Structural plasticity of Atg18 oligomers: organization of assembled tubes and scaffolds at the isolation membrane

Autophagy-related protein 18 (Atg18) participates in the elongation of early autophagosomal structures in concert with Atg2 and Atg9 complexes. How Atg18 contributes to the structural coordination of Atg2 and Atg9 at the isolation membrane remains to be understood. Here, we determined the cryo-EM structures of Atg18 organized in helical tubes as well as soluble oligomers. The helical assembly is composed of Atg18 tetramers forming a lozenge cylindrical lattice with remarkable structural similarity to the COPII outer coat. When reconstituted with lipid membranes, using subtomogram averaging we determined tilted Atg18 dimer structures bridging two juxtaposed lipid membranes spaced apart by 80 Å. Together with an AlphaFold Atg18-Atg2 model, we propose that Atg18 oligomers form a structural scaffold coordinating the Atg2 membrane bridge. The observed structural plasticity of Atg18’s oligomeric organization and membrane binding provide a molecular framework for the positioning of downstream components of the autophagy machinery.


Introduction
Macroautophagy (from here on referred to as autophagy) degrades long-lived proteins, macromolecular complexes and organelles in order to recuperate the molecular building blocks for the cell [1][2][3][4][5] . In this function, autophagy is critical for cellular maintenance and homeostasis in eukaryotes and dysregulation is implicated in neurodegeneration, cancer, inflammation and infection 6,7 . During autophagy a double-membrane organelle called autophagosome is formed de novo 8 , engulfs cytosolic contents and fuses with the lysosome 9 . In the meantime, more than 40 gene products have been elucidated, which are organized in six major functional protein complexes: Atg1 kinase 10 , class III phosphatidylinositol 3kinase (PI-3K) 11 , Atg9 lipid scramblase 12 , the Atg2-Atg18 lipid transfer complex 13 and two ubiquitin-like conjugation systems 14 .
In budding yeast Saccharomyces cerevisiae, the Atg1 complex initiates the preautophagosomal structure (PAS) by phase separation 15 . Subsequently, small lipid vesicles containing trimeric Atg9 are delivered to the early isolation membrane (IM) 16 . Initially, the IM was observed in direct contact with membranes of the endoplasmic reticulum (ER) 17,18 .
Moreover, COPII vesicles and its molecular components were also shown to be involved lipid provision to the autophagosome when the IM is in contact with the ER exit site 19,20 .
While the Atg2-Atg18 complex was found to tether the pre-autophagosomal membrane to the ER 21 including Atg9 22 , at later stages the Atg2-Atg18 complex was observed localized to the expanding tips of the IM 23,24 . A recent Atg9 cryo-EM structure revealed the trimeric organization and biochemical activity as a lipid scramblase 25,26 . Atg2 is a very large 1592 aa rod-shaped protein that was shown to function as an intermembrane lipid transfer protein [27][28][29] and has homology to a recently resolved VPS13 structure revealing the principal architecture of a lipid slide 30 .
Closely linked to Atg2 is Atg18 that early on was shown to be essential for the progression of autophagy 31,32 . Atg18 belongs to the protein family of β-propellers that bind polyphosphoinositides (PROPPIN). The corresponding primary structure contains a series of conserved WD40 repeats forming the 6 or 7 blades of a β-propeller fold. Initially, the Xray crystal structures of Hsv2, which is a close relative of the PROPPIN family 33,34 and the Pichia angusta homolog of Atg18 had been determined 35 . More recently, the structure of S. cerevisiae Atg18 was elucidated in the presence of phosphate and citrate, revealing binding sites for phosphatidylinositol-3-phophate (PI3P) and phosphatidylinositol-3,5-bisphophate PI (3,5)P2 (PDB-IDs 5LTD, 5LTG) 36 . Mutations in conserved binding sites for PI3P and PI (3,5)P2 (FRRG motif), affect the progression of the Cvt pathway and autophagy as well as vacuolar morphology 37,38 supporting the specific membrane adapter role of Atg18.
Moreover, Atg18 contains three up to 120 aa loop regions (i.e. 6AB, 6CD, 7AB) that possess high content of predicted structural disorder and emanate from the blades of the β-propeller fold. The 6CD loop (324-406) of Atg18, for instance, participates in membrane binding 35,39 and a subsegment was proposed to form an amphipathic helix in the membrane environment 40 . Moreover, lipid reconstitution experiments of Atg18 with giant unilamellar vesicles were shown to remodel and tubulate membranes thus mediating membrane scission 40 . Once bound to membranes, Atg18 was found to undergo oligomer formation as observed by cross-linking mass spectrometry 35 . In addition to P72R73, the 7AB loop (430-460) of Atg18 was shown to bind to the rod-shaped Atg2 molecule 36 that is thought to bridge the PI3P-rich IM on one end with a membrane lacking PI3P like the ER at the opposite end 27 .
In addition to Atg18, Atg21 and Hsv2 are structurally related yeast PROPPIN family members involved in membrane trafficking 41 . It was found that Atg18 is essential for autophagy and the related cytoplasm to vacuole targeting (Cvt) pathway whereas Atg21 solely for Cvt 13,37 . The specific PI3P binding and downstream Atg recruitment function of Atg18 and Atg21, however, can be compensated by one another 42 . Hsv2 was shown to be required for microautophagy of the nucleus 43 and was the first of the PROPPINs to be structurally resolved 33,34 . In higher eukaryotes, Atg18 family proteins are also known as WD-repeat protein interacting with phosphoinositides (PIP) (WIPI) with a total of four isoforms occurring in human cell lines. WIPI1, WIPI2 and WIPI4 have been shown to be directly linked to autophagic progression 44 . Overexpression of WIPI1 led to large and elongated light microscopic punctae in human cell lines 45 . WIPI2 recruits the ATG12-ATG5-ATG16L1 LC3 conjugation machinery 46 and WIPI4 was demonstrated to biochemically interact with ATG2A and assist in the lipid transfer activity 28 . The Atg18/WIPI family constitute a complementary network of proteins involved in autophagy and autophagyrelated cellular processes.
Although various aspects of Atg18 contributions to the autophagy pathway have been investigated, the involvement and structural role of oligomeric species including Atg2 at the IM remain poorly understood. In this study, we applied electron cryo-microscopy (cryo-EM) to solve structures of helically assembled, soluble and membrane-bound Atg18 using highresolution cryo-EM structure determination as well as subtomogram averaging. Together with structural models the observed structural plasticity indicates that oligomeric Atg18 dimers when bound to lipid membranes may define the spatial coordination of Atg2 with respect to the IM.

Atg18 forms higher-order helical assemblies
In order to structurally characterize Atg18 oligomers and their interaction with lipid membranes, we recombinantly overexpressed Saccharomyces cerevisiae Atg18 and subsequently purified it under high-salt conditions. Interestingly, once dialyzed into low salt buffer (100 mM KCl), Atg18 formed large supramolecular structures when observed in negatively stained EM samples. Using cryo-EM of vitrified Atg18, we identified helical tubes of 260 Å diameter including a diamond-shape repeat pattern along the helix (Fig. 1A). We acquired cryo-EM micrographs of wildtype Atg18 (Atg18-WT) and Atg18-PR72AA, an Atg2 binding mutant 47 , and subjected them to segmented helical image processing. 2D class averages of the helical tubes revealed individual Atg18 propeller disc-shaped structures including individual blade separation (Fig. 1B). The presence of helical layer lines beyond 1/10 Å in the corresponding Fourier transforms indicated an ordered helical repeat of the assembly. The transforms of Atg18-WT and Atg18-PR72AA 47 supported the same structural organization of both helical assemblies (Fig. 1C), with an improved helical ordering in the Atg18-PR72AA mutant indicated by the higher diffracting layer line at 1/6 Å.
The assemblies exhibited a pitch of approximately 100 Å with 5.1 helical units per turn. Using helical parameters of 19.4 Å helical rise and 70.0° rotation, we determined the helical structures to 3.8 Å and 3.3 Å, respectively ( Fig. 1D-E; Suppl. Fig. 1A-C). Interestingly, Atg18-WT tubes exhibited diameters of 210 Å whereas Atg18-PR72AA tubes are slightly wider at 220 Å (Suppl. Fig. 1D-E). In both structures, the basic helical asymmetric unit contained four copies of Atg18 assembled in a diamond outline with an internal channel of 110 Å diameter.  When we inspected the cryo-EM maps, they showed the presence of expected density features at the determined resolution such as β-strand separation and side chain densities.
The observed features allowed for atomic model building of Atg18 into the cryo-EM density (Table 1, Fig. 1F-G). In comparison with the Atg18 crystal structure (PDB-ID 6KYB) 36 , we found high structural overlap with only minor differences on the rotamer level (RMSD=1.18 Å with 2471 atoms) and slightly extended density at the loop 7AB (Suppl. Fig. 2A-D).
Residues beyond the β-propeller structural scaffold including the loops 4CD (157-221), 6CD (322-408) and parts of 7AB (446-457) were not resolved in the tube structures of Atg18-WT and Atg18-PR72AA, which is in agreement with both the AlphaFold2 prediction 49 (Uniprot ID P43601) and PDB-ID 6KYB. Although the Atg18-PR72AA tube structure was determined at slightly higher resolution when compared with the Atg18-WT (3.3 Å over 3.8 Å), the differences between the atomic models are only detectable as minor positional changes within the assembled tubular structure (Suppl. Fig. 1C).
Common to both structures Atg18-PR72AA and Atg18-WT are two major Atg18 interfaces in the assembly: first, inside the diamond the interface is formed by two perpendicularly arranged Atg18 discs resulting in a "-|" or "T" configuration and, second, the Atg18 interface between diamonds is formed by two Atg18 discs arranged in line giving rise to "--" or "I" configuration ( Fig. 2A-C). The T-interface includes R285 and R286 of the FRRG motif that are in direct contact with blade 6 of the adjacent molecule (Fig. 2D). The I-interface is located between blade 3 and blade 4, has due to a 180º disc rotation a pseudo dihedral symmetry including in a small tilt of the two discs with respect to each other.

Tubular assembly is stabilized by FRRG motif
In order to better understand lipid adaptor function of Atg18 in the context of the newly observed helical assembly, we mapped two binding sites of PI3P and PI (3,5)P2 separated by the FRRG-motif onto the presently determined Atg18 assembly structure (Fig. 3A, B).
Interestingly, the lipid binding sites appear buried in the T-interface of the helical tube suggesting a stabilizing role of the involved residues. In reference to previous mutation studies 34,37,38 that abolished binding to PI3P and PI(3,5)P2, we replaced the positively charged arginines in the FRRG motif to FGGG by neutral glycines. After purification, we found a complete solubilization of Atg18-FGGG over Atg18-WT tubes when subjected to a pelletation assay ( Fig. 3C and Suppl. Fig. 3). Furthermore, Atg18-FGGG did not form tubes in comparison with Atg18-WT when imaged by negative stain EM. To further experimentally test the influence of the PIP-binding site on the tubular assembly, we added the soluble diC8-PIP derivates of PI3P, PI(3,5)P2 and PI(4,5)P2 in four-fold molar excess to Atg18-WT prior to low-salt buffer change and visualized these samples by negative stain EM. In contrast to the control-containing formed Atg18 tubes, only smaller particulate aggregates could be observed in the presence of PI3P, PI(3,5)P2 and PI(4,5)P2, respectively (Fig. 4D).

Pelletation assays and subsequent separation by SDS-PAGE confirmed that the intensity of
Atg18 bands moved to the supernatant fraction for PI (3,5)P2 and PI(4,5)P2 incubations (Fig.   4E). In the case of PI3P, however, we observed a decrease in the pellet fraction in comparison with the untreated control as well as elongated aggregates in negative stain EM.
Subsequently, we investigated whether pre-formed tubes were affected by the supplemented lipids. Interestingly, only PI(3,5)P2 showed disassembled elongated aggregated tubes when visualized by negative stain EM (Fig. 4F), whereas the addition of PI3P or PI(4,5)P2 did not affect the helical assemblies visibly. These observations are in line with the highest measured affinities of PI (3,5)P2 to the FRRG site 39 . Together, these experiments show that the FRRG motif of Atg18 takes up a stabilizing role in the formation of the helical Atg18 assembly, which is further supported by binding studies of PI3P derivates that can prevent the assembly of Atg18 tubes and keep Atg18 soluble.

Soluble Atg18 is composed of smaller oligomers
In order to further investigate the structures of soluble Atg18 fractions, we set out to determine structural intermediates of the Atg18 assembly. As indicated by the pelletation assay of Atg18-WT, an equal-share fraction was not pelleted and remained in the supernatant. Using these preparations in high salt buffer, we determined the single-particle cryo-EM structure of soluble Atg18-WT fractions. Due to the very small particle size, i.e. the Atg18 monomer corresponds to 55 kDa, we imaged over 1000 particles on a single cryomicrograph (Fig. 4A). Classification analysis revealed separate views of single Atg18 displaying characteristic structural features of blade separation of the β-propeller (Fig. 4B).
Alongside Atg18 monomer particles, we were also able to classify a selection of Atg18 dimers (Fig. 4C). Comparison with projections of the T and I-interface dimers extracted from atomic model of the Atg18 tube matched some of the 2D classes (Fig. 4D). Moreover, we failed to reconstruct coherent dimer 3D structures of these Atg18 dimers presumably due to higher flexibility in solution than observed in the helical assembly. Nevertheless, using monomeric class members we determined the 3D structure of Atg18-WT at 4.8 Å resolution according to the FSC 0.143 criterion (Fig. 4E). Next, we successfully docked a single refined atomic model taken from the helically assembled Atg18-P72R73 into the density of monomeric Atg18-WT (Fig. 4G). In accordance to the helical assembly structure, three major loop regions were not resolved and disordered in monomeric Atg18. Together, structure determination of soluble Atg18 support the occurrence of monomers and dimers in agreement with the T and I-interfaces observed in the helically assembled lattice structures.

Visualization of membrane-associated Atg18 oligomers
To further investigate the binding mode of Atg18 to lipid membranes, we mixed soluble Atg18 with PI(3,5)P2-doped large unilamellar vesicles (LUVs). Next, we imaged the corresponding plunge-frozen samples by electron cryo-tomography. After acquiring a tilt series with the Volta phase plate, we reconstructed a total of 16 tomograms displaying strong low-resolution contrast. In comparison with common LUV controls that exhibited spherical vesicles of an average 30 nm diameter, the images showed tightly tethered deformed vesicles of square, pentagonal or higher polygonal-like shapes with several 100s nm long stretches of straight and parallel membrane paths next to other vesicles (Fig. 5A,   B; Suppl. Fig. 4 and Suppl. Movie 1). The membrane surfaces are typically coated by discshaped particles corresponding to Atg18 β-propellers. In between parallel membrane stretches, we found density that we interpreted as Atg18. In order to further analyze these membrane-associated structures, we applied membrane segmentation followed by membrane-guided particle picking using the PySeg package 50 . When generating rotational 2D class averages along the membrane plane from approx. 150,000 extracted subtomograms, some classes showed the presence of additional density between two 60 Å thick bilayers leaving an intra-membrane space distance of 80 Å (Fig. 5C). Other classes containing a single membrane only or a second more blurred membrane were excluded from further image processing. Further 3D subclassification resulted in a structure of two Sshaped densities packed against each other made up from eight disc-shaped densities with approx. 50 Å diameter each (Fig. 5D-F). A total of 8,300 subtomograms resulted in a structure at 26 Å estimated by mask-less FDR-FSC 48 (Suppl. Fig. 5). In the density, we found Atg18 molecules in four pairs of dimers with a slightly twisted interface matching the observed I-interface of the tubular Atg18 assembly determined above (Fig. 5G). Notably in the cryo-EM density, the basic Atg18 dimer unit was found to assume a 45º tilt angle with respect to both membrane planes, which could be independently confirmed by measurements in raw tomogram slices (Fig. 5H). The determined subtomogram structure reveals that tilted Atg18 dimers in I-configuration can establish contact between two opposing bilayers and can further align longer stretches of two opposing lipid membranes in constant distance of approx. 80 Å.

Model of Atg2-Atg18 complex on lipid membranes
In order to further interpret the experimentally obtained subtomogram average structure in the context of the Atg2 complex, we extended the determined Atg18 structures with AlphaFold (AF)-predictions. As a starting point, we verified that the rigid body fit of the Atg18 dimer was consistent with the location of the PI3P binding sites and loop 6CD facing opposite membranes (Fig. 6A, B). Based on this structure, we computed the protein complex prediction of N-terminally truncated Atg2 (541-1592) with Atg18 using the AFmultimer approach 51 (Fig. 6C). The lowest energy structure prediction of the Atg2 (541-1592)-Atg18 complex showed two major interfaces that are consistent with previously determined binding interfaces while leaving the I-dimer interface available for other Atg18 molecules (Fig. 6D): first, Atg18-P72R73 had been mutated to disrupt Atg2 binding 47 and second, a WIR motif containing peptide of human ATG2A had been crystallized with WIPI3 52 . This x-Φ-x-Φ-x-x-x-φ-F PROPPIN interaction motif identified in human ATG2A corresponds to residues 921-938 in Atg2 that bind between Atg18's blade 1 and 2 (Fig. 6E).
The full-length Atg2 model is available from the AF-EBI databank (Uniprot ID P53855). The Atg2 molecule consists of a 200 Å long β-helix cylinder with a hydrophobic channel inside predicted at high confidence with accessory loops and α-helices predicted at lower confidence. When extending the truncated Atg2 (541-1592) model with the predicted fulllength Atg2 and the determined Atg18 dimer, the conserved basic N-terminal residues of Atg2 are in contact with the opposite membrane 27 (Fig. 6F). In this rigidly extended complex, Atg2 assumes a tilted orientation with respect to the lipid membrane (Fig. 6 G-H). Together, based on the experimentally determined Atg18 dimer membrane scaffold, we modeled a Atg2-Atg18 dimer complex in which Atg2 emanates at a tilted angle with respect to the lipid membrane.

Discussion
In order to study the structural role of the Atg18 autophagy membrane adapter, we determined the structures of tubular and soluble Atg18 oligomers using high-resolution cryo-EM. Common to these Atg18 structures are two principal oligomeric binding modes that are mediated by so-called T and I-interfaces, respectively. The FRRG motif that is known to bind PIP constitutes a critical part of the T-interface thereby enabling the formation of a large tubular lozenge Atg18 lattice. When Atg18 is reconstituted with lipid membranes containing (PI)3,5P2, subtomogram averaging reveals Atg18 oligomers including an elongated tilted Atg18 dimer at the core, which is capable of juxtaposing two opposite membrane bilayers.
In the context of a modeled Atg2 complex, the determined membrane-bound Atg18 dimer subtomogram average suggests a scaffolding role of coordinating the structural organization of the Atg2 lipid membrane bridge. The autophagic membrane adapter Atg18 reveals an unexpected structural plasticity in multiple modes of oligomeric organization.
Initial structural characterization of purified Atg18 revealed the formation of higher-order helical assemblies under low-salt conditions (Fig. 1). The determined cryo-EM structures of Atg18-WT and Atg18-PR72AA constitute two main propeller interactions mediated by the T and I-interface between neighboring monomers (Fig. 2). Analysis of crystal symmetry contacts in a recent X-ray structure revealed identical T and I-interfaces in low-salt buffer, even though the protein was trypsinized to remove the disordered loop regions 36 .
Proteolytically truncated Atg18 forms a tightly packed lattice suitable for 3D crystallization but does not form a hollow cylindrical lozenge lattice built from diamond-shaped tetramers, which we observed for full-length Atg18 in the electron micrographs. Interestingly, the outer membrane-trafficking COPII coat of Sec13-31 displays a very similar structurally related diamond arrangement built from four T-shaped β-propeller units forming a scaffold around lipid membrane tubules 53 (EMD-11194, PDB-ID 6GZ6) (Suppl. Fig. 6). Similar roles of coat formation around lipids can also be envisioned for the observed Atg18 tubes in particular as previous studies showed that Atg18 is capable of tubulating GUVs 40 . Moreover, recently two groups identified Atg18 independently as a component of the CROP complex in a membrane-associated retromer complex 54,55 . This architectural conservation of proteinaceous coats suggests functional parallels between different modules of COPII and autophagy trafficking.
The structural similarity of Atg18 tubes to COPII outer membrane coat is a particularly noteworthy observation considering that COPII vesicles have been shown to deliver lipids to the autophagosome IM 20 . Moreover, phosphorylation states of COPII components affect autophagosome abundance in the cell presumably by modulating coat assembly 19 .
Similarly, when Atg18 was incubated in the presence PIP, we observed the disassembly of the tubular structures into soluble structures suggesting functional precursor role of the resolved tubular assemblies. In fact, the T-interface buries the PI3P/PI(3,5)P2 FRRG binding interface and showed direct competition between PIP binding and tube formation (Fig. 3).
Therefore, our data suggest that the characterized Atg18 tubes act functionally prior to engaging in its adaptor role at the phagophore. In yeast, filamentous assemblies have been shown to accumulate in light microscopic puncta under stress conditions and thereby regulating the activity of metabolic enzymes 56 . Therefore, a similar storage role may also be envisioned for the observed Atg18 assemblies.
Many other components of the core autophagy machinery such as Atg9 and Atg13 that were shown to be activated and de-activated by chemical modification of phosphorylation 57,58 .
More specifically for Atg18, it was shown for Pichia pastoris that Atg18 itself is subject to phospho-regulation by two distinct sites in loops 6CD and 7AB, both reducing PIP-binding affinities 59 . Under normal conditions in the cell, Atg18 is phosphorylated in loop 6CD and thereby the negative charge prevents the insertion of the amphipathic helix into the membrane 40 . Similarly, Atg18 phosphorylations in loop 7AB can be expected to affect the stability of the tubular assembly, in particular as the higher resolution structure of Atg2binding deficient mutant Atg18-PR72AA tubes suggested an increased rigidity over Atg18-WT. Conserved regulatory phosphorylation mechanisms of Atg18 will affect the tube stability and may prime smaller Atg18 oligomers to bind PIP-containing membranes.
Yeast and mammalian autophagosomes exhibit distinct PIP asymmetries, i.e. PI3P is present at the IM while almost absent in the ER membrane 60 whereas PI(3,5)P2 is found localized in late endosomes, autophagosomes and vacuoles 38 . The differential binding affinities of different PIP derivatives to Atg18 have been biochemically characterized and quantified before in detail 39 . In line with the highest affinity, we found that PI (3,5)P2 binding leads to disassembly of the helical tubes. The lower binding affinity of PI3P may require additional membrane tethers like the related PROPPIN Atg21 42 . Binding to Atg2 through PR72 in loop 7AB 47 may contribute further to the destabilization of the tubes in order to assist in the targeted localization of Atg18 to phagophore membranes.
When we investigated soluble Atg18 fractions at high-salt conditions by single-particle cryo-EM, we confirmed the presence of monomers and dimers in line with previous native mass spectrometry characterizations 35 (Fig. 4). Mass spectrometry cross-linking identified residues in loops 4 BC and loop 6CD engaged in Atg18 oligomer interactions 35 consistent with the here determined T and I-interface. Furthermore, the study also revealed that the interaction patterns between Atg18 oligomers changes in the presence and absence of liposomes 35 . In support, when we incubated Atg18 with PIP-containing membranes, we observed in subtomogram average structures of liposome decorated dimers and tetramers that lack the interaction via the T-interface due to PIP binding mode (Fig. 5). Together, Atg18 displays an unexpected structural plasticity of different multimer assemblies depending on buffer condition as well as binding partners, which is mechanistically accomplished by two distinct binding interfaces. In order to put the determined Atg18 dimer structure bound to the lipid membrane in the context of the Atg2 complex, we computed an AF model of the Atg18-Atg2 complex (Fig.   6). Here, Atg18 formed two contacts with Atg2 through P72R73 (loop 2BC) and blade 1/2 47,52 . In an expanded Atg2-Atg18 dimer complex, the I-interface as well as the FRRG motif in blade 5 including loop 6CD are spatially accessible to engage in oligomer and membrane binding. In the Atg18 scaffolding configuration, the rod-shaped 20-nm long Atg2 molecule assumes an angle of approx. 45º with respect to the membrane bilayer bridging membrane distance of ~10 nm. In vitro reconstitutions of Atg2 and ATG2A increased lipid transfer across vesicles in the presence of Atg18 and WIPI1/4, respectively 27,28 . Therefore, it is tempting to speculate that the derived geometric arrangement is energetically favorable to immerse Atg2's N-terminal and C-terminal ends of the hydrophobic lipid channel in the outer membrane leaflets. Moreover, participating complex partners such as lipid scramblase Atg9 may also require binding interfaces at this angle in order to channel lipids more efficiently across the membrane bridge to the growing IM. Table 1. Model report of four Atg18-PR72AA monomers (symmetry unit)

Soluble lipid and liposome binding experiments
To test the effect of soluble lipids on Atg18 tube formation, diC8-PI3P and diC8-PI(3,5)P2 (Echelon) were reconstituted in water (1 mM final concentration) and added in 4x molar excess to Atg18 in gel filtration buffer prior to the dialysis into tube buffer. As a control, the same procedure was performed with water instead of reconstituted lipids. To test the effect of soluble lipids on preformed Atg18 tubes, the soluble lipids in water were added to preformed Atg18 tubes in 4x molar excess. As a control, the same procedure was performed with water instead of reconstituted lipids. LUVs were prepared with a lipid content of 95%  Table 2.

Cryo-EM image processing
Micrographs containing Atg18 tubes and soluble monomer/oligomers were processed in classification and 3D variability analysis did not improve the obtained map resolutions.
Soluble monomers/oligomer data was picked by template matching against circular blobs (40-60 Å) in CryoSPARC live. Particle extraction was performed in 256 px / 215 Å boxes to include delocalized signal. 2D classification in 200 classes was performed with increased batch sizes (200) and 100 iterations to a high-resolution limit of 4 Å. Monomeric and dimeric classes were split and subjected to ab initio 3D map generation with increased resolution ranges (12 Å to 6 Å) and increased batch sizes (300 initial, 1000 final size) with K=3 classes.
The monomeric class was low-pass filtered to 8 Å and subjected to 3D refinement and nonuniform local refinement 68 , resulting in a map with FSC(0.143)=4.8 Å whereas dimeric classes could not be resolved further.
Tomograms were reconstructed using IMOD/eTomo software 71 including patch tracking and weighted back projection. Subsequently, they were subjected to the PySeg pipeline 50 for particle picking and classification. Briefly, lipid membranes were segmented by TomoSegMemTV software 72 and PySeg's tracing routines were used to pick subvolumes near the segmented membranes. A total of 150,000 picked subvolumes were averaged in three batches and subjected to rotational 2D classification focused by a cylinder mask using the affinity propagation clustering algorithm. Double membrane classes with protein signal in between the membranes were selected and all other classes were excluded (55 classes out of 300 classes included). This way, the majority of the subvolumes were excluded as they contained single membrane classes, double membrane classes with the second membrane outside of the focused mask as well as triple membranes with no connecting densities. 2D classification and class selection was repeated another time with a mask suppressing the membrane signal (32 out of 516 classes included) before extracted subtomograms were subjected to Relion v3.0 software for subtomogram averaging 73 . An ab initio 3D model was generated and refined and the PySeg tools were used for masked subtraction of the two membranes (100% dampening). Focused classification into four 3D classes showed one dominating class that was used for 3D refinement. FDR-FSC was used to determine the nominal resolution of 26 Å of the averaged subvolume map 48 . Atg18 dimers from helical assembly were docked into the map using ChimeraX software 74,75 .

Atomic model building and refinement
PDB-ID 6KYB was docked into the reconstructed cryo-EM density of Atg18-PR72AA after map auto-sharpening in Phenix v1.20.1 in ChimeraX software. The atomic model of Atg18-PR72AA was manually mutated and slightly modified in Coot v0.9 76 before real-space refinement in Phenix was performed with default parameters 77 . Validation parameters are displayed in Table 1 78,79 . Based on the refined atomic Atg18-PR72R73 model, we used the corresponding Atg18-WT model to rigidly place it into Atg18-WT tubules, Atg18 monomer as well as Atg18 subtomogram average dimer densities.