbaI restriction sites. The AP4E1 BioID construct was made using Gibson Assembly Master Mix (E2611, New England BioLabs) to introduce myc-BirA* into the flexible hinge region of AP4E1, between residues 730 and 731, inserted into the pLXINmod vector linearised with an HpaI restriction site. For AP4B1, AP4M1 and AP4S1 BioID constructs, a Cterminal myc-BirA* tagging construct was used. This was generated using Gibson Assembly to introduce myc-BirA*, preceded by a glycine-serine linker (10 amino acids) and an upstream BglII site, into the HpaI site of pLXINmod. The BglII site was used to linearise the myc-BirA* tagging construct and cDNAs for AP4B1/M1/S1 were added by Gibson Assembly.
RUSC2 cDNA was amplified from pCMV-SPORT6_RUSC2 (MHS6278-202800194, Thermo Fisher Scientific) and GFP cDNA was amplified from pEGFP-N2. Constructs for stable overexpression of GFP-tagged RUSC2, pQCXIH_GFP-RUSC2 and pQCXIH_RUSC2-GFP, were generated using Gibson Assembly to introduce RUSC2 and GFP cDNAs into the retroviral vector pQCXIH (Clontech). The AgeI site of pQCXIH was used for GFP-RUSC2 and the NotI site for RUSC2-GFP. pQCXIH_HA-RUSC2 was made by cutting out GFP from pQCXIH_GFP-RUSC2 with NotI and AgeI restriction sites and replacing it with a triple HA tag.  32 (the second mutation, which is at a splice donor site, was originally inaccurately reported as c.138_140delAATG). The control fibroblasts (from a healthy control donor) were a gift from Craig Blackstone (NIH) and the heterozygous AP4E1 WT /AP4E1* fibroblasts were from the unaffected mother of the homozygous AP4E1* patient.
Transient DNA transfections were carried out using a TransIT-HeLaMONSTER® kit
For the AP4E1 knockout HeLa cell line we inactivated all copies of the AP4E1 gene using the 'double nickase' CRISPR/Cas9 system 57,58 . Paired gRNAs targeting exon 6 of AP4E1 Clone x6C3 was negative for AP4E1 expression in both assays and was further validated by sequencing. Genomic DNA was harvested using a High Pure PCR Template Purification Kit (Roche) and PCR was used to amplify the region around the target sites. The PCR products were then blunt-end cloned (Zero Blunt PCR Cloning Kit; Invitrogen) and 17 clones were sent for Sanger sequencing with the M13_F primer (Beckman Coulter Genomics).
For knockout of AP4B1 and AP4E1 in SH-SY5Y cells, a lentiviral CRISPR/Cas9 system was used, as described 59  For CRISPR/Cas9-mediated endogenous tagging of SERINC1 and SERINC3, we used homology-directed repair to introduce a myc-Clover tag at the C-terminus of each protein.
Suitable sgRNA targets were selected to enable Cas9 to cut downstream and proximal to the Single cell clones were isolated and tested for knock-in of Clover by western blotting and immunofluorescence with an anti-GFP antibody. Clones SERINC1-Clover A3 and SERINC3-Clover B6 were positive for Clover expression in both assays and correct integration of the tag was confirmed by Sanger DNA sequencing.

Fluorescence microscopy
Cells were grown onto 13 mm glass coverslips and fixed in 3% formaldehyde in PBS membrane by wet transfer and membranes were blocked in 5% w/v milk in PBS with 0.1% v/v Tween-20 (PBS-T). Primary antibodies (diluted in 5% milk) were added for at least 1 hour at room temperature, followed by washing in PBS-T, incubation in secondary antibody (also in 5% milk) for 30 minutes at room temperature, washing in PBS-T and finally PBS.
Chemiluminescence detection of HRP-conjugated secondary antibody was carried out using Amersham ECL Prime Western Blotting Detection Reagent (GE Healthcare) and X-ray film.

GST pulldowns from cytosol
All steps were performed on ice with pre-chilled ice-cold buffers, unless otherwise noted.
To prepare cytosol for pulldowns, HeLa cells stably expressing EGFP or GFP-RUSC2

Immunoprecipitations
All steps were performed on ice with pre-chilled ice-cold buffers, unless otherwise noted.
For sensitive immunoprecipitation of TEPSIN-GFP, wild-type HeLa cells (control) and HeLa cells stably expressing TEPSIN-GFP 26 were grown in SILAC media. For two replicates (data shown in Figure 2E Vesicle-enriched fractions were prepared essentially as previously described 18

Generation of Dynamic Organellar Maps
Organellar maps were prepared as previously described in detail 19 , from wild-type (control), AP4B1 knockout and AP4E1 knockout HeLa cells, in duplicate (six maps in total).
Maps were prepared on two separate days, with a complete set of three on each occasion (one control, one AP4B1 knockout, and one AP4E1 knockout). In brief, HeLa cells were lysed mechanically, and post-nuclear supernatants were subfractionated into five fractions by a series of differential centrifugation steps. In parallel, a single membrane fraction was obtained from metabolically 'heavy' labelled cells (SILAC method). This fraction served as an internal reference, by spiking it into each of the "light" subfractions. Analysis by mass spectrometry provided a ratio of enrichment/depletion for each protein in each subfraction, relative to the standard. All five ratios combined yielded an abundance distribution profile for each protein across the subfractions. Principal component analysis revealed which proteins had similar fractionation profiles (apparent as organellar clusters in Fig. 1g-i)

In-solution digestion of proteins
Protein was precipitated by the addition of 5 volumes of ice-cold acetone, incubated at -20°C for 30 minutes and pelleted by centrifugation at 4°C for 5 minutes at 10,000 g.
Precipitated protein was rinsed in ice-cold 80% acetone and re-pelleted as above. All subsequent steps were performed at room temperature. Precipitated protein pellets were air-dried for 5 minutes, resuspended in digestion buffer (50 mM Tris pH 8.

Peptide purification and fractionation
Several different peptide fractionation and clean-up strategies were used in this study. For most mass spectrometric experiments, peptides were purified and fractionated on SDB-RPS (#66886-U, Sigma) stage tips as previously described 63  previously described 63 . Alternatively, protein samples were separated by SDS-PAGE, gels were cut into 5-10 slices, and proteins were digested with trypsin in-gel as described 64 .
Cleaned peptides were dried almost to completion in a centrifugal vacuum concentrator, and then volumes were adjusted to 10 µL with Buffer A* (0.1% (v/v) TFA, 2% (v/v) acetonitrile) and either immediately analysed by mass spectrometry, or first stored at -20C.

MS Instrumentation and configuration
Two different mass spectrometers were used (Q-Exactive HF and Q-Exactive; Thermo Fisher Scientific), as indicated in the summary table above. Mass spectrometric analyses were performed as previously described in detail (for Q-Exactive HF 19,20 and for Q-Exactive 64 ).

Processing of mass spectrometry data
Mass spectrometry raw files were processed in MaxQuant 65 version 1.5, using the human SwissProt canonical and isoform protein database, retrieved from UniProt (www.uniprot.org). For SILAC experiments (vesicle fractions; Dynamic Organellar Map subfractions; whole cell lysate analysis; TEPSIN-GFP immunoprecipitations) multiplicity was set to 2, with Lys8 and Arg10 selected as heavy labels; Re-quantify was enabled; minimum number of quantification events was set to 1. For label-free experiments (membrane fractions; BioID) multiplicity was set to 1; LFQ was enabled, with LFQ minimum ratio count set to 1. Membrane fractions were SILAC heavy labelled (Arg10, Lys8). Matching between runs was enabled. Default parameters were used for all other settings.

Proteomic data analysis
All analyses were performed on the 'protein groups' file output from MaxQuant. first filtered by removing matches to the reverse database, matches based on modified peptides only, and common contaminants ('standard filtering'). Further experiment-specific filtering, data transformation and analyses were performed as described below.

Dynamic Organellar Maps statistical analysis
To identify proteins with shifted subcellular localisation in response to AP-4 knockout, we applied our previously described rigorous statistical analysis 19,20 . We adapted the procedure to the experimental design of the present study as follows. Organellar maps were made in duplicate, from control, AP4B1 knockout, and AP4E1 knockout cells.
Abundance distribution profiles across all six maps were determined for 3,926 proteins.
First, the profiles obtained in the AP4B1 knockout and AP4E1 knockout cells were subtracted from the profiles obtained in the cognate control map, protein by protein, to obtain 2 x 2 sets of delta profiles (Con_1-AP4B1_1, Con_2-AP4B1_2; Con_1-AP4E1_1, Con_2-AP4E1_2). For proteins that do not shift, the delta profile should be close to zero. All delta profile sets were subjected to a robust multivariate outlier test, implemented in Perseus software 66 , to identify proteins with delta profiles significantly different from experimental scatter. The profile distance corresponds to a p-value reflecting how likely it is to observe this deviation by chance, assuming no real change. For each protein, four such pvalues were hence obtained, two from AP4E1 knockout, and two from AP4B1 knockout. For maximum stringency, we selected the least significant of these p-values as representative of a protein's shift. A shift of equal or greater significance was thus observed in all four comparisons. We did not treat the four delta maps as completely independent though, since both knockouts were compared to the same cognate control. Hence, as a very conservative measure of movement, the selected p-value was only squared (instead of being raised to the power of four), and then corrected for multiple hypothesis testing using the Benjamini Davies et al., 2017 Hochberg Method. The negative log10 of the corrected p-value corresponded to a protein's movement (M) score.
Next, the reproducibility of observed delta profiles across replicates was determined as the Pearson correlation (map (Con_1-AP4B1_1) vs map (Con_2-AP4B1_2); and map (Con_1-AP4E1_1) vs map (Con_2-AP4E1_2)). For maximum stringency, we chose the lower one of the two obtained correlations as representative of the protein's shift reproducibility, corresponding to its R score.
To control the false discovery rate (FDR), we then applied the same analysis to our previously published wild-type HeLa maps 19 (six untreated maps with no genuine protein shifts expected). In this mock experiment, we designated two maps as controls, two as 'mock knockout 1', and two as 'mock knockout 2'. As above, we calculated M and R scores from the lowest correlations and p values of movement. The estimated FDR at a given set of M and R score cut offs was then calculated as the number of hits obtained with the mock experiment data, divided by the number of hits obtained with the AP-4 maps data, scaled by the relative sizes of the datasets (which were almost identical). At the chosen high stringency cut-offs (M score >4, R score >0.81), not a single hit was obtained from the mock data. Hence, we estimate the FDR for the three hits obtained from the AP-4 maps at <1%.
Finally, as an additional criterion, we also evaluated the similarity of identified shifts across the two different knockouts (i.e. the correlation of map (Con_1-AP4B1_1) vs map (Con_1-AP4E1_1); and map (Con_2-AP4B1_2) vs map (Con_2-AP4E1_2)). All three hits showed a very high degree of shift correlation (>0.9) across the two AP-4 knockout lines, thus also passing the additional stringency filter. Davies et al., 2017

Membrane fraction analysis
Relative protein levels in membrane fractions from AP4B1 knockout and AP4E1 knockout HeLa cells (each in triplicate) were compared to those in membrane fractions from wild-type HeLa cells (in triplicate) using LFQ intensity data. The primary output was a list of identified proteins, and for each protein up to nine LFQ intensities across the wild-type and AP-4 knockout samples. Following standard data filtering, proteins were filtered to only leave those with nine LFQ intensities (no missing values allowed), leaving 6653 proteins. LFQ intensities were then log-transformed and comparison of knockout and wild-type membrane fractions performed with a two-tailed t-test. A permutation-based (1000 permutations) estimated FDR of 0.05 and an S0 parameter of 0.5 were set to define significance cut-offs (Perseus software).

Whole cell lysate analysis
Whole cell lysates from light-labelled AP4B1 knockout and AP4E1 knockout HeLa cells were compared to lysates from heavy-labelled wild-type HeLa cells by SILAC quantification, each in triplicate. The primary output was a list of identified proteins, and for each protein up to six H/L (Heavy/Light) ratios of relative abundance (three comparing AP4B1 knockout to wild-type and three comparing AP4E1 knockout to wild-type). Following standard data filtering, proteins were further filtered to require at least two H/L ratios for each knockout, leaving 6841 proteins. H/L ratios from each replicate were then normalised to the median H/L ratio for that replicate, log-transformed, and inverted to L/H so that a protein depleted from the whole cell lysate in the absence of AP-4 had a negative ratio. A one-sample t-test (two-tailed) was applied to compare the L/H ratios for each protein to zero (null hypothesis of no change between wild-type and knockout). To control the false discovery rate (FDR), an identical analysis of a mock experiment comparing light-and heavy-Davies et al., 2017 labelled wild-type HeLa lysates was performed (in triplicate; no genuine changes were expected here). The FDR was given by the number of hits observed in the mock experiment divided by the number of hits in the knockout experiment. Using the cut-offs p ≤ 0.02 and a minimum absolute fold change (log2) of 0.45, the estimated FDR was 25%. The t scores from the mock experiment were calculated from only three datapoints per protein (whereas up to six were used for the knockout vs control data), and hence were adjusted to emulate a 6 datapoint experiment. To this end, we assumed that the observed standard deviations and means had been observed from 6 datapoints, yielding much lower p values from the same t scores for the mock data. This procedure thus likely overestimates the number of false positives at a given cut-off, resulting in highly stringent FDR control.

Vesicle fraction analysis
Paired AP-4-depleted and control vesicle fractions were compared by SILAC quantification. The primary output was a list of identified proteins, and for each protein up to nine H/L ratios of relative abundance, and the number of quantification events (H/L ratio count) used to calculate each ratio. Following standard data filtering, proteins were further filtered to require at least one H/L ratio count in all experiments. This excluded AP4S1 due to it having a ratio count of 0 in one experiment, so the data for AP4S1 was manually added

AP-4 BioID
Relative protein levels were compared across samples using LFQ intensity data. LFQ intensities from pulldowns from the control cell lines (HeLa, HeLa BirA* and HeLa GFP-BirA*) were compressed from nine values (three cell lines in triplicate) to three using control compression as previously described 67  parameter of 0.5 were set to define significance cut-offs (Perseus software).

Data availability
All data supporting this work are available on reasonable request to the corresponding author.