Structural Organization of the Retriever-CCC Endosomal Recycling Complex

The recycling of membrane proteins from endosomes to the cell surface is vital for cell signaling and survival. Retriever, a trimeric complex of VPS35L, VPS26C and VPS29, together with the CCC complex comprising CCDC22, CCDC93, and COMMD proteins, plays a crucial role in this process. The precise mechanisms underlying Retriever assembly and its interaction with CCC have remained elusive. Here, we present the first high-resolution structure of Retriever determined using cryogenic electron microscopy. The structure reveals a unique assembly mechanism, distinguishing it from its remotely related paralog, Retromer. By combining AlphaFold predictions and biochemical, cellular, and proteomic analyses, we further elucidate the structural organization of the entire Retriever-CCC complex and uncover how cancer-associated mutations disrupt complex formation and impair membrane protein homeostasis. These findings provide a fundamental framework for understanding the biological and pathological implications associated with Retriever-CCC-mediated endosomal recycling.


121
Cryo-EM structure of Retriever reveals an assembly mechanism distinct from Retromer

122
To determine the structure of human Retriever, we co-expressed its three subunits, 123 VPS29, VPS26C and VPS35L, in Sf9 cells using individual baculoviruses, in which only VPS29 124 contained a His6 tag at its C-terminus (CT) to facilitate purification. The purified Retriever eluted 125 as a single peak in size-exclusion chromatography (Extended Data Fig. 1A), which gave rise to 126 high-quality cryo-EM grids with homogeneous single particle distributions (Extended Data Fig.   127 1B). We then determined the cryo-EM structure of Retriever to a resolution of 2.9 Å using single 128 particle reconstruction (Table 1). We applied local refinement to further improve the map quality

131
The overall structure of Retriever assumes a semicircular configuration measuring ~55 x 132 90 x 160 Å, which is dominated by the extended conformation of VPS35L adopting an / 133 solenoid fold that comprises 32 −helices. Within the complex, the globular structure of VPS29 134 is partially embraced within an extensive pocket formed by the CT region of VPS35L, while the 135 VPS26C binds to the outer ridge of the solenoid at the opposite end of VPS35L (Fig. 1A, B).

136
Another key feature of the complex is the first 37 amino acids (a.a.) at the N-terminus (NT) of 137 VPS35L. This NT peptide, hereafter referred to as the "belt" sequence due to its resemblance to 138 a seatbelt, wraps around the CT region of VPS35L and the bound VPS29 (Fig. 1, dark green).

139
Following the NT "belt" is a long, unstructured peptide linker of ~72 residues, which extends 140 toward the opposite end of the complex (Fig. 1A, B, dashed green line).

141
In many aspects, Retriever exhibits similar, yet distinct structural features compared to 142 Retromer ( Fig. 1B; Extended Data Fig. 2). The overall conformation of Retriever is more 143 compact and twisted than Retromer, shorter by ~40 Å in its longest dimension (Fig. 1B).

144
Moreover, the overall molecular surface of Retriever is less negatively charged than Retromer 145 (Extended Data Fig. 2A). Although VPS35L and VPS26C in Retriever share only ~15% and 146 23/24% sequence identity with VPS35 and VPS26A/B in Retromer, respectively, their 147 secondary structures exhibit remarkable similarities. Both VPS35L and VPS35 adopt the / 148 solenoid fold with a similar number and organization of helices ( Fig. 1B; Extended Data Fig. 2B).

149
Both VPS26C and VPS26A consist of two domains formed by a similar number and 150 arrangement of -strands, which pack into a deeply curved -sandwich resembling an arrestin 151 fold ( Fig. 1B; Extended Data Fig. 2B). Despite these similarities, however, VPS35L exhibits a 152 more compact structure than VPS35 and contains several unique features absent from VPS35.

153
These include the NT "belt" peptide, which makes extensive contacts with VPS29 and the CT 154 region of VPS35L, and several additional short helices and a -hairpin, which are inserted 155 between the common solenoid helices (Fig. 1B). Similarly, VPS26C is also more compact than 156 VPS26A and contains several distinct short -strand insertions compared to VPS26A ( Fig. 1B; larger than the interface between VPS35 and VPS26A in Retromer (~670 Å 2 ). This further 170 underscores the more compact nature of the Retriever complex (Extended Data Fig. 2D).

172
The NT "belt" sequence of VPS35L plays a key role in stabilizing Retriever

173
Given the distinctive feature of the NT "belt" sequence in VPS35L and its extensive 174 contact with both the CT region of VPS35L and the bound VPS29, we asked whether the "belt" 175 sequence could play an important role in stabilizing Retriever assembly. Close inspection of the 176 structure reveals two major anchoring points in the "belt" sequence ( Fig. 2A). First, the NT 11 177 residues of the "belt" sequence winds through a deep trough on the CT region of VPS35L 178 formed by the ends of helix 29, 30, 31, and 32, largely through structural complementarity 179 ( Fig. 2A, C). Consequently, this interaction makes the "belt" sequence an integral part of the CT 180 region of VPS35L. The interaction is centered around W6, a highly conserved residue in 181 orthologs ranging from amoeba to humans (Fig. 2B). W6 inserts into a deep pocket formed by 182 L825, L828, S829, C864, M868, I898, G902, and L909 from VPS35L (Fig. 2C). This interface is 183 stabilized by extensive van der Waals interactions and a few hydrogen bonds. At the boundary 184 of this extensive interaction surface, the conserved residue R11 of the "belt" sequence is 185 supported by salt bridges with two conserved residues, E16 of the "belt" sequence itself and 186 D99 from the bound VPS29 (Fig. 2B, C).

187
The second conserved anchoring point of the "belt" sequence is located at its C-188 terminus, where it interacts with VPS29 largely through a conserved "HPL" motif in Retriever 189 ( Fig. 2C-D). This interaction is unique to Retriever and absent between VPS29 and VPS35 in 190 Retromer (Fig. 1B). It is remarkable that the HPL motif is virtually 100% conserved in all 191 examined organisms (Fig. 2B). The motif adopts a type-I -turn structure through a network of 192 intrapeptide hydrogen bonds (Fig. 2D). At the tip of the -turn, P34 and L35 of the "HPL" motif 193 insert into a conserved and largely hydrophobic pocket on VPS29 formed by 1, 9, 10 and 194 the linker connecting 1 and 2, consisting of L6, L29, L30, K34, I35, F154, L156, Y167, and 195 Y169 (Fig. 2D, 3A). This interaction is further stabilized by a hydrogen bond network involving 196 K34 and Y169 from VPS29 and H33, P34, and L35 from VPS35L (Fig. 2D).

197
Consistent with the observation that the belt is an integral component of the CT region of 198 VPS35L and essential for VPS29 binding, deleting the first 10 amino acids of VPS35L was 199 sufficient to abrogate VPS35L-VPS29 interaction in cells, as noted in co-immunoprecipitation 200 (co-IP) experiments (Fig. 2E). In contrast, complete deletion of the "belt" sequence or even the 201 first 100 amino acids, which include the unstructured linker sequence, had no effect on the 202 binding between VPS35L and VPS26C (Fig. 2F), in agreement with the presented Retriever 203 structure. Surprisingly, disrupting the interaction between VPS29 and VPS35L eliminated the 204 interaction between VPS35L and the CCC subunits CCDC22, CCDC93, and COMMD1, as well 205 as DENND10 (Fig. 2E). This suggests an interdependence between VPS29-VPS35L and 206 Retriever-CCC interactions, as will be examined in later parts of the paper.

208
Conserved surfaces in VPS35L that bind to VPS26C and VPS29 are mutated in cancer

209
In addition to the NT "belt" sequence contacting VPS29, the CT region of VPS35L 210 interacts with VPS29 using a slightly concave and extensive surface (Fig. 3A

216
B). Many residues in the CT region of VPS35L establish extensive polar and non-polar contacts 217 with VPS29 through this broad interaction surface (Fig. 3B).

218
The similarity between VPS29-VPS35L interface in Retriever and VPS29-VPS35 219 interface in Retromer poses a challenge in the design of a mutation that can specifically disrupt 220 one interaction without affecting the other. To specifically disrupt VPS29-VPS35L interaction in 221 Retriever, instead of mutating this extensive surface, we introduced Y169A to VPS29 to disrupt 222 the interaction between VPS29 and the "HPL" motif in the "belt" sequence ( Fig. 3A). Y169 is 223 located at the base of the hydrophobic pocket forming hydrogen bonds and a - interaction 224 with the "HPL" -hairpin (Fig. 2D). As expected, Y169A significantly decreased the binding of 225 VPS29 to VPS35L (Fig. 3H). Interestingly, this mutation simultaneously increased the binding to 226 Retromer components VPS35 and VPS26A/B (Fig. 3H), suggesting a potential competition 227 between Retriever and Retromer for the same pool of VPS29 in cells. Next, we tested the effect 228 of a mutation in VPS29, I95S, which was previously shown to disrupt the VPS29-VPS35 229 interaction in Retromer 30 . Interestingly, although I95 in VPS29 has a close contact with both 230 VPS35 and VPS35L, this mutation selectively reduced the binding to VPS35, but preserved the 231 interaction with VPS35L (Fig. 3H). This result highlights the differences in the binding 232 mechanism of VPS29 between Retromer and Retriever.

233
The interaction between VPS35L and VPS26C is mediated by an extensive and 234 conserved interface involving 12 and 13 of VPS26C and 4, 5, 6, and 8 of VPS35L (

240
Previous studies reported that the rate of mutation in VPS35L exceeds random mutation 241 burden when the gene length is considered 31 . Our review of the COSMIC database 242 (https://cancer.sanger.ac.uk/cosmic) also indicates that the rate of somatic mutations in VPS35L 243 exceeds that of its closest paralog, VPS35, across all tumor types (Fig. 3E). We first used the 244 SNAP2 tool to assess the potential impact of the missense mutations 32 , through which we 245 identified a number of somatic mutations with high likelihood of functional impairment, 246 accounting for 25 -52% of total 235 missense mutations, depending on the evaluation 247 stringency (Extended Data Table 2). When projected onto the cryo-EM structure of Retriever, 248 several of these mutations were found to potentially disrupt the interaction between the NT "belt" 249 and the CT region of VPS35L, while others were clustered over the binding surfaces for VPS29 250 and VPS26C (Fig. 3F).

251
We then selected several mutations, including a few derived from our structural 252 analyses, to test how they may impact Retriever assembly. We found that mutations predicted 253 to disrupt the interaction between the NT "belt" and the CT region of VPS35L, including W6D, 254 S829E, and the cancer-derived mutation G902E, abolished the binding to VPS29 without 255 affecting VPS26C binding (Fig. 3G). In addition, these mutations simultaneously disrupted the 256 binding to CCC components, including CCDC93, CCDC22, and COMMD1, as well as the 257 binding to DENND10. The same effects were observed when the "belt" sequence was deleted, 258 as shown earlier (Fig. 2E, VPS35L 10). In contrast, the cancer-derived mutation G325E 259 specifically disrupted VPS35L binding to VPS26C, without affecting the binding to VPS29 or 260 CCC components (Fig. 3G). This suggests that, unlike VPS29, the association of VPS26C with 261 VPS35L does not contribute to the Retriever-CCC interaction. Other mutations in the NT "belt" 262 or the CT region of VPS35L did not exhibit appreciable effects on complex assembly under our 263 experimental conditions, when they were transiently expressed and mutated in isolation (

265
We proceeded with the four mutations that had profound effects on Retriever assembly 266 to further examine how they may impair Retriever function in cells. For this, we used 267 CRISPR/Cas9 mediated gene editing to knock out VPS35L from liver cancer Huh-7 cells and 268 then stably reconstituted VPS35L expression using an empty vector (EV), or VPS35L variants, 269 including wild-type (WT), W6D, S829E, G902E, and G325E (Extended Data Fig. 3A

304
The specific effects of structure-guided and cancer-associated mutations in VPS35L 305 allowed us to examine the physiological function of Retriever assembly in cell models. First, we 306 observed that all VPS35L variants maintained endosomal localization, irrespective of their ability 307 to interact with CCC, which is evident from their co-localization with the WASH subunit FAM21 308 (Extended Data Fig. 4A). We confirmed our prior observation that loss of the CCC complex, as 309 a result of COMMD3 or CCDC93 deficiency, increased VPS35L cytosolic staining 22 , but did not 310 completely abrogate endosomal recruitment of VPS35L (Extended Data Fig. 4B). Thus,

311
Retriever recruitment to endosomes, while enhanced by CCC, is not fully dependent on it, thus 312 explaining the similar localization of VPS35L mutants on FAM21-positive endosomes (Extended 313 Data Fig. 4A). While VPS35L is predominantly endosomal, we observed that a small amount of 314 the protein is detectable in LAMP1+ vesicles. Interestingly, mutants that lost the ability to bind to 315 VPS29 and CCC (i.e., W6D, S829E and G902E) had reduced localization to this compartment, 316 while the G325E mutation disrupting VPS26C binding had no significant effect (Fig. 4A, B).

317
Next, we assessed the impact of disrupting Retriever assembly on the trafficking of a 318 well-established cargo protein, Integrin-1 (ITGB1). Loss of VPS35L is known to impair ITGB1 319 endosomal recycling 22,23 , which was also observed in Huh-7 VPS35L knockout cells rescued by 320 empty vector (EV), where we observed significant endosomal trapping of ITGB1 (Fig. 4C, D).

321
Compared to EV, however, the impact on ITGB1 recycling was not as profound for other 322 mutants in VPS35L, with G902E showing significant endosomal trapping, while other mutants

323
showing a milder and statistically insignificant effect (Fig. 4C, D). Thus, these data suggest that 324 the mutations did not fully abrogate the function of Retriever.

325
To further delineate the functional effects of these mutations we used surface

341
Prominent in the latter group were several components of the Arp2/3 complex. It was previously 342 shown that Arp2/3 is more extensively recruited to endosomes in CCC and VPS35L deficient 343 cells 22 . This is consistent with our observations here that Arp2/3 was correspondingly reduced 344 from the PM (Fig. 4E). In agreement with the proteomic findings, we observed significant  combinations of subunits, we were able to obtain highly reliable models that depict the 361 architecture of the entire Retriever-CCC complex. These models were further validated using 362 our biochemical and cellular assays. In the following sections, we will describe the structural 363 models in separate segments.

365
The CCDC22-CCDC93 dimer binds to the outer ridge at the CT of VPS35L

366
We first evaluated the reliability of AFM predictions by examining its capability to predict 367 the structure of Retriever itself, for which no homologous structures were yet available.

368
Remarkably, all predicted models exhibited a near perfect alignment with our cryo-EM structure,

376
(not shown). Hence, AFM can reliably predict unknown structures of large complexes. In all our 377 AFM predictions, we applied three criteria to evaluate the reliability of the predicted models 33,34 .

378
These included the predicted local difference distance test (pLDDT) scores to assess the 379 accuracy of the local structure, the predicted aligned error (PAE) scores to evaluate the distance 380 error between every pair of residues, and the visual consistency of at least 25 solutions when 381 aligned to evaluate the convergence of predictions. In most cases, we found that the visual 382 consistency of the 25 aligned models agreed well with the PAE and pLDDT criteria.

383
In all the AFM predictions involving different subunits of CCC and Retriever, only 384 CCDC22 and CCDC93 always bound to VPS35L in a highly consistent manner, while none of 385 the other subunits were able to establish a reliable contact between CCC and Retriever. In all 386 solutions, CCDC22 and CCDC93 form an extended heterodimer containing four coiled coils.

387
The last two and a half-coiled coils (CC2b, CC3, and CC4) at the C-termini were consistently

397
The VBD interacts with VPS35L at two conserved surfaces. The first surface 398 encompasses helix 24 and the connecting loops between 25 and 26, 27 and 28, and 29 399 and 30 (Fig. 5B). The second surface is contributed by helix C, which precedes the solenoid 400 helix 1 (Fig. 5B). This C helix is absent in VPS35 and is not visible in our cryo-EM map of 401 Retriever. Interestingly, the first VBD binding surface is located at the opposite side of the same 402 solenoid region of VPS35L that binds to VPS29. In addition, the coiled coil CC2b is in close 403 proximity to the "belt" peptide ( Fig. 5B). This configuration provides a plausible explanation for 404 why disrupting the "belt" peptide or VPS29 binding impacted the binding to CCC (Fig. 3,

405
Extended Data Fig. 3). We propose that loss of the "belt" peptide or VPS29 disturbs the local 406 conformation of the CT region of VPS35L, which in turn allosterically destabilizes the binding of 407 CCDC22-CCDC93 binding to Retriever.

408
To validate the predicted model, we purified the recombinant CCDC22-CCDC93 VBD

463
The AFM model reveals that DENND10 consists of two closely packed domains, the N-

472
To validate the predicted structure, we purified the DBD heterodimer and DENND10 473 recombinantly and used size-exclusion chromatography to test whether they can directly 474 interact with each other. Individually, the untagged DBD dimer and DENND10 eluted at ~15 mL, 475 corresponding to their similar molecular weight of ~40 kDa. When the two components were 476 combined, a new peak appeared at ~13 mL, indicating the formation of a complex (Fig. 6B). The 477 peak contained all three proteins in near 1:1:1 stoichiometry, confirming the direct interaction 478 and stable complex formation between the DBD and DENND10 (Fig. 6B).

479
To further validate the predicted structure, we used MBP pull-down assays and co-480 immunoprecipitation to test if mutations in the conserved residues predicted to be crucial for the 481 interaction could disrupt the binding. Consistent with the AFM model, the W30D and Y32D 482 mutations in DENND10 completely abolished its binding to CCDC22-CCDC93 DBD (Fig. 6D).

483
Both residues are located at the center of the interaction surface between the NTD of DENND10

504
Remarkably, we could obtain a highly convergent model when we included one copy of 505 each of the ten COMMD proteins, with or without the CCDC22-CCDC93 heterodimer. This 506 model is consistent with our quantitative proteomic analyses of the native CCC-Retriever 507 complex isolated from blue native gels, where the ratio of all 10 COMMD proteins, except for 508 COMMD7, is nearly equimolar (Extended Data Table 1). The resulting AFM model consistently domains. The ten COMMDs are arranged in a highly organized manner, following the sequence 513 of (1/6)-(4/8)-(2/3)-(10/5)-(7/9). Subunits within the same parentheses form a heterodimer 514 through a face-to-face "hand shaking" interaction between their COMMD domains. These

542
To validate the AFM model, we extended our predictions of the COMMD ring in other 543 species, including fish and amoeba, which possess all the 10 COMMD proteins, as well as 544 CCDC22 and CCDC93. Strikingly, the positioning of COMMD orthologs within the ring follows 545 the exact sequence predicted for human proteins (Extended Data Fig. 7). Similarly, the regions 546 of CCDC22 and CCDC93 that interact with the ring and the points of contact on the ring itself 547 are also highly consistent across these species (Extended Data Fig. 7). It is interesting to note  and (5/10) (Fig. 7E). Notably, both mutations reduced the binding of CCDC22 to Retriever, 568 DENND10 and CCDC93 (Fig. 7E), even though the mutated residues are not expected to 569 directly interact with these proteins. These results suggest that the binding to the COMMD ring 570 creates a supra-structure that may be critical to support other protein-protein interactions

614
One key observation reported here is that cancer-associated missense mutations in 615 VPS35L can dramatically affect Retriever assembly. The precise mechanism by which the loss

623
The model of Retriever highlights key distinctions from Retromer, in both structure and 624 regulation. Retriever has a more compact structure and a unique mechanism of intramolecular 625 interaction between the NT and CT portions of VPS35L. In particular, the "belt" sequence at the 626 NT of VPS35L provides a direct binding interface for VPS29, a feature completely absent in 627 Retromer. Another feature unique to Retriever is the long unstructured peptide linker in VPS35L 628 that follows the "belt" sequence. The primary sequence of this serine-rich unstructured linker is 629 highly conserved in vertebrates, suggesting that it may be a site for regulatory interactions or 630 post-translational modifications, which remain to be elucidated.

631
Our study also uncovered how Retriever encounters CCC to assemble a larger complex.

632
We find that the CCDC22-CCDC93 heterodimer is the essential scaffold around which all the 633 components, including DENND10, are assembled. Rather than being "beads on a string", the 634 structure of Retriever-CCC is highly compact and selectively oriented, with only limited internal

642
DENN10 has been reported to localize to late endosomes and multivesicular bodies, and to act 643 on Rab27a/b 26 , which have not been implicated as cellular or molecular targets of Retriever-644 CCC. Therefore, these observations remain to be reconciled and expanded.

645
One crucial question to be addressed is the functional importance of the COMMD ring 646 and the exquisite conservation of its assembly through evolution. Our studies show that 647 mutations that prevent binding of CCDC22 to the ring also impaired its ability to dimerize with 648 CCDC93 or bind to DENND10 and VPS35L. Based on these observations, we propose that the 649 COMMD ring helps assemble or stabilize the CCDC22-CCDC93 heterodimer and is therefore 650 essential to the assembly of the entire complex. This could explain why the CCDC22-CCDC93 651 heterodimer is destabilized whenever any given COMMD protein is knocked out in cells or in    Table 1). Altogether, these 676 observations suggest that the complex may be dynamically regulated, rather than being a static 677 entity. The mechanism by which the ring is fully assembled from precursor complexes will likely 678 play a key role in the function of the CCC complex.

679
Another intriguing feature of the COMMD ring is the highly conserved order of assembly 680 of its ten COMMD proteins. Presumably, the high conservation of each of the COMMD family members is essential to yield this arrangement. In the structure, the central part of the ring is 682 created by COMM domain-mediated heterodimers, which are then the building blocks to 683 assemble the ring. Therefore, we postulate that unique sequence variability in the COMM 684 domains favors specific heterodimers over others. The model also shows that these 685 heterodimers are further "glued" together by the CBDs of CCDC22 and CCDC93, as well as the 686 N-terminal globular domains of each COMMD through conserved interfaces. Since the globular 687 domains of COMMD proteins provide most of the exposed surface of the ring, we speculate that 688 they likely provide key interfaces for regulatory interactions between the ring and other proteins.

689
Finally, concurrent with our effort, two other groups independently provided other 690 structural aspects of this assembly 55,56 . These studies are complementary to ours, as they did 691 not resolve the experimental structure of Retriever but determined the cryo-EM structure of the 692 CCC ring, which turned out to be highly similar to our predicted and tested models. Therefore, 693 our study is unique in providing the high-resolution cryo-EM structure of Retriever and

749
VPS35L Huh-7 knockout cells generated by CRISPR/Cas9 as detailed above were 750 reconstituted with either HA empty vector or HA-tagged VPS35L wild type or mutants, using a

775
After this, cells were rinsed by spinning and resuspension in FACS buffer, 3 times. For Villin 776 staining, cells were fixed and permeabilized using BD Cytofix/Cytoperm solution kit according to 777 the manufacturer's instructions (BD Biosciences). Thereafter, they were incubated with Villin 778 antibody overnight at 4⁰C in BD Perm/wash buffer, followed by 3 washes using the same buffer.
Thereafter, secondary antibody incubation was performed in BD Perm/wash buffer, followed by 780 3 washes using the same buffer. The primary and secondary antibodies used for staining are 781 detailed in Supplementary Table 5  transblotted to PVDF membranes. After transfer, the proteins were fixed by incubating the 798 membrane in 8% acetic acid for 15 minutes, followed by immunoblotting as described above.

799
For proteomic experiments, gels were stained with Coomassie blue and gel slices of specific 800 apparent mass were cut and submitted for analysis.

936
/Å2 over 60 frames. The defocus range of the images was set between -1.2 to -2.4 μm. In total, 937 3,594 movies were collected and used for data processing.

938
Electron Microscopy data processing: The cryo-EM data were processed using 939 cryoSPARC 62 v4.2.1. To correct for beam induced motion and compensate for radiation damage 940 over spatial frequencies, the patch motion correction algorithm was employed using a binning 941 factor of 2, resulting in a pixel size of 0.83 Å/pixel for the micrographs. Contrast Transfer 942 Function (CTF) parameters were estimated using patch CTF estimation. After manual curation, 943 a total of 2,892 micrographs were selected for further processing from the initial 3,594 944 micrographs. For particle picking, a 4.3 Å map of the Retriever complex obtained from the pilot 945 dataset was used as a template, resulting in the identification of 1,221,095 particles. After 2D 946 classification, 1,105,321 particles were selected and subjected to ab initio 3D reconstruction, 947 followed by heterogeneous refinement (Extended Data Fig. 1). The best resolved 3D class, 948 containing 426,624 particles, was selected for the final non-uniform refinement followed by the 949 CTF refinement, producing a full map with an overall resolution of 2.94 Å with a binned pixel 950 size of 1.0624 Å/pixel. DeepEMhancer 63 was then used with the two unfiltered half maps to 951 generate a locally sharpened map (EMD-40885/PDB-8SYM). To better resolve the interaction 952 between VPS29 and VPS35L, a mask was applied around VPS29 and the adjacent C-terminal 953 region of VPS35L and signals outside the mask were subtracted (Extended Data Fig. 1G). Next, 954 3D classification without alignment was applied to the subtracted particle stack, resulting in a 955 class containing 83,654 particles with better resolved density of the "belt" sequence. Local 956 refinement of this class resulted in a map with an overall resolution of 3.18 Å, which was further 957 sharpened by DeepEMhancer. This map was then aligned with the full map and combined using