Liquid-liquid phase separation and its mechanistic role in pathological protein aggregation

Abstract

Liquid-liquid phase separation (LLPS) of proteins underlies the formation of membrane-less organelles. While it has been recognized for some time that these organelles are of key importance for normal cellular functions, a growing number of recent observations indicates that LLPS may also play a role in disease. In particular, numerous proteins that form toxic aggregates in neurodegenerative diseases, such as amyotrophic lateral sclerosis, frontotemporal lobar degeneration, and Alzheimer’s disease, were found to be highly prone to phase separation, suggesting that there might be a strong link between LLPS and the pathogenic process in these disorders. This review aims to assess the molecular basis of this link through exploration of the intermolecular interactions that underlie LLPS and aggregation and the underlying mechanisms facilitating maturation of liquid droplets into more stable assemblies, including so-called labile fibrils, hydrogels, and pathological amyloids. Recent insights into the structural basis of labile fibrils and potential mechanisms by which these relatively unstable structures could transition into more stable pathogenic amyloids are also discussed. Finally, this review explores how the environment of liquid droplets could modulate protein aggregation by altering kinetics of protein self-association, affecting folding of protein monomers, or changing aggregation pathways.

Introduction

Membrane-less organelles (MLOs)¹ are intracellular assemblies of macromolecules that play host to a number of fundamental processes in the cell, including biochemical reactions as well as protein and RNA transport [1,2]. MLOs associate through liquid-liquid phase separation (LLPS), a metastable de-mixing of proteins and/or RNA mediated by transient, multivalent interactions [2–7]. The diversity of interactions that promote LLPS enables the formation of specific, tunable MLOs, each with separate functions, as exemplified by the nucleolus, stress granules, processing bodies, and many others [8,9].

While there are a number of MLO-associated proteins that undergo LLPS in the test tube, there is also significant overlap among proteins that undergo LLPS and those that aggregate in neurodegenerative diseases [10–12]. Many of these proteins contain intrinsically disordered regions (IDRs) that engender LLPS, but come at the cost of being highly aggregation-prone. This overlap suggests that the two phenomena—LLPS and aggregation—might be connected in disease pathogenesis [10–13]. Proposed models depict a progression from functional, phase-separated but highly dynamic liquid droplets to less dynamic or static assemblies. However, in some cases, such a progression has also been proposed as a cellular protective mechanism [14,15]. Consequently, defining the significance of and structural differences between reversible assemblies and irreversible aggregates has been a topic of great interest.

This review highlights recent advances in understanding the role of protein LLPS in pathological protein aggregation, especially in the context of neurodegenerative diseases. Because such a relationship has been most extensively demonstrated for RNA-binding proteins, key insights are drawn from studies of fused in sarcoma (FUS), transactive response DNA-binding protein of 43 kDa (TDP-43), and the heterogeneous ribonucleoproteins (hnRNPA1, hnRNPA2), but are also contextualized to studies of other proteins that do or do not aggregate in the neurodegenerative disease. We also critically discuss possible mechanisms by which LLPS might influence the protein aggregation landscape and how a progression from reversible liquid droplets to irreversible aggregates might occur.

Membrane-less organelles and neurodegenerative disease

Originally described for the segregation of germ cell P granules in C. elegans [16], biomolecular condensation (i.e., MLO formation) driven by liquid-liquid phase separation is now regarded as an important regulator of numerous biological processes [5,9,17,18]. Ribonucleoprotein (RNP) granules, such as RNA granules and stress granules (SG), are MLOs composed of networks of RNAs and RNA-binding proteins. Additionally comprised of ribosomal machinery and transport proteins, RNA granules are used for distal translocation of translational machinery to neuronal dendrites and axons [19]. Stress granules, on the other hand, act to minimize the negative impact of stressors on the cell, in part through localized translational regulation [20]. While the exact composition varies with stressor, SGs are largely composed of stalled preinitiation complexes, including additional components such as untranslated mRNAs, initiation factors, and ribosomal subunits [20,21]. Many of the RNA-binding proteins (RBPs) involved in RNP granules readily undergo LLPS in vitro, providing a model for studying the physicochemical basis for the assembly and maintenance of these MLOs in the cell. The proteins best characterized in this regard include hnRNPA1, hnRNPA2, FUS, and TDP-43 [22– 29].

Importantly, many of these LLPS-prone proteins have been found to form brain inclusions associated with a host of neurodegenerative diseases. For example, TDP-43 – that normally participates in pre-mRNA splicing, mRNA transport, and translation regulation [30] – is a major component of insoluble inclusions in amyotrophic lateral sclerosis (ALS), frontotemporal lobar degeneration with ubiquitination (FTLD-U), Alzheimer’s disease (AD), cerebral age-related TDP-43 with sclerosis, and chronic traumatic encephalopathy [31–35]. The RNA-binding protein FUS, which plays a role in transcriptional regulation, pre-mRNA processing, and DNA damage response [30], was found to be mutated and form inclusions in ALS and is the major component of ubiquitin-positive, filamentous inclusions in atypical FTLD-U [36–38]. Furthermore, dysregulation of hnRNPA1 and hnRNPA2, which normally play roles in mRNA translation, splicing, and trafficking [39], were observed to aggregate in diseases such as multisystem proteinopathy, ALS, inclusion body myositis, and others [39–41].

These aggregation-prone RBPs typically localize to a dense core within the SG [42]. When recapitulated in cell assays, SGs form rapidly in the presence of stress (i.e., sodium arsenite, heat shock), and fully dissociate soon after removal of the stressor [20]. However, any perturbation of this system could have a severe impact on the cell’s ability to dissociate SGs, in particular aggregation-prone RBPs within the dense core. As a result, persistent and elevated tendencies for SG assembly (and lack of disassembly) have been implicated in disease etiology [20,42]. For instance, pathogenic mutations can increase the aggregation propensity of RBPs such that SGs are less prone to dissociate after removal of the stressor [23,24,26]. Additionally, altered RNA-binding interactions may change dynamics of the granule itself, which can cause formation of new static and irreversible granules [43,44]. Finally, even in the absence of disease-associated mutations, repetitive head trauma correlates with aggregation of RBPs (and mimic diseases such as chronic traumatic encephalopathy), suggesting that cells can only withstand so much stress [45]. Importantly, many of these instances reflect altered interactions occurring at the molecular level. Therefore, an understanding of the specific types of intermolecular interactions involved in LLPS and aggregation is crucial to gaining insight into the relationship between the two phenomena.

Formation and evolution of liquid droplets

Weak, transient interactions mediate LLPS of IDRs

Proteins that undergo phase separation can be categorized as having repetitive modular domains or IDRs with weakly adhesive motifs, or a combination of both. Repetitive modular domains within proteins enable complex network formation with high valency of interaction, as exemplified by SH3-motif-containing protein systems that regulate actin polymerization [46–48]. The properties of IDRs that enable LLPS have been extensively reviewed [48,49]. Many contain repetitive amino acid motifs with weak adhesive properties, such as G/S-Y/F-G/S elements in hnRNP A1 [24], FUS [48], and TDP-43 [50]. While these motifs are common to FUS, they are few in TDP-43. Such a diversity of IDRs (with varying functional implications) has resulted in additional nomenclature for these regions, such as prion-like domains (PrLD, which refers to the sequence commonality to yeast prion proteins) [51], low complexity domains (LCD, which refers to the general lack of amino acid diversity) [52], amyloidogenic cores (which refers to the aggregation-prone regions that form amyloid fibrils in vitro), among others.

The enrichment of certain amino acid types in an IDR promotes specific intermolecular interactions that facilitate phase separation, though the same amino acid can be involved in multiple types of interactions [6]. One of the more common types of interactions facilitating LLPS is the electrostatic interaction. This is observed in highly charged proteins, such as tau, in which case LLPS is driven by attractive intermolecular interactions between positively and negatively charged protein regions [53]. Electrostatic interactions are also of key importance for complex coacervation of RNA and positively charged proteins that mediates RNP granule formation [54,55]. Charged residues like arginine or lysine can also make cation-π contacts with aromatic residues, as is seen in arginine-tyrosine interactions in FUS that facilitate LLPS [7,56]. Additionally, Ddx4 LLPS reportedly involves both π-π and cation-π interactions, also from arginine residues (among others) [57,58]. Depending on the amino acid composition of IDRs, LLPS can also be facilitated through dipolar and hydrophobic interactions [11,59]. Though limited evidence is currently available, interactions with cross-β character (a structural feature of amyloid aggregates, as discussed in detail in subsequent sections) has been proposed as a driving force for LLPS [25,60]. The nature of these varied interactions can be most directly studied in systems reconstituted in vitro by changing ionic strength [22,55], pH [61], or addition of hydrophobic agents such as 1,6-hexanediol [23,50]. Though specific amino acids and motifs are important, LLPS-promoting IDRs appear (to a certain extent) interchangeable, as exchanging IDRs or inserting artificial IDRs can grant phase separation capacity [24,62,63].

Coarsening and maturation of liquid droplets

Liquid-liquid phase separation is an inherently metastable state [64]. Accordingly, liquid-like droplets in vitro undergo a number of changes beginning immediately after preparation. While the protein concentration within the protein-dense phase remains constant for a given condition, droplets do begin to grow in size. This growth results in part from Ostwald ripening: larger droplets are inherently more stable than smaller ones, resulting in net movement of protein out of dissolving smaller droplets into those that are larger [64,65]. This coarsening is distinct from growth that results from fusion events between droplets. Nonetheless, both contribute to the appearance of much larger droplets with time.

Of key interest to a number of studies involving aggregation-prone proteins that also phase separate is the observation that droplets undergo maturation [24]. Here, we refer to maturation, or aging, specifically as the time-dependent coarsening and loss of dynamicity of a previously liquid-like droplet that does not necessarily imply loss of reversibility. This is most commonly assessed using fluorescence recovery after photobleaching (FRAP), which can measure the diffusion of protein molecules within an individual droplet (Figure 1). Importantly, the material state of a droplet formed by a given protein can be modulated by a number of factors, including the presence, type and concentration of RNA [66,67] or other polyanions [55], crowding agents [68], and divalent cations [69]. Consequently, a temporal comparison is key to assessing droplet maturation.

Figure 1. Maturation of liquid droplets.

Changes in the material properties of liquid droplets can be monitored using fluorescence recovery after photobleaching, which typically involves pre-bleach, bleach, and post-bleach (or recovery) stages. Initiation of liquid droplet formation classically results in highly dynamic assemblies that exhibit rapid and almost complete recovery after photobleaching (though these properties may vary between proteins and with the addition of co-factors). Removal of the LLPS-inducing conditions (i.e., change in temperature, addition or dilution of salt) results in complete dissolution of droplets, indicating high reversibility. Freshly-prepared droplets grow in size and undergo fusion events over time. The term maturation describes a decrease in the extent of recovery after photobleaching with time until recovery is no longer observed. Such loss of dynamicity has been thought of as a gelation or aggregation phenomenon; however, the exact nature of such species is not always entirely clear. Additionally, this loss of dynamicity often coincides with a loss of reversibility when LLPS-promoting conditions are removed.

Droplet maturation is not unique to aggregation-prone proteins, as recently exemplified by heterochromatin protein 1α condensates that mature over the course of a week [70]. In actuality, a gel-like, arrested state has been considered the final stage of LLPS maturation as a general phenomenon [71]. However, droplets formed by aggregation-prone proteins may mature at a much faster rate. When comparing LLPS of aggregation-prone IDRs, Rosen and colleagues found that the majority mature over the course of 24 hours [24]. Additional studies have revealed that these IDRs likely drive maturation, as the full-length FUS protein loses liquid-like character within 8 hours while the FUS LCD alone loses dynamicity within an hour [27,29,72]. Furthermore, at least some neurodegenerative disease-associated mutations in both TDP-43 and FUS can facilitate a faster maturation [22,27,29].

The exact structural underpinnings of droplet maturation is not yet clear (Figure 1). There is some evidence to suggest that maturation may involve formation of specific types of β-sheet-rich interactions, a hallmark of amyloid fibril formation (discussed in more detail in the following section) [25,56,73]. While the formation of fibrillar aggregates from previously liquid-like droplets has been reported, in many cases loss of dynamicity appears to occur before detection of such fibrils by microscopy or fluorescence-based methods [29,61,74]. Accordingly, there may be intermediate maturation steps that remain relatively elusive to current studies and caution should be taken when equating droplet maturation with irreversible aggregation.

Stable and labile aggregates

Networks of polymers: hydrogels and amyloids

An important development was the observation by McKnight and colleagues that the low complexity domains of hnRNPA2 and FUS can undergo a soluble-to-hydrogel phase transition [75]. These hydrogels were system-spanning, polymeric networks that formed from soluble protein at low temperatures. The authors hypothesized that this propensity for hydrogel formation stemmed from the ability of LCDs to assemble into β-sheet-rich structures [73,75]. Indeed, these polymers were found to have the same structural hallmark as classical amyloids, the cross-β structural motif.

HNRNPA2 and FUS both assemble into pathogenic fibrillar aggregates in disease and are only two of numerous proteins involved in neurodegeneration that are capable of forming amyloid fibrils [36,40]. These include, among many others, β-amyloid, tau, huntingtin, α-synuclein, prion protein, and TDP-43 (for a comprehensive list, see ref [76]). Amyloids characteristically have three major features: fibrillar morphology, cross-β structure, and tinctorial properties (the ability to be stained by amyloid-specific dyes) (Figure 2A) [76]. Consequently, amyloids can be identified ex vivo and kinetics of formation monitored in vitro using fluorescent dyes, such as Thioflavin-T (ThT) or Thioflavin-S, that exhibit enhanced fluorescence when binding to amyloid structures. Spectroscopic techniques, such as solid-state NMR [77] and EPR [78], have been quite useful in probing the architecture and conformational dynamics of amyloids. More recently, cryo-electron microscopy has emerged as a particularly valuable method for investigating atomic-resolution structures of not only synthetic amyloids (i.e., those formed from recombinant proteins in vitro) but also those derived from patient brains [79]. Although antiparallel [80,81] and solenoid-type [82,83] structures have been described, the vast majority of amyloids studied so far adopt parallel, in-register β-structures, in which unidirectionally oriented β-strands assemble perpendicular to the fibrillar axis and are stabilized by backbone hydrogen bonding [78,79,84]. β-sheets can also stack against each other via interactions between side chains. This stacked sheet conformation and spacing between β-strands produces the canonical X-ray fiber diffraction patterning, which involves 4.7 Å and ~10 Å reflections resulting from inter-strand and inter-sheet distances, respectively [85,86]. The specific amino acid segments that comprise this cross- β structure is commonly referred to as the core of an amyloid.

Figure 2. Characteristic features of classical amyloids.

(A) Pathological amyloids are characterized by three defining features: fibrillar morphology, cross-β architecture, and tinctorial properties (staining by amyloid-specific dyes). Fibrillar morphology can be assessed using atomic force microscopy (shown) or transmission electron microscopy. The cross-β structure is defined by the ~4.7 Å distance between individual β-strands of the same sheet, while the distance between sheets is variable, often ~10 Å. Thioflavin-T (ThT, an amyloid-specific dye) exhibits elevated fluorescence upon binding to fibrils and therefore can be used to monitor assembly. Typically, monitoring the ThT fluorescence intensity over time yields a sigmoidal curve with a nucleation/lag phase and an elongation/exponential phase. (B) The nucleation phase involves formation of an elongation-competent nucleus, which typically is oligomeric but can also be a misfolded monomer. This nucleus has the capacity to initiate templated assembly and therefore fibril growth. Additional interactions between pre-formed fibrils can facilitate gelation, though this is not necessarily the case for all amyloids.

The gelation phenomenon observed for hnRNPA2 and FUS is not unique to these two proteins, as many amyloid-forming proteins and peptides have the capacity to form hydrogels [87,88]. Conversely, the cross-β motif is only one structure among many by which hydrogels can form [89]. Some studies even indicate that, depending on experimental conditions, the same protein may have the capacity to form either amyloid-based or particulate, non-amyloid gels [87]. For classical amyloids, gelation appears to occur only after initial assembly of individual fibrils and such a transition from a Newtonian-like fluid (containing amyloids) to a viscoelastic gel relies upon formation of cross-links between individual, pre-formed fibrils [90,91] (Figure 2B). In some cases, such cross-links may be mediated by hydrophobic residues, such as phenylalanine, which can participate in π-π stacking [92]. From data presented by McKnight and colleagues [75] and a subsequent study [27], whether the transition of FUS and hnRNPA2 to fibrillar hydrogels followed the latter model or whether cross-β structure formed concomitant to gelation is difficult to unequivocally assess. However, a recent study suggests that fibril formation may accompany initiation of LLPS for hnRNPA1 [92].

Structural basis of stable and labile fibrils

Classical amyloids associated with neurological disorders and other pathologies are typically characterized by high stability, as evidenced by relatively strong resistance to dissociation by detergents, chaotropic agents, and elevated temperature [75,93–95]. This stability is engendered by two fundamental interactions: backbone hydrogen bonding between extended β-strands that form a single sheet and the interdigitation of side chains between two β-sheets through the steric zipper motif [96] (Figure 3A). These interdigitated side chains engage in van der Waals interactions that stabilize the structure and exclude water from the interaction interface [96,97].

Figure 3. Structures governing the stability of fibrils.

Atomic structures of microcrystals formed by short peptides can be classified as having extended β-strand, kinked β-strand, or non-cross-β architecture. For all cross-β structures, strands stack perpendicular to the fibril axis (arrow or point) via hydrogen bonds between backbone amides. (A) Peptides derived from Sup35 (left) and β-amyloid (right) form steric zippers with tight interdigitation of side chains and a dry interface between mating sheets. Interdigitation may involve polar (e.g., Asn in Sup35) or hydrophobic (e.g., Ile in β-amyloid) side chains. Fibrils formed by these classical amyloids are typically regarded as stable [101]. (B) Reversible amyloid cores (RACs) are another type of extended β-strand architecture observed for microcrystals of hnRNPA1- (left) and FUS-derived peptides (right). These structures do not contain very tight interdigitation of side chains and are more labile than classical amyloids. Microcrystals of hnRNPA1 peptides are loosely stabilized by a steric zipper with a polar sheet interface, but destabilized by the presence of negatively charged Asp side chains. These Asp side chains face away from the intersheet interface and stack along the fibril axis (referred to as a stacking-D). The labile structure with an extended β-strand architecture formed by the FUS-derived peptide contains exceptionally loose interdigitation and is weakly stabilized by hydrogen bonding between water and Tyr hydroxyl groups at the inter-sheet interface. (C) Another category of labile structures, LARKS, involves kinked β-strands that have been described for microcrystals formed by FUS, HNRNPA1, TDP-43, and nup98 peptides (FUS peptide is shown here). Kinks may be present at Gly or aromatic residues and enable close approach of peptide backbones at the intersheet interface without inter-digitation of side chains. (D) Microcrystals of the same FUS peptide shown in (C) may also form a labile non-cross-β structure. Consequently, this peptide sequence has been referred to both as a LARKS and RAC. In contrast to those structures containing β-strands, this non-cross-β structure is described as having a “coil with a sharp kink” at the central Gly40. The ordered-coil architecture is stabilized by hydrogen bonding between the Tyr38 and Ty41 in the same rung and between Tyr38 and Ser42 of neighboring segments. Cartoon depictions partially adapted from ref [92].

In stark contrast to the high stability of pathological amyloids, FUS, hnRNPA2, and, more recently, hnRNPA1 fibrils have been reported to readily dissociate in the presence of low concentrations of a detergent (SDS), upon incubation at elevated temperatures, and even upon dilution [27,73,75,92]. Examination of these so-called reversible (or labile) fibrils formed by the FUS and hnRNPA2 LCDs by solid state NMR [60,98] revealed that their general architecture was not unlike the majority of classical amyloids – a parallel, in register cross-β organization stabilized by intra-sheet backbone hydrogen bonding. However, a striking difference was observed between FUS LCD fibrils and pathogenic amyloid fibrils with regard to the nature of sidechain-sidechain interactions within the fibril cores. While the cores of pathogenic fibrils formed by proteins such as β-amyloid or α-synuclein are stabilized by multiple hydrophobic interactions between the abundant non-polar side chains, the core of FUS fibrils contains a high density of polar side chains and is devoid of any hydrophobic residues. These characteristics were postulated to contribute to their low stability [98]. Whether similar features are also present within the core of hnRNPA2 LCD fibrils is not clear, as the structural model has not yet been determined [60].

Atomic resolution studies with microcrystals formed by short peptide fragments of FUS, hnRNPA2, and hnRNPA1 LCDs have provided additional insight into the structural basis of low fibril stability. The structures observed for these peptides can be grouped into those containing extended β-strand, kinked β-strand, or non-cross-β motifs (Figure 3B–D). Extended β-strand structures (Figure 3B) have been found for several hnRNPA1 and FUS peptides and these structures have been referred to as reversible amyloid cores (RACs) [92,99]. The hnRNPA1 RAC ²⁰⁹GFGGNDNFG²¹⁷ forms a steric zipper structure in which a polar interface between sheets is formed by Asn side chains. An additional feature of this structure is the presence of stacked, negatively charged Asp side chains on the outer surface (referred to as stacking-D). Even though polar interfaces or the presence of stacked charged residues on the outer surface are not uncommon in classical steric zippers [96,100], the authors argued that the combination of these two features in hnRNPA1 might destabilize the fibrillar structure [92]. A second type of extended cross- β structure was found for the SYSSYG peptide derived from the FUS LCD (residues 54–59) [99]. Unlike the canonical dry steric zipper interface, this latter structure contains a wet interface between β-sheets that is stabilized by hydrogen bonding with water. Additionally, no interdigitation of side chains was observed.

In contrast to the extended β-strand architecture observed for classical amyloids and RACs described above, some peptides derived from LCDs of FUS, hnRNPA1 and TDP-43 can form parallel, in-register stacked β-strands with a central “kink” that commonly occurs at glycine or aromatic residues [101]. These structures are referred to as low-complexity aromatic-rich kinked segments (LARKS) (Figure 3C). In contrast to the interfaces observed for RACs that are based on side chain interactions, the β-strand kink enables close approach of peptide backbones at the interface, enabling favorable van der Waals interactions or hydrogen bonding [101]. LARKS are additionally stabilized by aromatic residues within a single sheet or between β-sheets. However, the lack of side chain interdigitation is proposed to underlie the relative instability of fibers based on these structures.

Intriguingly, the SYSGYS peptide (residues 37–42) from the FUS LCD has been referred to both as a LARKS and a RAC, as two different structures have been determined for microcrystals formed by this peptide [99,101].The RAC-type structure described for this peptide was not based on β-strand architecture, but rather resembled a “coil with a sharp kink” at Gly40 (interestingly, the same residue also made a kink in the LARKS structure) (Figure 3D). Like cross-β-based labile fibrils, fibrils based on this coiled structure were observed to be highly unstable.

Fibrillation in the context of LLPS

Early studies examining time-dependent changes associated with LLPS revealed that aggregates with a fibrillar morphology are a common end-stage product [23,24,27,29,74]. This liquid-to-solid phase transition provided initial evidence suggesting that amyloid aggregation could occur within the context of a crowded liquid droplet environment. However, the significance of labile fibrils and what role they play in the transition from dynamic droplets to more stable amyloid aggregates still remains elusive (Figure 4), despite significant efforts to discern the structures behind them.

Figure 4. Models depicting formation of reversible and irreversible fibrils within the context of LLPS.

(Left) Dynamic droplets undergo maturation with time. In some cases, this maturation leads to a reversible gel-like state that may contain labile fibrils. Upon prolonged incubation, more stable fibrillar aggregates are formed. (Right) Three models can describe the relationship between reversible and irreversible fibrils. On one extreme (Model 1), fibrillar structure (and therefore stability) may be predetermined by amino acid sequence. Thus, a protein that is rich in LARKS or RACs will be prone to only form reversible fibrils (i.e., hnRNPA1 or FUS), whereas a protein rich in steric zipper-forming sequences will only form irreversible fibrils (i.e., TDP-43). In a larger protein, the balance of these motifs could be a key regulator of stability that could also be modulated by pathogenic mutations. If multiple motifs are present, lability may be lost with time if a threshold number of RACs- or LARKS-based interactions is reached as more motifs sequentially associate. On the opposite extreme (Model 2), the amino acid sequence may not strongly regulate the type of fibril structure observed. In this latter scenario, the same sequence could form different types of RACs or LARKS (as is seen for the ³⁷SYSGYS⁴² FUS peptide) or even steric zipper. It is unlikely that both models 1 and 2 adequately describe all scenarios and some combination of both may be more realistic, at least in some cases. A natural sequelae of these two models is the possibility that, while a specific sequence may have a predisposition to form one type of structure (i.e., a labile LARKS or RACs structure), this structure may become more stable with time and might even transition entirely to a steric zipper (Model 3). Post-translational modifications (PTMs) could also regulate such a transition. Cartoon depictions partially adapted from ref [102].

On the one hand, labile (reversible) fibrils may represent an intermediate stage between dynamic, liquid-like droplets and irreversible, pathological amyloid aggregates [12]. Indeed, prolonged incubation of reversible FUS hydrogels (presumably formed by labile fibrils) at low temperature was found to be associated with loss of reversibility [27,92]. Similarly, loss of fluorescence recovery after photobleaching suggests that liquid droplets do become more “gel-like” before formation of amyloids [61]. On the other hand, labile fibrils have also been proposed to reinforce, mediate, or even drive liquid droplet formation, as opposed to purely being a byproduct of LLPS. The latter scenario implies a functional importance for labile amyloids. Indeed, limited evidence does support this possibility, as N-acetylimidazole footprinting experiments demonstrated that hnRNPA2 exists in similar conformation within liquid droplets and hydrogels in vitro as well as in the nucleus in vivo [25]. Additionally, transmission electron microscopy imaging has suggested that, for hnRNPA1, labile fibrillar assembly accompanies liquid droplet formation [92], even though time resolution of this study is not sufficient to rule out other possibilities. Furthermore, this may be a protein-specific phenomenon, as droplet formation and fibrillation appear to be temporally distinct for the TDP-43 LCD [61].

Models depicting formation of labile and more stable fibrils within the context of LLPS can be organized into three general frameworks. In one scenario (Model 1 in Figure 4), the structural motifs that govern reversibility/irreversibility (i.e., LARKS, RACs, or steric zippers) are pre-defined by the amino acid sequence and do not change. Consequently, reversibility is regulated by the total number of LARKS/RACs and, additionally, by the ratio of these two motifs to steric zippers [92,102]. Although direct evidence for this model at the fibril level is still incomplete, in a study of hnRNPA1 (where labile amyloid assembly was suggested to accompany liquid droplet formation), deletion of RACs suppressed droplet formation while artificial insertion of additional RACs promoted LLPS [92]. Furthermore, a study by Eisenberg and colleagues revealed that the majority of short peptide fragments of the TDP-43 LCD form steric zipper motifs, with only a single LARKS motif identified in this protein [102]. This propensity for steric zippers may explain the relative stability of TDP-43 fibrils formed by longer protein segments [103]. Additionally, a pathogenic mutation within the sole LARK segment of TDP-43 was found to enhance stability [102]. In disease, such a conversion from LARK segment to a steric zipper could tip the balance of TDP-43 assemblies toward higher stability and lower reversibility. Similar findings have been echoed in studies of FUS RACs, whereby a point mutation that disables the peptide from forming stacking-D was found to greatly increase the thermostability of microcrystals [92]. Accordingly, the balance of these interactions and the total number of RAC or LARKS motifs in a protein appear to be crucial to regulating fibril stability.

On the opposite side of the spectrum is a scenario in which the same protein fragment could either form a LARKS or RAC – or possibly also a steric zipper (Model 2 in Figure 4). Classical cross-β architectures can involve a number of different types of interactions [97], suggesting that a number of structures can be adopted by amyloid fibrils formed from the same protein—a term referred to as polymorphism. Indeed, structural polymorphism has been observed experimentally and appears to be a general property of classical amyloids [104,105]. Consistent with the possibility that structural polymorphism may also exist for labile fibrils, the FUS SYSGYS segment (residues 37–42) appears to have the capacity to form both LARKS and RACs (as discussed in the previous section) [99,101]. The question remains, however, whether this structural heterogeneity observed for microcrystals formed by small peptides translates to fibrillary aggregates formed by the much longer FUS LCD. An apparent lack of structural heterogeneity for FUS LCD fibrils was inferred from solid-state NMR data [98]. However, an important caveat must be mentioned: in this study, ¹³C-labelled FUS fibrils appear to have been produced via a seeded reaction [98], a method that can bias the creation of a structurally homogenous population of amyloids [106]. Thus, the potential for structural polymorphism in labile fibrils remains to be fully ascertained. Additionally, there is also the possibility that a protein, such as FUS, that typically forms highly labile fibrils may also adopt a more stable fibrillar structure based on a steric zipper-like motif. While there is no direct evidence for this scenario yet, fibrillar inclusions of wild-type FUS have been observed in sporadic neurodegenerative diseases and therefore such a possibility cannot be ruled out [107].

Finally, under some conditions, less stable LARKS and RACs could potentially transition to more stable steric zipper-like motifs with time (Model 3 in Figure 4). Much less evidence exists for this hypothetical scenario, even though it is a natural sequelae from Models 1 and 2. In particular, such a transition could be regulated by post-translational modifications. Indeed, phosphorylation of the TDP-43 pathogenic mutant described in Model 1 was shown to result in greater interdigitation of opposing β-strands in a zipper-like fashion [102]. Furthermore, the nature of FUS assemblies has been shown to be modulated by methylation of Arg residues. Given the potential for the same amino acid region to form multiple polymorphs and the relative dynamicity of residues observed in structures of labile amyloids [60,98], one should also consider the possibility that both RACs and LARKS might adopt more stable conformations with time, even in the absence of post translational modifications. However, such a transition has not yet been observed experimentally.

LLPS as a modulator of pathological protein aggregation

Kinetic impacts of LLPS

While it is well established that amyloid formation involves nucleation and growth (elongation) steps, two major but fundamentally distinct models have been proposed to describe this process. In the nucleated polymerization mechanism, the rate-limiting step is formation of the nucleus (which typically is oligomeric but can also be a misfolded monomer).This first step is followed by fibril elongation, whereby the nucleus recruits additional protein monomers and fibril extension occurs until free protein substrate becomes limiting [76]. This model predicts a very strong dependence of nucleation time on protein concentration. By contrast, the nucleated conformational conversion model proposes that monomers rapidly convert to loosely structured oligomers [76,108]. In this latter model, the rate-limiting step is the slow conformational rearrangement of these oligomers into more organized, growth-competent nuclei. Thus, this model predicts markedly weaker concentration dependence of the overall fibrillation rate as compared to the nucleated polymerization mechanism.

Perhaps intuitively, the high protein concentrations found in the liquid droplet environment would have a major influence on the kinetics of amyloid formation, particularly if aggregation follows a nucleated polymerization mechanism. However, few studies have investigated in detail the kinetic aspects of amyloid formation under conditions of LLPS. In a study with TDP-43 LCD, salt-induced liquid droplet assembly was shown to promote amyloid formation with considerably faster kinetics than in the absence of liquid droplets [61]. LLPS of the TDP-43 LCD is regulated by hydrophobic interactions and electrostatic repulsion [22,50], the latter of which also considerably influences aggregation kinetics in the absence of LLPS [61]. To circumvent this complication, LLPS was disrupted through temperature elevation without changes in salt concentration. This resulted in significantly longer aggregation times, indicating that the conditions of LLPS per se lead to faster fibrillation of TDP-43 LCD [61]. While aggregation within the context of LLPS has been shown for a number of other proteins [23,29,74,109,110], detailed examination of the kinetic impact has not been well described and remains an area in need of significant experimental investment.

Characterization of the influence of LLPS on protein aggregation is hampered by technical challenges. Turbidity measurements have been used extensively to develop phase boundaries and investigate regulators of LLPS; however, caution should be taken in attributing an increase in turbidity solely to LLPS, as protein aggregation outside the context of LLPS can have the same effect. Thus, droplet formation should always be verified directly by optical microscopy and fibril formation by TEM or AFM. Developing kinetic assays that can distinguish between aggregation-associated and LLPS-associated turbidity changes will also be helpful in the future. Fluorescent dyes such as ThT appear to be highly sensitive to amyloids and less so to droplets [61,92], providing a technique for distinguishing both. However, ThT fluorescence is dependent on local viscosity, which is high within droplets, and its fluorescence is also elevated in the presence of RNAs, a significant complication to studies of RNA-induced LLPS [111]. Alternative amyloid-sensitive dyes, such as derivatives of naphthalene sulfonates (e.g., ANS, bis-ANS), may provide additional insight in this regard.

Conformational changes associated with LLPS

Conformational compaction or expansion of IDRs is a common theme in the liquid droplet environment and is often protein-specific. For example, nucleophosmin, a nucleolar protein, was recently shown via single-molecule fluorescence resonance energy transfer (FRET) and other methods to undergo salt-dependent conformational compaction/expansion, and these electrostatically-driven conformational changes were found to control switching between different LLPS mechanisms [112]. Similarly, the proximity-dependent fluorophore pyrene was used to demonstrate a conformational unwinding or expansion associated with LLPS of K18, the aggregation-prone region of tau protein [113]. Most recently, solution NMR experiments have strongly suggested that both the hnRNPA2 and FUS LCDs retain a predominantly disordered state in the liquid droplet environment [26,28]. As discussed above, misfolding of protein monomers and/or assembly of oligomeric nuclei are critical first steps in amyloid fibril formation [76]. Even though the mechanism by which the environment of liquid droplets could modulate these initial misfolding events still remains unclear, it is possible that the expanded state may be important for the nucleation process of these proteins.

In the case of TDP-43, significant self-association of the LCD mediated by the protein’s α-helical region has been observed via solution NMR and electron paramagnetic resonance in the context of LLPS [22,61,114]. Assuming that these small oligomers are on pathway towards amyloid formation (as suggested by the decrease in the lag phase of fibrillation reaction observed under LLPS conditions [61]), this would indicate that intramolecular interactions involving residues within the α-helical region are important for nucleation of this protein. Interestingly, a chemical chaperone, trimethylamine N-oxide (TMAO), has recently been shown to decouple LLPS and amyloid formation of TDP-43 LCD, and it was argued that this decoupling is related to the compaction of the protein in TMAO-induced droplets [115]. This opens an intriguing possibility that the environment of liquid droplets might not always be conducive to amyloid formation. However, a caveat to these experiments is that very high concentrations of TMAO were required to produce such a decoupling.

Alternative aggregation pathways under LLPS conditions

In neurodegenerative diseases, TDP-43 forms a number of different aggregates, some of which are amyloid-like [116,117], while others appear to be non-amyloid aggregates [118]. This structural heterogeneity is also observed for TDP-43 aggregates formed in vitro under LLPS conditions (for specific examples in the context of LLPS, see: amyloids [61], oligomers [119], tufted particles [120], granulo-filamentous aggregates [121]). Along these lines, studies of FUS aggregation in the context of LLPS have revealed that pathogenic mutations can accelerate the transition to laterally-aligned fibers that emerge from droplets and appear to have amyloid-like character [29]. However, apparent non-fibrillar, irregular aggregates of full-length FUS have also been observed to form in the context of LLPS [122]. These aggregates appear to be related to maturation of previously-dynamic, liquid-like droplets. The mechanism by which such distinct aggregate structures form, however, is not yet clear. Nevertheless, these findings do suggest that variable aggregation pathways may be accessible under LLPS conditions.

Recruitment of aggregation-prone proteins

Yet another possibility for the role of LLPS in protein aggregation may involve the recruitment of aggregation-prone proteins by a relatively benign, phase-separating protein. For instance, hydrogels formed by the hnRNPA2 or FUS LCD can recruit other LCD-containing proteins [75]. More recently, the recombinant human prion protein, PrP, was found to form liquid-like droplets in vitro [123]. Interestingly, the addition of Aβ oligomers to PrP resulted in rapid formation of a hydrogel, within which PrP remained relatively mobile while Aβ oligomers were motionally restricted. These interactions were also shown to alter the dynamicity of native PrP^C (that contains the glycophosphatidylinositol anchor) at the cell surface, a mechanism that was argued to play a potential role in the pathogenesis of Alzheimer’s disease that is believed to be mediated by Aβ oligomers [123].

A similar phenomenon might occur with RNA-binding proteins. TDP-43 has been observed to preferentially partition to and subsequently aggregate within droplets formed by Ddx4N1, a DEAD-box ATPase that can regulate MLOs in the cell [120,124]. Furthermore, the TDP-43 LCD can be recruited by hnRNPA2 LCD droplets [26]. This co-phase separation can subsequently lead to aggregation. These insights reveal that the interactions between multiple proteins may be critical to understanding and predicting how protein aggregation might be altered in the phase-separated environment.

Conclusions and perspectives

The recent series of exciting studies have provided important insights into the role of LLPS in normal cell biology. Furthermore, there is rapidly growing evidence that LLPS may be associated with a number of neurodegenerative diseases, acting as a modulator of pathological protein aggregation. Biophysical studies indicate that fibrils formed by some LLPS-prone proteins are fundamentally different from classical amyloids, both with regard to their stability as well as structural motifs. Though the complete relationship between labile and more stable pathogenic fibrils remains to be fully elucidated, these advances have provided insights into potential mechanisms by which pathological aggregates could develop within the context of liquid droplets. LLPS can also impact the aggregation process itself, resulting in acceleration of reaction kinetics, changes in conformational dynamics, recruitment of other aggregation-prone proteins, and the potential creation of new aggregation pathways.

This progress notwithstanding, some areas of study on the relationship between LLPS and protein aggregation are still incompletely developed. In particular, a standardization of “reversibility” is necessary to fully elucidate the molecular and structural basis for differences in stability between labile and pathological fibrils. Thus far, the majority of investigations of labile fibrils include an internal control comparing reversibility to that of a prototypical amyloid fibril [75,102]; however, comparisons across studies are much more difficult due in part to differing definitions of reversibility. For example, dissolution upon mere temperature elevation from 4 °C to room temperature is used in some cases as a defining feature of labile fibrils [75,99]. In other studies, however, dissolution upon incubation in the presence of ~1% SDS at 70°C for 15 min is used as a criterion of fibril lability [102] – a condition that would dissociate at least some pathological amyloids [93,94,125]. The picture is further complicated by observations that different polymorphs of classical amyloids formed from the same protein may exhibit large differences in stability [125,126]. Thus, the implementation of well-defined thermodynamic parameters, such as G (which has been used in the study of some pathological amyloids [127]), would be important for assessing stability across studies and ascertaining what is actually meant by the term “reversible fibrils.”

Second, more detailed insight into the temporal relationship between liquid droplets, labile fibrils, and more stable pathological fibrils is necessary. In some studies, labile fibrils have been proposed to reinforce or even drive formation of liquid droplets [92]. This is, however, certainly not a universal mechanism, as for some proteins (e.g., TDP-43 [61]) fibril formation is only observed after induction of LLPS, and for other proteins (e.g., full-length tau [53]), wholly different types of interactions drive phase separation. Additionally, not all phase separating proteins form labile fibrils (e.g., tau). As such, determining whether these apparently conflicting observations result from characteristics of specific proteins or varying temporal resolution of techniques used will be important. This area of investigation would also benefit from development of novel biophysical or spectroscopic techniques that could characterize with sufficient time resolution individual molecular events during droplet formation and their maturation.

• RNA-binding proteins undergo liquid phase separation and aggregate in disease

• Liquid droplets formed by these proteins mature with time

• These proteins commonly form labile fibrils, distinct from classical amyloids

• Fibril stability is based on specific structural characteristics

• Phase separation can modulate protein aggregation by several mechanisms