Small molecule modulation of a redox-sensitive stress granule protein dissolves stress granules with beneficial outcomes for familial amyotrophic lateral sclerosis models

Neurodegeneràve diseases such as amyotrophic lateral sclerosis (ALS) are oten associated with mutàons in proteins that are associated with stress granules. Stress granules are condensates formed by liquid-liquid phase separàon which, when aberrant, can lead to altered condensàon behaviours and disease phenotypes. Here, we identified lipoamide, a small molecule which specifically prevents cytoplasmic condensàon of stress granule proteins. Thermal proteome profiling showed that lipoamide preferentially stabilises intrinsically disordered domain-containing proteins. These include SRSF1 and SFPQ, stress granule proteins necessary for lipoamide activity. The redox state of SFPQ correlates with its condensate-dissolving behaviour, in concordance with the importance of the dithiolane ring for lipoamide activity. In animals, lipoamide ameliorates aging-associated aggregàon of a stress granule reporter, improves neuronal morphology, and recovers motor defects caused by expression of ALS-associated FUS and TDP-43 mutants. In conclusion, lipoamide is a well-tolerated small molecule modulator of stress granule condensàon and dissection of its molecular mechanism identified a cellular pathway for redox regulàon of stress granule formàon.


Introduction
ALS is a fatal neurodegenerative disease with poor prognosis and few options for therapy 1 . Most forms of ALS are sporadic but around 10% are monogenic disorders 2 . Currently only two FDA approved drugs are available, riluzole and edaravone, however both only slow disease progression by a few months 3,4 . The mechanisms of riluzole and edaravone are not well understood and most likely do not directly target the underlying pathomechanism 3 . New approaches are therefore needed.
The precise mechanism of ALS pathogenesis is not known. Recent work has focused on familial ALSassociated mutations that are frequently found in RNA-binding proteins (RBPs) such as FUS (fused in sarcoma) and TDP-43 (TAR DNA-binding protein 43) 5,6 which have characteristic low complexity domains (LCDs) and are part of a broader class of intrinsically disordered proteins. Mutated TDP-43 and FUS often mislocalise to the cytoplasm where they promote stress granule formation and sometimes form cytoplasmic aggregates associated with disease [7][8][9] .
Stress granules are inducible RNP granules that form in the cytoplasm of eukaryotic cells upon stress.
They contain non-translated mRNAs and are thought to be involved in the shut-down of translation during stress 10 . Stress granules have attracted considerable attention in recent years because of their relationship to human disease 6,11,12 . Evidence is accumulating that aberrant forms of stress granules can result from liquid-to-solid transitions, and that these aberrant forms of stress granules underlie disease 7,[13][14][15] . This suggests in turn that dissolving stress granules and/or aggregated stress granule proteins could help ameliorate disease.
Stress granules are liquid-like membraneless compartments that have been proposed to form by liquid-liquid phase separation 7,[16][17][18] . In principle, it might be possible to target the physical chemistry driving stress granule formation to modulate them 19 . Indeed, there are a few compounds known to disrupt stress granule phase separation: 1,6-hexanediol [20][21][22] and similar alcohols 23,24 . However, these compounds suffer from two problems: First, they require extremely high concentrations (1-10% v/v) and are toxic 23,25 . Second, the effects of these alcohols are not specific to stress granules and affect other membraneless liquid compartments that are also liquid-like, such as the nucleolus. This suggests that identifying compounds that more specifically modulate phase separation of certain compartments could be important in therapeutic intervention.
We screened a small library of medicinal compounds (1600 drugs and compounds with known biological targets) to discover those that affect stress granule formation, using FUS as a wellcharacterised stress granule component target. This identified lipoamide and lipoic acid, which are non-toxic and relieved the effects of ALS-associated FUS mutations in vivo in different systems, namely, stress granule protein aggregation in worms (Caenorhabditis elegans), motor dysfunction in a Drosophila melanogaster model of familial ALS, and the motor neuron die-back in vitro in patientderived iPSCs carrying a FUS mutant associated with familial ALS. Together these suggest a plausible novel route to ALS therapeutics.

Results
Multi-parameter cellular image analysis screening identified lipoamide as a molecule that reduces stress granule formation We developed a cell-based screen using a HeLa cell line that stably expresses a GFP-FUS fusion protein at near-endogenous levels 7 . While not required for stress granule formation, FUS is a wellcharacterised protein with a domain structure typical of stress granule proteins 26 . In the absence of cellular stress, FUS-GFP primarily localizes to the nucleus, where it is partially excluded from the nucleolus ( Figure 1A). It also localises to small puncta called paraspeckles -these nuclear condensates are implicated in retaining RNA in the nucleus for rapid stress response 27 . In stressed cells, FUS is partly exported to the cytoplasm where, in combination with other proteins and mRNA, it phase separates to concentrate in liquid-like stress granules surrounded by cytoplasm consequently depleted in FUS 7,28 ( Figure 1A). In our screen we pre-treated cells with 10 μM compound from a compound library for 1 h, then stressed the cells with 1 mM potassium arsenate (still in the presence of compound) ( Figure 1B) and monitored FUS localization to stress granules.
Arsenate disrupts antioxidant responses by reacting with thiol groups, blocks the tricarboxylic acid (TCA) cycle by reacting with the thiols in vital lipoyl moieties and causes general oxidative damage 29,30 .
We used multi-parameter image analysis to determine the effect of compounds on FUS localisation in stressed cells -an image-based approach was necessary to identify compounds affecting stress granule formation rather than using a biochemical readout. Compounds were ranked by strength of effect on FUS localisation ( Figure 1C) with cytoplasmic puncta number, nuclear puncta number and nuclear/cytoplasm partition typically showing the largest changes. Two compounds in the library were expected to give a reduction of stress granules. These were the polysome stabilising compound emetine 31 , which prevents release of mRNA, and the heavy metal chelating compound dimercaprol, which chelates arsenic and has beneficial effects in animals on arsenate toxicity 32 ( Figure 1C).
Notably edaravone, thought to be an antioxidant 33 and used as an ALS therapeutic in Japan and the USA 34 , had no effect on FUS localisation in arsenate-stressed cells. Compound classes which tended to have a large effect on FUS localisation, typically reducing stress granule number, included cardiac glycosides, heterotri-and tetracyclic compounds (anthraquinones and acridines), surfactants and benzimidazoles.
The compounds with the strongest effects on intracellular FUS localisation were then tested for direct interaction with FUS condensate droplets formed under low salt conditions in vitro, carried out in the presence of 1 mM DTT to mimic the reducing intracellular environment. In a moderatethroughput screen we determined the effect of the 47 strongest in cell hits on the number of FUS condensate droplets formed or the ratio of FUS GFP signal inside the droplets to outside (partition into the droplets) ( Figure 2A). Of these 47 compounds 7 significantly affected FUS in vitro ( Figure 2B) and fell into just three compound classes: heterotri-and tetracyclic compounds, surfactants and lipoamide ( Figure 2C,D). In this non-equilibrium snapshot, the heterotri-and tetracyclic compounds tended to reduce condensate formation in a dose-dependent manner and the resulting droplets were smaller. In contrast, surfactants and lipoamide tended to increase condensate droplet number, partition into droplets and the droplets were larger ( Figure 2B,D). Surfactants are not plausible therapeutics as they will permeabalise cell membranes, and were present in the library due to their use as topical antiseptics. Both lipoamide and its closely related compound lipoic acid are non-toxic and lipoic acid has well-characterised pharmacokinetics: 1,600 mg orally gives plasma concentrations of 8 to 30 μM in humans 35,36 and has a long history of use in diabetic neuropathy therapy at doses around 600 mg/day 37 . Through the remainder of this work we therefore selected the lipoamide class as of primary interest with some comparison to mitoxantrone as an example of a heterotricyclic compound.

Lipoamide action in cells is non-enzymatic and non-antioxidant
The stressor, arsenate, will react with the thiol groups of lipoamide. Therefore, to exclude the possibility that lipoamide is acting only by removing the stressor we tested whether it can prevent stress granule formation triggered by other non-arsenate stresses: Mitochondrial electron transport chain inhibition (rotenone), heat stress (42°C), hyperosmotic stress (sorbitol, a non-metabolisable sugar), glycolysis inhibition (6-deoxyglucose in the absence of glucose) or serum starvation. 10 μM lipoamide reduced stress granule formation in HeLa cells with mitochondrial, hyperosmotic or arsenate stress but not heat or glycolysis stress ( Figure 3A). Lipoamide is therefore not only reacting with arsenate but reduces stress granule formation under several cellular stresses.
Lipoic acid is a compound similar to lipoamide which plays a part in normal cellular metabolism and was a weaker hit in the cell screen ( Figure 1C). Of the two lipoic acid stereoisomers R-(+)-lipoic acid naturally occurs in cells and is synthesised in the mitochondrion while S-(−)-lipoic acid is not.
However R-(+)-lipoic acid is present at very low free concentrations in the cell -it is biosynthesised, and therefore normally found, covalently bonded to proteins as a lipoyl moiety attached to the sidechain of lysyl residues as a lipoamide. These compounds are cyclic disulfides and the corresponding reduced dithiol forms of the lipoyl moiety are used by several mitochondrial metabolic enzymes, including dihydrolipoyl transacetylase feeding into the TCA cycle and dihydrolipoamide succinyltransferase in the TCA cycle. The R-(+) isomer is reduced from the cyclic disulphide to the dithol state in cells by dihydrolipoamide dehydrogenase 38 , however its natural substrate is the lipoyl moiety rather than free lipoic acid. We determined whether 10 μM R-(+), S-(−) or racemic (±, a mix of both stereoisomers) lipoic acid also reduced stress granule formation in HeLa cells. This showed both lipoic acid isomers also reduce stress granule formation under mitochondrial, hyperosmotic or arsenate stress, as seen for lipoamide ( Figure 3A).
Next, we analysed the dose dependent activity of lipoamide, lipoic acid and related compounds.
Lipoamide gave a reduction of stress granule number with an EC50 around 20 µM in both HeLa cells and induced pluripotent stem cells (iPS) cells following 1 h of 1 mM arsenate stress ( Figure 3B,C).
Lipoamide also caused FUS to return to the nucleus, as shown by a dose-dependent increase in nuclear/cytoplasm partitioning of FUS ( Figure 3B,C). We noted that the heterotri-and tetracyclic compounds, including mitoxantrone, had the inverse effect on nuclear/cytoplasm partition ( Figure   S1). Return of FUS to the nucleus is likely beneficial, mimicking the unstressed cell state 39,40 , indicating lipoamide is a more promising candidate.
In HeLa cells (±)-dihydrolipoic acid and (±), R-(+) and S-(−)-lipoic acid had a similar EC50, a little higher than (±)-lipoamide ( Figure 3D,E). Valeric acid and 1,3-propanedithiol had no effect up to 100 μM ( Figure 3D,E). In the cellular reducing environment it is likely that the lipoamide/lipoic acid cyclic disulfide is reduced to a dithol and the similar EC50 of (±)-dihydrolipoic acid and (±)-lipoic acid is consistent with this. Lipoic acid has been proposed to be an antioxidant 41 , although evidence for direct action as an antioxidant is disputed 42 . Cellular antioxidant effects involving an enzymatic redox cycle will involve dihydrolipoamide dehydrogenase, which is more active for the naturally occurring R-(+) isomer although normally reduces lipoyl moieties 38,43 . However S-(−)-lipoic acid had very similar activity to R-(+)-lipoic acid ( Figure 3C,D). We also note that classic antioxidants such as menadione (pro-vitamin K) and edaravone did not reduce stress granule number in the primary HeLa cell screen.
Any enzymatic role in the activity of lipoic acid as a coenzyme/addition to an apoenzyme requires a lipoate-protein ligase 44 , likely specific to R-(+)-lipoic acid. The similar activity of S-(−) and R-(+)-lipoic acid and absence of a clear human lipoate-protein ligase indicates this is unlikely ( Figure 3C,D).
Together, this indicates a non-antioxidant and non-metabolic mechanism of action; perhaps a stress signaling mechanism or a direct effect on the physical chemistry of phase separated stress granule proteins.

Lipoamide/lipoic acid dissolves stress granules
To gain insight into the breadth of action of lipoamide and lipoic acid, we comprehensively characterised their action in HeLa cells. We first determined that these compounds can lead to dissolution of existing stress granules by showing that addition of fresh medium after stress, containing 10 μM lipoamide and 1 mM arsenate, dissolved 80-90% of cytoplasmic FUS condensates within 20 minutes ( Figure 4A). This response is likely too rapid to represent a transcriptional response. 1 h pre-treatment of HeLa cells with 10 μM lipoamide followed by 1 h arsenate stress without lipoamide did not prevent stress granule formation, indicating no persistent lipoamideinduced cellular adaptation to resist stress ( Figure 4B).
Because several FUS-like proteins in stress granules are implicated in ALS, we anticipate a useful therapeutic would prevent other PLD-containing stress granule proteins from localising to stress granules. We used 1 mM arsenate to stress a panel of HeLa cell lines expressing GFP fusions of stress granule proteins (EWSR1, TIAL1, PABPC1, G3BP1) with either no treatment or 10 μM (±), S-(−) or R-(+)-lipoic acid or (±)-lipoamide. All proteins localised to stress granules in the absence of treatment.
In the presence of lipoic acid or lipoamide the proteins either did not localise to stress granules or localised to a reduced number of stress granules ( Figure 4C). All tested LCD/RBP proteins are therefore affected by lipoamide and lipoic acid treatment. As FUS is not required for stress granule formation this suggests that FUS is not the only nor necessarily a primary target of lipoamide or lipoic acid in cells. It also suggests a possible therapeutic effect on pathology arising from mutation of other stress granule proteins.

Lipoamide does not affect other cytoplasmic condensates
Stress granules are one of many important biomolecular condensates. We anticipate that a useful therapeutic would not affect all these compartments therefore we asked to what extent lipoamide was specific to affecting stress granule dissolution, using a panel of cell lines expressing GFP fusions 45 . For this analysis, we included mitoxantrone as a representative of the heterotri-/tetracyclic compound class of hits ( Figure 2C). Lipoamide, under conditions that dissolve stress granules, did not affect the localization of RNA processing body (DCP1A, P-body, cytoplasmic), Cajal body (COIL, nuclear), DNA damage focus (TRP53BP1, nuclear) or nucleolus (NCB1, nuclear) proteins ( Figure 5A).
In contrast, mitoxantrone affected all of these compartments to some extent, while histone localisation suggested nuclear structure overall was not dramatically affected ( Figure 5A). Different non-mixing membraneless organelles likely have differing intermolecular interactions underlying their phase separation. The mode of action of lipoamide may therefore be specific to the physical chemistry driving stress granule phase separation and/or a specific signaling route. We note that stress granules normally form and dissolve rapidly in comparison to many other compartments.
In addition to testing nuclear compartments formed by other proteins we tested how the compounds influenced de novo formation of FUS-containing nuclear compartments by the recruitment of FUS to sites of DNA damage. Defective recruitment of FUS to sites of DNA damage is detrimental: FUS nuclear localization (NLS) mutations (e.g. P525L) cause altered DNA damage signalling 46 and are strongly associated of with familial ALS mutations 39 . We therefore tested whether lipoamide affect FUS GFP recruitment to sites of DNA damage in iPSCs induced by focused UV laser irradiation. We used arsenate-stressed iPSCs so that stress granules were present as an internal control for compound activity in the cytoplasm ( Figure 5B). 20 μM lipoamide (which dissolved stress granules) had no significant effect on FUS GFP recruitment to sites of DNA damage.
1 μM lipoamide (which did not affect stress granules), increased recruitment of FUS GFP to sites of DNA damage, likely to be beneficial. In contrast, mitoxantrone blocked recruitment of FUS GFP to sites of DNA damage, at concentrations which were not sufficient to dissolve cytoplasmic FUS condensates.

Lipoamide affects FUS condensate behaviour in vitro.
We next sought to analyse the nature of direct lipoamide effects on FUS using FUS condensates in vitro. As for the in vitro screen, all assays contained excess (1 mM) of the reductant DTT which will reduce the lipoamide disulfide to free dithiols. The prior screen (Figure2, see above) had suggested drop enlargement in the presence of lipoamide. However, lipoamide at 100 µM had little effect on phase separation of FUS GFP -the resulting condensates had very similar threshold salt concentration and temperature for droplet formation (data not shown). One possibility is that lipoamide made the condensates more liquid-like. Following phase separation, average condensate droplet size increases while condensate droplet number decreases as large droplets grow at the expense of small droplets (Ostwald's ripening) and droplets fuse -both of which occur more quickly with more liquid droplets 47 . We therefore measured lipoamide effect on FUS eGFP condensate liquidity (surface tension and viscosity). Here, condensate droplets were brought together by optical tweezers and the rate of fusion and relaxation to a sphere was measured 7 ( Figure 6A). Lipoamide increased the liquidity by a factor of three ( Figure 6B). Together, the combined result of in vitro analyses suggests that the observed ( Figure 2B) increases in droplet size and number arise from faster generation of larger droplets, due to increased liquidity, followed by sedimentation of the larger droplets (see materials and methods for more details).
In vitro FUS condensates 'age' over time, first hardening then forming amyloid/prion-like fibres 7,48,49 , even in the presence of 1 mM DTT as a reducing agent. This is likely represents a pathological process as it is accelerated for G156E FUS, a mutation associated with familial ALS 7 . As lipoamide/lipoic acid did not cause dissolution of FUS droplets in vitro we were able to examine the effect of lipoamide and lipoic acid on fibre formation and hardening, initially analysing FUS condensates formed under conditions of (dextran-induced) crowding (as previously published 7 ) ( Figure 6C Finally, we analysed mechanistic detail at the sub-molecular level by using NMR of the FUS prion-like N-terminal LCD 18 to determine putative sites of lipoamide interaction. The LCD is able to phase separate to form condensates in vitro and the individual residues can be resolved and assigned in a

Lipoamide becomes greatly concentrated in cells
The concentration of lipoamide relative to protein required for an effect on condensate liquidity in vitro was higher than the EC50 on cells, on the order of 100 µM lipoamide with 1 µM FUS ( Figure   2A,B, Figure 6). In cells, FUS concentrations are high for a protein (low µM) 50 and the cellular EC50 was on the order of 10 µM ( Figure 3A,B). We also saw inverse effects of lipoamide on FUS-containing condensates in vitro (more, larger condensates) and in cells (fewer stress granules). To reconcile this difference and to understand the relevance of the effects in vitro in cells, we wanted to know the actual concentrations of compounds in cells. In principle, isotopic labelling allows direct monitoring of isotopically-labelled compounds even in complex environment through appropriate spectroscopy 51 . To use this approach, 15 N-labelled lipoamide ( Figure 7A) was synthesised. Solution state NMR experiments were then used to quantitatively detect the NH2 protons covalently attached to 15 N in cells to determine lipoamide concentration from the complex mixture inside cells 51 , while also revealing any chemical modification of the amide group (manifesting as chemical shift changes or spectral alterations) ( Figure S3).
Uptake of 15 N lipoamide by cells from the medium was quantified by NMR following incubation of lipoamide at 37 o C for 1 h either in the absence or presence of HeLa cells ( Figure 7B). Cellular uptake was determined by calculating the difference between medium incubated with cells or without cells for combinations of either R-(+) or (±)-lipoamide and either unstressed or stressed cells ( Figure 7C).
For one sample, stressed cells with R-(+)-lipoamide, we confirmed that strong signal from the cell fraction was consistent with uptake of a large proportion of lipoamide ( Figure 7C). Both R-(+) and ±lipoamide measurements indicated uptake of 35±11% (n=3 and 2, respectively) of the lipoamide present in the medium. No significant difference was observed either for uptake of R-(+) compared to (±)-lipoamide or stressed compared to unstressed cells ( Figure 7D). There was no evidence for metabolism or any other chemical modification of lipoamides: the NMR signal from the cell sample indicated that lipoamide was present in an unmodified form in the cells ( Figure 7B).

Lipoamide/lipoic acid prevents stress granule protein aggregation in vivo
As lipoamide and lipoic acid had a large effect on FUS aggregation in vitro we used a filter-trap retention assay (in which aggregated protein from cell lysate tends to be retained on a membrane) 52 to test whether lipoamide had a beneficial effect on spontaneous aggregation of wild-type or P525L FUS GFP in iPSCs ( Figure S4). Both cell lines had evidence for some FUS aggregation, which was reduced following treatment with lipoamide ( Figure S4B).
To look for evidence of in vivo effects on stress granule formation and protein aggregation we turned to Caenorhabditis elegans. In C. elegans, ageing or chronic stress are associated with aggregation of stress granule proteins with an LCD/RBP domain structure (including the orthologs of TIAL1 and PABC1) 53 . This has parallels with the observation of aggregated stress granule proteins in ALS pathogenesis. Notably, C. elegans grown in liquid culture with R-(+) or S-(−)-lipoic acid showed a dose-dependent decrease in the proportion of animals with aggregation of PAB-1 ( Figure 8A,B), the ortholog of PABC1 (analysed in HeLa cells in Figure 4C). At the highest concentration tested (2 mM) there was some toxicity leading to 6 to 8% worm death. Testing of lipoamide was not possible due to precipitation in the worm culture medium.
To investigate whether this action of lipoic acid is specific to RBPs with LCDs, we tested the aggregation of two globular proteins KIN-19 and RHO-1, which have been previously shown to aggregate with age in C. elegans 54 . Neither protein has an RNA-binding domain or LCD. We found no effect of 1.5 mM lipoic acid on RHO-1 aggregation and a slight decrease in KIN-19 aggregation ( Figure 8A). Thus, lipoic acid can affect stress granule solidification in vivo on the timescale of an organism lifespan, correlating with its behavior on short time scales in vitro. This is consistent with its direct interaction with stress granule proteins to reduce stress granule formation and/or stress granule protein aggregation.

Lipoamide/lipoic acid recover neuron and organism FUS-associated defects
P525L is a well-characterised example of an ALS-associated FUS NLS mutation and iPSC-derived motor neurons (iPSC MNs) expressing P525L from ALS patients show defects consistent with motor neuron defects in the patient 46 . iPSC MNs can be grown through silicone channels, positioning the soma on one side with the axons protruding through the channels to the far side. Such cultures can be maintained for long periods (>60 days). However, in this system, P525L FUS MNs exhibit cellular defects associated with neurodegeneration, including axonal die-back and reduced axonal transport.
This occurs even without exogenous cell stressors, although P525L FUS MNs have a greater propensity to form stress granules 46 , and has similarities to the axon retraction leading to motor dysfunction in ALS patients.
We therefore tested whether lipoamide or lipoic acid was able to improve these phenotypes. FUS P525L neurons initially appear morphologically normal ( Figure 8D) but by 60 days in culture diedback axonal material has accumulated around the exit of neurons from the silicone channels ( Figure   8D). Inclusion of 2 μM lipoic acid or lipoamide in the culture medium prevented the die-back ( Figure   8C-E). As defective axonal transport is thought to lead to axon die-back 55 we analysed the transport of lysosomes within the axons of iPS MNs expressing P525L FUS with or without lipoamide in comparison to iPS MNs expressing WT FUS ( Figure 8F). Lipoamide recovered axonal transport in P525L FUS iPS MNs to the same level as WT neurons. This is a greatly beneficial effect of lipoamide in a minimal ex vivo model using human cells in a long-term treatment regime. As no stressor is needed to induce axon die back, we suggest that lipoic acid either helps iPSC MNs handle stochastic stress in culture or facilitates the amelioration of a defect caused by P525L.
Next, we tested whether lipoic acid could have a similar beneficial effect on motor neurons in vivo using a fly model. The Drosophila menalogaster FUS ortholog, cabeza, is required for normal neuronal development 56

Discussion
In this paper, we identify lipoamide and lipoic acid as promising therapeutics for ALS. These compounds have a long and complex history as bioactive molecules however their mechanism of action is ambiguous. Stress granules have been proposed to be crucibles of ALS pathogenesis 6 therefore we screened for compounds that affect the physical chemistry of a model stress granule protein, FUS. Here, we showed lipoamide altered the properties of phase separated FUS condensates in vitro ( Figure 6A Stress granules form by phase separation of proteins within the cytoplasm and a few small molecules have previously been identified which modulate phase separation (notably 1,6hexanediol), however lipoic acid and lipoamide have far more plausibility as a therapeutic.
Concentrations of 1,6-hexanediol between 1 to 10% (several hundred mM) are required for in vitro or cellular activity, and this is rapidly toxic 23,25 . The molecules we identified in our screen (both heterotri-and tetracyclic compounds and lipoamide and related compounds) were active in vitro (Figure 2,6) and on cells (Figure 1,3,S1) at vastly lower concentrations (tens to hundreds of µM), three to four orders of magnitude lower than 1,6-hexanediol. The effect of lipoamide on stress granules is also specific, in the sense that lipoamide did not affect other membraneless liquid like compartments in cells -it did not affect nuclear compartments formed by FUS ( Figure 5B) or formed by other proteins either in the cytoplasm or nucleus ( Figure 5A) and was well-tolerated by HeLa, iPS and motor neuron cells in culture.
The necessary properties of a compound affecting protein phase separation are likely different to those of a canonical drug binding a well-defined structured protein site. Condensates are formed by protein liquid-liquid phase separation involving many transient interactions 59 , unlike strong enzymesubstrate or protein-protein interactions typically targeted by drugs. Small molecules can interfere with phase separation and alter the properties of phase boundaries (surfactants). For instance, ATP has been identified as a hydrotrope which helps keep proteins soluble 60 (Figure 6), appearing to affect the kinetics rather than thermodynamics of FUS phase separation. One likely reason is that lipoamide accumulates to remarkably high concentrations in cells (Figure 7) where it is likely reduced but is not otherwise metabolised ( Figure S3). The estimated cellular concentration (approximately 5 mM) is one order of magnitude higher than we were able to reach in vitro (300 µM), and much higher than the concentration necessary to increase the liquidity and reduce the hardening of phase separated FUS in vitro (Figure 6), however it may also affect cell volume through osmotic effects. It is possible that 5 mM lipoamide would dissolve FUS condensate droplets in vitro. There are also other potential explanations. Firstly, higher liquidity of stress granules could make them more sensitive to cellular stress granule dissolving factors. Secondly, that cells have stress granule, cytoplasm and nuclear environments, making it a more complex three phase system (perhaps related to the lipoamide effect on FUS nuclear/cytoplasm partition, Figure 3). Finally, lipoamide has a strong effect on the solidification of FUS (Figure 6), and this is thought to be driven in part by LCD-LCD interactions that are not required for phase separation in vitro 26 . If LCD-LCD interactions are more important for phase separation inside cells then this may manifest as an increased sensitivity stress granules to lipoamide relative to FUS in vitro. The utility of lipoamide/lipoic acid as a therapeutic presumably depends on the balance between stress granules promoting ALS pathogenesis and stress granule formation as a beneficial cellular response to stress. Similarly, whether stress granule protein aggregation is a means of sequestering a harmful protein, whether it is intrinsically harmful, or if it is detrimental to other (for example, nuclear 46 ) stress granule protein functions. The strong association of FUS NLS mutations and the capacity for nuclear import of FUS with ALS gives particular plasubility to the latter 39,40,46,61 . Our motor neuron and D. melanogaster models of ALS showed that the pattern of action of lipoamide and lipoic -dissolving stress granules, returning FUS to the nucleus, not affecting FUS recruitment to DNA damage and not affecting other compartments -was beneficial. In humans, a 600 mg daily dose of lipoic acid gives plasma concentrations of 8 to 30 µM, comparable to the concentrations used in our cell-based assays, suggesting that lipoic acid has surprising plausibility as a therapeutic.
We were unable to unambiguously link the short time-scale effects of lipoamide/lipoic acid on cells and purified protein to the longer time-scale effects on motor neurons, C. elegans stress granule protein aggregation and D. melanogaster motor function. There is evidence that lipoamide and lipoic acid promote mitochondrial biogenesis 62,63 as part of longer timescale 'indirect' antioxidant effects 64 .
These mechanisms do not plausibly contribute to the short timescale cell-based assays, but are likely to contribute to the effects of long-term lipoamide/lipoic acid treatment. This is important, as there are strong links with mitochondrial health and ALS 65 . However, we saw that lipoamide and lipoic acid could recover pathology caused by expression of FUS NLS mutants which have no direct link to mitochondrial biology ( Figure 8). One possibility is that lipoamide improves mitochondrial health and helps cells overcome the FUS-associated defect. Alternatively, lipoamide modulates cellular stress responses overcoming the FUS-associated defect and leading to improved mitochondrial health. It is notable that lipoamide, on dissolution of stress granules, promotes a return of FUS to the nucleus, which may be a key upstream mechanism for FUS NLS mutant pathology 46  Stress granules are just one example, and many other condensates are also linked to disease. This places our question of whether it is possible to interfere with the physical of stress granule formation using small molecules within a wider question of whether phase separation of IDPs is a possible drug target. Targeting individual IDPs associated with disease is not feasible as many proteins are implicated in a single neurodegenerative disease. Analogously, within ALS, many different mutations cause pathology and it is likely not feasible to target each mutation separately.
We have shown that compounds which affect phase separation can be found and are beneficial. This points to a new class of drugs -affecting the physical chemistry of phase separating proteins-and suggests the term physicochemical drug. Perhaps the solution to protein aggregation in disease is to support the cell's ability to maintain a dissolving environment, and future screens may identify further small molecule able to do so.

Methods
Stable Kyoto HeLa BAC cell lines expressing proteins with a C-terminal GFP fluorescent marker were generated using BAC recombineering 45 . This gives near-endogenous expression level of the fusion protein 7,66 . In these lines, GFP is part of a modified localisation and affinity purification (LAP) tag 67 , providing a short linker. HeLa cells were grown in high glucose DMEM supplemented with 10% FCS.
Human iPS cells lines, derived from three different donors, expressing FUS with a C-terminal GFP fluorescent marker were used. All were generated using CRISPR/Cas9 assisted tagging and mutagenesis and have been previously described 46 iPS cells were grown in TeSR E8 medium (Stem Cell Technologies) at 37°C with 5% CO2 69,73,74 .
MNs were generated from AH-ALS1-F58 iPS cells expressing P525L FUS in Matrigel coated plates with silicone channels for axons by inducing differentiation as previously described 74 . This yields clusters of soma on one side of the channels with axons which extend through the channels and protrude out of the far side of the channels. AH-ALS1-F58 were used to generate motor neurons (MNs) as they have previously been characterised in axonal transport assays 46 . iPS MNs used for assays within 4 weeks of the completion of differentiation, unless otherwise stated.
All procedures using human cell samples were in accordance with the Helsinki convention and approved by the Ethical Committee of the Technische Universität Dresden (EK45022009, EK393122012).

Recombinant protein
For in vitro condensate formation screening and solidification assays recombinant GFP FUS and GFP G156E FUS were purified using baculovirus/insect cell expression system, exactly as previously For the initial screen FUS GFP signal was analysed using KNIME 75  Appropriate solvent controls were used.
For Western blots and analysis of intracellular FUS aggregates iPS cells were lysed with RIPA buffer.
Western blots were performed using standard methods and the following antibodies: Mouse anti-FUS (AMAB90549 Sigma Aldrich, 1:500 dilution), rabbit anti-GFP (sc-8334 Santa Cruz, 1:400) or rabbit anti-GAPDH (2118S NEB, 1:5000) primary antibody with horseradish peroxidase-conjugated anti-mouse or anti-rabbit (Dianova 1:10,000) secondary antibodies. The filter retardation assay for intracellular FUS aggregates was performed as previously described 52 . Briefly, protein extracts were loaded on 0.2 μm cellulose acetate membrane then subject to microfiltration, leaving aggregated protein on the membrane. Aggregated FUS was detected as above.  Ex vivo DNA cut assays UV micro-irradiation of live cells to induce DNA damage was performed as previously described 7 .

In vitro FUS GFP solidification assays
KOLF iPS MNs expressing wild-type GFP were stressed by addition of 1 mM potassium arsenate for 1 h, then were treated with lipoamide, mitoxantrone or an equal volume of DMSO for 1 h. A single point in the nucleus was subject to 3 UV pulses as described for FRAP, but at 10% laser power. GFP fluorescence was imaged at 1 Hz, and intensity of response was analysed in ImageJ 76 .

Protein aggregation in C. elegans
The effect of lipoic acid on stress granule protein aggregation in vivo was analysed using a C. elegans model for stress granule formation and aggregation. As previously described 53 were maintained at at 20°C.
The animals were exposed to lipoic acid in liquid culture in a 96 well plate starting from larval stage

Motor defects in D. melanogaster
All Drosophila stocks were maintained on standard cornmeal at 25°C in light/dark controlled incubator. The w1118, UAS-eGFP, and D42-GAL4 were obtained from Bloomington stock center. The UAS-FUS WT, UAS-FUS P525L, and UAS-FUS R521C were previously described 80 .
Climbing assays were performed as previously described 80     no effect no effect 1,3-propanedithiol Fragment 2 no effect no effect aging. e) Quantitation of FRAP in c). Error bars represent starndard deviation. Aged condensates treated with lipoamide maintain large FUS GFP mobile fraction and short FRAP half-life while untreated condensates harden Effect of lipoamide and lipoic acid on G156E FUS GFP condensate 'aging' while shaking, relative to an equivalent DMSO solvent control (0.3%). Both compounds delay fibre formation. f-h) As for d-e), except using condensates formed under low salt conditions (more physiologically relevant) treated with 30 μM lipoamide or lipoic acid or an equivalent DMSO solvent control (0.3%). a c 6.7 6.9 7.1 6.7 6.9 7.1 6.7 6.9 7.1 6.7 6.9 7.1 6.7 6.9 7.1 6.7 6.9 7.1 6.7 6.9 7.1 6.7 6.9 7.1  with H2O (100 ml) and saturated aqueous NaHCO3 (100 ml). The organic layer was dried over MgSO4, filtered off and concentrated under reduced pressure. The NHS ester obtained, trimethylamine (1.1 ml, 7.89 mmol) and 15 NH4Cl (500 mg, 9.36 mmol) were dissolved in DCM (20 ml) and the mixture was stirred for 20 h. The solution was diluted with DCM (100 ml), washed with H2O (100 ml), saturated aqueous NaHCO3 (100 ml) and again with H2O (100 ml, 2 times). The organic layer was dried over MgSO4, filtered off and concentrated under reduced pressure to give crude 15 15 N lipoamide (including covalent attachment to an apoenzyme) would give a substantial change in the ( 15 N) 1 H NMR spectrum. Similarly, dissolution of lipoamide in a phospholipid membrane would give substantial peak broadening in the cell samples. We observed neither, consistent with freely diffusing lipoamide.
Integrated NMR signal intensity is proportional to concentration, provided conditions (including ionic strength, buffer composition, temperature and pH) are identical 82  Cellular uptake was measured by comparing the signal intensity of the trans-amide proton of lipoamide acquired in the absence (− , sample i, Figure 3B) and presence (+ , sample ii, Figure 3B) of HeLa cells. The measured fractional uptake was given by: The quantity (moles) of lipoamide added ( ) becomes distributed between the intracellular ( ) and extracellular ( ) environments following uptake. This can be expressed in terms of concentrations and volumes :

= +
The total volume of cells is given by = 1 where 1 is the volume of a single cell and is the number of cells.
≪ so we assume = . The fractional uptake can also be written in terms of these concentrations and volumes: N-(2-hydroxyethyl) ethylenediamine Fragment 1 no effect no effect 1,4-dihydroxy anthraquinone Fragment 2 no effect no effect  Larger shifts tended to occur needed near tyrosine (Y) residues and light grey bars indicate tyrosine residues and residues neighbouring a tyrosine. c,d) Average 1 H and 15 N shifts across residues zero, one, two, three or more than three residues from a tyrosine in the presence of either mitoxantrone or lipoamide. In the presence of mitoxantrone, chemical shifts correlated with distance from tyrosine. No other residue showed a similar correlation.  (14) approaches a plateau, indicating a slow exchange regime where signal intensity is an unambiguous measure of concentration. e) At 10°C, the intensity of the cis-and trans-amide proton resonances increased with decreasing pH indicating the presence of ms to µs second dynamics. Below pH 8.6, the intensity of the trans-amide proton was constant, indicating a slow exchange regime. Together, d) and f) indicate at 10°C and below pH 8.6 integrated signal intensity of the trans-amide proton of lipoamide in 15 N edited 1 H NMR experiments is a reliable proxy for concentration. f) Signal intensity of the transamide proton of lipoamide, when dissolved in growth medium, decreased over time at 37°C but not at 10°C. At 10°C signal intensity is stable for >10 h experiments. g) The signal intensity of both the cis-and trans-amide protons under different experimental conditions. This is an expanded version of Figure 7B, which only shows the trans-amide proton, and includes one additional condition: v) Cells (from iii) disrupted with Triton X-100 and DNaseI. Taken together, i) to iv) imply lipoamide is taken up by HeLa cells in a mobile form while most of the molecules were unmodified (with uptake quantified in Figure 7C-D). h) An expanded view of the spectra in g) i) and v). On disruption of cells with Triton X-100 and DNaseI the spectrum changed significantly, indicating chemical modification of lipoamide when cellular compartments are disrupted.