The ESCRT (endosomal sorting complex required for transport) machinery is known to sort ubiquitinated transmembrane proteins into vesicles that bud into the lumen of multivesicular bodies (MVBs). Although the ESCRTs themselves are ubiquitinated they are excluded from the intraluminal vesicles and recycle back to the cytoplasm for further rounds of sorting. To obtain insights into the rules that distinguish ESCRT machinery from cargo we analyzed the trafficking of artificial ESCRT-like protein fusions. These studies showed that lowering ESCRT-binding affinity converts a protein from behaving like ESCRT machinery into cargo of the MVB pathway, highlighting the close relationship between machinery and the cargoes they sort. Furthermore, our findings give insights into the targeting of soluble proteins into the MVB pathway and show that binding to any of the ESCRTs can mediate ubiquitin-independent MVB sorting.
The multivesicular body (MVB) pathway is part of the endosomal system of eukaryotic cells. It is responsible for the delivery of transmembrane proteins into the lumen of lysosomes/vacuoles where these proteins are degraded (reviewed in ). As such, the MVB pathway represents the major degradation pathway for transmembrane proteins in eukaryotes. In addition, the MVB pathway traffics lysosomal/vacuolar hydrolases and thus plays an important role in maintaining the function of lysosomal compartments. With a few exceptions, the sorting of cargo into the MVB pathway has been shown to be dependent on ubiquitination (reviewed in ). On the endosome, ubiquitinated transmembrane cargoes are recognized by a group of protein complexes called ESCRTs (endosomal sorting complex required for transport), which sort these cargo proteins into the intraluminal vesicles (ILVs) of the MVB (reviewed in ).
The ESCRT machinery consists of five protein complexes: ESCRT-0, ESCRT-I, ESCRT-II, ESCRT-III and the Vps4 complex. These protein complexes are recruited from the cytoplasm and assemble on the MVB into a protein network that is involved both in ILV formation as well as sorting of cargo into the ILVs. Deletion analyses suggested a linear order in both assembly and function of the ESCRT machinery: ESCRT-0 is acting upstream of ESCRT-I, which then activates ESCRT-II, which subsequently recruits ESCRT-III and the Vps4 complex (Figure 1A, reviewed in ). Several ubiquitin-binding sites have been identified in ESCRT-0, ESCRT-I and ESCRT-II, indicating that these ‘early’ ESCRTs are the main cargo-binding complexes of the MVB sorting machinery . ESCRT-III subunits oligomerize on membranes into spiral-shaped polymers that are able to deform the underlying membrane into tubular structures [5-9]. Therefore, ESCRT-III has been suggested to function in the invagination of the membrane and the abscission reaction that severs the neck between the forming ILV and the limiting membrane .
Even though the ESCRT machinery is directly involved in both the formation of ILVs as well as the sorting of membrane proteins into these vesicles, the ESCRT proteins remain cytoplasmic and are not packed into the lumen of ILVs. This observation is surprising considering that the ESCRTs localize to endosomes in part by binding to phosphatidylinositol 3-phosphate (PI3P), a lipid that enters ILVs [11-15]. Furthermore, studies have shown that ESCRTs are ubiquitinated [16, 17], a modification that is thought to cause efficient sorting into ILVs. These findings pose the question: What differentiates cargo from machinery? In fact, several studies have indicated that the ESCRTs are, at least to some extent, treated similarly to cargo proteins. These studies showed the regulated degradation of ESCRT-I via the MVB pathway . Furthermore, analysis of exosomes, extracellular vesicles that originate from MVBs, identified ESCRT proteins as constituents of these vesicular carriers . Finally, the deletion of the gene encoding for the ESCRT-I subunit Mvb12 resulted in the delivery of mutant ESCRT-I to the lumen of the vacuole, suggesting that this mutation caused ESCRT-I to behave more like cargo than machinery . Together, these observations confirm the close relationship between the ESCRTs and the cargo they sort.
The three-dimensional structure of most of the ESCRT machinery has been elucidated. The interactions between the ESCRTs are also well studied. However, the spatial and temporal organization of the ESCRT network that allows for sorting of cargo into ILVs without entrapment of the machinery remains enigmatic. Therefore, the goal of this study was to obtain insights into the rules that govern ESCRT interactions and cargo sorting. We analyzed the behavior of artificial protein fusions designed to mimic ESCRTs. The results from these experiments revealed a surprisingly flexible and simple ESCRT system that differentiates between cargo and machinery simply by interaction strength.
Like cargo, ESCRTs are ubiquitinated and localize to the endosomal membrane but remain at the limiting membrane of the MVB, and upon completion of MVB formation, recycle back to the cytoplasm for additional rounds of sorting. To identify the parameters for this behavior, we constructed a series of fusion proteins that mimic both ESCRT localization to the endosome and interaction with other ESCRT machinery. These fusion proteins contained well-defined membrane and ESCRT-interaction domains (EIDs), thereby minimizing the possibility of unknown interactions that might skew the results. Because of the complexity of interactions within the ESCRT machinery, we considered the use of ESCRT-like proteins to be more informative than mutagenesis of ESCRT components.
For endosome association, the ESCRT-like constructs contained the C-terminal region of EEA1, a region that consists of a FYVE domain and a dimerization domain (Figure 1B). The FYVE domain, which is also present in the ESCRT-0 subunit Vps27, binds specifically to the head group of PI3P, a lipid enriched in endosomal membranes [13, 14, 21, 22]. Furthermore, the dimerization of the FYVE domain increases the affinity of the fusion protein to endosomes . As expected from previous studies , GFP-FYVE expressed in yeast colocalized with the ESCRT-III-associated factor Ist1-RFP to small structures adjacent to the vacuole, which represent MVBs or late endosomes (Figure 2A). But unlike Ist1-RFP, GFP-FYVE also localized to the vacuolar membrane suggesting that GFP-FYVE was not recycled like ESCRTs but remained associated with the membrane for a prolonged time, even after fusion of MVBs to the vacuole (Figure 2A). However, we are not able to exclude the possibility that some of the GFP-FYVE is directly recruited to the vacuolar membrane. The vacuolar lumen was mostly devoid of GFP, indicating that GFP-FYVE was not a cargo of the MVB pathway and thus did not contain an MVB sorting signal. Together, the microscopy data suggested that GFP-FYVE associated with endosomes but did not interact with the ESCRT machinery, properties that are essential for our studies of the ESCRT mimics.
We fused a set of different EIDs to the N-terminus of GFP-FYVE. These domains belong to ESCRT proteins and are known to interact with other proteins of the ESCRT machinery (Figure 1A,B and Table 1). The resulting ESCRT mimics were expressed in yeast and CPY-invertase sorting assays indicated that none of the fusion proteins exhibited a dominant-negative effect on the MVB pathway, suggesting that the fusion proteins did not interfere with normal ESCRT function (Table 2). However, when overexpressed from a high-copy plasmid the EID-GFP-FYVE proteins caused MVB sorting defects (data not shown).
Table 1. Fusion proteins
Observed behavior (majority of cells)
ESCRT-III (MIM1, MIM2)
ESCRT-0, -I, -II
ESCRT-III (MIM1, MIM2)
Table 2. CPY-invertase secretion assay
% CPY-invertase secretion (normalized values, average of three experiments)
0 ± 0.4
100 ± 4.8
6 ± 0.2
6 ± 0.2
10 ± 0.3
10 ± 0.3
9 ± 0.3
−6 ± 0.4
−3 ± 0.6
−6 ± 0.4
−3 ± 0.5
18 ± 0.9
−2 ± 0.5
−7 ± 0.2
Quantitative analysis of fluorescence microscopy pictures was used to determine the effects of the EIDs on GFP-FYVE localization. The localization of each fusion protein was compared to those of the well-studied proteins GFP-Dap2, GFP-Cps1 and Ist1-GFP. Dap2 is a vacuolar peptidase (also called DPAP B) that upon synthesis traffics via the MVB to the vacuolar membrane without interacting with the ESCRT machinery (referred to as ‘non-interacting’;  ). In contrast, GFP-Cps1, a vacuolar protein that is efficiently sorted via the MVB pathway into the lumen of the vacuole , served as an example of a cargo protein. Finally, the ESCRT protein Ist1-GFP represented the ESCRT machinery. This fusion protein localizes to MVBs by binding to ESCRT-III and upon completion of MVB formation recycles back to cytoplasm . As a consequence, Ist1-GFP is found on late endosomes and in the cytoplasm of the cell. Together, these three proteins represented the three possible outcomes for an MVB-associated protein, namely ‘non-interacting’, ‘cargo’ and ‘machinery’.
For each EID-GFP-FYVE, pictures of approximately 50 cells were taken and the GFP intensity of late endosomes, vacuolar membrane and vacuolar lumen was determined. In the cases where the vacuolar membrane signal was too low to be separable from vacuolar lumen (membrane staining is less or equal to that of the lumen), the value of the vacuolar membrane was set to be equal to that of the lumen. The resulting data were used to calculate for each cell the ratios of ‘vacuolar lumen-to-vacuolar membrane’ and ‘vacuole-to-total signal’ (Figure 2B). An example of these calculations is shown in Figure 2C. Notice that because the vacuolar membrane signal is never lower than that of the lumen, the ‘vacuolar lumen-to-vacuolar membrane’ ratio cannot be above 1. The ratios were then compared to those of GFP-Dap2, GFP-Cps1 and Ist1-GFP. To simplify the presentation of the data, the ‘vacuolar lumen-to-vacuolar membrane’ and ‘vacuole-to-total signal’ ratios of the standard proteins (GFP-Ist1, GFP-Cps1 and GFP-Dap2) were plotted on a scatter plot and areas that contained at least 90% of the cells expressing a particular standard fusion protein were defined (Figure 2B,D). These defined areas were then used to determine to which group (non-interacting, cargo and machinery) a particular ESCRT-mimic-expressing cell belonged (summarized in Figure 1A and Table 1).
Binding to ESCRT-III can mediate recycling or MVB sorting
The first ESCRT mimic we tested was MIT-GFP-FYVE, a fusion protein containing the ESCRT-III-binding domain of Vps4 . MIT-GFP-FYVE localized almost exclusively to late endosomal compartments, which is in contrast to GFP-FYVE that localized to both endosomes and vacuolar membranes (Figure 3A). This difference in localization suggested that MIT-GFP-FYVE was removed from the MVBs before fusion to the vacuole occurred, a behavior that mimicked the recycling observed with ESCRT components. Therefore, MIT-GFP-FYVE behaved similar to proteins of the ESCRT machinery, a result that is reflected in the quantitative analysis, which indicated that >90% of cells belong to the ‘machinery’ group (Figure 3A). This observation also indicated that the MIT–ESCRT-III interaction was strong enough to compete with the FYVE–PI3P interaction, causing MIT-GFP-FYVE together with ESCRT-III subunits to be pulled off of the endosomal membrane in a Vps4-dependent manner.
The MIT domain was mutated to test if recycling of MIT-GFP-FYVE is indeed dependent on the interaction with ESCRT-III. The MIT domain is able to bind simultaneously to two distinct motifs in ESCRT-III, called MIM1 (MIT interaction motif 1) and MIM2 . Mutating isoleucine at position 18 of the MIT domain to an aspartate [MIT(I18D)] has been shown to impair the MIM2 interaction with ESCRT-III . The I18D mutation caused a slight decrease in the recycling of MIT-GFP-FYVE from the endosome, consistent with the idea that the interaction with ESCRT-III is important for the machinery behavior of MIT-GFP-FYVE (Figure 3A). The rather weak phenotype of the I18D mutation is consistent with previous studies that showed a mild defect in MVB sorting in cells expressing vps4(I18D) .
Surprisingly, mutating the MIM1-interaction site of the MIT domain [MIT(L64D)] resulted in a fusion protein that was not recycled from the endosome but efficiently sorted into the MVB pathway, mimicking cargo (Figure 3A). Expressing MIT(L64D)-GFP-FYVE in rsp5-1, a strain mutated for the E3 enzyme responsible for ubiquitination of MVB cargoes , caused only a slight defect in the sorting of the fusion protein into the lumen of the vacuole (Figure 3B), suggesting that MVB sorting of MIT(L64D)-GFP-FYVE was not due to ubiquitination of the fusion protein. This notion was further supported by the finding that expressing Hse1-DUB did not block the sorting of MIT(L64D)-GFP-FYVE into the MVB pathway (Figure 3B). Hse1-DUB is a fusion protein that has been shown to efficiently deubiquitinate endosomal cargoes thereby inhibiting their entry into the MVB pathway . Because a complete lack of ubiquitinated cargo interferes with ESCRT function , a Mup1-ubiquitin fusion was coexpressed with Hse1-DUB in cells used for MVB trafficking experiments. Together, the data indicated that sorting of MIT(L64D)-GFP-FYVE into ILVs is largely independent of ubiquitination but mediated by binding to ESCRT-III. In contrast, proper sorting of a GFP-Dap2 fusion protein containing a ubiquitination site (US-GFP-Dap2) was found to be strongly defective in the rsp5-1 mutant strain or in the presence of Hse1-DUB, a result that is consistent with a defect in ubiquitin-dependent sorting in these strains (Figure 4D). Sorting of MIT(L64D)-GFP-FYVE into ILVs was dependent on a functional ESCRT machinery because deletion of the essential ESCRT factor VPS4 blocked delivery into the vacuolar lumen (Figure 3B).
The idea that MIT(L64D)-GFP-FYVE is sorted into ILVs by binding to ESCRT-III was further tested by mutating both ESCRT-III-interaction sites of the MIT domain. The resulting fusion protein MIT(I18D,L64D)-GFP-FYVE showed in most cells a localization pattern of a ‘non-interacting’ protein, consistent with the idea that this fusion protein lost most of its interactions with ESCRT-III and thus was no longer sorted by the MVB pathway (Figure 3A). However, in approximately 35% of the cells, MIT(I18D,L64D)-GFP-FYVE did sort into the vacuolar lumen, albeit inefficiently, suggesting that this fusion protein maintained some interaction with the ESCRT machinery, either with ESCRT-III or with another unknown binding partner.
The difference in behavior of wild-type and mutant MIT-GFP-FYVE constructs (machinery versus cargo) could not be explained by the different types of ESCRT-III interaction, because mutating either the MIM1- or MIM2-binding site caused reduced recycling of the fusion protein. In contrast, changes in the strength of ESCRT-III interaction were able to explain the different behaviors of the MIT-GFP-FYVE fusions. Previous studies suggested that the ESCRT-III affinity of the different MIT domains follows the order: MIT>MIT(I18D)>MIT(L64D) > MIT(I18D,L64D) . Our data indicated that lower ESCRT-III affinities result in decreasing efficiency of MIT-GFP-FYVE recycling and an increasing chance that the fusion protein is sorted into ILVs. On the other hand, if the ESCRT-III-binding affinity of the MIT domain drops too low, the sorting efficiency drops and the fusion protein will behave more like GFP-FYVE, a non-interacting protein. This effect could explain why MIT(L64D)-GFP-FYVE was found to be a better cargo than MIT(I18D,L64D)-GFP-FYVE. In summary, the data supported the notion that, depending on the interaction strength, binding to ESCRT-III resulted either in MVB sorting (weak interaction) or recycling from the endosome (strong interaction).
We analyzed the localization of other ESCRT-III-interacting GFP-FYVE constructs. The domains tested were the N-terminal region of Bro1 (Bro1 domain), the C-terminal catalytic domain of the deubiquitinating enzyme Doa4 (localizes via Bro1 to ESCRT-III) and the N-terminal region of Did2 [31, 32].
Both Bro1(N)-GFP-FYVE and Doa4(C)-GFP-FYVE were efficiently sorted into the lumen of the vacuole and thus mimicked cargo proteins (Figure 3C). When expressed in strains containing either a mutant RSP5 allele (rsp5-1) or the deubiquitinating Hse1-DUB protein the GFP signal of both Bro1(N)-GFP-FYVE and Doa4(C)-GFP-FYVE localized mainly to the lumen of the vacuole, indicating that MVB sorting of these proteins was ubiquitin-independent (Figure 3D). In contrast, the delivery to the vacuolar lumen was dependent on functioning ESCRT machinery (vps4Δ, Figure 3D). Mutating the Bro1-binding site in Doa4(C)-GFP-FYVE  blocked delivery of the fusion protein into the vacuolar lumen [Doa4(C,AAFA)-GFP-FYVE, Figure 3C] further demonstrating that MVB sorting of Doa4(C)-GFP-FYVE is independent of ubiquitination but requires binding to ESCRT-III via Bro1.
The localization of Did2(N)-GFP-FYVE mimicked ESCRT machinery (Figure 3C), suggesting that this ESCRT-III-binding protein was removed from the MVB prior to fusion with the vacuole. One explanation for the efficient recycling of the fusion protein could be a strong interaction of the Did2(N) domain to ESCRT-III, thus mimicking the behavior of MIT-GFP-FYVE.
Together, the data of GFP-FYVE fusions indicated that strong binding to ESCRT-III could indeed result in a protein mimicking ESCRT machinery. The recycling of these fusion proteins from the MVB was most likely mediated by the Vps4-dependent disassembly of ESCRT-III. Not expected was the observation that weaker interactions with ESCRT-III would cause sorting of the fusion protein into ILVs. This result was particularly surprising considering that ESCRT-III has been postulated to function late in the formation of ILVs, mainly acting together with Vps4 in the membrane fission reaction that releases the forming vesicles into the lumen of the MVB (Figure 1A).
Binding to the early ESCRT machinery mediates MVB sorting
We tested two ESCRT mimics that were designed to bind to the early ESCRT machinery (Figure 1A). The first fusion protein contained the PSDP domain, the region of Vps27 that has been shown to interact with ESCRT-I . The second construct contained the C-terminal region of Vps28, the region of ESCRT-I that is known to bind to ESCRT-II . Both of these constructs, Vps27(PSDP)-GFP-FYVE and Vps28(C)-GFP-FYVE, did not recycle to the cytoplasm but instead sorted into the MVB pathway. The sorting efficiency was rather low, suggesting that these fusion proteins behaved as poor MVB cargo (Figure 4A). Mutating RSP5 or expressing HSE1-DUB had weak or no effects on the sorting of these fusion proteins, indicating that they entered the MVB pathway in a ubiquitin-independent manner (Figure 4B). However, Vps4, and thus ESCRT activity, was necessary for the delivery of the fusion protein to the vacuolar lumen (Figure 4B).
To compare the sorting efficiency of Vps27(PSDP)-GFP-FYVE and Vps28(C)-GFP-FYVE with that of a corresponding ubiquitinated cargo, a fusion protein of ubiquitin and GFP-FYVE was constructed. Surprisingly, when expressed in yeast, Ub-GFP-FYVE localized to endosomes and the vacuolar membrane, indicating that this fusion protein was not recognized by the ESCRT machinery as ubiquitinated cargo but, in contrast, mimicked a non-interacting protein (data not shown). One explanation for this result might be that the ubiquitin domain of Ub-GFP-FYVE was inaccessible to the ESCRTs. Therefore, a fusion protein was constructed that contained a second ubiquitin domain at the membrane-proximal C-terminus of the fusion protein. The resulting 2Ub-GFP-FYVE fusion was partially sorted into the lumen of the vacuole, mimicking the sorting of poor MVB cargo (Figure 4A). The poor sorting of 2Ub-GFP-FYVE, Vps27(PSDP)-GFP-FYVE and Vps28(C)-GFP-FYVE was not the result of proteolytic clipping of the EID in the cytosol because western blot analysis showed that the majority of the fusion proteins present in a vps4Δ strain were full length (the strain contained additional deletions of vacuolar peptidases to minimize proteolysis during sample preparation, Figure 4C). Together, the data suggested that the ESCRT-I- and ESCRT-II-binding domains used [Vps27(PSDP) and Vps28(C)] acted as MVB sorting signals that were as efficient as ubiquitin itself.
MVB sorting is independent of the type of membrane association
To determine if the type of membrane association plays a role in the ubiquitin-independent sorting of cargo into the MVB pathway, we fused the MIT and Vps28(C) domains to the N-terminus of the type II transmembrane protein GFP-Dap2 (Figure 1B and Table 1). We observed that both the ESCRT-III-binding domain (MIT) and the ESCRT-II-binding domain (Vps28(C)) caused sorting of the corresponding fusion proteins into the lumen of the vacuole (Figure 4D). Unlike MIT-GFP-FYVE, which is recycled together with the ESCRT machinery, MIT-GFP-Dap2 is a transmembrane protein that cannot be removed from the MVB and thus is efficiently packaged into ILVs. The MVB sorting of these fusion proteins was mainly independent of ubiquitination, because mutating RSP5 (rsp5-1 strain) or expressing HSE1-DUB caused only minor drop in the localization to the vacuolar lumen (Figure 4D). The MVB sorting of MIT-GFP-Dap2 was more efficient than sorting of Vps28(C)-GFP-Dap2, a fusion protein that showed partial vacuolar membrane localization (Figure 4D). The poor sorting of Vps28(C)-GFP-Dap2 was not caused by proteolytic clipping of the fusion proteins because western blot analysis showed that the majority of the fusion proteins present in a vps4Δ strain were full length (Figure 4E). This observation was consistent with the data obtained from the GFP-FYVE fusions, demonstrating more efficient MVB sorting of ESCRT-III-binding proteins compared to early ESCRT-binding proteins. Together, these experiments indicated that, independent of the type of membrane association, binding to ESCRT-II or ESCRT-III could serve as a ubiquitin-independent sorting signal for the MVB pathway. However, previous studies have shown that the fusion protein MIT-GFP does not localize to the vacuole , indicating that ESCRT binding alone is not sufficient for MVB sorting. Therefore, we conclude that membrane association is necessary in order for a protein to be sorted into ILVs.
In silico analysis of the MVB sorting model
Previous models of ESCRT function suggested that the ESCRTs assemble into a temporally and spatially highly ordered structure in which cargo is sorted by the early ESCRTs (ESCRT-0, -I and -II) and membrane deformation and membrane fission is executed by ESCRT-III and Vps4 function (Figure 1A). In contrast, our data suggested that all the ESCRTs are able to mediate cargo sorting and we found that a simple change in ESCRT affinity converted a protein that behaved like an ESCRT component to a protein that was efficiently sorted into ILVs. The ESCRT mimics contained EIDs that compete with endogenous ESCRTs for binding to the ESCRT network. Even though the mimics were expressed at higher levels than the endogenous ESCRTs (based on GFP signals) they did not interfere with normal ESCRT function (Table 2). Similarly, the fusion proteins that behaved like cargoes were binding to sites of the ESCRT machinery that normally are not interacting with cargo.
To explain our data we propose a model in which cargoes trigger the assembly of ESCRTs on the endosomal membrane into a protein network with great flexibility. Although each of the ESCRT complexes has a defined quarternary structure with defined interaction sites, the position of the ESCRTs within the network is not fixed and each particular position can be occupied by either machinery or cargo. A simplified view of this model would be a system of just two different components, cargo and ESCRT, that polymerize together into a two-dimensional (2D) lattice. The question remains if such a simple system would be able to specifically sort cargo and leave non-cargo (or non-interacting proteins) behind. To answer this question we turned to computer modeling. For the in silico model we used a 2D hexagonal grid to represent the membrane surface on which membrane-associated proteins were able to randomly move and interact. The sorting simulation contained three types of elements: (i) ESCRTs containing one cargo-binding site and five ESCRT–ESCRT-interaction sites, (ii) cargoes and (iii) non-cargoes (not to be sorted by the system). The starting condition was a random distribution of specified proportions of cargoes and non-cargoes. Binding and movement of proteins were randomly determined based on probabilities (described in more detail in Materials and Methods). The goal of the simulation was to test the hypotheses that the protein interactions were sufficient to achieve high ESCRT-dependent sorting efficiency and fidelity, namely that all cargoes should be sorted into vesicles whereas all non-cargoes should be excluded.
A simplified example of the cargo sorting simulation is shown in Figure 5. The basic rules for the computer model are as follows. (i) ESCRTs are recruited to the membrane from the cytoplasm by binding to cargo (appear next to a cargo on the grid), diffuse on the grid together with the bound cargo, can bind to other ESCRTs or can break the interaction with cargo that releases the ESCRT into the cytoplasm. We assume that ESCRTs are not stable on the membrane without interacting with other proteins and therefore are removed from the membrane if unbound (disappear from the grid). (ii) All proteins and protein complexes can diffuse on the membrane and form new interactions with each other (ESCRT can bind one cargo and up to five ESCRTs). (iii) ESCRT–cargo and ESCRT–ESCRT interactions can break. However, a network of more than two proteins (ESCRT or cargo) is stable and cannot break. (iv) Networks and their bound cargoes and entrapped non-cargoes are removed from the grid (simulating vesicle formation) with probability depending on the size of the network.
We found that stabilization of the ESCRTs by multiple interactions (ESCRT–cargo or ESCRT–ESCRT interactions, rules i and iii) was important to form larger networks and to prevent ESCRTs alone to form networks. This rule is consistent with recent data demonstrating the necessity of ubiquitinated cargo for ESCRT function .
Running the computer simulation starting with different conditions found that sorting efficiency and fidelity was very high and not dependent on the concentration of cargo or non-cargo (Figure 6A,B and Movie S1, Supporting Information). Only rarely non-cargo became entrapped by the ESCRT–cargo network and was removed during the vesicle formation step of the simulation. Furthermore, including a component with a single ESCRT-interaction site to the starting conditions, a component that mimicked the behavior of a EID-GFP-FYVE fusion construct, did not disrupt the formation of networks and resulted in efficient sorting of both cargo and the GFP-FYVE mimic (Figure 6A). Furthermore, changing the number of ESCRT–ESCRT-interaction sites of the ESCRT components from 5 to 4 or 3 did not diminish sorting efficiency and fidelity (Figure 6A). Together, the simulations showed that the proposed cargo-sorting system is robust and highly adaptable to various conditions.
Our study provided key insight into the rules that govern ESCRT recycling and MVB cargo sorting. The key conclusions are provided below:
Cargo sorting can be mediated by binding to any ESCRT complex
Our data suggested that binding to ESCRT-I, ESCRT-II or ESCRT-III leads to sorting into the MVB pathway. This observation is unexpected because previous models, such as the epistasis model shown in Figure 1A, suggested that cargo sorting is mediated mainly by the ‘early ESCRTs’ (ESCRT-0, ESCRT-I and ESCRT-II), whereas ESCRT-III and Vps4 function only in ILV formation.
ESCRT-III and Vps4 function concurrently with cargo sorting
ESCRT-III and Vps4 have been predicted to act solely as membrane fission machinery, severing the membrane neck after cargo sorting and membrane deformation has been completed. In contrast, our data indicated that ESCRT-III can participate in cargo sorting, suggesting that ESCRT-III assembles before a narrow neck inhibits diffusion of cargo in and out of the forming ILV. Furthermore, the fate of the MIT-GFP-FYVE depended on its interaction strength with ESCRT-III, suggesting that Vps4-mediated disassembly of ESCRT-III is concurrent with cargo sorting.
Cargo ubiquitination is not required for sorting into the MVB pathway
Every EID we tested was able to function as a MVB sorting signal, suggesting that ubiquitin is not a unique sorting signal that is essential for entry into the ILV. Rather, our data suggested that ubiquitin functions as temporary EID that can be added and removed from proteins in order to regulate their interaction with the ESCRT machinery. Recent studies showed that aquaporin AQP2 and the G protein-coupled receptor PAR1 are sorted into the MVB pathway independent of ubiquitination [36, 37]. In both cases, MVB sorting was mediated by direct interactions with ESCRT factors that are predicted to function together with ESCRT-III and Vps4 at later stages of ILV formation (LIP5 and ALIX). These studies are consistent with our findings that direct interaction with ESCRT-III or ESCRT-III-associated proteins can result in efficient delivery to the vacuolar lumen.
Soluble proteins can be sorted into the MVB pathway
A few examples of soluble MVB cargo have been described in the literature. However, it is not understood how these proteins are delivered into forming ILVs whereas all the other cytoplasmic proteins are excluded. Published examples include MVB-dependent degradation of soluble signaling proteins [38, 39] and the packaging of soluble proteins into exosomal vesicles , an ESCRT-mediated process similar to MVB cargo sorting. We conclude from our studies that soluble proteins can be sorted into the MVB pathway but membrane association is essential for this process. Similar to transmembrane proteins, binding to any ESCRT component seems to serve as a sorting signal for soluble proteins. However, the interaction strength between soluble cargo and ESCRTs must be weak; otherwise, the cargo protein will be removed together with the ESCRTs from the MVB and recycled back to the cytoplasm.
ESCRTs and their cargoes are alike
ESCRTs are endosome-associated proteins that interact with other ESCRTs and are modified by ubiquitination . Why then are ESCRTs not efficient cargoes? This question was answered by the analysis of MIT-GFP-FYVE. This ESCRT-III-interacting fusion protein cycled between MVBs and cytoplasm and thus resembled the localization of an ESCRT protein. However, mutating one of the ESCRT-III-binding sites of the MIT domain converted the fusion protein into an efficient MVB cargo [MIT(L64D)-GFP-FYVE, Figure 3]. This result suggested that a weak interaction with the ESCRT machinery resulted in efficient sorting into the MVB pathway, whereas a strong affinity to the ESCRTs was essential to promote recycling to the cytoplasm. Therefore, the key event that determines the fate of an MVB-associated protein seems to be the competition between the interaction of the protein with the membrane and ESCRT machinery network: if the membrane interaction wins, the protein becomes a cargo, if it loses, the protein is recycled and behaves like an ESCRT protein. Because transmembrane proteins are thoroughly anchored in the membrane, the membrane interaction always wins this competition and thus even strongly ESCRT-interacting proteins are efficiently sorted into the MVB pathway (e.g. MIT-GFP-Dap2, Figure 4D).
The protein network formed by early ESCRTs shows a high degree of variability and built-in redundancy
Two of the fusion proteins we constructed are sorted into the MVB pathway by binding to the ESCRT-0 – ESCRT-I and the ESCRT-I – ESCRT-II interactions sites. Surprisingly, the presence of these fusion proteins did not interfere with ESCRT function (Table 2), even though these proteins were competing with ESCRTs for the same interaction sites and were expressed at higher levels than the ESCRTs. For example, Vps28(C)-GFP-FYVE is predicted to compete with ESCRT-I for the interaction with ESCRT-II; however, no dominant-negative effect was observed. Furthermore, Vps28(C)-GFP-FYVE is sorted like a cargo even though it occupies the space in the ESCRT network that is normally reserved for ESCRT-I. Together, these observations support a model in which ESCRT-0, ESCRT-I and ESCRT-II assemble into a network that can adapt to various cargoes by changing the local structure. Furthermore, this system shows a high degree of redundancy; disruption of some of the ESCRT–ESCRT interactions does not impair the cargo sorting mechanism. Consistent with this model, we found no particular entry point for cargo into the ESCRT system. Even interaction with ESCRT-III, a complex thought to mainly act in the final membrane fission step of MVB vesicle formation, mediated efficient cargo sorting. Therefore, we found no evidence for a strict temporal order within the ESCRT system (as depicted in Figure 1A) that would limit the cargo sorting function to only the early ESCRTs. Similarly, no single ubiquitin-binding site of the ESCRT machinery was found to be essential for cargo sorting .
Together our data can be explained with a simple oligomerization model in which the ESCRTs together with membrane-associated ESCRT-binding proteins coassemble on the endosomal membrane (Figure 7). For the purpose of cargo sorting, the identity of the ESCRT complex binding to the cargo and the type of interaction (ubiquitin-dependent or not) seem to be irrelevant. The resulting protein network defines the area on the endosomal membrane that will form an ILV. Non-interacting proteins (or non-cargoes) are excluded from this area simply by steric hindrance. In silico studies showed that such a simple system is able to efficiently sort cargoes (Figure 6). The final structure and composition of the protein network is not defined but adapts to the size and structure of the components. However, it is likely that the ESCRT system has an overall organization. This order within the ESCRT system might be provided by ESCRT-III, a spiral-forming polymer that could act as the ‘backbone’ of the network, giving the ESCRT system directionality. After the oligomerization reaction, the vesicle formation phase is initiated by the disassembly of the protein network by Vps4 and the deubiquitinating enzyme Doa4. This stage decides which components of the network will behave like ESCRT machinery and which components will become cargoes of the MVB pathway. ESCRT components are recycled, whereas cargoes remain associated with a membrane region that invaginates by a lipid-driven mechanism and forms the ILV (discussed in ). In a final step, ESCRT-III and Vps4 are likely to drive the membrane fission reaction that completes the ILV formation.
Materials and Methods
Strains, media and plasmids
The strains and plasmids used in this study are described in Table 3. Yeast gene knockouts were constructed as previously described . Plasmids were constructed using conventional restriction enzymes or by Quikchange mutagenesis protocol using Phusion polymerase enzyme. pSK11 and pSD2 were constructed by inserting the previously described GFP-Dap2 and GFP-Cps1 constructs, respectively [25, 45], into pRS415 vector. Similar vector swap was performed for the previously published GFP-FYVE construct  to obtain pSK43. pSK122 was constructed by fusing the ubiquitination sequence from Cps1 to the N-terminus of GFP-Dap2. Yeast strains were grown in rich YPD (yeast extract–peptone–dextrose) medium or in the appropriate synthetic drop out media (YNB) as published . In every set of microscopy experiments, strains were grown in the same growth medium. A total of 100 μM CuSO4 was added to the medium to drive the expression of CUP1 promoter-containing constructs.
Yeast cell extracts for western blotting were obtained from strains grown to logarithmic phase. Cells were pelleted, resuspended in SDS-PAGE sample buffer (2% SDS, 0.1 M Tris, pH 6.8, 10% glycerol, 0.01% bromophenol blue and 5% β-mercaptoethanol) and lysed using glass beads. The anti-GFP antibody used in western blotting was purchased from Roche Diagnostics Corporation. The anti-Snf7 antibody used was previously described .
Fluorescence microscopy and quantification
Yeast strains were grown for 10–14 h in YNB to optical densities of 0.7 (at 600 nm); 100 μM CuSO4 was added 3.5 h before microscopy analysis to cells containing CUP1 promoter-driven constructs. Fluorescence microscopy was performed on a deconvolution microscope (DeltaVision; Applied Precision). Fifty cells of each strain were used to determine point values for the vacuolar limiting membrane, the vacuolar lumen and the endosome. Because of its small size the endosomal values were defined as the brightest pixel of the endosomal structure. The value for the vacuolar lumen was determined by calculating the average intensity of 150 pixels within the vacuole. Similarly, the value for the vacuolar membrane was determined by averaging the intensities of 50 pixels representing the limiting membrane of the vacuole. The background value was measured as the average of 150 pixels outside of and in close proximity to each cell. The background value was subtracted from each of the above measurements resulting in the values for endosome, vacuolar lumen and vacuolar membrane (see Figure 1C). In cells with efficient MVB cargo sorting, the vacuolar membrane was not clearly visible. In such cases, vacuolar membrane was given the same value as vacuolar lumen. In cells with multiple vacuoles, the most prominent vacuole was used for quantification. Cutoff values were defined to exclude cells with very high (causing sorting defects) or very low (low signal-to-noise ratio) fluorescence intensities from the analyses.
The liquid invertase assay was performed as previously described . Strains were grown in YNB medium to a density of 0.6–0.9 OD600; 0.5 ODV of each strain was washed with and resuspended in 1 mL of 0.1 M sodium acetate (pH 4.9). A 50 μL aliquot of each yeast suspension was incubated with 12.5 μL of 1 M sucrose at 30°C for 30 min. Simultaneously, glucose standard solutions containing 0, 0.4, 0.8, 1.2 and 1.6 mM glucose were made in 50 μL NaAc and incubated with 12.5 μL of 1 M sucrose. Reactions were quenched by adding 75 μL of 0.2 M dipotassium phosphate solution and placed on ice; 500 μL of Glucostat reagent [2.5 µg/mL horseradish peroxidase, 2 U/mL glucose oxidase, 100 μM N-ethylmaleimide and 614 μM o-dianisidine in 0.1 M dipotassium phosphate buffer (pH 7)] was added to each of the aliquots and incubated at 30°C for 30 min. Reactions were stopped with 500 μL of 6 M hydrochloric acid and OD540 was measured for each sample. All samples were done in triplicates. The averaged value for each strain was normalized with the negative control (BHY10 strain containing empty vector) at 0% and positive control (BHY10 strain containing vps4E233Q) at 100%.
MATLAB was used to run all computer simulations. A cellular automata model was employed on a 2D hexagonal grid. The grid size used in all simulations is 28 by 28. The simulation involves randomly selecting grid points and updating the grid according to a given set of rules for assigning probabilities of updates according to the current state of the grid. Inputs to the simulation include: (i) initial proportions of cargo, non-cargo and GFP-FYVE construct in the grid, (ii) number of ESCRT–ESCRT-binding sites and (iii) parameters specifying propensities for diffusion, ESCRT–ESCRT binding, ESCRT–cargo binding, ESCRT–GFP-FYVE binding, ESCRT removal, unbinding and critical vesicle size.
Every protein in the grid is associated with a network that includes all the proteins it is connected to through bonds. An unbound protein is simply considered a network of one protein. As proteins bind together, larger networks are formed. Updates may affect the entire network of the grid point selected and not just the individual grid point selected to update. Potential updates are determined by the following rules for each type of update.
Recruitment of ESCRTs
ESCRTs will be recruited to an empty grid point if there is at least one cargo in an adjacent grid point. The recruited ESCRT is bound to an adjacent cargo.
Removal of ESCRTs
Unbound ESCRTs may be removed from the grid. Upon removal, the current grid point becomes empty. Because of our choice of parameters, it is almost certain that an unbound ESCRT will be removed from the grid.
The hexagonal grid allows for up to six bindings per protein. However, cargo and GFP-FYVE are limited to one binding. ESCRTs have one cargo-binding site and up to five ESCRT-binding sites. The number of allowed ESCRT-binding sites is specified as an input. If GFP-FYVE is present in the simulation, it may occupy one of the ESCRT-binding sites of an ESCRT. Cargo may bind to one ESCRT. GFP-FYVE may bind to one ESCRT. ESCRTs may bind to one cargo, one GFP-FYVE (if present) and up to five other ESCRTS.
Cargo may only unbind if its network is an ESCRT–cargo pair. If a cargo is part of a network of three or more proteins, the cargo is prohibited from unbinding. An ESCRT may unbind from if it has only a single bond with another ESCRT and no other interactions. GFP-FYVE is prohibited from unbinding once it is bound to an ESCRT.
Networks of any size are allowed to move in any of six directions (up left, down right, up right, down left, left and right) if all the adjacent grid spaces to the network in that direction are empty. If any adjacent space is occupied by a protein, including adjacent spaces in the interior of the network, movement of the network is prohibited.
Any protein bound to only one other protein is allowed to rotate clockwise if the adjacent grid space in the clockwise direction is empty. Similarly, a protein bound to only one other protein is allowed to rotate in the counter-clockwise direction as well.
Removal of networks and trapped non-cargoes
As networks grow they may be removed from the system mimicking vesicle formation. An input to the simulation is critical vesicle size, which is defined as a sufficient number of proteins included in a network in order to form a vesicle. We calculate the value V(N) = 1/(1 + exp(−(N − C))), where N is the size of the current network defined by the number of proteins in the network and C is the critical vesicle size. As a function of N, V is a sigmoid function centered at C. If the computed value of V is greater than a randomly generated number, the entire network is removed from the grid and every grid point occupied by the network becomes empty. As a network grows it is more likely to be removed from the grid as V is an increasing function. A network rarely grows much larger than the critical vesicle size owing to the nature of V(N). In addition, small networks are almost never removed from the grid. It is possible that non-cargo may be trapped in a network that is selected to be removed from the grid. A non-cargo is considered trapped if (i) it is in an interior row of the current network and (ii) in each interior row of the network, it is in an interior column of that row. If a non-cargo is trapped when a network is removed, it is also removed with the network and the grid point occupied by that non-cargo becomes empty. This is the only way that non-cargo can be removed from the grid.
Determination of probabilities
The probability of each allowable update for a selected non-empty grid point is determined by the propensity for that update scaled by the sum of propensities for all allowable updates. Therefore, the probabilities of all possible updates sum to 1. Further, if only one update is allowable, that update will be selected with probability 1. If an empty grid point is selected, an ESCRT will be recruited with probability 1 if at least one cargo occupies an adjacent grid point. Otherwise, the empty grid space remains empty.
To begin each simulation, a random grid is generated based on the initial proportions of cargo, non-cargo and GFP-FYVE specified. Following the initial grid setup, the simulation repeats as follows until a specified number of runs are completed. Updates to the grid are selected based on the calculated probabilities.
Randomly select a grid point.
Determine the occupancy type of all the grid points adjacent to the current grid point.
Is the current grid point empty?
Yes: Move to step 4.
No: Move to step 5.
Recruit an ESCRT?
Yes: Add an ESCRT to the current grid point, bind the ESCRT to an adjacent cargo and begin again with step 1.
No: Begin again with step 1.
Remove network of current grid point?
Yes: Remove network (along with any trapped non-cargo) and begin again with step 1.
No: Move to step 6.
Determine what updates are possible for the current grid point or the network of the current grid point.
Is there at least one possible update?
Yes: Randomly select one of the possible updates and update the grid. Begin again with step 1.
No: Begin again with step 1.
While initial proportions and number of ESCRT–ESCRT bindings varied for the results shown, the propensity parameters used for all simulations are as follows:
Diffusion (for all proteins): 1,
ESCRT–ESCRT binding: 10,
ESCRT–cargo binding: 10,
ESCRT–GFP-FYVE binding: 10,
ESCRT removal: 1000,
Unbinding: 1 and
Critical vesicle size: 20.
We set the propensity parameters for binding to be equal, independent of protein type. However, as the updates are determined by the calculated probabilities, an ESCRT (with five ESCRT–ESCRT-interaction sites) is usually more likely to associate with another ESCRT than a cargo because it has five ESCRT-interaction sites and only one cargo-interaction site. In this type of probability-driven model, binding affinities are not explicitly specified; rather, the qualitative behavior of the protein interactions was accounted for.
For each set of initial proportions of proteins 20 simulations were completed. The amounts of each type of protein (cargo, non-cargo, ESCRT and GFP-FYVE) were recorded at 100 intervals of 2500 steps of each simulation. The data from each simulation were normalized by dividing by the total number of grid spaces (in this case 28^2) giving a grid concentration of each protein type. The average and standard deviation of the 20 simulations were calculated. Movie frames were captured every 250 runs of a simulation and include 700 frames. Furthermore, still images of the same simulation were recorded.
We thank Rob Piper for sharing plasmids and for helpful discussions. This work has been supported by grants NIH R01 GM074171 (to M. B.) and NSF-DMS 1122297 (to J. P. K.). The authors declare that they have no conflict of interest.