Phase separation of microtubule-binding proteins – implications for neuronal function and disease

ABSTRACT

Protein liquid–liquid phase separation (LLPS) is driven by intrinsically disordered regions and multivalent binding domains, both of which are common features of diverse microtubule (MT) regulators. Many in vitro studies have dissected the mechanisms by which MT-binding proteins (MBPs) regulate MT nucleation, stabilization and dynamics, and investigated whether LLPS plays a role in these processes. However, more recent in vivo studies have focused on how MBP LLPS affects biological functions throughout neuronal development. Dysregulation of MBP LLPS can lead to formation of aggregates – an underlying feature in many neurodegenerative diseases – such as the tau neurofibrillary tangles present in Alzheimer's disease. In this Review, we highlight progress towards understanding the regulation of MT dynamics through the lens of phase separation of MBPs and associated cytoskeletal regulators, from both in vitro and in vivo studies. We also discuss how LLPS of MBPs regulates neuronal development and maintains homeostasis in mature neurons.

Summary: Review discussing how phase separation of microtubule-binding proteins impacts their regulation of microtubule dynamics, and the significance for these mechanisms in neuronal development and brain diseases.

Introduction

Neurons are highly polarized cells optimized to sensitively detect and rapidly relay chemical and electrical signals (Bear, 2016). All neurons contain a single main axon, which can transmit action potentials over great distances, and numerous highly branched dendrites, which serve as functional contacts for neighboring axons. Neurotransmission begins when an action potential is propagated through the axon to the presynaptic terminal, leading to the release of neurotransmitters. These neurotransmitters are then detected by ligand-gated receptors that are concentrated within tiny protrusions from the dendritic shaft called dendritic spines. At every level of neurotransmission, the cytoskeleton serves as a platform for the dynamic movement of ligand-gated receptors and propagation of chemical and electrical signals.

The development, plasticity and stability of the elaborate neuronal morphology are determined by its underlying cytoskeleton, which is chiefly composed of microtubules (MTs), actin and intermediate filaments (Witte and Bradke, 2008). MTs are crucial for neuronal function and the maintenance and remodeling of the neuron – they provide mechanical support to serve as a structural framework, generate mechanical forces during neuronal migration and morphogenesis, and act as tracks for cargo delivery (Kapitein and Hoogenraad, 2015; Rolls, 2022). Recently, studies have proposed that microtubule-binding proteins (MBPs) can modulate the cytoskeleton by locally concentrating in the cytoplasm or on MTs through liquid–liquid phase separation (LLPS) (Tan et al., 2019; King and Petry, 2020; Hernández-Vega et al., 2017; Song et al., 2022; Meier et al., 2022; Maan et al., 2022). LLPS is a reversible mechanism through which macromolecules transition from a soluble state to form membrane-free condensates separated from the bulk solution of the cytoplasm. The ability of proteins to undergo LLPS is linked to intrinsically disordered regions (IDRs), protein sequences that typically lack a defined structure. LLPS of MBPs is proposed to be a mechanism by which these proteins recruit additional effectors at MT ends or lattices to promote MT stability and dynamics for neuronal function.

In this Review, we discuss how MTs are assembled in vitro and in neurons, how phase separation of MBPs impacts regulation of MT dynamics, how dysregulation of MBP phase separation leads to neurodegenerative diseases (NDs) and current methodologies to study MBP phase separation in cells.

MT assembly in neurons

Newly born neurons migrate to their final position in the brain along a substrate of radial glial fibers. Migrating neurons are led by one or more protrusive processes, which sense the environment and guide the migration path. MTs originating from the centrosome form a cage that surrounds the nucleus and emanates into the leading edge of the migrating neuron (Rivas and Hatten, 1995; Solecki et al., 2004; Sakakibara et al., 2014). In response to forces applied by MT-associated motors, the centrosome moves toward the leading edge, followed by the translocation of the nucleus – a phenomenon known as nucleokinesis (Solecki et al., 2004; Cooper, 2013).

Later in migration, neurons can have multiple protrusive processes. The centrosome moves to the base of the specific protrusion that is fated to become the axon (Sakakibara et al., 2014; de Anda et al., 2010). In cultured neurons, neurites initiate through actin-based protrusions, followed by the invasion of bundled MTs into the developing process (Rivas and Hatten, 1995; Dent et al., 2007; Rochlin et al., 1999). Axons quickly become distinguishable from other neurites as they show enrichment in markers of stable tubulin and increased resistance to MT-destabilizing agents (Witte et al., 2008). In fact, stabilization of MTs through local photo-uncaging of the MT-stabilizer taxol at an individual neurite is sufficient to induce axon differentiation (Witte et al., 2008). Thus, MT invasion and stabilization are crucial for differentiation and extension within the nascent axon.

Furthermore, MTs are distinctly organized in axons versus dendrites. In axons, the majority (>95%) of MTs are oriented with their rapidly growing plus ends facing outward (away from the soma), whereas dendrites contain MTs with mixed polarities (Heidemann et al., 1981; Burton and Paige, 1981; Baas et al., 1988, 1991). The plus-end-out organization in axons likely facilitates directional trafficking of membrane and other building blocks necessary for rapid neuronal extension. In contrast, the mixed polarity of dendritic MTs is essential for localization of Golgi outposts, ribosomes and other organelles important for dendritic function (Baas et al., 1988; Burton, 1988; Baas and Yu, 1996).

Even after polarization, MTs continue to play key roles in shaping the neuron. The formation of axon branches begins with local unbundling of axon MTs, MT severing and the coalescence of new MT bundles into the nascent branch (Dent et al., 1999). MTs also transiently invade dendritic spines, enabling the targeted delivery of materials to regulate synapse size and/or function in response to external cues or changes in neuronal activity (Hu et al., 2008, 2011; Merriam et al., 2013, 2011; Gu et al., 2008). These findings reinforce the importance of understanding how MTs are assembled to support these neuronal processes.

MT nucleation in neurons

In newly differentiated neurons, the centrosome, which is the primary MT-organizing center (MTOC), contains the γ-tubulin ring complex (γTuRC), a large multi-protein complex that acts as a template for MT formation and anchors the radial MT network. However, γTuRC alone has low nucleation efficiency and must be recruited to MTOCs or pre-existing MTs, where it undergoes activation by effectors, such as CDK5RAP2, the augmin (also known as HAUS) complex members and TPX2 (Choi et al., 2010; Zhang et al., 2017; Zupa et al., 2022; Xu et al., 2024). Additional proteins, including the MT polymerase XMAP215 (also known as CKAP5), synergize with γTuRC to promote highly efficient MT assembly (Thawani et al., 2018). Additionally, several MBPs, such as CAMSAP (also known as Patronin and Nezha) proteins (Goodwin and Vale, 2010; Yau et al., 2014; Rai et al., 2024), CLASPs (Efimov et al., 2007) and p150^Glued (also known as DCTN1) (Lazarus et al., 2013), promote MT assembly at non-centrosomal MTOCs by enhancing longitudinal and lateral tubulin–tubulin contacts, stabilizing the formation of a crucial nucleus for MT assembly. Many factors contribute to MT nucleation, and current research is focused on understanding the formation of stable nucleation intermediates. For further details, we refer readers to recent excellent reviews on MT nucleation mechanisms (Roostalu and Surrey, 2017; Weiner et al., 2021), and on the activation and recruitment of γTuRC to MTOCs during spindle assembly and beyond (Kraus et al., 2023a; Tovey and Conduit, 2018).

During neuronal differentiation, γ-tubulin and other key nucleation components become delocalized from the centrosome, which loses its role as the MTOC (Yau et al., 2014; Sánchez-Huertas et al., 2016; Weiner et al., 2020; van de Willige et al., 2016; Stiess et al., 2010). In the next sections, we highlight key studies that have begun to elucidate the principles, machinery and mechanisms by which MT nucleation occurs in axons and dendrites.

MT nucleation in axons

The earliest axonal MTs arise from a centrosome-directed population that polymerizes into the nascent axon and/or are severed by katanin from the centrosome for transport into the axon (Baas and Yu, 1996; Ahmad and Baas, 1995). In Drosophila neurons, MTs can also originate at the Golgi and soma and grow preferentially into axons, guided by the MT plus end-directed motor protein kinesin-2 (Mukherjee et al., 2020). Existing MTs also serve as platforms for nucleating daughter MTs. In axons, branching nucleation is initiated by the augmin complex, which binds to the parent MT lattice and subsequently recruits γTuRC. γTuRC nucleation activity is stimulated by the binding of activator proteins such as RanGTP and TPX2 (Petry et al., 2013; Kraus et al., 2023b; Scrofani et al., 2015). γ-tubulin and augmin complex members also colocalize at excitatory presynaptic boutons to nucleate MTs, a process essential for synaptic vesicle exocytosis and motility (Qu et al., 2019).

The MBP SSNA1 is also involved in axonal MT nucleation. SNNA1 self-clusters in vitro and promotes formation of aster-like MTs and nucleation from the sides of pre-existing MTs (Basnet et al., 2018). SSNA1 assembles longitudinal fibrils along MTs via interactions with tubulin tails that guide growth of new MTs branching from mother filaments. In primary neurons, SSNA1 localizes to and determines axon branching sites. Transfection of mouse hippocampal neurons with an SSNA1 mutant deficient in MT branching function completely abolishes axon branch formation (Basnet et al., 2018).

MT nucleation in dendrites

MT nucleation hotspots in dendrites were first proposed to originate from Golgi outposts, which were suggested to serve as mini-MTOCs at dendritic branch points in Drosophila sensory neurons (Mukherjee et al., 2020; Ori-McKenney et al., 2012). A later study found that the removal of dendritic Golgi outposts did not delocalize γ-tubulin from dendritic branch points and yielded only minor changes in dendritic MT organization, suggesting that another key element serves as a tether for dendritic MT nucleation (Nguyen et al., 2014).

Emerging research strongly suggests that dendritic MTs are more likely to originate from early endosomes rather than Golgi outposts. A recent study implicated Wnt signaling from endosomes as a key driver of dendritic MT nucleation in Drosophila sensory neurons (Weiner et al., 2020). Endosomes localized to most proximal dendritic branch points, where most growing MT plus ends initiated, whereas MTs only occasionally originated from the Golgi (Weiner et al., 2020). In fact, artificially tethering the Wnt signaling scaffold protein, Axin, to mitochondria was sufficient to direct γ-tubulin and Centrosomin (Cnn) to dendritic branch points (Weiner et al., 2020). Following axon injury, the endosomal Wnt signaling pathway promoting dendritic MT nucleation is upregulated to stabilize dendritic arbors. A similar increase in dendritic MT nucleation is observed following depletion of Ror, an orphan tyrosine kinase receptor and a key Wnt signaling partner and regulator of Axin localization (Nye et al., 2020). The molecular mechanisms by which Wnt signaling components recruit γ-tubulin and modulate nucleation and how these processes contribute to overall neuronal development remain to be elucidated.

Endosomes are also implicated in MT nucleation at dendritic growth cones. In developing C. elegans sensory neurons, the minus-end-out polarity in anterior dendrites is directed by a dendritic growth cone-localized non-centrosomal MTOC, which moves distally in concert with the extension of the growth cone (Liang et al., 2020). γTuRC also localizes to dendritic growth cones of sensory neurons in Drosophila during initial outgrowth, where it colocalizes with the endosomal marker Rab11 (Weiner et al., 2020). In this context, Rab11 endosomes might play a role similar to the one they fulfill in mitotic spindle assembly, where they transport MT-nucleating proteins to spindle poles to facilitate MTOC-directed MT assembly.

Regulation of MT dynamics and stability by MBPs

The regulation of MT network remodeling by MBPs is essential for building and maintaining MT infrastructure, especially in response to environmental cues that regulate neuronal morphogenesis (Baas et al., 2016; Hoogenraad and Bradke, 2009; Kapitein and Hoogenraad, 2011; Koleske, 2013). MTs exhibit various structural architectures that are differentially recognized by MBPs (discussed in Box 1), which then exert their functions on MT lattices (e.g. lattice repair) or at MT ends (e.g. promoting incorporation of new GTP-tubulin). For an excellent review of the mechanisms underlying MT dynamics, we direct readers to Brouhard and Rice (2018).

Box 1. MTs are dynamic polymers that adopt distinct structures in various nucleotide states and are differentially recognized by MBPs.

MTs are composed of GTP-binding α- and β-tubulin heterodimers, which assemble head-to-tail to form linear protofilament chains. Protofilaments interact laterally to form hollow MTs (Desai and Mitchison, 1997). MTs comprise growing ends and a lattice. New tubulin subunits polymerize while bound to GTP, which hydrolyzes to GDP, inducing structural changes in the MT. MTs exhibit a nonequilibrium behavior known as dynamic instability (Mitchison and Kirschner, 1984), characterized by a GTP hydrolysis-driven stochastic interconversion between growth and shrinkage (Hyman et al., 1992; Erickson and O'Brien, 1992). MT growth is biased towards plus-end addition of GTP-bound tubulin heterodimers, which form a stabilizing ‘GTP-cap’ (O'Brien et al., 1987; Walker et al., 1988) that is dynamically regulated based on cellular needs (Lansbergen and Akhmanova, 2006; Brouhard and Rice, 2018) (see figure).

MBPs recognize distinct MT conformations associated with assembly, disassembly or changes in lattice spacing based on nucleotide-bound state (LaFrance et al., 2022). Growing MT ends predominantly contain GTP- and GDP-P_i-bound tubulin and adopt tapered, blunt or flared configurations (McIntosh et al., 2018; Höög et al., 2011). In contrast, the MT lattice contains predominantly GDP-tubulin and displays a straight conformation. GTP hydrolysis causes an ∼2.5 Å (1 Å=0.1 nm) longitudinal compaction at the interdimer interface within the tubulin lattice (Alushin et al., 2014). EBs mark the GTP cap at growing ends and recruit other effectors to modulate MT dynamics, which is essential for processes like neurite outgrowth and dendritic branching (van de Willige et al., 2016; Rickman et al., 2017). The non-centrosomal MT nucleator CAMSAP3 preferentially binds to expanded GTP-bound MT lattices, stabilizing newly generated minus ends (Liu and Shima, 2023). MBPs can also alter MT conformation, thereby affecting the binding of other MBPs. For instance, kinesin-1 binds to GDP-bound MTs and induces a longer-pitch conformation, or a ‘high-affinity’ state, to cooperatively recruit more kinesins (Shima et al., 2018). For a comprehensive overview of MT lattice structure and plasticity, we redirect readers to the following reviews (Romeiro Motta et al., 2023; Cross, 2019). Figure created in BioRender, Koleske, A., 2024. https://BioRender.com/s23t142. This figure was sublicensed under CC-BY 4.0 terms.

In all cells, MT growth, stability and shrinkage are regulated by a large collection of MBPs. Both MT ends undergo growth and shrinkage, but plus ends experience greater dynamic instability than minus ends (see Box 1) (Howard and Hyman, 2009). The ends are recognized by distinct sets of plus- and minus-end-tracking proteins that regulate dynamics and stability at their respective ends (Akhmanova and Steinmetz, 2008). The plus-end-tracking protein (+TIP) end-binding (EB, also known as MAPRE) family protein preferentially binds to plus ends where polymerized tubulin is in an expanded lattice state and recruits other MBPs to growing MT plus ends (Akhmanova and Steinmetz, 2008). +TIPs also stimulate tubulin GTP hydrolysis, increasing both MT growth rate and catastrophe frequency (Duelberg et al., 2016; Zhang et al., 2015; Rickman et al., 2017; Maurer et al., 2014; Akhmanova and Steinmetz, 2011). Minus-end-binding proteins, including the CAMSAP family, autonomously bind and cap the minus ends, stabilizing the MT lattice, and restricting minus-end growth and shortening (see Box 1) (Akhmanova and Hoogenraad, 2015; Akhmanova and Steinmetz, 2015).

In addition to dynamic behaviors at MT ends, the internal MT lattice is a key site for regulating MT stability. Neurons are enriched for a subset of MT-associated proteins (MAPs) that regulate lattice stability. MAPs like tau and MAP2 bind and stabilize the MT lattice (Al-Bassam et al., 2002; Siahaan et al., 2022), protecting it from severing proteins, such as katanin and spastin (Errico et al., 2002; McNally and Vale, 1993). By contrast, kinesin-8 (Kip3; Kif18A, Kif18B and Kif19 in humans) induces lattice defects that make MTs more prone to depolymerization (Gardner et al., 2011). MT lattice defects can also result from mechanical stress, fast polymerization or strain caused by MT-based motor traffic (Triclin et al., 2021; Andreu-Carbó et al., 2022; Reid et al., 2017; Schaedel et al., 2015, 2019; de Forges et al., 2016; Arellano-Santoyo et al., 2017). The ability to detect and repair these defects is crucial for the maintenance of local MT stability. Incorporation of tubulin dimers at these damage sites is facilitated by a growing list of MBPs, including CSPP1 (van den Berg et al., 2023) and CLASP2 (Al-Bassam et al., 2010; Aher et al., 2020).

MT dynamics is regulated by MBP phase separation

The precise spatiotemporal control of MT dynamics during neuronal development relies on the local concentration of MT nucleators and regulators within specific cellular compartments. MTs themselves serve as platforms for local concentration and diffusion via electrostatic interactions of MBPs (Wang and Sheetz, 1999; Cooper and Wordeman, 2009; Hinrichs et al., 2012; Bigman and Levy, 2020). Here, we briefly discuss biological contexts in which LLPS is crucial for MBP function in cells and neurons, and refer readers to an excellent recent review on LLPS at MTs more broadly (Volkov and Akhmanova, 2024).

MT nucleation and phase separation of MBPs

The centrosome has been proposed to be a condensate enriched with MBPs that undergo LLPS to concentrate tubulin and promote MT nucleation. C. elegans Spindle-defective-5 (SPD-5) and its homolog Drosophila melanogaster Centrosomin (Cnn) both phase separate in vitro, and likely contribute to the formation of pericentriolar material (PCM) surrounding centrosomes (Woodruff et al., 2017, 2015; Feng et al., 2017). SPD-5 forms condensates that recruit the C. elegans homologs of the MT polymerase XMAP215 (ZYG-9) and TPX2 (TPXL-1) to nucleate aster-like MT arrays in vitro, even in the absence of γ-tubulin (Woodruff et al., 2017), suggesting that SPD-5 works in conjunction with γTuRC-mediated nucleation. LLPS of non-centrosomal MBPs also concentrates tubulin dimers to facilitate acentrosomal MT nucleation, even in the absence of MT templates. For instance, TPX2, a MT nucleation factor crucial for spindle assembly, undergoes LLPS to concentrate tubulin dimers (Fig. 1Ai) (King and Petry, 2020). TPX2 contains rigid folded regions flanked by stretches of IDRs that mediate dynamic MT-binding interactions (Guo et al., 2023). Although intrinsically disordered in solution, TPX2 becomes folded upon binding two adjacent MT protofilaments. LLPS of TPX2 is inhibited in vitro by the spindle assembly factors importin-α and importin-β, which suppress MT nucleation during the cell cycle until the onset of mitosis (Safari et al., 2021). TPX2 binds to MTs along the neurite shaft, and TPX2 depletion reduces nucleation MT frequency at the tip and base of neurites (Chen et al., 2017). It remains unclear whether TPX2 undergoes LLPS in cells, and how this process is regulated.

Fig. 1.

Regulation of MT nucleation, stability and growth through MBP LLPS. (A) LLPS of MBPs, such as (i) TPX2 can undergo LLPS, form condensates (in light purple) and co-condense tubulin dimers to form a critical nucleus. The critical nucleus is proposed to be the smallest tubulin oligomer that can structurally support the initial assembly of MTs. (ii) Severing proteins spastin and katanin harness energy from ATP hydrolysis to create new MTs by severing existing ones, generating ‘new’ shorter MTs in the process. CAMSAP2 and likely CAMSAP3 (in red) undergo LLPS to decorate minus ends of newly generated MTs to prevent their depolymerization, thereby stabilizing the minus ends to enable the continuation of MT growth. (B) MBPs such as (i) tau undergo LLPS to form envelopes (light green) on MT lattices to sterically inhibit binding of spastin and katanin. (ii) Both tau (top) and MAP2c (bottom) form envelopes on MTs to selectively gate trafficking, for example, by allowing activated dynein–dynactin–adaptor complexes (e.g. dynein–dynactin–Hook3) to enter envelopes while preventing dynein-mediated retrograde and kinesin-mediated anterograde transport from passing at envelope boundaries. (C) EB1 (purple) undergoes LLPS to recruit and concentrate +TIPs (e.g. MCAK and TIP150) at plus ends for MT growth. This +TIP machinery mediates correct MT–kinetochore interactions for chromosome segregation. Created in BioRender, Koleske, A., 2024. https://BioRender.com/l81n471. This figure was sublicensed under CC-BY 4.0 terms.

As neurons mature, the MT nucleation machinery relocates from the centrosome to axons and dendrites (Stiess et al., 2010) (see above) and LLPS likely plays a crucial organizational role in these compartments. Vertebrate CAMSAP1, 2 and 3 proteins (and their Drosophila ortholog Patronin) promote MT assembly by stabilizing MT minus ends for neuronal development (Feng et al., 2019; Wang et al., 2019; Zhou et al., 2020; Jiang et al., 2014; Goodwin and Vale, 2010; Yau et al., 2014). In Drosophila sensory neurons, Patronin knockdown yields misoriented minus-end-out MTs in dendritic branches as a result of the loss of protection of minus ends against kinesin-13 MT depolymerization, thereby attenuating dendritic pruning (Feng et al., 2019; Wang et al., 2019). CAMSAP2 undergoes LLPS and co-partitioning with tubulin to stimulate spontaneous MT nucleation independently of the γTuRC complex (Imasaki et al., 2022). Negative stain and cryo-electron microscopy have revealed that CAMSAP2 condensates nucleate ring-like MT nucleation intermediates (Fig. 1Aii) (Imasaki et al., 2022). Similarly, the Abl2 non-receptor tyrosine kinase, which is essential for dendritic arbor stability, undergoes phase separation both in vitro and in cells, forming coacervates with tubulin and promoting MT assembly (Duan et al., 2023). The next challenge is to understand when and where MBP nucleators undergo LLPS in neurons, how this process is regulated and how these mechanisms contribute to overall MT assembly and dynamics.

MT lattice stabilization and regulation of cargo trafficking by phase separation of MBPs

The MT stabilizer tau utilizes its MT-binding repeats to preferentially bind GDP-bound MT lattices, thereby stabilizing MTs. Tau oligomerizes and undergoes LLPS, forming ‘envelopes’ around MTs (Siahaan et al., 2022; Tan et al., 2019). This envelope-forming propensity is shared by the tau family member MAP2c, which is also enriched in neurons. Tau envelopes protect against severing of MTs by spastin in vitro (Fig. 1Bi) (Tan et al., 2019). In addition, these envelopes selectively regulate the movement of motor proteins along protected MTs: highly processive dynein–dynactin complexes with two dynein dimers readily traverse tau envelopes, whereas single dynein motor complexes and the neuronal cargo transporter kinesin family member 1A (KIF1A) are restricted from entering (Fig. 1Bii). Although it remains unclear whether tau and MAP2c form envelopes in cells, they have been shown to protect MTs against katanin severing in fibroblasts. Depletion of tau, but not MAP2c, coupled with overexpression of katanin in cultured hippocampal neurons results in a loss of axonal MT mass (Qiang et al., 2006).

MAP proteins might also govern the specificity of motor trafficking on associated MT substrates. For example, MAP7 recruits and activates kinesin-1, modulating its motility and organelle transport in axons (Tymanskyj et al., 2018). MAP7 also inhibits dynein motility (Ferro et al., 2022). Interestingly, MAP7 and tau exhibit divergent effects on kinesin-1 – whereas MAP7 recruits kinesin-1 to the MT lattice and activates its ATPase activity, tau inhibits kinesin-1 ATPase activity and transport (Monroy et al., 2018). Clearly, MAPs can undergo LLPS to form protective envelopes on stable MT lattices and selectively gate motors for cargo trafficking. Future research should determine whether, where and how envelopes selectively gate cargo trafficking within neurons.

Interestingly, LLPS of some MT motors might be crucial for their function. For example, in COS-7 cells, kinesin-3 KIF1C undergoes LLPS via its disordered C-terminal tail, deletion of which both abolishes its ability to phase separate and disrupts the delivery of mRNA to the cell periphery (Geng et al., 2024). As KIF1C is highly expressed in neurons and ensures fast transport of dense core vesicles (DCVs) for regulation of dendrite morphology (Lipka et al., 2016), it will be interesting to explore whether and how KIF1C undergoes LLPS in neurons and how it contributes to the regulation of cargo-activated transport of integrins, mRNAs, and other secretory vesicles.

MT plus-end tracking protein body formation and phase separation of MBPs

The EB family of +TIPs and associated MT regulators undergo condensation that is crucial for their regulation of MT growth. The yeast MT-actin crosslinking factor Kar9, Bim1 (an ortholog of vertebrate EB1, also known as MAPRE1) and Bik1 (an ortholog of CLIP-170, also known as CLIP1) co-condense to form +TIP bodies, both in vitro and in yeast. These +TIP bodies assemble on MT plus ends to mediate MTs crosslinking to actin cables during nuclear positioning (Meier et al., 2022). Similarly, the Schizosaccharomyces pombe EB homolog Mal3, the kinesin homolog Tea2 and the CLIP-170 homolog Tip1 coalesce via LLPS in vitro to associate with the MT plus end to form ‘comets’ (a feature of growing MT ends revealed in fluorescence microscopy) that track on MT tips (Maan et al., 2022). Cryo-electron tomography (cryo-ET) has revealed that Mal3, Tea2 and Tip1 form large, disordered structures on the MT tip, spanning hundreds of nanometers from the MT lattice. However, it remains unclear whether Mal3 mediates the co-condensation of Tea2 and Tip1 at MT ends in cells.

As EB1 tracks growing MT tips in cells, it forms comets that exhibit liquid-like properties (Song et al., 2022). Its paralog EB3 (also known as MAPRE3), together with CLIP-170, also undergoes LLPS in cells, and recruitment of tubulin into these condensates promotes MT growth (Miesch et al., 2023). An EB1 mutant in which lysine and arginine residues – both charged amino acids that mediate electrostatics important for LLPS – within the IDR were mutated (EB1-KR6Q) failed to co-partition with tubulin dimers, track on MT plus end tips or recruit additional MT trackers in HeLa cells (Fig. 1C) (Song et al., 2022). Furthermore, EB1 comet liquidity regulates MT dynamics in vitro – an EB1 chimera engineered to undergo a ‘liquid-to-gel’ transition was compromised in its ability to promote MT growth and reduce catastrophe frequency (Song et al., 2022). Although the roles of LLPS of EBs have only been recently studied in vitro, previous studies have characterized key roles for EB1 and EB3 in axon initial segment maintenance (Leterrier et al., 2011), motor-based delivery of dense core vesicles (DCVs) (Park et al., 2023), neuronal morphogenesis (Dema et al., 2023; Geraldo et al., 2008; Poobalasingam et al., 2022) and dendritic spine morphology (Pchitskaya et al., 2017; Sweet et al., 2011). Future work is required to understand how LLPS of EB proteins is spatiotemporally regulated in these processes. Altogether, these studies indicate that LLPS by +TIPs is directly linked to their ability to modulate MT dynamics.

Phase separation of MBPs in neurons

Maintenance of the highly polarized neuronal morphology is crucial for normal neuronal function (Wu et al., 2020). Rapidly growing research suggests that LLPS of MBPs plays a significant role in MT dynamics in vitro, warranting further exploration of how these mechanisms contribute to neuronal development. Recent reviews have examined LLPS in non-MBP proteins at the synapse (Chen et al., 2020) and during neuronal development (Wu et al., 2020). Here, we focus primarily on studies employing biochemical and cell-based assays to elucidate how the MBP phase separation contributes to axonal and dendritic processes and highlight potential MBP candidates whose phase separation capabilities are critical for their function.

Phase separation of MBPs in axons

Tau is a key modulator of MT assembly, dynamics and spatial organization in neurons. Consistent with previous studies on tau envelope formation in vitro (Siahaan et al., 2022; Tan et al., 2019), single-molecule tracking reveals nanoscale tau clusters within both the axon and presynaptic terminal (Fig. 2A) (Longfield et al., 2023). Presynaptic tau nanocluster formation was stimulated by synaptic activity and helped tether the recycling synaptic vesicle pool within the presynaptic compartment. LLPS-deficient tau mutants exhibit increased mobility at the presynaptic terminal, leading to heightened synaptic vesicle mobility (Longfield et al., 2023). It remains unclear how this increased mobility impacts synaptic transmission.

Fig. 2.

LLPS of MBPs in neurons. (A) Tau and MAP7 form envelopes (light green and red, respectively) around axonal MTs to sterically inhibit the severing proteins katanin and spastin and to selectively filter motor proteins, such as kinesin, dynein or both. (B) The actin nucleator Cobl undergoes LLPS on MTs and likely nucleates actin filaments upon reaching the cell periphery of axonal growth cones. (C) (i) Phase-separated Cobl likely nucleates actin filaments to also promote dendritic arbor growth. (ii) EB3 is hypothesized to undergo LLPS at MT plus ends to promote MT growth into spines. In response to NMDA receptor activation, MAP2 serves as a local MT reservoir that binds to EB3, redistributing it from MT plus ends to MT shafts. The post-synaptic density (PSD, shown in dark purple cloud) is a condensate formed by the scaffold protein PSD-95 (blue), which binds to SynGAP and GKAP. GKAP directs binding to Shank3, which recruits Homer3. Together, these PSD proteins condense to form a network that contributes to the structural integrity of dendritic spines. (D) In the synaptic bouton at axon terminals, DCLK1 might undergo LLPS on MTs to recruit kinesin-3 KIF1A molecules for targeted delivery of cargoes, including Syt-4-associated DCVs. The asterisks denote hypothesized LLPS, inferred through in vitro and in cellulo studies. Created in BioRender, Koleske, A., 2024. https://BioRender.com/c59n838. This figure was sublicensed under CC-BY 4.0 terms.

The actin nucleator cordon bleu (Cobl) colocalizes with both α-tubulin and β-actin in mouse hippocampal neurons, forming clusters through LLPS in vitro and in epithelial cells (Tsukita et al., 2023). Cobl contains three Wiskott–Aldrich syndrome protein homology 2 (WH2) domains, which are required for Arp2/3 complex-dependent actin nucleation. Cobl promotes dendritic arborization in several neuronal cell types – which requires all of the WH2 domains (Ahuja et al., 2007; Haag et al., 2012). In vitro reconstituted MTs serve as platforms for two-dimensional diffusion of Cobl, enabling Cobl molecules to condense into LLPS-mediated clusters that robustly enhance actin nucleation. MT-bound Cobl condensates employ this mechanism to potently stimulate actin polymerization in axonal growth cones and facilitate axon initiation and branching (Fig. 2B,Ci). Notably, MTs promote Cobl condensation to enhance actin polymerization at apical junctions in epithelial cells (Tsukita et al., 2023). Similarly, in developing primary hippocampal neurons, nanoclusters of syndapin 1 form at membrane surfaces, where they recruit Cobl and the related protein Cobl-like to promote actin nucleation at dendritic branch initiation sites (Izadi et al., 2021).

The difficulty of monitoring LLPS in cells has limited studies that test the importance of MBP phase separation in neurons (Alberti et al., 2019; Musacchio, 2022). However, several candidates have been suggested to exhibit LLPS in axons. For example, MAP7 is intrinsically disordered and recruits kinesin-1 to MTs to promote organelle transport in dorsal rat ganglion neurons (Tymanskyj et al., 2018; Hooikaas et al., 2019) (see above). The MAP7 projection domain tethers kinesin-1 to the MT, allowing the motor protein to diffuse locally to sites adjacent to MAP7 (Ferro et al., 2022). High MAP7 concentrations on MTs decrease kinesin recruitment, as kinesin-1 and MAP7 compete for binding sites on the MT. Hence, the density of MT-bound MAP7 must be tightly regulated in axons to fine-tune motor processivity for transporting mitochondria, endoplasmic reticulum (ER) and peroxisomes. A tau concentration gradient might also regulate kinesin motility in axons – low tau concentration at the cell body allows kinesin to bind to MTs and undergo anterograde transport, whereas high synaptic tau levels promotes kinesin dissociation and cargo release (Dixit et al., 2008). As MAP7 and tau also compete for the same MT-binding sites (Monroy et al., 2018), it will be interesting to explore how the LLPS of tau and/or MAP7 determines the correct distribution and balance of motor activity during axonogenesis.

CAMSAP3 is preferentially enriched in axons where it might undergo LLPS to stabilize axonal MTs, as its paralog CAMSAP2 has been shown to undergo phase separation in vitro. CAMSAP3 preferentially binds expanded MT lattices and decorates minus ends via its CKK domain, reducing the minus end growth and stabilizing newly severed MTs in vitro (Liu and Shima, 2023). In epithelial cells, LLPS of CAMSAP3 stabilizes the minus ends of MTs newly released from centrosomes (Dong et al., 2017). It is possible that LLPS of CAMSAP3 stabilizes minus ends of MTs destined for axons in neurons.

LLPS in dendritic spine maintenance

Several observations support the hypothesis that MBP condensates at MT plus ends might deliver actin remodeling factors or cargoes that are crucial for shaping the dendritic architecture. MT-based motors transport synaptic vesicle precursors, mRNAs, membrane proteins and organelles, some of which are destined for actin-rich postsynaptic sites (Kennedy and Ehlers, 2006, 2011; Martin and Zukin, 2006). MT-based motors or their cargoes are likely captured by actin or actin-associated factors and LLPS might serve as an efficient mechanism to mediate MT–actin crosstalk in dendrites and dendritic spines. Here, we will focus on the contributions of LLPS to dendritic arbor development and remodeling and discuss how LLPS promotes MT–actin crosstalk to regulate synaptic activity.

Transient dynamic invasions of single MTs from the dendritic shaft into spines are crucial regulators of dendritic spine morphology and synaptic function (Gu et al., 2008; Dent, 2017; Jaworski et al., 2009; Hu et al., 2008, 2011). The direct invasion of MTs into spines provides highways for MT-based motors to deliver cargoes containing key building blocks to the spine head (Dent, 2020). For example, the anterograde motor protein kinesin-3 (KIF1A) directly deposits synaptotagmin-4 (Syt-4) cargo into dendritic spines (McVicker et al., 2016). Depletion of KIF1A results in increased Syt-4 entry and exocytosis in spines, suggesting that MTs not only mediate the transport of materials, but might also help sequester vesicles to prevent inappropriate fusion with the plasma membrane. Interactions between MAP2 and EB3 – both of which undergo LLPS in vitro (Miesch et al., 2023; Siahaan et al., 2022) – regulate MT invasion of spines in response to synaptic activity (Fig. 2Cii). Activation of the ligand-gated NMDA receptors significantly reduces the number of EB3-containing MT plus-end comets, causing EB3 to be rapidly redistributed along MTs in the dendritic shaft via direct MAP2–EB3 interactions (Kapitein et al., 2011). Thus, MAP2 serves as a ‘sink’ for EB3 on MT bundles, the availability of EB3 to promote MT growth into dendritic spines. A key remaining question is whether EB3 and MAP2 form coacervates in neurons and how this process may be regulated by the downstream events from Ca²⁺ influx through NMDA receptor activation.

The post-synaptic density is a hotspot for MT capture in dendrites

The post-synaptic density (PSD), which clusters ligand-gated receptors along with their associated scaffolds and signaling partners (Cohen et al., 1977), is a dense interactome of proteins that behaves as a condensate (Chen et al., 2023) (Fig. 2Cii). Actin filaments make direct contact with the PSD and can nucleate at this site to support spine growth (Chazeau et al., 2014). Indeed, a purified PSD condensate containing the scaffolding and signaling proteins synaptic Ras GTPase-activating protein 1 (SynGAP), postsynaptic density protein 95 (PSD-95, also known as DLG4), Shank3, guanylate kinase-associated protein (GKAP, also known as DLGAP1) and Homer3 can promote actin polymerization, despite lacking any known actin-binding proteins (Chen et al., 2023).

The kinesin-4 motor KIF21B localizes to actin-enriched dendritic spine protrusions in cultured hippocampal neurons, where it colocalizes with PSD-95 (Gromova et al., 2023). KIF21B, typically associated with MTs, can ‘hitchhike’ on the actin-based motor myosin Va to localize to the PSD via its binding to GKAP. A separate study has revealed that KIF21A, which shares 61% protein sequence identity with KIF21B, but is distributed throughout neurons (Marszalek et al., 1999), promotes MT dynamics and regulates MT-based mitochondrial transport to support spine formation and morphology (Muhia et al., 2016). It remains an open question whether or how GKAP–KIF21B interactions contribute to the size of PSD condensates, which regulate the strength of synaptic transmission, and whether KIF21B directly interacts with PSD-95 via MT capture at the PSD.

Aside from MT-based motor proteins, MBPs like doublecortin-like kinase 1 (DCLK1) indirectly modulate MT-based cargo transport. DCLK1 is a conserved paralog of doublecortin (DCX), and mutations in the DCLK1 gene are associated with lissencephaly, a lack of gyri on the brain surface yielding a ‘smooth brain’ phenotype (Gleeson et al., 1999). DCLK1 has two MT-binding domains, a C-terminal serine/threonine kinase domain and an unstructured N-terminal tail that auto-inhibits MT binding via intramolecular interactions. Static TIRF images have revealed the formation of GFP-tagged DCLK1 puncta on MTs, which are reminiscent of tau and MAP2c envelopes on MTs (Agulto et al., 2021). DCLK1 uses its N-terminal MT-binding domain to interact with members of the kinesin-3 family to mediate DCV transport into dendrites and regulate dendritic outgrowth (Lipka et al., 2016). Depletion of both DCX and DCLK1 in developing neurons not only mislocalized KIF1A, but also the associated synaptic vesicle (SV) cargo, Vamp2 (Liu et al., 2012). Additionally, neurons deficient in DCX and DCLK1 displayed decreased run lengths of KIF1A-mediated Vamp2 transport in axons. Collectively, these data demonstrate that DCLK1 both regulate KIF1A motor-dependent trafficking of SV-associated proteins and DCVs into dendrites and axons (Fig. 2D). Future studies will determine whether DCKL1 undergoes LLPS in cells and how this process impacts MT binding to regulate KIF1A-mediated vesicle trafficking in different neuronal compartments. The next section will discuss how dysregulation of LLPS of MBPs contributes to the pathogenesis of neurodegenerative diseases.

Phase separation of MBPs in neurodegenerative diseases

Although LLPS enhances reaction kinetics and buffers or sequesters substrates, liquid-like condensates can transition into significantly less dynamic gel- or solid-like states. These states can evolve into protein aggregation associated with NDs (hypothetical model shown in Fig. 3) (Zbinden et al., 2020). In this section, we discuss contexts in which LLPS of MBPs becomes dysregulated and leads to NDs. For a more comprehensive overview, we direct readers to two general reviews on phase separation in NDs (Zbinden et al., 2020; Boyko and Surewicz, 2022).

Fig. 3.

Hypothetical model for the formation of NFTs implicated in neurodegenerative diseases, such as AD and FTD. Tau (green) binds and stabilizes axonal MTs. Tau phosphorylation (shown as ‘P’ in magenta) might increase the propensity for tau to self-associate, undergo LLPS, and form cohesive envelopes on MTs (light green). However, dysregulated and hyperphosphorylated tau molecules dissociate from MTs and form insoluble oligomers in solution. Tau oligomers associate into neurofibrillary tangles, which are hallmarks of many neurodegenerative diseases including Alzheimer's disease and frontotemporal dementia. Created in BioRender, Koleske, A., 2024. https://BioRender.com/y17c906. This figure was sublicensed under CC-BY 4.0 terms.

Under normal conditions, kinase and phosphatase signaling pathways converge on tau to govern its propensity to undergo LLPS and regulate organelle and mRNA transport along axons (Morfini et al., 2004). Dozens of mutations in the MAPT gene encoding tau have been identified in frontotemporal dementia (FTD), a disease characterized by language loss and personality changes. FTD is just one of several tauopathies that ultimately lead to dementia – a set of NDs that includes Alzheimer's disease (AD). AD is characterized by memory loss and deterioration of intellectual capacity. In AD, hyperphosphorylation of tau promotes tau droplet formation, depleting tau from its role in maintaining MT stability and inducing ER and unfolded protein response stress pathways (Wegmann et al., 2018). In addition, droplets formed from hyperphosphorylated tau can harden into irregular bodies that are less diffusive and fusogenic and stain with thioflavin S – a key marker for neurofibrillary tangles (NFTs), which are pathological hallmarks of AD. Importantly, some missense mutations in MAPT enhance droplet formation and alter protein dynamics within these droplets (Kanaan et al., 2020; Wegmann et al., 2018; Boyko et al., 2020). Together, these data strongly suggest that tau droplet formation by tau can shift the equilibrium toward pathological tau accumulation. Intraneuronal tau droplets can become pathogenic upon hyperphosphorylation (Biernat et al., 1992), by decreasing tau binding affinity to axonal MTs and reducing MT stability. The formation of pathogenic tau droplets likely promotes aggregation and the hardening of these aggregates into NFTs in the brains of individuals with AD and FTD (Wegmann et al., 2018).

Similarly, hyperphosphorylation of collapsin response mediator protein 2 (CRMP2, also known as DPYSL2) has been implicated in AD, HD and amyotrophic lateral sclerosis (ALS). Normally, CRMP2 promotes axonogenesis by stabilizing axonal MTs, preferentially binding to GTP-bound MTs and promoting their growth (Niwa et al., 2017). Although CRMP2 does not aggregate on its own, hyperphosphorylation reduces its binding to soluble tubulin, leading to axon retraction and eventual degeneration. Interestingly, hyperphosphorylated CRMP2 has been observed alongside tau in NFTs, and both proteins are targets of increased phosphorylation by GSK3β and CDK5 in AD (Cole et al., 2007). A key question is whether or how modified CRMP2 and tau form coacervates, and how phosphorylation tips the balance between the formation of liquid-like droplets and NFTs.

Like tau and CRMP2, prion protein (PrP), a cell surface glycoprotein, is highly expressed in the nervous system; however, PrP binds tubulin but not MTs (Nieznanski et al., 2006, 2005; Dong et al., 2008). PrP is normally localized in the outer membrane of neurons but can be translocated into the cytosol as cytosolic PrP (PrP^C). PrP^C can undergo conversion into the pathogenic scrapie conformation (PrP^Sc), triggering the onset of prion disease. PrP^Sc aggregation initiates a chain reaction of PrP misfolding that ultimately leads to transmissible spongiform encephalopathies (TSEs) – a set of fatal neurodegenerative diseases that induces neuronal loss and motor and cognitive impairments. Higher-order PrP aggregates recruit tubulin, thereby inducing it to adopt irregular structures not amenable for MT nucleation (Dong et al., 2008; Nieznanski et al., 2006). PrP undergoes phase separation in vitro mediated by its unstructured octarepeat region, which is prone to disease-associated mutations (Kamps et al., 2024; Tange et al., 2021; Kamps et al., 2021; do Amaral et al., 2023; Perera and Hooper, 2001). Clustering of PrP^C on the cell surface precedes its conversion into PrP^Sc (Rouvinski et al., 2014), suggesting that LLPS of PrP^C is involved in the aberrant liquid–solid phase transition. Interestingly, tau inhibits PrP-induced tubulin oligomerization in vitro, likely by competitive binding for tubulin and/or by stabilizing a tubulin dimer structure that is less susceptible to oligomerization (Osiecka et al., 2011). Future studies are required to determine what distinguishes physiological PrP condensation from PrP aggregation implicated in TSEs and explore the potential interplay between LLPS of tau and PrP in regulating MT stability and tipping the balance towards critical pathogenesis.

How do non-pathological liquid condensates evolve into pathological aggregates? Under physiological conditions, TDP-43 (also known as TARDBP) functions as a nuclear protein that regulates transcription, pre-mRNA splicing, and translation. However, under oxidative stress, TDP-43 can leak into the cytoplasm and aggregate in uniformly distributed stress granules. Then, TDP-43 begins to form puncta within the granules that exhibit reduced dynamics and recruit additional nuclear TDP-43 (Yan et al., 2024 preprint). This intra-condensate de-mixing promotes a gain-of-toxicity that might underlie ALS progression. A key focus for future research will be to understand how liquid de-mixing promotes the formation of ND-associated MBP aggregates and whether this process can be targeted therapeutically to impede the progression of NDs.

Remaining questions and future perspectives

LLPS of MBPs has been increasingly well-studied in vitro, where it promotes localized macromolecular functions. However, in most cases, the specific mechanisms and regulatory factors governing LLPS in cells and neurons remain unclear. Significant questions persist – what cues or factors initiate LLPS to ensure the proper spatiotemporal concentration of key MT regulators? For example, when and how do tau and MAP2 molecules form envelopes on axonal and dendritic MTs to protect against severing and regulate motor protein and cargo trafficking in neurons? Do phase-separated MT regulators perform similar functions in neurons as they do in vitro? Additionally, how do other proteins or post-translational modifications impact the dissipation of MBP condensates when LLPS is no longer necessary? Addressing these crucial questions is essential for ongoing and future studies.

Rigorous tests have been developed to measure LLPS in vitro (Alberti et al., 2019; Musacchio, 2022), but assessing whether proteins undergo LLPS in cells presents more challenges. Qualitative measures have relied on measurements of shape (roundness) of prospective condensates and/or their ability to undergo optogenetically controlled fusion and fission (Shin et al., 2017). However, these methods are insufficient for definitively confirming LLPS (McSwiggen et al., 2019; Mittag and Pappu, 2022; Musacchio, 2022). Moreover, these approaches presume that condensates are spherical, as they exist in three-dimensional compartments (e.g. cytoplasm) and can undergo fusion and fission. However, this is not the case for all condensates, as some exist on the two-dimensional lipid bilayer (Case et al., 2019). Whereas chemical fixation can preserve cellular architecture, paraformaldehyde fixation can alter the appearance of condensates, raising concerns about its utility in studying LLPS in cells (Irgen-Gioro et al., 2022). Selective photobleaching can characterize the liquidity of putative condensates within cells, but its effectiveness is limited for condensates smaller than the resolution of light microscopy – for example, presynaptic active zones and PSDs typically range from 200 to 500 nm in diameter (Südhof, 2012).

Emerging technologies, such as super-resolution microscopy and cryo-ET, have been used in proof-of-principle studies to demonstrate the feasibility of measuring macromolecular diffusivity in neurons and resolving both phase-separated internal structures of reconstituted cellular assemblies (see Box 2). Ultimately, more cell-based studies incorporating these technologies are required to dissect how MBP LLPS is controlled in different neuronal compartments. Such studies will uncover the fundamental cellular biophysical principles by which LLPS regulates MT dynamics and stability, which will enable the development of effective therapeutic strategies for NDs. For example, it has been demonstrated that the small-molecule drug SM875 can recognize and target a PrP folding intermediate for proteasomal degradation, thereby lowering PrP levels, blocking the conversion of PrP^C into PrP^Sc and limiting prion replication in murine fibroblasts (Spagnolli et al., 2021). Understanding the principles underlying LLPS might pave the way for developing therapies to inhibit pathological MBP aggregation, potentially preventing the development and progression of NDs.

Box 2. Emerging technologies to study LLPS behavior in cells.

Improvements in fluorescent dye-labeling strategies and single-molecule imaging techniques now enable high-speed tracking of individual proteins entering diffraction-limited compartments, such as synaptic boutons. For example, synapsin 1 (Syn1) can undergo LLPS in vitro (Hoffmann et al., 2023). To assess its diffusivity in neurons, Halo7- and mEos3.2-tagged Syn1 molecules were tracked using a super-resolution confocal microscope (top panel). This identified two distinct populations of Syn1 – one slow-diffusing population confined to synaptic boutons, presumably within a phase condensate, and another that exhibited high diffusivity between boutons. The IDR of Syn1 is required for the protein to adopt the slow-diffusing state, which promotes clustering of SVs into boutons and modulates their diffusion in hippocampal neurons. The rapid dynamics of Syn1 within condensates, alongside its confinement in boutons, strongly suggest that synapsins use LLPS to regulate SV clustering for neurotransmitter release during exocytosis (see figure, upper panels).

Cryo-ET has also emerged as a leading structural biology tool for visualizing condensates of heterogeneous sizes − even for those in transition from liquid-like to solid-like physical phases (Goetz and Mahamid, 2020). Cryo-ET has revealed the molecular organization of condensates at various stages; for example, in tetranucleosome (a structural unit of a chromatin fiber) LLPS (see figure, lower panels) (Zhang et al., 2022). Neurons are particularly well-suited for cryo-ET applications due to the negligible volumes of axons and dendrites (Zuber and Lučić, 2022). Indeed, cryo-ET has facilitated the visualization of subcellular structures, including axon branching sites (Martinez-Sanchez et al., 2021; Nedozralova et al., 2022), providing high resolution snapshots of cellular assemblies in rodent cerebrocortical synaptosomes and neurons (Martinez-Sanchez et al., 2021), as well as in mammalian dorsal root ganglion neurons and Drosophila neurons (Foster et al., 2021). Specifically, cryo-ET has been used to understand the structure and composition of protein aggregates, as demonstrated in the case of inclusion bodies associated with Huntington's disease (HD) (Bäuerlein et al., 2020). It has also been used to investigate changes in organelles in iPSC-derived neurons from individuals with HD (Wu et al., 2023).

Upper panels from this figure are reprinted from Hoffmann et al. (2023) where they were published under a CC-BY 4.0 license; lower panels from this figure are reprinted from Zhang et al. (2022) with permission from Elsevier.