Hairpin containing RNA phase separates in vitro into gel-like granules
To test whether hairpin containing RNA can phase separate in vitro we designed six synthetic long non-coding RNA (slncRNA) binding-site cassettes using our binding site resource19–21. We divided our slncRNAs into two groups. For the first group (class I slncRNAs), we designed three cassettes consisting of three, four, or eight hairpins that encode for PCP binding sites (PCP-3x, PCP-4x, and PCP-8x, respectively). In this group, hairpins were spaced by a randomized sequence that did not encode for a particular structure. For the second group (class II slncRNAs), we encoded three cassettes that consisted of three, four, and fourteen PCP binding that were each spaced by hairpin structures that do not bind PCP (PCP-3x/MCP-3x, PCP-4x/MCP-4x, and PCP-14x/MCP-15x, respectively). The sequences encoding for the slncRNAs were cloned downstream to a pT7 promoter and transcribed in vitro to generate the corresponding RNA. To visualize the RNA, we incorporated fluorescent nucleotides in the transcription reaction such that an estimated 30% of uracil bases were tagged by Atto-488 fluorescent dye. Each slncRNA-type was separately mixed with granule forming buffer (see methods and Figure. 1a) at equal concentration (8.5 nM final concentration) and incubated for 1 hour at room temperature. 2 µl of the granule reaction were then deposited on a glass slide and imaged using an epi-fluorescent microscope.
The images show formation of a multitude of bright localized fluorescent condensates for all slncRNA types except for the PCP-3x case, where no such structures were detected (Figure 1b). In addition, the longer slncRNA molecules (e.g., PCP-8x and PCP-14x/MCP-15x) also exhibit larger structures, consistent with a gel like solid network, in addition to the smaller condensates or puncta. An examination of the median fluorescence obtained for each slncRNA type condensate (Figure 1c) reveals a dependence on the number of fluorescently labelled uracil nucleotides or the number of hairpins encoded into the slncRNA. However, the relationship between the median fluorescence values obtained for each species is not consistent with a linear dependence on hairpins, and instead suggests a more complex set of structures.
To further analyze the condensate structure, we fitted the measured condensate fluorescence intensity distributions to a modified Poisson distribution (see Figure 1d, Figure S1 and Supplementary methods). The panels reveal three characteristic distributions. For PCP-4x, an exponential distribution is recorded (i.e., λ=0). For PCP-3x/MCP-3x, PCP-4x/MCP-4x, and PCP-8x, a Poisson distribution of λ~1-2 seems to be the best fit. Finally, for the PCP-14x/MCP-15x, a Poisson distribution of λ~3-5 fit best. These results are consistent with the formation of condensates that are characterized by an increasing number of slncRNA molecules that are cross-linked to form a gel-like "granule", where the number of hairpins encoded into the slncRNA determines the average number of molecules or cross-links within the observed field of granules. Moreover, the interpretation suggested by the shape of the distribution is contrasted by the counter-intuitive observation of decreasing value of the of the fitting parameter K0 as a function of an increasing number of hairpins (Figure 1e). In this particular context, this observation is manifested by a significantly more gradual increase in mean or median granule fluorescence as compared to what would be naively expected by a simple rescale that takes into account the number of hairpins. Together, these observations suggest that slncRNA granules form via cross-linking interaction of multiple slncRNA molecules, and that an increasing number of hairpins and cross-links lead to a denser condensate. Denser granules, in turn, may result in fluorescence quenching of the labelled uracils22 leading only to a gradual and disproportionate increase in fluorescence observed.
RNA-based granules co-localize with protein-binding partners
To test if the hairpins retain their ability to bind the PP7 phage coat protein while in the granule state, we added recombinant tandem dimer PP7 coat protein fused to mCherry (tdPCP-mCherry) to the granule formation reaction in excess amount (final concentration 25 nM) to account for the multiple binding sites present on one slncRNA molecule (Figure 2a). The tdPCP-mCherry version used lacks the necessary moiety to form the wildtype viral capsid23. The images (Figure 2b) show colocalization between the 488 nm channel (Atto-488) and the 585 nm channel (mCherry) for all slncRNA designs used in the experiment implying that PP7 coat proteins are able to bind the RNA hairpins in the condensate state. Hence, the slncRNA and their protein partners form synthetic RNA-protein (SRNP) granules. Unexpectedly, PP7-3x granules were witnessed in the presence of the protein, implying that tdPCP-mCherry adds a measure of multivalency to the system, and thus triggers condensation of RNA molecules that do not phase separate on their own. To check that this condensation was hairpin dependent, we tested whether a control RNA (of the same length and GC content as PCP-8x) containing no designed hairpin condensed either on its own or in the presence of tdPCP-mCherry forms granules. In both cases, no condensates were detected in either the 488 nm or 585 nm channels (data not shown). Finally, unlike for the slncRNA only case, SRNP granules (particularly for high number of hairpins) show an increased propensity to form large-scale extended structures, suggesting a more complex structure formation and condensation for the SRNP granules as compared with the slncRNA-only case.
Next, we measured the median fluorescence intensity of the mCherry protein in different SRNP granules. The distributions of median values (Figure 2c) show a clear dependence on the number of binding sites available for protein binding. First, the PCP-3x/MCP-3x and PCP-4x granules appear to have a similar number of proteins in the granule and are both weaker than PCP-4x/MCP-4x granules, suggesting that PCP-4x slncRNAs inside the granules are not fully occupied by proteins. In addition, the PCP-14x/MCP-15x granules seems to be >2-fold brighter as compared with the PCP-8x granules, despite having <2-fold the number of hairpins. This stands in contrast to the observation that PCP-14x/MCP-15x granules appear to be ~3 times brighter than PCP-4x/MCP-4x granules, reflecting the difference in the number of binding sites available for binding. Finally, PCP-3x granules appear to be half as bright as PCP-14x/MCP-15x granules, providing more evidence that the former are not RNA-dependent entities. We also observe that when the spacing regions within the slncRNA encode for the MCP hairpins, the formed granules contain a larger protein cargo.
To confirm this observation, we also observed the SRNP granules in the 488 nm channel. Here a similar image emerges, whereby the median fluorescence values for each granule are percentage-wise more differentiated as compared with the slncRNA-only case, reflecting a more proportional increase in fluorescence (Figure. 2d). Together, the observations in both channels indicate that SRNP granules are less dense gel-like structures as compared with the slncRNA-only granules. To authenticate the granules as being solid-like RNA-protein structures, we imaged them using a SIM super resolution microscope with 120 nm resolution. Figure 2e shows a sample image of a PCP-14x/MCP-15x granule containing the tdPCP-mCherry protein. The image shows that slncRNA is found mainly in the periphery of the granule, with filaments protruding into its core, where a high amount of protein is amassed in a network like configuration. The RNA seems to encase the protein cargo in a reduced density structure.
Finally, we explored the phase space of SRNP granule formation. To do so, we characterized formation of the PCP-14x/MCP-15x SRNP granules as a function of both slncRNA and protein concentration. For this we produced non-fluorescent RNA molecules (for higher concentrations) and mixed different titers of slncRNA and tdPCP-mCherry protein, each varied over two orders of magnitude. Puncta like structures were detected only for slncRNA and proteins concentrations of 100 ng/μl and above (Figure. 2f). The images display bright puncta that are embedded within a filamentous structure. Quantification of the maximal intensity of the puncta both at time T=0 (i.e., beginning of the reaction) and time T=1 [hr] (Figure. 2g) reveals a fluorescent intensity distribution which declines by two orders of magnitude (i.e., from ~105 to ~103) in a step-like function as the RNA concentration is reduced from 1000 to 10 ng/μl, providing further indication that RNA is essential for granule formation. Likewise, the intensity distribution of the puncta declines in a more gradual fashion as the protein concentration is reduced, but overall, a similar disappearance of puncta is observed.
Temporal tracking of individual SRNP granules reveals that granules function as protein capacitors
A hallmark of phase separation is the exchange of molecules between the dilute phase and the dense phase. This is also true for gels with non-permanent intermolecular interactions, wherein random breaks and rearrangement of the connections which form the inner network allow macromolecules (monomers and small polymers) to diffuse in and out of the gel phase8–11, albeit at a slow rate. These exchange events are predicted to occur independently of one another, at a rate which depends on multiple parameters: the probability of cross linking within the gel network (i.e., number of hairpins), the transient concentration of the molecules in the surrounding solution, and the average diffusion rate of the monomers. The movement of molecules (fluorescent CPs, slncRNA, and CP-bound slncRNA complexes) between the different phases should be reflected by changes in granule fluorescence intensity.
To test whether the synthetic granules display this characteristic, we tracked the fluorescence intensity of each granule in a given field-of-view for 60 minutes. We analyzed the brightness of each granule at every time point using a customized analysis algorithm (see Supplementary Methods). The resulting signals are either decreasing or increasing in overall intensity, and dispersed within them are sharp variations in brightness, that are also either increasing or decreasing. Next, we employed a statistical threshold which flagged these signal variation events, or “signal bursts”, whose amplitude was determined to not be part of the underlying signal distribution (p-value<1e-3) (See Supplementary Methods for definitions of bursts, algorithm details, and relevant numerical controls). The events were classified as either increasing bursts (green), decreasing bursts (red), or non-classified segments (blue), which are segments where molecular movement cannot be discerned from the noise (Figure. 3a). For each detected burst, we measure its amplitude (Δ intensity) and duration (Δ time), in addition to measuring the time between bursts and the order of their appearance. In Figure. 3b we plot the distributions of amplitudes for all three event types, obtained from ~156 signal traces, each gathered from a different granule composed of PCP-14x/MCP-15x and tdPCP-mCherry. We observe a bias towards negative burst or shedding events. Assuming an interpretation that fluorescent burst events correspond to insertion and shedding events of slncRNA-CP complexes into or out of the synthetic granules, the amplitude bias towards negative events is consistent with RNA degradation and lack of transcription within the in vitro suspension, leading to a net shedding of slncRNA molecules out of the granules over time.
We repeated the tracking process for granules produced from all previously-described slncRNA designs (including the PP7-3x which does not phase separate on its own). Comparison of the amplitude distributions per design (class I vs. II), (Figure. 3c) reveals a dependence on the number of hairpins available for protein binding, where more protein binding sites translate directly into larger amplitudes. Next, we also measured burst amplitude in the green-channel to confirm that bursts indeed correspond to the shedding of a slncRNA-protein complex from the granule. Figure. 3d shows a sample signal for the PCP-14x/MCP-15x granules showing concomitant occurrence of bursts in both the red and green channels, supporting our interpretation of this signal. Using the burst distributions, we then computed the ratio between the mean granule fluorescence and the mean burst amplitude, providing a measure for the number of slncRNA molecules within the granule. The results (Figure. 3e-f) show that with the exception of the PCP-4x based SRNP granules, the ratio in the green channel is 5, suggesting that a typical granule contains five slncRNAs. The ratio computed for the red channel is typically smaller, and for PCP-8x is ~2. This means that in every burst approximately half the protein content is released with the RNA. This further implies that PCP-8x may be permeable to proteins diffusing out of the granules due to reduced cross-linking as a result of a lack of hairpin spacing structures. For the PCP-3x/MCP-3x, PCP-4x/MCP-4x, and PCP-14x/MCP-15x the ratio in the red channel is approximately equal to that of the green channel, suggesting that these granules have a better protein storage capacity. Hence, the granules composed of class II slncRNA seem to form more robust and better insulated granules from the perspective of their protein storage capacity.
To provide further evidence for this interpretation, we measured the time duration between events for each granule type. For the granules, this rate (~5 minutes) is two orders of magnitude above the typical rate observed in liquid phase separated condensates24, but is in line with the measurements performed on RNA gels by Vale et. al.11, providing additional confirmation that the SRNP granules are gel-like particles. A closer examination of the duration boxplots (Figure. 3g) obtained for each granule-type reveals that more binding sites lead to longer durations on average, for both negative and positive bursts. Additionally, there appears to be a difference between the slncRNA designs themselves. While the granules composed of class I slncRNA granules present on average longer durations between positive bursts, compared to negative bursts, the opposite is true for the class II slncRNAs. Assuming a roughly uniform distribution of molecules outside of the granule (given enough time to equilibrate), this may mean that on average, a protein-bound slncRNA molecule has a higher probability of leaving a class I slncRNA granule than entering it, and vice versa for the class II slncRNA granules. This result confirms the interpretation of the burst ratio analysis, and together these results imply that class II granules are characterized by a highly-cross-linked slncRNA network which prevents the diffusion away of molecules (leading to a longer time between negative bursts), while the granule boundary still contains free cross-linking points that can latch on to incoming molecules more easily, increasing the chances of molecular entry (leading to a shorter time between positive bursts). Together, these SRNP granule characteristics are reminiscent of data and energy storage devices (e.g., capacitors), with the protein cargo replacing the electric charge in the biochemical analog.
Expression of slncRNAs and protein in bacteria yields puncta-like condensates
Given the capacitor analogy, we hypothesized that in vivo the granules can be used as devices that store the granule-bound proteins. This is due to the steady state production of slncRNAs and proteins via the cellular transcriptional and translational machinery, that ensures a constant flux of proteins into the granules. To show this, we first proceeded to test whether the granule material characteristics that are measured in vivo match the in vitro measurements. To do so, we decided to utilize two previously reported slncRNA designs which were shown to yield bright localized puncta in vivo in earlier work19. The first slncRNA is of a class II design, PCP-4x/ QCP-5x, consisting of four native PCP binding sites and five native Qβ coat protein (QCP) hairpins used as spacers in an interlaced manner. The second slncRNA is the ubiquitous PCP-24x cassette25, which from the perspective of this work can be regarded as a class I design slncRNA.
To confirm the granules form conditions in vivo, we encoded the slncRNA component under the control of a T7 promoter, and the tdPP7-mCherry under the control of an inducible pRhlR promoter (Figure. 4a). We first wanted to test whether puncta develop in vivo and whether they are dependent on the existence of hairpins in the RNA. For this we co-transformed plasmids encoding either the negative control RNA or the PCP-4x/QCP-5x slncRNA, together with a plasmid encoding for the tdPCP-mCherry protein, into BL21-DE3 E. coli cells. Examination of cells expressing the slncRNA and protein following overnight induction of all components revealed the formation of bright puncta at the cell poles (Figure. 4b), which were absent in cells expressing the control RNA which lacks hairpins (Figure. 4c).
Next, to test whether cellular concentration of slncRNA influences the formation of the granules, we quantified the fraction of puncta per cell for cells expressing the PCP-4x/QCP-5x from a multicopy expression vector, and cells expressing the same slncRNA from a bacterial artificial chromosome (BAC) expression vector which is maintained at a single copy level in cells. We found that cells containing the multicopy plasmid frequently present puncta in at least one of the poles, while cells containing the single copy generally show between zero and one punctum (Figure. 4d). Given that cells expressing the slncRNAs from single copy vectors still present puncta, we decided to continue using this expression vector in follow-up experiments to reduce variability stemming from copy number differences.
We compared cells expressing the PCP-4x/ QCP-5x or the PCP-24x (expressed from a BAC vector) in terms of the spot per cell fraction. Much like in the in vitro experiments, we found a dependence on the number of binding sites in accordance with the in vitro results and the cross linking model of gel phase formation6,26 (Figure. 4d). Finally, to test whether the polar localization of the granules is a consequence of nucleoid exclusion27, we grew the cells in starvation conditions for several hours, triggering a transition to stationary phase. In stationary phase the nucleoid is known to condense28–30, thus increasing the amount of cellular volume which is likely to be molecularly dilute. This, in turn, generates a larger accessible cellular volume for granule formation, which should lead to different presentation of the phase-separation phenomena as compared with exponentially growing cells. In Figure. 4e, we show images of bacteria displaying ‘bridging’ (the formation of a high intensity streak between the spots) of puncta (left), whereby granules seem to fill out the available dilute volume, and the emergence of a third puncta at the center of the cell (right). Both behaviors are substantially different than the puncta appearing under normal conditions. Such behavior was observed in >40% of the fluorescent cells and was not detected in non-stationary growth conditions. Thus, SRNP granules with characteristics that are consistent with the in vitro observations form in vivo, in a semi-dilute bacterial cytosolic environment and independent of cell-state.
slncRNA expression increases cellular protein concentration
To investigate the dynamic properties of granules formed in vivo, we utilized the same analysis approach as was used in the in vitro experiments, with minor differences. Normalizing the fluorescence of the granule by that of the cell (see methods) for every time point results in a signal vs. time trace largely independent from the effects of photobleaching and cellular background noise, allowing us to search for and measure burst events, as was done previously. In Figure. 5a, we plot the distributions of amplitude (Δ intensity) of all three event types (positive, negative, and non-classified), obtained from 255 traces gathered from cells expressing the PCP-4x/QCP-5x slncRNA together with the tdPCP-mCherry protein. The symmetry in both shape and spread of the negative and positive distributions indicates that both are measurements of the same type of macromolecule, distinguished only by the direction in which it travels (into or out of the granule). Moreover, a similarly symmetric burst distribution is recorded for the PCP-24x slncRNA (Figure S1). This result contrasts with the in vitro amplitude distribution data (Figure 3b), which presented a skewness towards negative bursts. This implies that in vivo, the transcriptional and translational processes in the cell balance the loss of granule components due to degradation.
Next, we measured the amplitudes of the bursts for both slncRNAs and found that positive and negative amplitudes are proportional to the number of binding sites within the encoded cassette (Figure 5b). In addition, a more quantitative analysis of these distributions (Figure S2) reveals that a single burst for the 24x cassette is ~2.5-3x more fluorescent as compared with the 4x cassette, indicating that the molecules transitioning in and out of the 24x granules are slncRNAs partially or fully bound rather than lone proteins. Moreover, estimations of the positive and negative amplitudes are practically equal per slncRNA, providing additional evidence that these are in fact representations of one physical process, with the difference being the directionality of the transitioning slncRNA-protein molecule. Finally, we measured the duration between burst events, revealing that slow shedding and absorption processes on the order of minutes are taking place for the in vivo granules as well (Figure 5c). Altogether, the non-existence of puncta in cells expressing the negative control RNA, the slow exit/entry rate of molecules, and the dependence on the number of binding sites, suggest that synthetic RNA protein granules are phase separated condensates in vivo and possess the same gel-like characteristics that were observed for the in vitro suspensions. Consequently, in vivo burst analysis is consistent with the capacitor model, where the amount of protein stored within the SRNP granule seems to be in steady state when there is a steady supply of protein and slncRNA.
Next, to ascertain whether the granules facilitate increased protein titers in vivo in accordance with the capacitor model predictions, we measured for each bright granule the mean fluorescence intensity (Figure 5d), and the mean intensity of the cell which contains it (Figure 5e). We observed a dramatic increase in mean cellular fluorescence between cells which express only tdPCP-mCherry and cells which express it together with a slncRNA, suggesting that slncRNA molecules have some effect in the cytosol, regardless of the granules. To quantify this phenomenon more accurately, we measured the total fluorescence of the population using flow cytometry. For this, we grew cells expressing only the protein component (tdPCP-mCherry), and cells expressing both protein and a slncRNA (PCP-4x/QCP-5x or PCP-24x), with different combinations of induction: IPTG (induces the slncRNA) and C4HSL (induces the protein). The data (Figure 5f) shows that cells expressing a slncRNA, regardless of induction (due to T7 leakiness), show higher fluorescence than cells expressing the protein only. In addition, induction of slncRNA expression with IPTG results in an increase in fluorescence, indicating that slncRNA is a deciding factor in this behavior. Finally, cells expressing the PCP-24x slncRNA show higher fluorescence than cells expressing PCP-4x/QCP-5x, displaying a dependence of the cellular protein titer on number of binding sites available for protein binding.