Strategies for identifying new prions in yeast

Abstract

The unexpected discovery of two prions, [URE3] and [PSI⁺], in Saccharomyces cerevisiae led to questions about how many other proteins could undergo similar prion-based structural conversions. However, [URE3] and [PSI⁺] were discovered by serendipity in genetic screens. Cataloging the full range of prions in yeast or in other organisms will therefore require more systematic search methods. Taking advantage of some of the unique features of prions, various researchers have developed bioinformatic and experimental methods for identifying novel prion proteins. These methods have generated long lists of prion candidates. The systematic testing of some of these prion candidates has led to notable successes; however, even in yeast, where rapid growth rate and ease of genetic manipulation aid in testing for prion activity, such candidate testing is laborious. Development of better methods to winnow the field of prion candidates will greatly aid in the discovery of new prions, both in yeast and in other organisms, and help us to better understand the role of prions in biology.

Introduction

Infectious proteins, absent any nucleic acid component, are known as prions, and include both the causative agents of spongiform encephalopathies in mammals and the prions of yeast.^1,2 Prions typically result from the structural conversion of protein to an amyloid form. Amyloid fibrils are organized, aggregated protein with high β-sheet content, resistance to protease digestion, filamentous morphology and yellow-green birefringence in the presence of Congo red stain.^3,4 Although the various prion and non-prion amyloid-forming proteins have a wide range of amino acid sequences, they share many common features, including similar overall amyloid structure. Therefore, model systems have been useful for studying the general properties of amyloids and prions. This review will focus on work in the yeast Saccharomyces cerevisiae, which is highly tractable for identifying and characterizing prions. Over the past few years, the list of yeast prions has rapidly grown (Table 1), providing a diverse data set for exploring the causes and consequences of prion formation and propagation. We overview the methods used historically for identification of prions in yeast and discuss ways in which this work may be extended.

Table 1.

Yeast prion proteins

Protein	Prion Form	Initial identificationa	Prion confirmationb	Function
Ure2	[URE3]	Genetic screen for defects in nitrogen regulation6	Wickner, 19941	Regulator of nitrogen metabolism
Sup35	[PSI+]	Genetic screen for nonsense suppressors5	Wickner, 19941	Translation termination factor
Rnq1	[PIN+]/[RNQ+]	Examination of factors affecting [PSI+]23Similarity to Ure2/Sup35 12	Sondheimer and Lindquist, 200012; Derkatch et al. 2001.17	Unknown
Swi1	[SWI+]	Bioinformatic screens for Q/N-richness,28,30 Able to substitute for [PIN+]17	Du et al. 200825	Transcriptional regulator
Cyc8	[OCT+]	Bioinformatic screens for Q/N-richness,28,30 Able to substitute for [PIN+]17	Patel et al. 200926	Transcriptional co-repressor
Mot3	[MOT3+]	Bioinformatic screens for Q/N-richness,28,30 Aggregation activity18	Alberti et al. 200918	Transcription factor
Sfp1	[ISP+]	Screen for antisuppressors of sup35 mutants56	Rogoza et al. 201029	Transcription factor
New1	[NU+]c	Similarity to Ure2/Sup35, 10 Able to substitute for [PIN+]17	Not applicablec	Required for ribosomal biogenesis

^aThat is, either the identification of the prion phenotype or identification of the protein as a prion candidate. For [URE3], [PSI⁺], [PIN⁺] and [ISP⁺], the discovery of the prion phenotypes led to subsequent prion characterization, although for [PIN⁺], independent bioinformatics identification of Rnq1 as a prion candidate played an important role. For Cyc8, Mot3, New1 and Swi1, the indicated studies identified the proteins as candidate prion proteins, leading to rigorous studies to characterize the prion phenotype.

^bEither when an observed phenotype was shown to be prion-based, or when a candidate prion protein was shown to form prions.

^cWhile the PFD of New1 can substitute for the Sup35 PFD in supporting prion formation, full-length New1 has not been verified to form prions.

The first observations of yeast prions were serendipitous and initially thought to merely represent either DNA elements with unusual genetic inheritance or viruses. [PSI⁺] was identified in yeast when a genetic screen for nonsense suppression uncovered a cytoplasmic suppressor with non-Mendelian inheritance.⁵ [URE3], another yeast prion, was discovered when a nitrogen uptake screen revealed strains with defects in regulation of nitrogen uptake in which the phenotype showed non-Mendelian heritability.⁶

More than two decades later, [PSI⁺] and [URE3] were proposed to be the prion forms of the S. cerevisiae Sup35 and Ure2 proteins, respectively, based on three criteria.¹ First, in both cases overproduction of the prion protein increases the frequency of prion formation—presumably because increasing the concentration of the prion protein increases the probability of the misfolding event that initiates prion formation. Second, although both [PSI⁺] and [URE3] can be cured by low concentrations of guanidine HCl, the prions can arise again de novo in the cured strain because the prion protein is still present in the cells. This is in contrast to viruses, which should only reappear if reintroduced from outside. Because guanidine treatment cures almost all yeast prions, it has become a critical tool for distinguishing prions from other genetic elements. Third, the URE2 and SUP35 genes are required for [URE3] and [PSI⁺] formation, respectively, and the prion phenotype mimics that of mutation of the genes.

The discovery of two yeast prions led to the search for additional prions in S. cerevisiae. However, [URE3] and [PSI⁺] were both discovered in screens for specific phenotypes, not in specific searches for prions. Therefore, identifying the full range of yeast prions will clearly require more systematic search methods. Taking advantage of some of the unique characteristics of prion proteins, various labs have developed creative methods of identifying new prions. Here we will examine these unique characteristics, and discuss the successes and limitations of the resulting methods.

The Sup35 and Ure2 Prion Domains are Glutamine/Asparagine-Rich

The early bioinformatics approach to the identification of new prions relied on the obvious common features of Sup35 and Ure2. Ure2 and Sup35 each contain an N-terminal prion-forming domain (PFD) that is dispensable for the normal function of the protein and a C-terminal functional domain (Fig. 1).^7,8 In each case, the PFD is highly enriched for glutamine (Q) and asparagine (N) and lacking in hydrophobic and charged residues. A random mutagenesis study of Sup35 found that most mutations that block addition to wild-type [PSI⁺] prion aggregates involved substitutions of charged residues for Q or N.⁹ Additionally, analysis of chimeric proteins demonstrated the importance of a short, highly Q/N-rich segment within the Sup35 PFD for prion aggregate growth.¹⁰

Figure 1.

Schematic of yeast prion protein domain structures for Sup35, Ure2, New1 and Rnq1. Hatching indicates the reported prion-forming domains (PFDs) for each protein. Glutamine and asparagine content of each PFD is indicated (%).

Therefore, two early bioinformatics studies looked for proteins with similar sequence features. Santoso et al. identified an uncharacterized ORF (YPL226W, NEW1) with high Q/N content and few charged residues.¹⁰ Sondheimer and Lindquist searched yeast ORFs with BLAST¹¹ using Ure2/Sup35 query sequences.¹² Three genes were identified as possible prion encoding genes in the screen: RNQ1, YBR016W and HRP1.

Yeast Prion Domains are Modular and Transferable

The identification of potential PFDs necessitated methods for testing prion activity. For both [PSI⁺] and [URE3], prion formation is a very rare event (occurring in approximately one cell per million), and the only significant phenotype is a loss of function of the prion protein. Both [PSI⁺] and [URE3] were discovered in the context of genetic screens where, unusually, a loss of function of the protein conferred a selective advantage to the organism.^5,6

For most proteins, such a rare loss of function event would be difficult to detect. However, the PFDs of Sup35 and Ure2 can retain prion activity when fused to other proteins.^13,14 This modular nature of prion proteins has provided a useful system for testing prospective PFDs; by inserting prospective PFDs in the place of the Sup35 PFD, the same genetic assays used to detect [PSI⁺] formation can be used to screen for prion-forming ability. This method was first used to test prospective PFDs from New1 and Rnq1.^10,12 The PFDs of both proteins were able to form prions when fused in place of the PFD in Sup35^10,12 and to aggregate when fused to GFP.^12,15,16 Rnq1 was later shown to be the protein responsible for the [PIN⁺] prion.¹⁷ However, full length New1 has not been shown to have prion activity. It is unclear whether this reflects an inability of the New1 PFD to support prion formation and/or propagation in its native context, or whether such prions have just not yet been recognized.

This method has since been used in a more systematic study of Q/N-rich domains in yeast. Alberti et al. identified the 100 yeast domains with greatest compositional similarity to known PFDs (their algorithm is discussed in detail below), and tested each in four assays for prion-like activity.¹⁸ One of these assays involved testing fusions to the Sup35 functional domain to see if the domains could support prion formation. Twenty-three proteins were identified that showed prion-like activity in this assay, providing a substantial list of potential prion candidates. However, two known prion proteins, Cyc8 and Mot3 failed to show prion activity when fused to Sup35. This highlights a key limitation of such assays—prion activity is somewhat context-dependent. Even for Sup35 and Ure2, prion formation is substantially affected by the presence of, or mutations in, the proper catalytic domains for each protein.^7,19,20 Because of this, some domains that show prion activity when fused to Sup35 may not show prion activity in their native contexts, and vice versa.

Similar experiments could also be used in library screening experiments. For example, in a recent study, the yeast proteome-GFP fusion protein strain collection was screened for proteins that assembled into visible cellular structures under light microscopy.²¹ While the screen found some interesting cytoplasmic filaments that bear some resemblance to prion filaments, the identified prion candidates were determined unlikely to be prion proteins based on several other characteristics.²¹

Yeast Prions Form Homomultimers and Interact with Other Prions

Another method used to identify prion proteins in yeast evolved from study of the interactions between unrelated prion proteins. Different yeast prion proteins have been shown to interact, both positively and negatively influencing prion formation and propagation.²² The most studied of these interactions involves the prion [PIN⁺], which is necessary for [PSI⁺] formation.^23,24 In an attempt to identify the protein responsible for [PIN⁺], Derkatch et al. performed a screen for proteins that, when overexpressed, could allow for [PSI⁺] formation in the absence of [PIN⁺].¹⁷ They identified 11 candidates: New1, Ure2, Lsm4, Swi1, Nup116, Yck1, Pin2/Yor104w, Cyc8, Pin3/Ypr154w, Pin4/Ybl051c and Ste18. Interestingly, none of these candidates turned out to be the [PIN⁺] prion protein; using a candidate gene approach, [PIN⁺] was shown to be the prion form of the previously identified prion protein Rnq1.¹⁷ However, two of these 11 candidates—Swi1²⁵ and Cyc8²⁶—were later verified to form prions. In both cases, researchers subsequently developed assays that would allow for detection of prion-based loss-of-function phenotypes, and then used genetic criteria, similar to those used for Sup35 and Ure2, to demonstrate prion formation.

Despite these successes, this method has its limitations for identification of new prions. The exact mechanism by which proteins can substitute for [PIN⁺] is not known. The simplest explanation is a direct cross-seeding mechanism. However, indirect mechanisms such as titration of an inhibitor of prion formation are also possible. Even if the mechanism is direct cross-seeding, this assay will still have limitations. Presumably, the ability to cross-seed [PSI⁺] formation requires two activities: the protein must form aggregates, and these aggregates must be able to seed [PSI⁺] formation. However, many proteins that aggregate are unable to propagate as prions, so success in this assay does not guarantee that a protein can act as a prion. Additionally, because the proteins need to interact with Sup35, some prions will likely be excluded. This may introduce some biases into the screen, such as selection for proteins with compositional similarities to Sup35; notably, all 11 proteins identified in this screen were Q/N-rich. Even with these caveats, this strategy may prove helpful in future genetic screens. For example, [URE3] formation was recently shown to be strongly promoted by overexpression of certain Q/N-rich domains,²⁷ suggesting a similar strategy could be employed using Ure2.

Yeast Prions are Defined by Similar Amino Acid Composition

The early bioinformatics successes in identifying Rnq1 and New1 led to more systematic screening. A larger screen for hypothetical yeast prion proteins using high Q/N content as an identifier yielded a long list of potential candidates.²⁸ An algorithm was developed that used a sliding-window approach to scan for the most Q/N-rich region for every protein with at least 30 Q/N residues per 80 residue window. In S. cerevisiae, 107 Q/N-rich ORFs were found. Most of these proteins did not have any significant similarity beyond high Q/N content. Previously identified prion proteins noted in this study included Sup35, Ure2, New1 and Rnq1. Several proteins identified on this basis were later shown to form prions: Swi1,²⁵ Cyc8,²⁶ Mot3,¹⁸ and Sfp1.²⁹ Several others were subsequently shown to have prion-like activity in some assays.¹⁸ In addition, 8 out of the 11 proteins that were able to substitute for the [PIN⁺] prion in promoting [PSI⁺] formation were identified in the screen.¹⁷ However, the vast majority of the candidates have not yet been demonstrated to form prions.

A similar bioinformatics approach was used by Harrison and Gerstein to identify additional yeast prion candidates.³⁰ In finding lowest probability subsequences with various residue biases (e.g., high Q or N content), 172 yeast prion candidates were identified with bias toward the presence of Q, N or both. Adding in subsidiary biases reduced the number of candidates: (1) additional bias against charged or non-aromatic hydrophobic residues yielded 96 candidates and (2) additional bias in favor of G, Y or S residues yielded 31 candidates. Interestingly, hydrophobic residues were shown in subsequent studies to promote prion formation,³¹ rather than inhibit as suggested by Harrison and Gerstein, making the prion predictions for those 96 candidates less likely to be useful. Although the list of Q/N-rich proteins was provided by the authors, the list of those sequences with subsidiary biases was not. However, among the 172 Q/N-rich domains identified in the study, 101 of them were previously identified as candidate prion proteins by Michelitsch and Weissman (Fig. 2)²⁸ and 9 of 11 were identified by Derkatch et al.¹⁷ This list included each of the known PFDs (Sup35, Ure2, New1, Rnq1), as well as Sfp1, Swi1, Cyc8 and Mot3, each of which were subsequently shown to form prions.^18,25,26,29

Figure 2.

Bioinformatics screens based on Q/N-richness identified a largely overlapping group of prion candidates.¹⁸,²⁸,³⁰ The dashed circle indicates a subset of the Alberti et al. set that passed all four prion assays or have been demonstrated to form prions.

Such bioinformatics methods are clearly effective at identifying potential prion candidates. However, these methods showed no ability to distinguish among these candidates. Therefore, Alberti et al. employed a more systematic approach to define the sequence basis for prion formation.¹⁸ Randomizing the order of the amino acids in the Ure2 and Sup35 PFDs does not block prion formation, demonstrating that prion formation is driven primarily by amino acid composition, not primary sequence.^32,33 Therefore, Alberti et al. used compositional similarity to Sup35, Rnq1, Ure2 and New1 in the context of a hidden Markov model to identify the top 100 proteins with at least 60 amino acids of compositional similarity to existing PFDs.¹⁸ Many of these had been identified in previous bioinformatics screens (Fig. 2). They employed four separate assays to look for prion activity among this large predicted data set: semi-denaturing gel electrophoresis (SDD-AGE) to monitor the formation of SDS-resistant aggregates in vivo; an in vitro aggregation assay; fusion of the potential PFD to GFP to monitor formation of fluorescent foci; and substitution of the potential PFD for the Sup35 PFD to test for the ability to support [PSI⁺] formation and propagation. Eighteen domains passed all four tests, including New1, and the known prions Ure2, Sup35, Rnq1 and Swi1; most of the rest had been previously predicted in the two previous bioinformatic searches (Asm4, Cbk1, Ksp1, Lsm4, Nrp1, Nsp1, Pub1, Puf2, Rlm1, Ybl081w, Ybr016w and Ypr022c). Several proteins passed three of four tests, including the known prion Cyc8. One such protein, Mot3, was examined further, and shown to act as a prion.¹⁸ Interestingly, analysis of the full data set, and subsequent follow-up studies, suggested that Q and N are not completely interchangeable, highlighting a potential limitation of search algorithms that treat these residues equally.^18,34

While the work of Alberti et al. did identify many potential prion candidates, the algorithm was not effective at distinguishing among these candidates. There was almost no correlation between compositional similarity to existing PFDs and observed prion-like activity.^18,31,35 This might seem surprising if composition is the predominant determinant of prion propensity. However, the ability of this algorithm to rank prion candidates is predicated on the assumption that all compositional deviations from that of the known yeast PFDs will reduce prion propensity. In reality, known PFDs are likely not optimized for maximal prion propensity, so some compositional changes will increase prion propensity. Although it has been suggested that prions may be advantageous to yeast,^36,37 others have noted that the absence of most prions in wild strains likely indicates a negative effect on yeast.³⁸ If yeast prions are deleterious to host cells, there would obviously be selection away from maximal prion propensity. However, even if prions can have beneficial effects, it is still unlikely that prion proteins would be evolved for maximal prion propensity; rates of prion loss and formation would likely be optimized to provide a balance between the prion and non-prion states.³⁹

Therefore, a method that scored amino acid residues based on their prion propensity, not merely similarity to known prions, would be expected to make better predictions about new yeast prions. Recently, a preliminary attempt was made to do exactly that.³¹ A segment from a scrambled version of Sup35 was replaced with a random sequence to generate a library of mutants. By comparing the amino acid composition of the starting library with the subset that maintained the ability to form prion, a prion-propensity score was developed for each amino acid. The prion propensity of a region could then be scored based on the sum of the prion propensity scores for the individual amino acids across the window. Preliminary analysis suggested that the yeast PFDs were characterized by extended disordered regions of modest prion propensity. Therefore, a sliding window approach was used, where each window was scored for predicted prion propensity using the experimentally derived prion propensity values and for order/disorder propensity using FoldIndex.⁴⁰ This method was reasonably effective at predicting the prion propensity for the proteins examined by Alberti et al.^18,31

However, there are some caveats with this study. Although it appears that distinct compositional features promote prion formation versus prion propagation,^41,42 this study did not consider how composition separately affects these two activities. More significantly, this prediction algorithm is likely to be less accurate for more divergent sequences. The prion propensity estimates for each amino acid currently have large confidence windows associated with them. Therefore, the further a protein's composition deviates from that of the training set, the more these errors will compound and the less accurate the algorithm will likely be. This issue would be exacerbated if there were a non-linear relationship between the prevalence of a specific amino acid and prion propensity. For example, the algorithm assigns relatively low prion propensities to Q and N. Within the context of a Q/N-rich PFD, small changes in Q/N content have a small effect;³¹ however, there may be a threshold below which changes in Q/N content exert a larger effect.

Therefore, while in theory this algorithm could be applied to non-Q/N-rich prions (in contrast to the other previously described bioinformatic methods that are designed solely to identify Q/N-rich prions), it is unclear whether it will be accurate for non-Q/N-rich proteins. The Podospora prion, [Het-s], is not Q/N-rich, yet can propagate as a prion in yeast;^43,44 likewise the PrP mammalian prion is not enriched for Q or N, although Q is found in the tandem octarepeats of that protein. Neither PrP nor Het-s is predicted to have high prion propensity with this algorithm.

In addition to the challenges of accurately predicting more divergent sequences, another possible reason for the low predicted prion propensity of PrP and Het-s using this algorithm is that there may be different classes of amyloids, driven by different sequence features, and different algorithms may only be well suited for specific classes. Numerous prediction algorithms have been developed to score aggregation propensity, including TANGO,⁴⁵ Zyggregator,⁴⁶ BETASCAN,⁴⁷ Waltz⁴⁸ and ZipperDB.⁴⁹ However, it is not clear how effective these algorithms are for the yeast prions. The statistical dynamics-based algorithm TANGO did not identify any β-aggregation nuclei in the Sup35 and Ure2 PFDs.⁵⁰ Other methods such as Waltz and ZipperDB identify amyloid-prone segments in some of the yeast prions; however, none of these methods have yet been demonstrated to be able to distinguish between domains with and without prion-like activity. It appears as though there are at least two classes of amyloid proteins—those for which amyloid formation is driven by short, highly amyloidogenic stretches,⁵¹ and those (including the Q/N-rich yeast PFDs) for which amyloid formation is driven by larger disordered segments of modest amyloid propensity.³⁵ Many of the common algorithms seem specific for the first class of amyloid proteins (those driven by short stretches), and are ineffective for the yeast prions. By contrast, the yeast prion prediction method based on experimentally derived prion propensity scores³¹ seems specific for the second class of amyloid proteins.

Other Potential Methods

Other bioinformatics methods show promise, but have not yet been shown to be sufficient for PFD identification. Molecular dynamics simulations of short peptides in solution have provided some insight into yeast nucleation by Q/N-rich proteins,^34,52 but these methods have not yet been applied to full-length PFDs. Another recent bioinformatic analysis of amyloidogenic proteins, searching for the presence of ‘chameleon’ and ‘discordant’ sequences that might predict amyloidogenicity found statistical relationships between these three protein structure phenomena.⁵³ However, it is unclear if this analysis would be useful for identification of new prion proteins.

Conceivably, high-throughput proteomics methods could be applied to the search for new prions as well. Techniques such as two-dimensional gel electrophoresis and mass spectrometry, while potentially technically difficult with aggregated amyloid fibrils, may prove useful. For example, to identify naturally-occurring prions, one could perform proteomic analysis on a yeast strain before and after treatment with guanidine HCl. However, one caveat is that because prions are so rarely formed in yeast cells, the detection of naturally occurring prions at sufficiently high levels for biochemical methods may prove problematic. Additionally, because the presence of a prion can cause widespread proteomic changes (either due to loss of activity of the prion protein⁵⁴ or due to a cellular reaction to the prion), separating signal from noise could be challenging.

Summary

The initial identification of yeast prions by loss-of-function assays required a combination of luck and diligence in setting up appropriate tests to verify the prion hypothesis. Although researchers have attempted to develop more systematic search methods, most current bioinformatics and genetic screening methods can generate long lists of prion candidates, but cannot effectively distinguish among them. Because negative results are rarely published (and difficult to prove), it is unclear how many of these candidates have already been tested and found not to form prions. Developing methods to better hone in on good candidates will clearly accelerate yeast prion discovery, and may aid in the discovery of novel prions in other species. Newer evaluation methods that rely on more sophisticated analysis, such as molecular dynamics, biophysical modeling and composition-based prion propensity, may prove useful for prion discovery. Additionally, accurate identification of new prions will require a better understanding of how factors such as context, expression level and localization affect prion activity.

Ultimately, while many of the techniques developed in yeast could eventually be useful for identifying prion candidates in other systems, such studies will still present challenges. In particular, the difficulty of developing phenotypic assays in other (particularly multicellular) systems will hinder prion identification. One possibility is to continue to take advantage of the yeast system as a testing system, as was done for the CPEB protein from Aplysia.⁵⁵ However, because of differences in cellular environment between organisms, it will still be necessary to demonstrate prion activity in a protein's native context. Therefore, developing detailed biochemical methods or genetic assays (as were used to demonstrate prion function in mammalian PrP and yeast prions) will still be essential for confirming the prion nature of candidate proteins.

Abbreviations

PFD

prion-forming domain

Q

glutamine

N

asparagines