Compositional Determinants of Prion Formation in Yeast

Numerous prions (infectious proteins) have been identified in yeast that result from the conversion of soluble proteins into β-sheet-rich amyloid-like protein aggregates. Yeast prion formation is driven primarily by amino acid composition. However, yeast prion domains are generally lacking in the bulky hydrophobic residues most strongly associated with amyloid formation and are instead enriched in glutamines and asparagines. Glutamine/asparagine-rich domains are thought to be involved in both disease-related and beneficial amyloid formation. These domains are overrepresented in eukaryotic genomes, but predictive methods have not yet been developed to efficiently distinguish between prion and nonprion glutamine/asparagine-rich domains. We have developed a novel in vivo assay to quantitatively assess how composition affects prion formation. Using our results, we have defined the compositional features that promote prion formation, allowing us to accurately distinguish between glutamine/asparagine-rich domains that can form prion-like aggregates and those that cannot. Additionally, our results explain why traditional amyloid prediction algorithms fail to accurately predict amyloid formation by the glutamine/asparagine-rich yeast prion domains.Amyloid fibers are associated with a large number of neurodegenerative diseases and systemic amyloidoses. Amyloid fibrils are rich in a cross-beta quaternary structure in which β-strands are perpendicular to the long axis of the fibril (8).URE3] and PSI⁺] are the prion (infectious protein) forms of the Saccharomyces cerevisiae proteins Ure2 and Sup35, respectively (61). Formation of both prions involves conversion of the native proteins into an infectious, amyloid form. Ure2 and Sup35 have served as powerful model systems for examining the basis for amyloid formation and propagation. Both proteins possess a well-ordered functional domain responsible for the normal function of the protein, while a functionally and structurally separate glutamine/asparagine (Q/N)-rich intrinsically disordered domain is necessary and sufficient for prion aggregation and propagation (4, 26, 27, 52, 53). Both proteins can form multiple prion variants, which are distinguished by the efficiency of prion propagation and by the precise structure of the amyloid core (14, 54).Five other prion proteins have also been identified in yeast: Rnq1 (13, 46), Swi1 (15), Cyc8 (33), Mca1 (30), and Mot3 (1). Numerous other proteins, including New1, contain domains that show prion activity when inserted in place of the Sup35 prion-forming domain (PFD) (1, 42). Each of these prion proteins contains a Q/N-rich PFD. Similar Q/N-rich domains are overrepresented in eukaryotic genomes (28), raising the intriguing possibility that prion-like structural conversions by Q/N-rich domains may be common in other eukaryotes. However, we currently have little ability to predict whether a given Q/N-rich domain can form prions.A variety of algorithms have been developed to predict a peptide''s propensity to form amyloid fibrils based on its amino acid sequence, including BETASCAN (6), TANGO (17), Zyggregator (51), SALSA (62), and PASTA (55). These algorithms have been successful at identifying regions prone to amyloid aggregation and predicting the effects of mutations on aggregation propensity for many amyloid-forming proteins. However, they have generally been quite ineffective for Q/N-rich amyloid domains such as the yeast PFDs. For example, using the statistical mechanics-based algorithm TANGO (17), which predicts aggregation propensity based on a peptide''s physicochemical properties, Linding et al. found that the Sup35 and Ure2 PFDs both completely lack predicted β-aggregation nuclei (24). Similarly, yeast PFDs are generally lacking in the hydrophobic residues predicted by algorithms such as Zyggregator to nucleate amyloid formation.Why are these algorithms so effective for many amyloid-forming proteins but not for yeast PFDs? For most amyloid proteins, amyloid formation is driven by short hydrophobic protein stretches, and increased hydrophobicity is correlated with an increased amyloid aggregation propensity (34). In contrast, the yeast PFDs are all highly polar domains, due largely to the high concentration of Q/N residues and the lack of hydrophobic residues. High Q/N content is clearly not a requirement for a domain to act as a prion in yeast, since neither the mammalian prion protein PrP nor the Podospora anserina prion protein HET-s is Q/N rich, yet fragments from both proteins can act as prions in yeast (49, 50). However, the significant compositional differences between the yeast PFDs and most other amyloid/prion proteins suggest that there may be two distinct classes of amyloid-forming proteins driven by different types of interactions. Specifically, Q/N residues, which are predicted to have a relatively low amyloid propensity in the context of hydrophobic amyloid domains (34), may promote amyloid formation when present at sufficiently high density. Stacking of Q/N residues to form polar zippers has been proposed to stabilize amyloid fibrils (35). Consistent with this hypothesis, mutational studies of Sup35 indicate that Q/N residues are critical for driving PSI⁺] formation (12), and expanded poly-Q or poly-N tracts are sufficient to drive amyloid aggregation (36, 63). Therefore, this paper examines the sequence features that allow the polar, Q/N-rich yeast PFDs to form prions.Mutational studies of the PFDs of Ure2 and Sup35 have shown that amino acid composition is the predominant feature driving prion formation (40, 41). Due to the unique compositional biases observed in the yeast PFDs, algorithms have been developed to identify potential PFDs based solely on amino acid composition (19, 28, 42). These algorithms are designed to produce a list of potential prion proteins that meet a specific set of criteria (such as high Q/N content) but are not able to predict the prion propensity of each member of the list or to predict the effects of mutations on prion formation. A recent study by Alberti et al. was the first to systematically test whether compositional similarity to known PFDs is sufficient to distinguish between Q/N-rich proteins that form prions and those that do not. They developed a hidden Markov model to identify domains that are compositionally similar to known PFDs and then analyzed the 100 highest-scoring Q/N-rich domains in a series of in vivo and in vitro assays (1). Remarkably, they discovered 18 proteins with prion-like activity in all assays. However, an equal number, including some of the domains with greatest compositional similarity to known PFDs, showed no prion-like activity.This inability to distinguish between Q/N-rich proteins that form prions and those that do not might seem to suggest that amino acid composition is not an accurate predictor of prion propensity. However, an alternative explanation is that known yeast PFDs are not an ideal training set for a composition-based prediction algorithm, since yeast prions are likely not optimized for maximal prion propensity. It is unclear whether yeast prion formation is a beneficial phenomenon providing a mechanism to regulate protein activity or a detrimental phenomenon analogous to human amyloid disease. PSI⁺] can increase resistance to certain stress conditions (56), but the failure to observe PSI⁺] in wild yeast strains (29) argues that beneficial PSI⁺] formation is at most a rare event. If yeast prions are diseases, the PFDs certainly would not be optimized for maximum prion potential. If prion formation is a beneficial event allowing for rapid conversion between active and inactive states, the prion potential of the PFD would be optimized such that the frequencies of prion formation and loss would yield the optimal balance of prion and nonprion cells (25). Thus, specific residues might be excluded from yeast PFDs either because they inhibit prion formation or because they too strongly promote prion formation; bioinformatic analysis can reveal which residues are excluded from yeast PFDs but not why they are excluded. Accurate prediction of prion propensity requires understanding which deviations from known prion-forming compositions will promote prion formation and which will inhibit.We have therefore developed the first in vivo method to quantitatively determine the prion propensity for each amino acid in the context of a Q/N-rich PFD. As expected, we found proline and charged residues to be strongly inhibitory to prion formation; but surprisingly, despite being largely underrepresented in yeast PFDs, hydrophobic residues strongly promoted prion formation. Furthermore, although Q/N residues dominate yeast PFDs, prion propensity appears relatively insensitive to the exact number of Q/N residues. Using these data, we were able to distinguish with approximately 90% accuracy between Q/N-rich domains that can form prion-like aggregates and those that cannot. These experiments provide the first detailed insight into the compositional requirements for yeast prion formation and illuminate the different methods by which Q/N- and non-Q/N-rich amyloidogenic proteins aggregate.