Perfecting precision of predicting prion propensity

Amyloid is a highly structured fibrous aggregate of a single type of protein. It propagates by recruiting the normal form of the protein and converting it into the amyloid conformation as it joins the growing fiber. All proteins appear to have some capacity to form amyloid under denaturing conditions (1), and some amyloid is functional (2), but many proteins that form amyloid under physiological conditions accumulate as aggregates associated with tissue pathology in many human diseases. The latter have a higher than normal propensity to form amyloid, and the common structural features shared by the amyloid fibers they form had suggested they possess similar amino acid motifs that contribute to this propensity. With an interest in identifying such motifs, various algorithms have been developed that analyze amino acid sequences of proteins to predict their tendency to aggregate or to form amyloid (3–6). A paper in PNAS by Toombs et al. (7), from Eric Ross's group at Colorado State University, describes an algorithm they call the prion aggregation prediction algorithm (PAPA) that significantly improves the precision of predicting the ability of a protein to propagate as a prion in yeast.

Prions are infectious proteins that typically propagate as amyloid. In yeast, infectivity is shown by transmission between strains through cytoplasmic mixing during mating or by direct transfection of amyloid that can be either extracted from prion-containing cells or assembled in vitro from purified protein. The yeast prions [URE3] and [PSI⁺] are composed of the amyloid forms of the nitrogen catabolism repressor Ure2 and the translation termination factor Sup35, respectively (8). As with other yeast prion proteins, Ure2 and Sup35 have intrinsically unstructured glutamine/asparagine (Q/N)-rich regions that drive assembly of the proteins into amyloid. This region on each protein, sometimes referred to as the “prion-forming domain” (PFD), is necessary and sufficient for establishment and propagation of the prion (amyloid) form of the full-length protein in cells. This subregion forms the core of the amyloid, whereas the remaining portion of the protein retains its native fold despite being incorporated into the fibers (9). PFDs of yeast prions are modular and can confer prion behavior when attached to other proteins, even if the other protein does not normally form a prion in yeast (10).

In earlier work as a postdoctoral fellow with Reed Wickner at the National Institutes of Health, Ross and his colleagues (11) generated five versions of the Ure2 PFD by randomly shuffling the amino acids within this sequence without altering the overall constitution of amino acids. Remarkably, all the Ure2 proteins with the randomized PFDs formed amyloid in vitro when purified, and all these rearranged segments functioned as a PFD for [URE3] prions. In a demonstration of the generality of this effect, similar results were obtained with regard to [PSI⁺] prions when the PFD of Sup35 was shuffled (12). These findings imply that the overall composition of amino acids within these regions is more important than the way the amino acids are ordered in sequence, which provided the insight that led to the development of PAPA.

Having established that the composition of amino acids was a more important factor for prion character than linear amino acid sequence, Ross and colleagues (13) went on to develop an in vivo assay for quantitatively testing how amino acid composition influences the ability of a polypeptide to form prions. To identify which portion of the PFD was most important for it to confer prion characteristics, variants of Sup35 containing PFDs with a series of deletions were assessed for loss of prion-forming capacity. After they identified a region particularly sensitive to deletion, libraries encoding Sup35 with various peptides inserted in place of this region were then screened to identify peptides that restored prion-forming ability. The amino acids found most often in the peptides identified in this screen have a tendency to reside in disordered regions and are similar to those found in known yeast prions.

Using this information, Toombs et al. (7) go on to use this in vivo system to evaluate strategies for analyzing regions to identify those with a high propensity to form prions. They confirmed that scanning small stretches of residues would identify areas with a high tendency to form amyloid, but proteins with short, highly amyloidogenic regions do not necessarily form prions readily. They also found that scanning longer regions was more effective at identifying proteins with prion-forming ability. Ultimately, they concluded that the regions that work best as PFDs, including known PFDs, are large, with a strong tendency toward intrinsic disorder and modest propensity to form amyloid, rather than short with highly amyloid-prone stretches of amino acids. Although the peptides that function as PFDs form amyloid readily when purified, they are not optimized for maximal amyloid-forming ability. Values obtained from these experiments were combined with the disorder-predicting algorithm FoldIndex (14) to generate PAPA, which is based entirely on amino acid composition.

An important point clarified by this work is that Q/N richness or the compositional similarity of a Q/N-rich region to known PFDs alone is not a good predictor of the propensity of such a domain to confer ability to form a prion. Earlier algorithms ranking proteins as possible prions on the basis of high Q/N content did well at identifying potential candidates (15, 16), but none of them were adept at clearly distinguishing which among the pool of proteins they identified were able to form prions. Additionally, roughly half of the proteins identified by a rather successful algorithm that ranked potential prion proteins on the basis of shared similarity with known yeast PFDs did not propagate as stable prions (17). By taking into account the more clearly defined elements that confer prion behavior, PAPA has improved the precision for predicting the propensity of a protein to form prions to better than 90%.

As solid confirmation of the power that PAPA has at identifying polypeptides with prion-forming ability, Toombs et al. (7) use PAPA to design two completely synthetic polypeptides predicted to be functional as PFDs for Sup35 (Fig. 1). These peptides were designed by PAPA to be

Fig. 1.

PFD sequences of Ure2, Sup35, shuffled Sup35-27, and PAPA-designed synthetic prion and control (nonprion forming) peptides used to replace the Sup35 PFD. Asparagine and glutamine residues are shown in red and green, respectively.

Toombs et al. use PAPA to design two completely synthetic polypeptides predicted to be functional as PFDs for Sup35.

entirely disordered, to possess the same numbers of Q and N residues as Sup35, and to have a high probability of forming prions. When used in place of the native Sup35 PFD, both synthetic PFDs allowed Sup35 to form stable prions. In contrast, three polypeptides of similar length designed by PAPA to possess only the first two criteria failed to provide PFD function. In line with the observation that compositional similarity to known PFDs is not a critical parameter, both synthetic PFDs differed from WT Sup35 in that they contained charged or hydrophobic amino acids that are not found in the WT Sup35 PFD.

Why is composition more important than linear sequence? Prion propagation requires near-identity in sequence between donor and recipient, implying a specific interaction between amino acid side chains in the amyloid fiber that enforces this requirement. When Ross et al. (18) found that shuffled prion domains could still form prions, they inferred that the prion domains must be in-register, parallel, β-sheet structures, because among possible β-sheet types, only in this one are identical amino acid side chains interacting and in a shuffled sequence, the same interactions could occur, only in a different order. This inference was confirmed when solid-state NMR studies showed that amyloids of both the original prion domains and shuffled prion domains have an in-register, parallel, β-sheet architecture (19, 20).