Revisiting the evolution of the yeast Atg1 complex

Atg101 is distributed in budding yeast species

It is unclear why Atg101 and Atg29/Atg31 are mutually exclusive, given that they play different proposed functions in the context of the Atg1 complex. Atg101 is thought to be only present in fission yeast and higher eukaryotes, but at what point during evolution this protein is gained or lost is unknown.^[2,23] To address this question, we began by probing the putative subunit composition of yeast species from the Saccharomycetes (budding yeast) and Schizosaccharomycetes (fission yeast) classes, specifically focusing on the 149 species with annotated entries of Atg17 in the UniProt database. We queried for Atg29, Atg31, and Atg101 from each curated yeast species. Our systematic search revealed that Atg101 is not only present in the 4 fission yeast strains of the Schizosaccharomycetes class, but it is also found in 79 budding yeast species in the Saccharomycetes class (Figure 2). Interestingly, we discovered that two budding yeast species contain annotated Atg101 and Atg29/Atg31: Nadsonia fulvescens and Cyberlindnera fabianii. Briefly, N. fulvescens is a member of the genus that was originally discovered from tree sap in Russia and is characterized by bipolar budding,^[30] while C. fabianii is an opportunistic yeast that can cause bloodstream infections.^[31] Whether or not these two yeast species indeed form Atg1 complex with both of these sets of subunits needs to be validated experimentally; however, we may have captured a point in evolution when Atg29/Atg31 emerged or when Atg101 was about to be lost. Importantly, our analysis showed that Atg101 is more broadly distributed among yeast than previously thought, and this subunit is present in budding yeast species related to Saccharomyces cerevisiae.

AlphaFold3 accurately predicts the overall architecture of S. cerevisiae and S. pombe Atg17

In the fission yeast S. pombe, the change in subunit composition of its Atg1 complex is accompanied by changes in the molecular structures of the inherently dimeric Atg17 scaffolding subunit and the HORMA domain of Atg13.^[23] Notably, S. pombe Atg17 appears to change from the S-shaped architecture, as observed for S. cerevisiae Atg17, to a rod-shaped architecture.^[21] However, it is not known if such structural change also occurred in Atg101-containing Atg1 kinase complexes from other yeast species. AlphaFold3, a machine learning-based algorithm that accurately predicts the structures of a wide range of proteins and protein complexes with the backbone Cα arrangement matching those determined de novo by X-ray crystallography and other experimental techniques, offers a potential approach to systematically model the structures of Atg17 from different yeast strains.^[29] To evaluate the feasibility of this approach, we performed AlphaFold3 modeling of dimeric S. cerevisiae Atg17 and dimeric S. pombe Atg17. The top-ranked models from our AlphaFold3 runs, which had pTM (predicted template modeling) scores of 0.53 and 0.50, respectively, showed that dimeric S. cerevisiae formed a double-crescent S-shaped overall architecture (which we termed Atg17_S) while dimeric S. pombe Atg17 adopted a linear, rod-shaped overall architecture (which we termed Atg17_R) (Figure 3A,B). The overall lengths of these predicted structural models were also consistent with that measured by negative stain EM (~350 Å compared to ~ 345 Å for S. cerevisiae Atg17; ~376 Å compared to ~ 350 Å for S. pombe Atg17 dimer).^[21] Despite only being trained on experimental structures with a single S-shaped Atg17 orthologue (the crystal structures of L. thermotolerans Atg17 with its interacting subunits (PDB: 4HPQ, 4P1W, 5JHF)), AlphaFold3 accurately predicted the rod-like structure of S. pombe Atg17. Inspection of the AlphaFold3 models revealed the molecular basis of how the structural transition may have occurred. Both S. cerevisiae Atg17 and S. pombe Atg17 were constructed from helical bundles that contained different numbers of α-helices: four for S. cerevisiae Atg17 and eight for S. pombe Atg17 (Figure 3A,B). In S. cerevisiae Atg17, the longest helix α5 (234–391) established the crescent-shaped curvature, which is also observed in the crystal structure of L. thermotolerans Atg17-Atg31-Atg29 (PDB: 4HPQ). By contrast, in S. pombe Atg17, the corresponding curvature-determining helix encompassed three helices, α6 to α8 (204–265, 267–284, 286–358), which were separated by short linkers that preclude the adoption of a curved structure. AlphaFold3 also predicted that S. pombe Atg17 contains an additional N-terminal α-helix that succeeds the interfacial β-strand but is absent in S. cerevisiae Atg17. In summary, the AlphaFold3 modeling results we obtained on S. cerevisiae Atg17 and S. pombe Atg17 demonstrated the power of AlphaFold3 and the suitability of this in silico approach to model the structures of Atg17 from other yeast species.

Phylogenetic analysis of Atg17 dimers revealed a dichotomy in predicted structure architecture

We next applied AlphaFold3 to systematically explore the architecture of dimeric Atg17 from other curated yeast species of the Saccharomycetes and Schizosaccharomycetes classes. Our initial hypothesis is that Atg17 from the Saccharomycetes class would have similar curved architecture to S. cerevisiae Atg17 while those from the Schizosaccharomycetes class would resemble the rod-like S. pombe Atg17. To accomplish our analysis, we collected Atg17 sequences from different yeast species from Uniprot and then applied AlphaFold3 to predict the corresponding dimeric Atg17 structures. In total, 149 structural models were generated. We excluded 6 of these from the final results as they did not meet the criteria of pTM ≥ 0.5. Of the remaining 143 structural models, 139 belonged to species from Saccharomycetes and 4 belonged to species from Schizosaccharomycetes. As an additional validation, we compared the predicted L. thermotolerans Atg17 structural model, which has a pTM score of 0.55, to the crystal structure of the Atg17-Atg31-Atg29 ternary complex from the same species. The predicted model of L. thermotolerans Atg17 aligned well with the Atg17 component of the Atg17-Atg31-Atg29 crystal structure over 623 Cα atoms with an RMSD of 2.176 Å (Figure 3C). The increased RMSD was caused by a slight curvature difference between the predicted structural model and the crystal structure, though both displayed an overall curved architecture.

Next, we classified the overall architecture of predicted models into Atg17_S or Atg17_R, with the species names anonymized to minimize bias. Of the 143 predicted models, 54 were designated Atg17_S, while 89 were designated Atg17_R. We also conducted multiple sequence alignments of the different protein sequences using Clustal Omega and constructed a phylogenetic tree based on the overall shape of dimeric Atg17 (Figure 2).^[28] This composite tree showed a broad distribution of Atg17_R architecture and a clear division between Atg17_S and Atg17_R. As predicted, all species from the Schizosaccharomyces class had predicted Atg17_R architecture similar to S. pombe. Interestingly, 85 of the 139 species from the curated Saccharomycetes class encoded Atg17 that showed a higher level of sequence similarity to S. pombe Atg17 than to S. cerevisiae Atg17. The Atg17 from these species, as expected, were predicted to adopt Atg17_R. These observations suggest that Atg17_R was likely retained in selected species of the Saccharomycetes class and that the change to the Atg17_S architecture and the evolution of Atg29/Atg31 subunits took place within the Saccharomycetes class. Another intriguing observation was that the transition from Atg17_R to Atg17_S correlated with a slight reduction in protein size. Atg17_R had an average protein length of 452 residues, while Atg17_S had an average protein length of 421 amino acid residues.

Correlation between Atg17 architecture and Atg1 complex subunit composition

The composite phylogenetic tree allowed us to further examine the relationship between Atg17 architecture and Atg1 complex composition (Figure 2). Yeast species with Atg17_S contained only Atg31 and/or Atg29, but not Atg101, confirming our previous investigation that showed Atg29 and Atg31 stabilize the curved architecture of S. cerevisiae Atg17. On the other hand, most yeast species that had Atg17_R possessed Atg101. Our observation that some yeast species possess either Atg29 without Atg31 or Atg31 without Atg29 or have Atg17_R with no Atg101 may stem from gaps in UniProt data rather than the true absence of these proteins. Both Atg29 and Atg31 have intrinsically disordered regions, leading to higher sequence divergence across species, which may affect the automatic annotation of these proteins for certain yeast species. We hypothesize that yeast species with Atg29 likely also possess Atg31, and vice versa. Nevertheless, gaps in UniProt data have minimal impact on our finding that some yeast species contain both Atg29/Atg31 and Atg101. For the two budding yeast species that have annotated Atg31, Atg29, and Atg101 (Nadsonia fulvescens and Cyberlindnera fabianii), Atg17_R was observed. N. fulvescens reproduces through an atypical mechanism termed “bud-fission.” As N. fulvescens is evolutionarily farther from S. cerevisiae than C. fabianii while still bearing Atg31, Atg29, and Atg101, we decided to further investigate N. fulvescens Atg17. The predicted overall architecture of N. fulvescens Atg17_R was similar to S. pombe (Figure 3D), but the root mean square deviation (RMSD) over 604 Cα atoms is 7.413 Å, showcasing that there were differences between the structures (Figure 3E). Looking closer at the structure, N. fulvescens had four α-helices making up the helical bundle, similar to S. cerevisiae. However, the corresponding “curvature-determining helix” for the N. fulvescens Atg17 dimer had two helix kinks between residues 247–248 and 295–296 that resembled the short linkers of S. pombe and caused the architecture to be Atg17_R. N. fulvescens Atg17 shared 20.9% and 25.9% sequence similarity with S. cerevisiae Atg17 and S. pombe Atg17, respectively. On the other hand, N. fulvescens Atg29 and Atg101 shared 38.1% and 16.9% sequence similarity with S. cerevisiae Atg29 and S. pombe Atg101. No sequence similarity was detected between the annotated N. fulvescens Atg31 and S. cerevisiae Atg31.

To further validate the AlphaFold3-predicted structure of Atg17 dimers, we experimentally assessed S. cerevisiae, S. pombe, and N. fulvescens Atg17 (Figure 3F–I, Supplementary Figure S1A–C). Our goal was to see if N. fulvescens forms the Atg17_R structure, as predicted by AlphaFold3, or if it behaves like S. cerevisiae Atg17, adopting the Atg17_S structure, especially considering that N. fulvescens still possesses Atg31 and Atg29. We overexpressed and purified recombinant S. cerevisiae, S. pombe, and N. fulvescens Atg17 in E. coli. Atg17 was fused to an N-terminal histidine-maltose binding protein (His-MBP) to maintain the stability of the overexpressed dimers, an approach used in previous characterization of S. pombe Atg17. We analyzed the purified His-MBP-Atg17 dimers using negative stain single-particle EM and observed that His-MBP-tagged S. cerevisiae and S. pombe Atg17 adopted the expected Atg17_S and Atg17_R dimer structure respectively (Figure 3F,G). His-MBP-tagged N. fulvescens Atg17 was dimeric and adopted the Atg17_R architecture (Figure 3H). The junction-to-junction length of N. fulvescens His-MBP-Atg17 was measured to be ~ 337Å, which falls within the range of that observed for S. pombe Atg17. In addition, there were shorter 2D averages of His-MBP-Atg17 that were found to be ~ 253Å long, similar to the small population found in negative stain EM analysis of S. pombe His-MBP-Atg17 (Figure 3I).^[21]

We next used AlphaFold3 to test if N. fulvescens Atg17 could interact with the Atg31 and Atg29 orthologs. For this prediction, we supplied the sequence of the N-terminal domain of N. fulvescens Atg29 (1–140) devoid of its IDR. We also focused on modeling monomeric Atg17 with the Atg29 and Atg31 subunits. The top-ranked model of the N. fulvescens Atg17-Atg31-Atg29 monomer had a pTM of 0.63 (Figure 3J). The AlphaFold3-generated model showed that N. fulvescens Atg17 still maintained a helical fold composed of four α-helices, and was rod-shaped with no substantial curvature. Atg31 formed a β-sandwich and a C-terminal α-helix (239–260). For one of the β-sheets of the sandwich, one β-strand was from the N-terminus of Atg29. Following the β-strand of Atg29 were three α-helices. The observed subunit interactions were consistent with those found in the crystal structure of L. thermotolerans Atg17-Atg31-Atg29, with Atg31 bridging the ternary complex: Atg31 attached to Atg17 through its C-terminal helix which interacted with the Atg17 helical bundle, and Atg31 interacted with Atg29 from the Atg29 N-terminus β-strand folding into the Atg31 β-sandwich.^[18] There were no observed interactions between Atg17 and Atg29 in the predicted model. The predicted N. fulvescens Atg17-Atg31-Atg29 monomer model aligned with the predicted structural model of dimeric N. fulvescens Atg17, suggesting that N. fulvescens Atg17 adopts Atg17_R even in complex with Atg29 and Atg31, and that complex formation is unlikely to induce curvature of the N. fulvescens Atg17 dimer (Figure 3K). However, we were unable to obtain an accurate prediction of N. fulvescens Atg17 in complex with Atg13 and Atg101, as the predicted Atg17 binding site is located in a disordered region of Atg101.

AlphaFold3 model of N. fulvescens and S. pombe Atg13 HORMA domain suggests the potential presence of a stabilizing “cap”

The lack of a stabilizing β-sheet “cap” comprising three β-strands in the S. pombe Atg13 HORMA structure was thought to be a reason for the emergence of Atg101.^[25] Based on this theory, one would expect that the Atg13 HORMA domain of N. fulvescens, which possesses Atg101, would lack this cap. To test this, we carried out AlphaFold3 to model the structure of N. fulvescens Atg13 HORMA (residues 1 to 336) in complex with Atg101. The top-ranked model, with pTM score of 0.75, resembled the crystal structure of the S. pombe Atg13 HORMA-Atg101 complex, with Atg13 HORMA and Atg101 adopting the O-Mad2 and C-Mad2 conformations, respectively (Figure 4A). Furthermore, the model showed that the Atg101 subunit interacted with the β-hairpin and two α-helices of N. fulvescens Atg13 HORMA, similar to what was observed in S. pombe Atg13 HORMA-Atg101. To our surprise, this structural model showed that N. fulvescens Atg13 contained a stabilizing cap similar to that found in the crystal structure and AlphaFold3-predicted model of L. thermotolerans Atg13 HORMA (Figure 4B). The three β-strand cap of N. fulvescens Atg13 HORMA was composed of residues 123–129, 133–137, and 324–328 and the cap was observed to interact with a loop region opposite of the Atg101-interacting region called the “safety belt.”

Out of curiosity, we conducted Alphafold3 runs to predict the structures of S. pombe Atg13 HORMA in a complex with Atg101 and of S. cerevisiae Atg13 HORMA alone. The resulting model of S. pombe Atg13 HORMA-Atg101, which had a pTM score of 0.78, showed that Atg13 HORMA contains a β-sheet “cap,” composed of residues 124–129, 134–136, and 265–269, that was not observed in the previously reported crystal structure (Figure 4C). The pLDDT (predicted local distance difference test) confidence score of this local structure is moderately confident (90 > pLDDT > 70). Upon further investigation, two of the strands in this putative stabilizing cap were not modeled in previous x-ray crystallographical studies due to poorly resolved electron density in this region. Additionally, S. cerevisiae Atg13 HORMA predicted model had a pTM of 0.88 and, as expected, corresponded to the C-Mad2 state and carried a cap (Figure 4D). We observed that AlphaFold3’s ability to predict a cap on the Atg13 HORMA of different yeast species is not dependent on the presence of Atg101 (Figure 4E). Previous studies suggested that the key tryptophan and valine residues located in the second β-strand in the cap region of S. cerevisiae Atg13 HORMA are absent in S. pombe and higher eukaryotes.^[25] However, our expanded MSA showed that the aforementioned tryptophan is conserved in many yeast species containing Atg101. Furthermore, the neighboring valine is also highly conserved, except for N. fulvescens. Interestingly, for N. fulvescens and C. fabianii, the two species that contain all three Atg31, Atg29, and Atg101, their Atg13 HORMA show more sequence divergence in the second β-strand of the cap. And similar to S. pombe, this strand is shorter compared to that of S. cerevisiae.

The identification of a cap in the Atg13 HORMA in S. pombe and other yeast species containing Atg101 raises questions about the proposed role of Atg101. The poorly resolved electron density for the cap in the crystal structure of S. pombe Atg13 HORMA, which is substantiated by the lower confidence score in this region of the Alpha3Fold-predicted model, may indicate that the cap is not associated with stabilization. This might explain why such cap is absent in the Atg13 HORMA from other higher eukaryotes. On the other hand, the presence of a cap with a self-stabilizing function may facilitate the loss of Atg101 and acquisition of Atg31/Atg29 during the evolution of different yeast species, and in particular those in the Saccharomycetes class of budding yeast.

In summary, our investigations generated a more complete picture of the structural differences in Atg17 and tracked the evolutionary trajectory of the Atg31, Atg29, and Atg101 accessory subunits in different yeast species. Our findings also provide the basis to reevaluate previous proposals that the curvature of Atg17 is a required structural feature for its tethering function and that Atg31/Atg29 and Atg101 are mutually exclusive. Additionally, the observation that N. fulvescens and S. pombe Atg13 HORMA domains were predicted to contain a β-sheet cap may represent an emergence/remnant of the Atg1 complex composition divergence, where Atg101 is lost and Atg31/Atg29 is gained. The presence of Atg101 in species with predicted Atg13 β-sheet caps may also support the hypothesis that the presence of Atg101 has to do with functions beyond structural stabilization, such as coordinating the recruitment of downstream autophagic factors. This suggests that the functional divergence of Atg101 and Atg31/29 may reflect a broader shift in the regulatory aspects of the Atg1 complex. Our in silico investigation provides a framework for future investigations of the unique yeast species that carry Atg31, Atg29, and Atg101, verifying the protein interactions with Atg17 and determining their function in conjunction with one another. Finally, our systematic AlphaFold3-based studies showed that this modeling tool could be applied to understand the complexity of protein evolution and interaction networks in the autophagy pathway.

Alphafold3 modeling and bioinformatics

Atg17 (and succeeding annotated Atg31, Atg29, and/or Atg101) sequences from the Saccharomycetes and Schizosaccharomycetes class were retrieved from UniProt and Alphafold3 was used to predict the Atg17 dimers’ structure and architecture type. Atg17 sequences were aligned using Clustal Omega Multiple Sequence Alignment and the resulting phylogenetic tree was visualized using the Interactive Tree Of Life (iTOL) online program.^[32,33] BLAST (basic local alignment search tool) was used to compare protein sequence similarity.^[34]

Molecular cloning of Atg17 constructs

S. cerevisiae and S. pombe Atg17 proteins was constructed as previously described.^[12,21] The coding region of N. fulvescens Atg17 proteins was synthesized by Thermo Fisher Scientific and subcloned into the pET28bHMT vector, a modified version of pET28a vector (Novagen 69,864), using FastDigest NotI (Thermo Scientific, FD0596) and FastDigest NdeI (Thermo Scientific, FD0583) restriction sites.

His-MBP-Atg17 expression and protein purification

Atg17 constructs were expressed in T7 Express Escherichia coli (NEB, C2566I), grown in 2x YT medium at 37°C to an OD₆₀₀ of ~ 0.6, induced with 0.5 mM IPTG (Gold Biotechnology, I2481) and further grown at 16°C for 18 h. Bacterial cells were harvested, flash frozen in liquid nitrogen, and stored at − 70°C until use. Cell pellets were resuspended in lysis buffer (50 mM Tris-HCl pH 8.0, 150 mM NaCl, 2 mM phenylmethylsulfonyl fluoride [PMSF]) and were lysed by sonication using a Branson Sonicator 450 set to 50% duty and 5 output control for 30s, followed by 120 s of cooling, repeated 4 times. The resulting cell lysate was centrifuged using a Beckman JA-25.50 rotor at 31,000 g for 15 min at 4°C. The supernatant was then incubated with HisPur Ni-NTA resin (ThermoScientific 88,222) for 1 h at 4°C while rocking. The resin was washed with lysis buffer supplemented with 10 mM imidazole and bound proteins were eluted with lysis buffer with 250 mM imidazole. The eluate was then incubated with amylose resin (NEB, E8021S) for 1 h at 4°C while rocking, washed with lysis buffer and eluted with lysis buffer containing 10 mM maltose. Protein fractions were pooled and concentrated using an n Amicon 30 kDa concentrator (Millipore, UFC903008). Proteins were loaded onto a Superose 6 (10/300) gel filtration chromatography column (Cytvia 17,517,201) equilibrated with 50 mM Tris-HCl pH 8.0, 150 mM NaCl. Target Atg17 protein dimers were collected in flow-through for SDS-PAGE analysis and to be examined by negative stain electron microscopy.

Negative stain electron microscopy and image processing

Negatively stained specimens were prepared as previously described.^[35] Briefly, each Atg17 protein sample from the peak fraction of the size exclusion chromatography was adsorbed onto a carbon-coated copper grid for 2s, washed with ddH₂O and stained with uranyl formate for 30 s. Micrographs were collected using the Talos L120C Transmission Electron Microscope (ThermoFisher) equipped with a Ceta camera and operating at an accelerating voltage of 120 kV, nominal magnification of 49,000 and a defocus of 2 µm. The micrographs were imported and processed using Relion 3.0.8. Particles were manually selected and extracted with a box size of 256 pixels. Particles were then subjected to multiple rounds of 2D classification to discard poorly defined particles, resulting in representative 2D class averages of the Atg17 dimer architecture.