Aberrant phase separation is a common killing strategy of positively charged peptides in biology and human disease

Abstract

Positively charged repeat peptides are emerging as key players in neurodegenerative diseases. These peptides can perturb diverse cellular pathways but a unifying framework for how such promiscuous toxicity arises has remained elusive. We used mass-spectrometry-based proteomics to define the protein targets of these neurotoxic peptides and found that they all share similar sequence features that drive their aberrant condensation with these positively charged peptides. We trained a machine learning algorithm to detect such sequence features and unexpectedly discovered that this mode of toxicity is not limited to human repeat expansion disorders but has evolved countless times across the tree of life in the form of cationic antimicrobial and venom peptides. We demonstrate that an excess in positive charge is necessary and sufficient for this killer activity, which we name ‘polycation poisoning’. These findings reveal an ancient and conserved mechanism and inform ways to leverage its design rules for new generations of bioactive peptides.

INTRODUCTION

Nucleotide repeat expansions are implicated in an expanding group of human diseases^1,2. There are several different subclasses. Polyglutamine (polyQ) diseases, at least 9 different diseases each caused by the expansion of an exonic CAG repeat, are perhaps the most well-known³. These disease genes harboring the CAG repeat expansions get translated into proteins with an expanded polyQ tract, which makes them aggregation-prone, resulting in the formation of inclusion bodies as the hallmark pathology. Besides polyQ, expansions of exonic GCG and CGG repeats produce aggregation-prone polyalanine and polyglycine proteins, and cause a series of other diseases^4,5. Repeat expansions in non-coding parts of genes can also cause disease. For example, in Fragile X-syndrome, large GC-rich expansions within promoter or 5’ untranslated regions (UTR) become hypermethylated, inhibiting transcription and resulting in haploinsufficiency⁶. When expanded nucleotide repeats are transcribed into RNA, the repeat RNAs themselves can be toxic–for example, CUG/CCUG intronic/3’UTR RNA repeats in myotonic dystrophies⁷. These repeat RNAs have a propensity to condense and can trap important RNA-binding proteins in the process, impairing their function. Even though, these distinctions between coding and non-coding repeats seemed simple at first, the picture has become more complicated. The repeat RNA of the exonic CAG repeats can be just as toxic as the intronic CUG ones⁸. Expanded repeats also can generate antisense transcripts, implicating additional types of repeat RNAs and proteins⁹. Moreover, repeat RNAs (even non-coding ones) can be aberrantly translated in an AUG-independent manner (referred to as RAN translation)¹⁰, greatly expanding the range of proteins that can be encoded from these nucleotide repeats.

An archetype of this complexity is the repeat expansion in the C9orf72 gene. A hexanucleotide GGGGCC expansion in the first intron of a previously undescribed gene was found to be the leading genetic cause of amyotrophic lateral sclerosis (ALS) and frontotemporal dementia (FTD)^11,12. Upon the discovery of this repeat expansion, an immediate question was one of mechanism—how does this expansion drive human disease? Evidence has emerged linking all of the aforementioned mechanisms to C9orf72 pathogenesis. Hypermethylation of the expanded repeat affects transcription of the gene¹³, which seems to be implicated in autophagy¹⁴, vesicle trafficking¹⁵ and immune function¹⁶. Repeat RNA foci (both sense and antisense versions) accumulate in patient cells^12,17. RAN translation of the sense and antisense repeat generates so-called dipeptide repeats (DPRs) that are detected in the brain and spinal cord of patients harboring this mutation^17–19. These are poly-GA, poly-GP, poly-PA, poly-GR and poly-PR (abbreviated as their amino acid codes from here on). GP and PA do not seem to be toxic, whereas GA has mixed effects based on the model being used. The two arginine-rich DPRs, GR and PR, are potently toxic in diverse model systems—ranging from yeast to flies and worms, to vertebrates and iPSC-derived motor neurons^20–23.

Because GR and PR seem to be so toxic, there have been many different studies investigating the mechanisms of this toxicity. These studies have implicated disturbances in nucleocytoplasmic transport^24–26, defects in stress granules^27,28, mitochondrial defects²⁹, perturbation of axonal transport³⁰, proteasome deficits³¹, and inhibition of translation^28,32. All of these mechanisms could help explain the etiology of the disease, which is undoubtedly complex, but we wondered how can two amino acids repeated head to tail cause such widespread cellular defects? In other words, why are PR and GR such promiscuous killers that seemingly hit many unrelated biochemical pathways at once?

Since their discovery a decade ago, the apparent “aspecificity” of PR and GR toxicity has confused the field. A unifying framework to explain the biological connection between their myriad effects and the observed target promiscuity has remained elusive. Here, we unexpectedly discover that the interactome of GR and PR is actually encoded in their biophysical behavior. We and others previously showed that PR and GR can undergo biomolecular condensation in vitro and target biomolecular condensates in cells^22,27,28,33. Using new unbiased proteomics studies, we find that even the non-condensate effects of PR and GR can be explained by their “sticky” multivalency. Moreover, we find that nucleic acids are key modulators of their target space (i.e., the set of proteins that interact with GR and PR), explaining previously unresolved discrepancies between nuclear and cytoplasmic effects, and between disease models and patient material. Further, the recent identification of arginine-rich repeat peptides in at least three other degenerative conditions points at the emergence of a new class of repeat disorders centering around cationic peptides^34–36, with potentially a shared pathophysiology.

Since our framework was able to connect pathological condensation events in a divergent set of biological pathways, we wondered whether we could harness it to predict novel physiological condensate proteins. Using machine learning approaches, we discovered that the promiscuous condensation-based toxicity of PR and GR is not limited to human disease but is an integral part of ancient natural anti-microbial defense systems and spider, scorpion, and rattlesnake venoms. Similar cationic peptides have evolved countless times independently throughout the tree of life, and all can drive aberrant condensation via electrostatic interactions. Finally, leveraging protein design, we show that a surplus of cationic charge is sufficient to turn an inert protein into a toxic killer peptide. Thus, a mechanism we name polycation poisoning is emerging as a powerful and unifying framework that guides our understanding of promiscuous killer peptides in health and disease, and may aid in the design of new generations of therapeutic bioactive peptides.

RESULTS

RNA concentration tunes the target space of arginine-rich dipeptides.

In addition to C9orf72 ALS/FTD, other diseases also exhibit the same or similar arginine-rich peptides (Fig. 1A). RAN translation of GGCCTG (NOP56) and AGAGGG (TAF1) repeats, linked to spinocerebellar ataxia type 36 (SCA36) and X-linked dystonia and parkinsonism (XDP), generate PR and GR DPRs, respectively^34,35. QAGR tetrapeptide repeats are produced from the CCUG repeat in myotonic dystrophy type 2 (DM2)³⁶. As we have shown previously for PR and GR, these QAGR peptides similarly condense in vitro when added to cell lysate (U2OS, Fig. 1B). We previously showed that PR and GR form liquid droplets in the test tube when combined with nucleic acids²⁷, and these assemblies—or ones using other cationic peptides—have been since used as very simple in vitro models to study the biophysics of more complex RNA-centric liquid condensates in the cell^37–40. Surprisingly, when we add these peptides to cell lysate, we observe the formation of irregular networked condensates (Fig. 1C, Fig. S1A)—more reminiscent of the pathological gel-like condensates found in neurodegenerative disease^41,42. Adding additional RNA into the mixture (i.e., total yeast RNA), caused these assemblies to change morphology and acquire spherical shapes more similar to the liquid-like behavior of some physiological condensates (e.g., nuclear bodies like nucleoli⁴³). Because the addition of RNA dramatically affected the appearance of these test tube assemblies, we wondered what this would mean for their protein composition. As expected, its negative charge allows RNA to compete with proteins in the cell lysate for binding to the positively charged PR peptide (Fig. 1D, Fig. S1B-D). Using unlabeled quantitative mass spectrometry, we dissected what this competition looks like at the level of individual proteins (Table S1). We calculated the relative solubility of proteins in the test tube after spinning down the PR-induced condensates—as a measure of condensate partitioning—and clustered proteins according to their respective solubility profile as a function of the added RNA concentration (Fig. 1E-F). Around 50% of proteins in the mixture never partitioned into PR condensates and remained soluble throughout the experiment (cluster 1; grey proteins). For those proteins that condense with PR, 25% (clusters 2–5; red proteins) were outcompeted by low concentrations of RNA; the other 25% (clusters 6–9; blue proteins) were only outcompeted by high RNA concentrations or partitioned more strongly into the pellet upon the addition of RNA. Of note, this blue behavior is strikingly similar to the RNA-dependent re-entrant condensation behavior that has been observed for several RNA-binding proteins^38,44,45. Thus, these results demonstrate that PR-interacting proteins can have distinct RNA-dependent interaction profiles. Interestingly, these three clusters—simply based on the in vitro condensation behavior of proteins with PR and RNA—each specifically enriched different functional categories of proteins (Fig. 1G). The blue class of proteins included canonical members of physiological condensates (e.g., nuclear speckles, nucleoli) or cellular assemblies that have been proposed to form via similar biophysical principles (i.e., heterochromatin^46,47) (Fig. S2). Red proteins consisted of proteins not typically associated with physiological condensates but that have been implicated in protein aggregation or its modulation (Fig. 1H, Fig. S1E-H). Consistent with these findings, several of these distinct biological pathways—both red and blue—have been directly implicated in the toxicity observed in C9orf72 disease models and/or patients. These include nuclear import receptors^24,48, protein arginine methyl transferases^24,49,50, microtubules³⁰, the proteasome³¹, the translational machinery^{28,32,51–53}, splicing factors^28,54, nucleoli^27,28,55,56, and heterochromatin⁵⁷, among others. Thus, our simple test tube model recapitulates the target promiscuity of arginine-rich DPRs and provides a framework to understand how local RNA concentration modulates this target space, with important implications for pathophysiology (see Discussion).

Figure 1: RNA modulates protein composition of ex vivo PR30 condensates.

(A) Scheme indicating toxicity of cationic RAN peptides. (B) Cationic RAN peptides drive condensation of U2OS cell lysate. Heatmap showing quantification of RNA-dependent solubility for 1324 proteins. (C) Morphology or PR30 condensates is modulated by exogenous RNA addition (2 µg/µl RNA). (D) Silver stain indicating differential protein composition of PR30-induced condensates in response to RNA addition. (E) Heatmap showing quantification of RNA-dependent solubility for 1324 proteins. (F) Proteins cluster according to solubility responses as a function of RNA concentration. % indicates fraction of total proteins in the cluster. (G) Clusters differentially enrich for GO terms relating to subcellular compartments. Font size = −log(p-value). (H) Overview indicating that the morphology and composition of PR30 condensates are a function of the ratio of RNA concentration over basic protein concentration. All of the highlighted protein assemblies and factors have been directly implicated in C9orf72 ALS/FTD pathogenesis. 3D reconstructions of PR condensates were obtained via soft x-ray tomography. See also Fig. S1.

RNA-dependent behavior is encoded in sequence and follows an electrostatics framework.

PR and GR’s apparent target promiscuity has remained enigmatic; how are they able to target such a wide range of unrelated biological pathways? Because we now have in hand an in vitro model of this target space, we could ask whether this behavior and its modulation by RNA are encoded by a protein’s sequence. First, overlapping our data with a set of experimentally validated RNA-binding proteins⁵⁸ showed that, as expected, blue proteins are enriched for RNA-binding activity compared to red and grey proteins (Fig. 2A). Second, based on amino acid composition, blue and red proteins separate based on their charge (Fig. 2A-B); blue proteins are on average more basic, red ones more acidic, and grey ones more neutral. We can even observe this effect when looking within protein families. 60S ribosomal proteins are on average more positively charged than the ones making up the 40S subunit, and this correlated with lower solubility profiles in our mass spectrometry experiment (Fig. 2C).

Figure 2: Electrostatic charge is a sequence feature modulating PR30 condensate partitioning.

(A) Blue proteins are enriched for experimentally validated RNA-binding proteins and cationic amino acids. Red proteins have higher acidic amino acid composition. Grey proteins are enriched for neither. (B) On average, red proteins are more acidic and blue proteins are more basic compared to grey proteins (left). Moreover, on average, blue proteins have more clusters of basic residues in their IDRs (middle), while red proteins have more clusters of acidic residues in their folded domains (right) compared to grey proteins Kruskal-Wallis. (C) 40S and 60S ribosomal subunits show differential solubility profiles that correlate with changes in average pI. Two-way ANOVA and Mann-Whitney. (D) Positive charge clustering on the IDRs of DDX helicases correlates with their partitioning response to RNA (change in solubility upon addition of 0.4 µg/µl RNA. Linear regression. (E) Two example folded proteins with mirrored convex-concave solubility curves show opposite surface charge clustering (NAID = negative average inverse distance) and pI. Diablo (PDB: 1FEW)¹⁰⁸, SnrpD2 (PDB: 1B34)¹⁰⁹. See also Fig. S3-4. * p-value < 0.05, ** p-value < 0.01, *** p-value < 0.001, **** p-value < 0.0001.

While average charge can be informative, how charged residues are distributed across intrinsically disordered regions (IDRs) or on the surface of folded domains can substantially impact molecular interactions^59–62. Therefore, we co-opted and extended our previously applied sequence clustering analysis⁶³ to evaluate the position of acidic and basic residues in both IDRs and on the surface of neighboring fold domains (in linear space and three-dimensional space, respectively) (Fig. S3). By integrating proteome-wide structural predictions, this analysis revealed that—on average—blue proteins possess clusters of basic residues within IDRs, while red proteins possess clusters of acidic residues on the surface of folded domains (Fig. 2B, Fig. S4). We highlight two examples to further illustrate these observations. First, DEAD-box helicases (DDX) are a family of RNA chaperones that share the same type of folded domain typically flanked by IDRs. We identified red, blue, and grey DDX proteins in our assay, and, importantly, the positive charge clustering of their respective IDRs dictates their RNA-dependent condensation (Fig. 2D, Fig. S1G-H). Second, we identified two proteins that had virtually mirrored solubility profiles. SnrpD2 (a spliceosome component) shows very little acidic clustering on its folded domain but has a strongly basic pI, while Diablo (a mitochondrial protein) shows specific subregions with clusters of acidic residues (Fig. 2E). Despite having the same initial and final solubility in our mass spectrometry experiment, these proteins show opposite RNA-dependence, in line with their respective charge profiles. These analyses indicate that the differential protein-partitioning we observed in our mass spec experiment—on average—follows simple rules relating to the density and positioning of acidic and basic residues that tune protein-protein and protein-RNA interactions. We stress that these analyses report on average properties. Charge or the ability to bind RNA is predictive, but there are many examples of proteins that enrich differently. This argues that other—yet unresolved and probably more complex—sequence features and protein-protein interactions modulate protein partitioning in our assay.

Generating a machine learning algorithm to predict protein condensation.

Because the condensate partitioning behavior in our mass spectrometry experiment correlated with certain sequence features, we reasoned that we could in principle use our experimental data set to learn to predict such behavior de novo. We trained a machine-learning algorithm⁶⁴ (see Material & Methods) to predict blue proteins from their red and grey counterparts (Fig. 3A). Next, we asked the algorithm to rank intracellular proteins based on their “color” (Fig. 3B, Table S2). As expected, top predicted blue proteins enriched for gene ontology terms associated with nuclear bodies, RNA and chromatin metabolism (Fig. 3C). Interestingly, the algorithm also enriched intermediate filament proteins, which have been shown previously to engage with arginine-rich DPRs via their disordered head domains³³. None of these condensation-related gene ontology terms came up for the red and grey class of proteins nominated by the algorithm.

Figure 3: Machine learning predicts RNA-modulated phase separation of uncharacterized proteins.

(A) Scheme illustrating workflow to predict red (concave) vs blue (convex) proteins. (B) The intracellular proteome ranked according to normalized confidence scores (Z score). (C) GO categories associated with predicted red and blue proteins. (D) Scheme highlighting breakdown of top 10 predicted blue unknown proteins. (E) 9 out of 10 predicted blue proteins target membraneless organelles. 8 do so in a manner that is dependent on nuclear or cytoplasmic RNA availability (2 µg/ml actinomycin D inhibits all transcription, 0.5 mM arsenite induces polysome disassembly and increases free cytoplasmic RNA levels). White arrowhead indicates nucleolar stress cap targeting of C3orf49. C9orf85 targets nucleoli and stress granules upon arsenite treatment. C19orf47 demixes from nuclear speckles upon transcription inhibition. C4orf50 condenses into nuclear bodies upon arsenite treatment. C20orf96 forms cytoplasmic bodies unresponsive to changes in RNA availability. C4orf54 binds the actin cytoskeleton (see Fig. S5).

Given blue proteins appear to consistently be associated with intracellular condensates, we decided to directly test our algorithm’s ability to predict condensation behavior for previously uncharacterized proteins. We tested the top ten uncharacterized proteins from our predicted blue protein list and expressed them in cells (Fig. 3D, Table S2). Seven of these formed or targeted condensates under normal conditions: C3orf49, C11orf98, C16orf87, C18orf21, and C19orf53 targeted nucleoli, C19orf47 localized to nuclear speckles, and C20orf96 formed unknown cytoplasmic condensates. This result clearly demonstrates that our machine-learning model can predict cellular and biophysical behavior. Since our initial mass spectrometry experiment investigated condensation as a function of RNA concentration, we wondered how changing RNA availability would influence the behavior of our seven condensate-forming proteins. To do this, we decided to stress cells in two ways to change nuclear and cytoplasmic RNA availability–transcriptional shutdown and polysome disassembly. First, we shut down all transcription with a high dose of actinomycin D⁶⁵. This made four out of five of the nucleolar proteins demix from the nucleolus into the nucleoplasm—a well-known response of several nucleolar proteins^66,67. C3orf49 also showed localization to the nucleolar periphery in assemblies reminiscent of nucleolar stress caps⁶⁸. The nuclear speckle protein C19orf47 remained in a condensed state upon transcription inhibition, but demixed from the nuclear speckle, showing that its condensation behavior was also modulated by RNA. Second, stressing the cells with arsenite—known to cause polysome disassembly in the cytoplasm and rewire transcription in the nucleus⁶⁹—made C9orf85 partition into stress granules. Peculiarly, in addition to stress granule partitioning, C9orf85 also translocated into nucleoli in parallel. C4orf50, previously found in the nucleoplasm, now formed abundant small nuclear granules that were clearly distinct from nuclear speckles. Transcription inhibition or arsenite stress did not affect C20orf96 condensates. Finally, C4orf54 localized to the actin cytoskeleton under all tested conditions (Fig. S4).

Combined, our algorithm predicted nine out of ten proteins correctly as condensate proteins, of which eight did so in a manner that was modulated by RNA availability. The only “incorrect” prediction, turns out to be a novel actin-binding protein. Interestingly, we and others have previously found that PR and GR can interact with actin⁵⁰, microtubules³⁰, and intermediate filaments³³—highlighting that there may be more generally shared sequence features between cytoskeleton-binding and condensate proteins. Future studies will be required to investigate the functions of these new proteins and the functional significance of their condensate formation. In all, these findings indicate that our algorithm can predict condensate formation properties of proteins with high accuracy, allowing us and others to start systematically interrogating diverse proteomes.

Condensation is a novel mode of action of bioactive killer peptides.

In the above analysis, we restricted our search for novel condensate proteins to intracellular proteins since biomolecular condensates have been almost exclusively studied within the confines of the plasma membrane. But given the accuracy of our prediction tool, we reasoned that we could explore the possible existence of condensation events outside of the cellular context. When we fed sequences of secreted proteins into the algorithm, we unexpectedly found that it predicted several antimicrobial peptides (AMPs) to condense (Fig. 4A, Table S2). Animals, including humans, have evolved a wide array of AMPs⁷⁰. A well-studied example is LL-37 (Fig. 4B). Some cationic AMPs have been shown to kill bacteria by permeabilizing their membranes, but for LL-37 and several other AMPs this does not seem to be their main mode of action⁷¹. Because of their positive charge, such peptides are expected to interact with membranes, hence, potentially crossing them. Indeed, cationic peptides are often called cell-penetrant since they are spontaneously taken up by both prokaryote and eukaryote cells⁷². This raised the question: what would happen once a cationic peptide like LL-37 enters the bacterial cytoplasm? To answer this, we first performed a similar ex vivo lysate assay as described above. Adding LL-37, or PR as a positive control, to E. coli lysate was able to potently drive condensate formation (Fig. 4C). Second, we treated live bacteria with LL-37 at a concentration corresponding to its minimum inhibitory concentration against a range of pathogens⁷³. We subsequently imaged these bacteria using soft x-ray tomography, allowing us to gain information on their subcellular organization (Fig. 4D, Fig. S5A). LL-37 treated bacteria showed compacted nucleoids and evidence for the formation of dense cytoplasmic structures (Fig. 4E). Because DNA and ribosomes are the main components of the bacterial nucleoid and cytoplasm, respectively, we tested whether LL-37 could condense these biomolecules in the test tube; it did (Fig. 4F). Thus, we suggest that the intracellular condensation caused by LL-37 is responsible for its bactericidal activity. Since LL-37 condenses with negatively charged biological structures, similar to PR, this condensation can be described according to the framework of so-called complex coacervation—originally conceptualized by the polymer physics field to describe the condensation of oppositely charged (bio)polymers^74,75.

Figure 4: Killer peptides drive coacervation of cellular components and target biomolecular condensates.

(A) The intracellular + secreted proteome ranked according to normalized confidence scores (Z score). GO categories associated with predicted red and blue proteins. (B) Scheme illustrating LL-37 as an AMP. (C) PR20 and LL-37 induce condensation of E. coli lysate. (D) Soft x-ray tomography of sham and LL-37-treated E. coli cells. n = nucleoid, c = cytoplasm. (E) Quantification of nucleoid compaction and cytoplasmic condensation. Mann-Whitney. ** p-value < 0.01, **** p-value < 0.0001. See also Fig. S6. (F) LL-37 condenses both E. coli ribosomes and plasmid DNA. (G) Positive charge is a common feature among AMPs. (H-I) Examples of natural AMPs and venom peptides that induce cell lysate condensation. (J) Crotamine and PR30 (500 nM) target endogenous biomolecular condensates in human cells.

Is LL-37 a sole example of such a mechanism, or could this be more widespread? We analyzed thousands of curated AMPs⁷⁶, revealing that positive charge is a common feature among them (Fig. 4G). We tested a set of cationic AMPs from various vertebrates and invertebrates and found that several induce condensation in our ex vivo assay (Fig. 4H). These results support our emerging hypothesis that positively charged peptides could offer broad-spectrum antibiotic functionality through aberrant condensation of essential cellular components.

While AMPs target microbes, nature has also evolved biochemical tools to target large multicellular organisms in the form of venoms. Although several well-known (paralytic) venom peptides have a clear mode of action—usually specifically binding and modulating ion channels—venoms are complex mixtures containing a range of other peptides. Among them, so-called necrotic peptides are cationic and seem mainly responsible for tissue damage and cell death⁷⁷. Indeed, we found that cationic peptides from the venoms of scorpions, spiders, and rattlesnakes drive condensation in our ex vivo lysate assay (Fig. 4I). These results suggest that they have the biophysical capacity to condense, but do they exhibit this behavior in cells? Crotamine is a well-studied cell-penetrant rattlesnake peptide whose mode of action is incompletely resolved⁷⁸. We added fluorescently-labeled crotamine to the medium of human cells (Fig. 4J). As a control, we added fluorescently labeled PR, which has been documented to traffic to nucleoli^55,79. PR targeted the granular center of the nucleolus, consistent with other reports^28,56, while crotamine targeted nucleolar subcompartments. PR did not target nuclear speckles but crotamine did. Both these potent killer peptides—one aberrantly produced from an expanded repeat in disease, one evolved in a venomous animal—similarly targeted RNA-rich nuclear bodies. Yet despite these similarities, distinct structural and sequence features determined condensate partitioning, providing evidence that condensate target specificity can be encoded in the sequences of even small peptides.

Cationic charge is necessary and sufficient for killer activity.

Both the arginine-rich repeat peptides in human disease and the natural killer peptides have one main feature in common: positive charge. Could it be that charge is necessary and sufficient for killer activity? To address the first question, we looked more deeply at the specific amino acid enrichments we had seen above. While AMPs are enriched for arginine- and lysine-rich sequences, we did not see any enrichment for histidine (Fig. 4G)⁷⁶. This observation is not unexpected, given histidine is expected to be largely neutral at a pH of 7.0 – 7.4. Nonetheless, two of our top ten secreted predicted blue proteins belonged to the histatin family—a histidine-rich group of AMPs (Fig. S5B). When tested in the same assay as LL-37, histatin-3 (HTN3) did not induce condensation in our ex vivo assay. However, histatins are expressed specifically in saliva, which is slightly acidic (≈ pH 6.5). Switching assay conditions from pH 7.5 to pH 6.5—coinciding with a net charge gain of +3—drove the strong condensation of HTN3 in E. coli lysate. This finding provides evidence that cationic charge is required for condensation, and is consistent with the enhanced bactericidal activity of histatins under mildly acidic conditions first observed three decades ago⁸⁰. Thus, evolution seems to have used histidine residues to generate pH-dependent condensation switches that tune the activity of AMPs to their respective biofluid of origin.

Our work thus far suggests a net positive charge is necessary for toxicity of numerous AMPs, but is it sufficient? In other words, can we make a non-toxic protein toxic simply by altering its charge? GFP is probably the best-studied protein, with predictable behavior and folding, and is generally non-toxic to cells. By mutating residues facing out from the β-barrel, one can generate GFP mutants that still fold properly (i.e., remain fluorescent) but have an altered surface charge (Fig. 5A)⁸¹. When we expressed a library of super folder GFP (sfGFP) variants—with net charges ranging from −24 to +24—in E. coli, we observed that cationic GFP variants caused bacterial growth arrest (Fig. 5B). Imaging these cells, we observed that the cationic GFP variants, but not the anionic ones, formed aberrant condensates with ribosomal proteins (Fig. 5C) and RNA⁸² at the bacterial cell pole. Similarly, when we combined recombinantly purified GFP variants with E. coli ribosomes in the test tube, only the cationic ones formed condensates (Fig. 5D). Thus, like we saw for LL-37, these GFP variants can drive charge-based coacervation of ribosomes—killing the cells.

Figure 5: Net positive charge is sufficient for protein toxicity.

(A) Engineering supercharged GFP versions for toxicity assays. (B) Expression of cationic GFP variants drives toxicity in E. coli. One-way ANOVA. (C-D) Cationic GFPs condenses with ribosomal proteins in E. coli and the test tube. (E) GFP partitions into nucleoli of U2OS cells in a positive charge-dependent manner. One-way ANOVA. (F) Exogenous recombinant cationic GFP is taken up, targets nucleoli and kills mouse cortical neurons in the dish. Student’s t-test. (G) Cells with low levels of GFP (+) have a normal nuclear organization, whereas high levels (+++) perturb nucleoli and heterochromatin. Adjacent cells that did not take up GFP (−) are normal. (H) Scheme depicting polycation poisoning in pro- and eukaryote systems. * p-value < 0.05, *** p-value < 0.001, **** p-value < 0.0001.

We next tested the effect of these GFP variants in human cells. The cationic GFP variants spontaneously condensed in nucleoli and nuclear speckles in a charge-dependent fashion (Fig. 5E). Given cationic peptides often undergo spontaneous cellular entry, we wondered if our synthetic positively charged GFPs might also be sufficient for cellular uptake. To test this, we also added recombinant anionic and cationic GFP to the medium of cultured mouse cortical neurons. Virtually identical to our observations with PR and crotamine, cationic GFP was spontaneously taken up by cells and targeted the nucleolus, killing the neurons (Fig. 5F). Moreover, this cationic version of GFP did so at concentrations similar to or lower than the ones we and others have found GR to be toxic^49,55,79. Importantly, this observation challenges that canonical expectation that cell-penetrant peptides must have evolved a specific and highly constrained mechanism for cellular uptake, instead implicating cellular entry as being driven in part if not entirely by a strong net positive charge. Finally, neuronal uptake of cationic GFP was not uniform; some cells rapidly accumulated large quantities of GFP, whereas in others the accumulation was slower or even absent. Note, this same heterogeneity has been observed for other cell-penetrant peptides as well^83–85 and is linked to cellular tolerance against their toxic effects⁸³. Cells with low levels of nuclear GFP showed nucleolar localization as expected (Fig. 5G). Surprisingly, cells with high levels of GFP had non-detectable nucleoli (based on NPM staining) and showed marked condensation of DNA. The oversaturation of the nuclear compartment with these charged molecules seems to upset the physiological balance of electrostatic interactions, dramatically perturbing nuclear organization and leading to the demise of the cell (Fig. 5H).

Taken together, our results suggest that positively charged peptides and proteins offer broad-scale cellular toxicity. In some cases, this toxicity is an evolved function, while in others, it may be an unfortunate mishap through spontaneous repeat expansions that underlie human disease. Our results demonstrate that while precise target selectivity can clearly be encoded in a killer peptide’s sequence (e.g., crotamine vs PR), we conclude that cationic charge is a unifying sequence feature that is sufficient to drive killer activity. This process, which we name polycation poisoning, follows the simple framework of complex coacervation and has independently evolved countless times across the tree of life.

DISCUSSION

For a decade, the observed target promiscuity of C9orf72 arginine-rich DPRs has puzzled the field. Compelling data has connected their toxicity to a bafflingly diverse array of biological pathways^20–22. But a unifying framework has so far remained elusive. Were these studies monitoring the cause of cell death or simply the downstream consequences of dying cells? We and others had previously shown that these peptides can condense RNA and other biomolecules, explaining their effect on physiological condensates^22,27,28,33. Here, we used proteomics to explore the target space of cationic dipeptide repeats and unexpectedly found that even their non-condensate targets fit into a simple electrostatics-based framework of complex coacervation. We propose that key actors in all these different biological pathways share the right sequence features that drive their aberrant interaction with these positively charged peptides.

Our framework also allows us to address the most common criticism against the pathological role for arginine-rich DPRs in patients. In every model system tested so far, PR—and to a lesser extent GR—targets the nucleolus, making it potently toxic. Yet, this has never been observed in any pathology report using patient post-mortem material. For reasons that are currently unresolved, but a topic of intense study, both PR and GR seem to get trapped in the cytoplasmic GA inclusions in patients, thus preventing them from entering the nucleus^17–19. Our results unambiguously demonstrate that RNA tunes the target space of PR (Fig. 1). As such, which proteins PR (and, by extension GR) interact with within the cell will depend on the local RNA concentration around those DPRs. The concentration of RNA in the nucleus is estimated to be 36 times higher than the cytoplasm⁴⁴, where most RNA is also trapped in polysomes. As such, our test tube experiments provide lysate mimics of either a nucleus-like (high RNA) or cytoplasm-like (low RNA) environment, enabling us to dissect which proteins are PR targets in these two compartments. Under high RNA conditions, we observe PR interacting with proteins associated with nuclear bodies and chromatin. This mirrors the subcellular localization observed in model systems, where PR interacts with nucleoli^27,28,55,56 and heterochromatin⁵⁷. Under low RNA concentrations, PR interacts with the translation initiation complex, ribosomes, axonal transport factors, and protein arginine methyl transferases. Analysis of patient post-mortem samples provides evidence that these kinds of proteins do associate with cytoplasmic GR/PR inclusions in patients^{24,30,49,52,53}. Although defects in condensation seemed initially limited to disease models that failed to fully recapitulate human pathology, we now show that the same framework of electrostatic condensation is still applicable but is tuned by the local RNA concentration. To our knowledge, this is the first model that is able to explain a decade of disconnect between most C9orf72 disease model-based studies and human pathology reports.

Pathogenesis does not exist in a vacuum, but instead follows already-established biochemical principles. Can we use a framework designed to understand pathology to infer novel functional biology? We used our proteomics dataset to train a machine learning algorithm to detect similar electrostatics-based condensation behavior. When tested on the human proteome, we found already known condensate proteins but, importantly, also predicted this behavior in completely uncharacterized proteins. We experimentally validated this condensate behavior in nine out of the ten predicted proteins—showing that our approach allows for the discovery of new biology. The complete functional characterization of this set of novel condensate proteins is beyond the scope of this study, but we note that some of these proteins have substantial phenotypic effects in high-throughput knock-out studies⁸⁶, arguing that their yet unknown function may carry cellular importance. It will be interesting to test if their condensation propensity is involved in their function, and whether they are implicated in human disease.

The (re)discovery of biomolecular condensation as an organizing principle of the cell besides membranes has revolutionized cell biology^87,88. Ironically though, our exploitation of these insights has typically remained delimited by the plasma membrane. With our sequence-based condensate protein predictor in hand, we could now test for the condensation properties of extracellular proteins in a direct and unbiased way. To our initial surprise, we discovered that cationic killer peptides found in innate immune systems and venom cocktails are potent drivers of condensation, following the exact same set of rules we delineated for the human disease repeat peptides. Not impeded by GA inclusions like PR, venom peptides are able to gain entry to mammalian cells and readily target the nuclear centers of RNA production. AMPs condense both the nucleoid and ribosomes in bacteria. Our results demonstrate that two very different cellular systems (eukaryotes versus prokaryotes) with completely unrelated peptides follow the exact same framework of polycation poisoning. Indeed, using GFP as a scaffold we can engineer killer activity by simply making it positively charged, with its toxicity in neuronal culture rivaling that of the neurotoxic GR and PR peptides.

One question remains though: why positive and not negative charge? Although we do not rule out the existence of polyanion-based toxicity, there are some simple reasons why cationic peptides could be more disruptive. If we consider the cell as a collection of lipids, nucleic acids and proteins, a clear trend emerges. Lipids tend to have negatively charged head domains, nucleic acids are inherently anionic, but proteins can span the entire charge spectrum. Given all the negative charge on lipids, nucleic acids, and even the cytoskeleton (both actin filaments and microtubules have negatively charged surfaces), a cationic-biased proteome would promiscuously interact with all of these exposed biomolecular surfaces. Hence, it may come as no surprise that for organisms living at mesophilic pH and salt concentrations, proteomes tend to be—on average—more negatively charged^89,90. Yet, we know many examples of highly cationic proteins in the cell, raising the question of how cells protect themselves from their own. By analyzing six model organisms, we indeed find no evidence for selection against positively charged proteins (Supplemental text, Fig. S7), despite our data suggesting that polycationic folded and disordered proteins drive cellular toxicity across the kingdoms of life. To better understand these seemingly contradictory observations, we performed an integrative bioinformatic analysis showing that while proteomes do possess many examples of proteins with a net positive charge, these are either expressed at low copy numbers, sequestered from the cellular environment via constitutive binding to anionic biomolecules, or neutralized through posttranslational modifications (Supplemental text, Fig. S8-S11, Table S3). Remarkably, we find that for essentially every single example of an abundant, positively charged protein, its charge is neutralized by a phosphate moiety—either in trans (nucleic acids or phospholipid headgroups) or in cis (through phosphorylation). We argue that this neutralization is essential to keep endogenous polycationic proteins in check to prevent them from poisoning the cell. Intriguingly, similar approaches are used by several human pathogens to neutralize cationic antimicrobial peptides and antibiotics^91–93.

The above observations point at a seemingly universal proteome ‘design principle’. The net polarity for the large cellular surfaces ensures cellular proteomes do not aspecifically stick to lipids and nucleic acids. At the same time, it becomes relatively easy to evolve high-affinity domains that engage with these acidic moieties if needed. Cationic surface patches or disordered domains are pervasive interaction motifs in the proteome. For example, countless RNA binding proteins possess cationic disordered domains adjacent to their folded RNA binding domains^27,94,95. Recent work has shown that even modest clusters of positively charged residues in IDRs from RNA binding proteins can have a profound impact on RNA binding affinity, mirroring our bioinformatic analyses⁹⁶. Similarly, positively charged microtubule-binding proteins, such as tau, do so in part by interacting with the anionic tubulin tails that coat the microtubule lattice⁹⁷. Membrane-binding proteins, like α-synuclein, do so via the exact same electrostatic interactions⁹⁸. The use of cationic interaction and targeting motifs in a sea of anionic biomolecules is a powerful one but highlights the inherent weakness of cells against killer systems that corrupt this principle. It also shows why the cell has a variety of fail-safe mechanisms in place to prevent toxicity of its own cationic proteins. When we tested the cationic disordered domains of some human nucleic acid binding proteins in isolation, we found they act virtually identical to the killer peptides described in this study (Fig. S12). So how does the cell tame these sequences? Above we highlight the role that phosphorylation could have in regulating these proteins. Recent studies by us and others have also shown that these cationic domains are often paired with anionic or aromatic domains that can potently quench or regulate their activity^99–103. Additionally, nuclear import factors that bind proteins translated in the cytoplasm via their nuclear localization sequence—usually cationic in nature—were recently shown to moonlight as bona fide protein chaperones^104–107. Although philosophical, it is worth considering which function came first when transcription and translation became uncoupled during eukaryogenesis: chaperoning a condensation-prone protein in an RNA-poor environment (i.e., the cytosol) or translocating it to an RNA-rich environment (i.e., the nucleus)?

In this study, we provide a simple electrostatics-based framework that unifies the toxicity of cationic peptides in human disease and killer peptides in biology. McKnight and co-workers speculated that RAN translation of the C9orf72 repeat could be considered the “failed birth of a gene”⁵⁵. Upon repeat expansion, a non-coding part of the genome all of a sudden becomes protein-coding. Yet, the gene product is toxic to the cell producing it. When venomous animals evolved their killer peptides, a requirement was that the venom did not kill its host. Unsurprisingly, venom peptide production is a tightly regulated process often involving extracellular processing and cleavage of quenching domains together with signal peptides correctly targeting the protein for secretion. Note that in the lumen of a secretory vesicle these peptides never “see” the cytoplasm of the cell they are produced in. With this in mind, one can only wonder whether we have witnessed the failed birth of a venom peptide that lacked its regulatory sequences.