Fluorescence lifetime sorting reveals tunable enzyme interactions within cytoplasmic condensates

The RNA life cycle is associated with an array of membraneless organelles known as ribonucleoprotein (RNP) condensates (Banani et al., 2016; Alberti and Hyman, 2021). Evolutionarily conserved RNP condensates impact RNA metabolism by compartmentalizing the aqueous spaces of membrane-bound organelles. Nucleoli, Cajal bodies, and nuclear speckles catalyze ribosome assembly, transcription, and splicing in the nucleus, respectively (Banani et al., 2016; Alberti and Hyman, 2021). Mitochondrial granules maintain RNA stability in the mitochondria (Jourdain et al., 2013; Antonicka et al., 2013). In the cytoplasm, mRNA processing (P)-bodies are enriched with factors associated with mRNA sequestration, silencing, editing, and decay (Decker and Parker, 2012; Hubstenberger et al., 2017). Despite the reconstitution and functional characterization of many RNA processing enzyme complexes in vitro (Jonas and Izaurralde, 2013; Chang et al., 2014; Tibble et al., 2021; Schirle et al., 2014; Cho et al., 2003), the co-existence of condensate-enriched enzyme complexes within the surrounding milieu at dilute, non-negligible concentrations has obscured the full impact of many RNP condensates on cellular function. Furthermore, there are technical hurdles associated with isolating or reconstituting RNP condensates for functional assays (Hubstenberger et al., 2017; Freibaum et al., 2021; Currie et al., 2023). Therefore, distinguishing RNP condensate functions from dilute phase functions requires the development of new tools and approaches to bridge existing biochemical studies on enzyme function with condensate-level resolution in live cells.

Conventional fluorescence microscopy approaches measure fluorescence as a function of photon counts, thus relying upon intensity as a single parameter for separating desired signals from the background. The sole reliance on intensity for studying condensate-enriched proteins is suboptimal because the dilute fraction of fluorescently labeled condensate proteins elevates the amplitude of background signals, thus decreasing signal-to-noise ratios. Fluorescence lifetime imaging microscopy (FLIM) measures fluorescence with two parameters: intensity and fluorescence lifetime. Fluorescence lifetime is sensitive to changes in microenvironment, such as pH (Lin et al., 2003), ion and water density (Malacrida and Gratton, 2018; Mangiarotti et al., 2023), and the proximity of acceptor molecules suitable for Förster resonance energy transfer (FRET) (Chen et al., 2003; Hinde et al., 2012). Two recent studies also detected fluorescence lifetime shifts when fluorescent proteins were fused to factors enriched in nuclear condensates (Pliss et al., 2019) and stress granules (Perego et al., 2023), thus making fluorescence lifetime a promising parameter to exploit for condensate studies. Additionally, FLIM provides investigators with additional signal processing approaches to potentially distinguish condensed and dilute phases independent of fluorescence intensity, such as phasor plot analyses (Weber, 1981; Digman et al., 2008; Malacrida et al., 2021).

FLIM excites samples with a pulsed light source to capture time-correlated-single-photon-counting (TCSPC) histogram data and generate fluorescence decay curves for each pixel. Pixels are moved from the time domain to frequency domain by performing a Fourier transform on fluorescence decay curves to yield G and S phasor coordinates, such that pixels with similar fluorescence lifetimes cluster together on a phasor plot (Fig. 1 A). Pixel clusters on a phasor plot can be segmented to provide a threshold-independent approach for resolving desired signals from undesired signals, such as spectral bleed-through, autofluorescence, and light source noise (Digman et al., 2008; Malacrida et al., 2021). Here, we evaluated the relationship between fluorescence lifetime and condensates by analyzing TCSPC FLIM data of fluorescently tagged proteins that localized to either stress granules or P-bodies. The detection of condensate-specific fluorescence lifetime signatures provided the basis for developing phasor plot-based approaches to isolate condensate pixels with the purpose of resolving key proteinprotein interactions within condensates using FLIM-FRET.

Stress granules are non-spontaneous RNP condensates that are absent under basal conditions and form within minutes of exposure to oxidative stress (Sanders et al., 2020; Kedersha et al., 1999). The inducible property of stress granules provides ideal experimental conditions to evaluate the sensitivity of fluorescent proteins in condensate microenvironments. Human osteosarcoma (U-2 OS) cells stably expressing monomeric NeonGreen (mNG) fused to a core stress granule protein, Ras GTPase-activating protein-binding protein 1 (mNG-G3BP1), were captured using FLIM before and during the oxidative stress response. As expected, fluorescence intensity images showed cells reorganizing mNG-G3BP1 from a diffuse dilute phase to a mix of stress granules and a dilute phase after the addition of sodium arsenite (NaAsO₂) (Fig. 1 B). Since unstressed cells displayed no detectable stress granules, we compared mNG-G3BP1 mean fluorescence lifetimes from the dilute phase of unstressed cells to stress granules of stressed cells. A traditional organelle segmentation algorithm using background subtraction (Otsu, 1979; Bray and Carpenter, 2018), referred to herein as intensity-threshold segmentation, was used to generate a stress granule mask and obtain stress granule-localized fluorescence lifetime histograms (Fig. 1 C).

The fluorescence lifetimes of mNG excited inside stress granules were longer than the lifetimes of signals emitted from the dilute phase of unstressed cells (Fig. 1 and Fig. S1). Notably, we observed similar changes in stress granule–localized fluorescence lifetime with an enhanced green fluorescent protein (G3BP1-eGFP) when analyzing data captured from a different FLIM system (Perego et al., 2023; Rossetta et al., 2022) (Fig. S2). Detectable fluorescence lifetime differences between stress granules and dilute phase suggest that condensate-localized pixels have fluorescence lifetime signatures, which can be exploited for a broad range of condensate-relevant analyses, such as calculating condensate size and abundance or tracking condensate-enriched interactions.

We further tested whether fluorescence lifetime can be used to distinguish condensate microenvironments from dilute phase by artificially recruiting a non-stress granule resident protein to stress granules using rapamycin-inducible FKBP and FRB (Fujisawa Kaihatsu-506 binding protein and FKBP-rapamycin binding protein) dimerization domains (Csordás et al., 2010) (Fig. 1 E). U-2 OS cells stably co-expressing mTurquoise2 (mTQ2)-G3BP1-FKBP and valosin-containing protein (mNG-VCP-FRB) were first exposed to oxidative stress to form stress granules, then FLIM images were captured before and after a 10-min rapamycin treatment (Fig. 1 F). The rapid recruitment of VCP to stress granules induced a shift in mNG-VCP-FRB fluorescence lifetimes (Fig. 1 G). Taken together, these data demonstrate the potential of using fluorescence lifetime to isolate condensate pixels for particle and time-lapse FLIM-FRET analyses, which enables the visualization of protein–protein interaction network dynamics within RNP condensates.

Organelle segmentation by intensity-threshold segmentation remains challenging for organelles with a broad range of fluorescence intensities, such as the endoplasmic reticulum (ER) (Fig. S3), or small condensates, such as P-bodies or nuclear speckles. Moreover, intensity-threshold segmentation is particularly inefficient at parsing desired from undesired signals in tissue samples with high background, or when studying condensate proteins with weak phase separation, which yields a bright dilute phase (Riback et al., 2020). Phasor plots cluster pixels with similar fluorescence lifetimes and spatially separate pixels with different lifetimes, thus providing the potential to sort condensate pixels from dilute phase pixels without background subtraction. Here, we tested whether phasor plots can resolve stress granule–specific fluorescence lifetime shifts to generate condensate masks for particle and lifetime analyses (Fig. 2). Unfiltered phasor coordinates from mNG-G3BP1 and G3BP1-eGFP FLIM data (Perego et al., 2023) displayed a single cluster on phasor plots, which was not amenable for stress granule segmentation (Fig. S2 C and Fig. S4 A). Since signal processing and filtering of phasor coordinates have both been shown to improve lifetime cluster separation on phasor plots (Digman et al., 2008; Wang et al., 2021), we tested the effectiveness of median- and wavelet-filtered phasor coordinates for separating stress granule and dilute phase fluorescence lifetime clusters (Fig. 2 A; Fig. S2, D and E; and Fig. S4, A and B).

We used open-source, Python-based software to transform TCSPC data into unfiltered and median-filtered phasor coordinates (Gottlieb et al., 2023). A previous study combined a dual-tree complex wavelet filter with an Anscombe transform to improve the resolution of FLIM data captured from samples with low signal-to-noise (Wang et al., 2021). Thus, we developed a Python-based wavelet filter optimized for resolving condensates from the dilute phase on phasor plots to process unfiltered phasor coordinates into wavelet-filtered coordinates. Under basal conditions, mNG-G3BP1 signals were enriched within a single dilute phase lifetime cluster on all phasor plots (Fig. 2 A; and Fig. S4, A and B, top panels). Strikingly, after cells were exposed to oxidative stress, phasor plots with median- and wavelet-filtered phasor coordinates revealed two distinct mNG-G3BP1 lifetime clusters (Fig. 2 A and Fig. S4 B, bottom panels). Importantly, analogous FLIM data captured from oxidative stress-exposed cells expressing G3BP1-eGFP on a different FLIM system responded similarly to median and wavelet filtering (Fig. S2, D and E).

These data show that stress granule pixels can be distinguished from dilute phase pixels on filtered phasor plots, thus demonstrating the feasibility of using filtered phasor plots to segment condensates without signal subtraction. Furthermore, the wavelet filter outperforms the median filter by retaining spatial resolution in time domain and by increasing stress granule lifetime cluster separation in the frequency domain (Fig. 2 A compared to Fig. S4 B). Thus, we focused herein on optimizing phasor plot segmentation using the wavelet filter.

Next, we evaluated whether wavelet-filtered phasor plots are capable of sensing and sorting P-body pixels, which are significantly smaller condensates than stress granules. P-bodies are evolutionarily conserved RNP condensates that are present under basal conditions and responsive to oxidative stress. Time-lapse FLIM images were captured in human alveolar basal epithelial (A549) cells stably expressing the core P-body protein, mNG-DDX6 (dead-box helicase 6), before and during oxidative stress (Fig. 2 B). TCSPC data were used to generate wavelet-filtered phasor plots (Fig. 2 C). DDX6-marked P-body phasor plots displayed multiple lifetime clusters like stress granule phasor plots. Strikingly, P-body pixels displayed a larger separation from dilute phase pixels on phasor plots compared to stress granules, which could reflect differences in the partitioning of G3BP1 and DDX6, or a difference in the photo-physics of mNG when tagging different proteins (Fig. 2 C).

Manual phasor plot segmentation is subjective, time-consuming, and inaccessible to most FLIM systems. Therefore, we developed an unsupervised machine-learning approach to segment condensate lifetime clusters on wavelet-filtered phasor plots using Gaussian mixture modeling (GMM) (Vallmitjana et al., 2021), which analyzes covariance to identify gaussian centers within spatially distributed data. Thus, we used GMM calculated covariances around the center coordinates of the condensate cluster to define region of interest (ROI) size. Additionally, most stress granule factors display weaker partitioning compared with many P-body proteins (Sanders et al., 2020). Here, we show that differences in partitioning strengths appear to be proportional to the distances between the condensed and dilute phase lifetime clusters (Fig. 2 D). We exploited the separation between condensed and dilute phase clusters to optimize ROI size, such that stress granules labeled with mNG-G3BP1 and P-bodies labeled with mNG-DDX6 were efficiently segmented on phasor plots with elliptical ROI sizes determined by 2.5 and 3.0 times the covariance, respectively. The segmentation of wavelet-filtered phasor plots produced stress granule and P-body masks, which precisely captured pixels within condensates of a wide range of size and brightness (Fig. 3, A and D; and Fig. S2 E).

Next, the performance of phasor plot segmentation was evaluated by calculating the extent of overlap between intensity-threshold-generated masks of stress granules and P-bodies (Fig. 1 C, bottom, and Fig. 3 E) and their respective wavelet-phasor plot masks (Fig. 3, A and D). Scatterplot analysis of mask overlap as a function of stress granule area reveals that phasor plot segmentation matches the performance of intensity-threshold segmentation for larger structures and significantly outperforms for smaller structures (Fig. 3, B and C). For P-bodies, phasor plot segmentation significantly outperforms intensity-threshold segmentation for quantifying both P-body size and abundance (Fig. 3 E). Together, these experiments demonstrate the utility and precision of phasor plot filtering and automated segmentation to precisely distinguish RNP condensates from their respective dilute phases across multiple FLIM systems.

The ability to resolve RNP condensates by two parameters, intensity and fluorescence lifetime, provides a foundation for developing condensate-specific FLIM-FRET assays. Briefly, FLIM-FRET is calculated by measuring the extent of donor lifetime shortening, which is a function of the distance between donor and acceptor fluorophores (Malacrida et al., 2021; Chen et al., 2003) (Fig. 4 A). A major advantage of a donor FRET readout compared with the acceptor readout of classic FRET is that FLIM-FRET assays can be centered around donor fluorophores engineered to be bright, photostable, and monomeric, such as mNG (Shaner et al., 2013; Heim and Tsien, 1996). Additionally, FLIM-FRET is not affected by spectral bleed-through, which enables the use of efficient FRET acceptors while circumventing the impact of direct acceptor excitation with donor excitation laser pulses. Here, we describe an automated condensate FLIM-FRET analysis pipeline founded on the ability to extract condensate-specific pixels from wavelet-filtered phasor plots. Moreover, we outline two experimental approaches using two different classes of acceptor fluorophores to increase accessibility and rigor.

The basis for condensate FLIM-FRET relies on reference cells, which only emit donor fluorescence, and experimental cells capable of emitting donor and acceptor fluorescence. Donor TCSPC data from reference and experimental cells are processed to generate wavelet-filtered phasor plots, which are then segmented to sort condensate-specific pixels and generate donor condensate masks from the two cell lines (Fig. 4 A). Reference and experimental condensate masks are then used to compile reference and experimental lifetime histograms to calculate mean FRET efficiency (Clegg, 1995) with single condensate resolution.

To demonstrate the feasibility of condensate FLIM-FRET, we evaluated known interactions between DDX6 and Lsm14a in P-bodies, and G3BP1 homodimers in stress granules (Sanders et al., 2020; Brandmann et al., 2018). P-body–localized fluorescence lifetimes were calculated from A549 reference cells expressing mNG-DDX6 (donor) (Fig. 3 A) and from A549 experimental cells co-expressing mNG-DDX6 and mScarlet-Lsm14a (acceptor) (Fig. 4 B). The two cell lines were captured with FLIM on the same day using identical capture and stress-inducing conditions. P-body–specific lifetime histograms (Fig. 4 C) and mean FRET efficiencies of individual P-bodies (Fig. 4 D) revealed FRET efficiencies of 9.0% and 10.4% before and during stress, respectively. A similar FRET efficiency (11.3%) was observed in stress granules with dimers consisting of mNG-G3BP1 and mScarlet-G3BP1 (Fig. S5). Together, condensate FLIM-FRET analyses of core protein interactions within stress granules and P-bodies provide a benchmark for interpreting the significance of FRET efficiency measurements between dynamic and potentially tunable interactions.

The regulated assembly of biomolecular complexes constitutes a fundamental cellular mechanism for dictating where and when cellular processes occur. P-bodies enrich for enzyme complexes that drive mRNA decapping and decay (Jonas and Izaurralde, 2013; Chang et al., 2014; Tibble et al., 2021), mRNA silencing (Schirle et al., 2014), and mRNA editing (Cho et al., 2003). However, a major question is whether P-bodies concentrate mRNA decay factors to inhibit or to facilitate their assembly. RNA sequencing performed on P-bodies isolated by fluorescence particle sorting identified largely intact transcripts (Hubstenberger et al., 2017), thus supporting a P-body model of mRNA sequestration and an inhibition of mRNA decay. Conversely, purified subunits of the yeast mRNA decapping complex, Dcp2, Dcp1, and EDC3 (enhancer of decapping 3) readily formed condensates in vitro to increase decapping rates of mRNA substrates suggesting P-bodies might enrich mRNA decapping complex subunits to enhance Dcp2 enzyme function (Tibble et al., 2021). However, both approaches were conducted outside of a live cellular environment (Tibble et al., 2021; Hubstenberger et al., 2017). Therefore, we used condensate FLIM-FRET to determine whether key subunit interactions of the human mRNA decapping complex occur within P-bodies.

Using mNG and mScarlet as FRET pairs, we evaluated interactions between Dcp2-Dcp1A, Dcp2-EDC3, and EDC3-Dcp1A in A549 cells (Fig. S6). Under basal conditions, P-body–enriched Dcp2-Dcp1A and EDC3-Dcp1A yielded FRET efficiencies of 8.5% and 9.9%, respectively (Fig. S6), which are in line with DDX6-Lsm14a FRET efficiency calculations (Fig. 4 D). Dcp2-EDC3 yielded a significantly lower FRET efficiency (3.3%), which suggests EDC3 might primarily interact with Dcp1A and not Dcp2 in humans. After 45 min of oxidative stress, Dcp1A-EDC3 FRET remains unchanged, whereas FRET efficiency between Dcp2-Dcp1A and Dcp2-EDC3 drops to 2.8% and 1.8%, respectively. These data reveal context-dependent plasticity of the mRNA decapping complex within human P-bodies, whereby Dcp2 and Dcp1A interactions are reduced during stress despite their co-enrichment with P-bodies. Additionally, human Dcp1A stably interacts with EDC3 and Dcp2 only under basal conditions (Fig. S6), which is contrary to yeast where yeast Dcp2 has been shown to interact with Dcp1 and EDC3 (Charenton et al., 2016). Thus, our P-body FLIM-FRET data of the human decapping complex supports a model, which attributes differences between yeast and human decapping complex interactions to the transfer of a long intrinsically disordered region (IDR) from yeast Dcp2 to human Dcp1A/Dcp1B (Jonas and Izaurralde, 2013).

Technically, the dual cell line approach is an efficient and cost-effective approach for FLIM-FRET experiments, especially if expression constructs for candidate proteins fused to FRET pairs already exist. It is important to note that using two cell lines will have intrinsically different condensate compositions because reference cell lines only express mNG-tagged condensate proteins, whereas experimental cell lines express mNG- and mScarlet-tagged proteins. Below, we describe a condensate FLIM-FRET approach, which relies on a single cell line to report reference and experimental lifetimes from condensates with intrinsically identical compositions (Fig. 5 A).

The compositional and material properties of biomolecular condensates can be sensitive to changes in condensate factor expression, as well as the expression of certain fluorescent proteins, such as green and red fluorescent protein (GFP, RFP) (Currie et al., 2023; Barkley et al., 2024). HaloTag and mNG displayed the best performance for studying viral condensates because cells expressing Halo-, mNG-tagged, or untagged viral proteins yielded condensates with similar size, shape, number, and diffusion rates (Barkley et al., 2024). Thus, we aimed to overcome two technical challenges of the mNG-mScarlet FLIM-FRET pair by replacing mScarlet with HaloTag (Fig. 5 A). First, potential condensate artifacts associated with expressing an RFP (mScarlet) are circumvented. Second, compositional and material differences between reference and experimental condensates assayed from different cell lines are avoided by developing a single-cell line approach to capture reference and experimental condensate lifetimes. Here, we tested a commercially available, cell-permeable dye (JaneliaFluor JF549), which, when conjugated to Halo, has a predicted Förster radius (R_o) of 58.65Å with mNG (Lambert, 2019).

To evaluate the validity of a mNG-HaloTag FLIM-FRET approach, we stably expressed mNG-Dcp2 in DCP1A/DCP1B double knockout (1A/1B DKO) U-2 OS cells (Fig. S7) and acutely transduced the expression of either Halo-Dcp1A or Halo-DDX6. Reference and experimental FLIM images of mNG-Dcp2 and HaloTag-expressing 1A/1B DKO cells were captured prior to and 30 min after JF549 addition to the cell culture media, respectively (Fig. 5 B). At 0 min, Halo-Dcp1A– and Halo-DDX6–expressing cells displayed a single mNG-Dcp2 lifetime cluster (Fig. 5 C: left phasor plot). Strikingly, after 30 min of JF549, only cells expressing Halo-Dcp1A yielded a decrease in mNG-Dcp2 fluorescence lifetimes (Fig. 5 C: top-right phasor plot). Furthermore, JF549 labeled a subset of cells to reveal HaloTag expressors (magenta/green boundaries) and non-HaloTag expressors (white boundaries) (Fig. 5 B). The two subpopulations of cells from the Halo-Dcp1A condition were reflected by two distinct P-body lifetime clusters on phasor plots (Fig. 5 C) and P-bodies with ∼0% and ∼7% mean FRET efficiencies (Fig. 5 D). Dcp2 and DDX6 have not been reported to directly interact in humans, which is reflected on phasor plots and a near-zero percent mean P-body FRET efficiency (Fig. 5, C and D). Together, these data demonstrate the viability and sensitivity of using mNG and JF549-HaloTag as condensate FLIM-FRET pairs.

Mammalian cell lines re-engage bulk translation and disassemble stress granules within 1 to 2 h of stress recovery (Wheeler et al., 2016; Moon et al., 2019; Lee et al., 2020); therefore, we chose the 2-h timepoint to assess whether Dcp2-Dcp1A FRET returns to prestressed levels (Fig. 6 A). 1A/1B DKO U-2 OS cells stably expressing mNG-Dcp2 and acutely expressing Halo-Dcp1A were assessed for condensate FLIM-FRET before stress, during stress, and after 2 h of stress recovery. Experimental P-body lifetime clusters displayed a clear FLIM-FRET shift on wavelet-filtered phasor plots under basal conditions and a movement closer to the reference cluster during stress, which corresponded to mean Dcp2-Dcp1A FRET efficiencies of 9.0% and 4.7% within P-bodies, respectively (Fig. 6, B and C). 2 h into stress recovery, experimental P-body lifetime clusters surprisingly maintained a tight overlap with reference clusters and reported mean Dcp2-Dcp1A FRET efficiencies of 3.0% within P-bodies, which suggests mRNA decapping remains suppressed even after the stressor was removed (Fig. 6, B and C).

To test whether the absence of Dcp2-Dcp1A FLIM-FRET during recovery was due to the loss of endogenous Dcp1B in 1A/1B DKO U-2 OS cells or if the lack of Dcp2-Dcp1A FLIM-FRET recovery is cell-type specific, an analogous experiment was performed in A549 cells stably expressing mNG-Dcp2 and acutely expressing Halo-Dcp1A or Halo-DDX6. Cells co-expressing Halo-DDX6 yielded near-zero mean P-body FRET efficiencies for all treatment conditions (Fig. 6 D). Conversely, Halo-Dcp1A co-expressing cells registered mean P-body FRET efficiencies of 12.0%, 4.8%, and 4.3% under basal, stressed, and stress recovery conditions, respectively (Fig. 6 D). Collectively, these data demonstrate the tunability of a critical interaction for mRNA decapping under basal and stressed conditions. Furthermore, these data suggest that mRNA decapping activity lags behind the resumption of mRNA translation during stress recovery.

FRET detection is evaded when the orientation of two interacting proteins position donor and acceptor fluorophores beyond their Förster radius. Thus, any structural insight into how proteins interact can be useful for determining whether to place fluorophores on the N- or C-terminus. As protein structure prediction algorithms continue to advance, algorithms, such as AlphaFold (Abramson et al., 2024) and RoseTTAFold (Krishna et al., 2024), will become increasingly useful FLIM-FRET experimental design tools for predicting optimal donor and acceptor fluorophore placement. Below, we provide a conservative approach to using protein interaction structural maps generated by AlphaFold3 for predicting of the distance between donor and acceptor fluorophores, and for connecting our FLIM-FRET experimental results with AlphaFold3-calculated confidence levels using the predicted local distance difference test (pLDDT) (Mariani et al., 2013).

We modeled interactions between mNG-Dcp2 and Halo-Dcp1A using AlphaFold3 and measured a distance of 30.58 Å between the two fluorophores, which is within the Förster radius of mNG and JF549 (Fig. 7, A and B). Next, we cataloged potentially interacting atoms by their pLDDT confidence level. Atoms of two proteins are considered potential interactors or contacting when they are ≤4Å apart. Highlighted is a contact between high-confidence atoms of Dcp2 and Dcp1A with pLDDTs ≥0.9 (Fig. 7 A inset), which has been experimentally verified (Chang et al., 2014). Therefore, contacting atoms between Dcp2 and Dcp1A were binned and filtered between high (pLDDT ≥ 0.9 red), medium (pLDDT ≥ 0.7 white), and low (pLDDT ≥ 0.5 blue) confidence levels of atoms on both sides of the contact site. Dcp2-Dcp1A and Dcp1A-EDC3 interaction models yielded 21 and 27 high-confidence contacts, respectively (Fig. 7 C). Modeled interactions between the G3BP1 homodimer and DDX6-Lsm14a both showed the presence of high confidence contacts. Conversely, no high-confidence contacts were detected for Dcp2-EDC3 or Dcp2-DDX6 interaction models, such that the presence or absence of high-confidence contacts are consistent with our FLIM-FRET results (Fig. 7 C). Since many proteins enriched within condensates contain IDRs, the utility of protein structure algorithms for condensate FLIM-FRET will grow as algorithms improve their ability to predict the atomic positions of IDRs within complexes.

Biomolecular condensates provide cells with the reflexes to acutely concentrate hundreds of molecules critical for cellular and organismal health within minutes in response to intrinsic and extrinsic stimuli. However, condensates remain a black box with respect to how their molecular networks are wired and dynamically regulated in different cellular contexts. This study reveals that subtle fluorescence lifetime differences between condensed and dilute phases can be exploited by phasor plot filtering to faithfully isolate condensate-specific pixels for rigorous FLIM-FRET analyses across tens to hundreds of cells. We used condensate FLIM-FRET to provide the first evidence that human mRNA decapping complex subunits interact within P-bodies. The tracking of P-body interactions under basal, stressed, and stress recovery conditions also revealed an unexpected layer of regulation, which converged on interactions between the mRNA decapping enzyme and a key coactivator. Our results raise new questions about how P-bodies and other condensates coordinate the assembly of enzyme complexes and re-ignites existing questions of whether complexes within P-bodies are catalytically active.

Short-read RNA-sequencing (RNA-seq) of P-body– or stress granule–enriched fractions has demonstrated that these condensates are sites of mRNA sequestration (Hubstenberger et al., 2017; Matheny et al., 2019). Our live-cell results and others (Blake et al., 2024) support the possibility that mRNA decay might occur in parallel to mRNA sequestration within P-bodies. Blake et al. developed a rapamycin-inducible non-sense mediated RNA decay (NMD) assay to show that P-bodies coalesce around RNA substrates upon recruitment of the NMD machinery. P-body-localized NMD was attenuated during NaAsO₂-induced oxidative stress (Blake et al., 2024), which is consistent with our results showing that Dcp2-Dcp1A FRET was significantly decreased under the same stress condition. Consistent with DCP2 knockdown studies in drosophila cells (Eulalio et al., 2007), inhibiting Dcp2 enzyme function by gene ablation of DCP1A and DCP1B in U-2 OS cells resulted in significant increases in P-body size and abundance (Fig. 6 and S7). Re-expression of Dcp1A took several days to restore P-body abundance to lower wild-type U-2 OS numbers suggesting that DCP1A/1B DKO cells are primed for P-body–localized RNA decay. Thus, short-read and true-end-to-end RNA-seq (Ibrahim et al., 2021) performed on P-bodies isolated from DCP1A/1B DKO cells with and without re-expression of Dcp1A would provide insight into which transcripts are accumulating within P-bodies from mRNA decapping-deficient cells, and whether transcript read densities in rescue cells will differ from published P-body–enriched transcriptomes, respectively.

Condensate FLIM-FRET provides the ability to connect descriptive live-cell studies of condensate dynamics with functional studies from reconstituted systems. Here, Dcp2 and Dcp1A were targeted because previous studies showed inhibition of decapping activity with immunoprecipitated human Dcp2 harboring mutations that disrupt Dcp1A binding (Chang et al., 2014). Our results from P-bodies in live cells showed decapping-critical interactions between Dcp2 and Dcp1A were present before stress but disrupted during stress, despite interactions between Dcp1A and EDC3 remaining stable under both conditions. The continued suppression of Dcp2 and Dcp1A interactions at a stress recovery timepoint where most cells have restarted bulk mRNA translation, suggests that the potential return of mRNA decapping might represent a decision checkpoint for cell survival or death. Moreover, EDC4/Hedls appears to be important for the interaction of Dcp2 with Dcp1A and other P-body–enriched enzymes (Fenger-Grøn et al., 2005; Brothers et al., 2023). Future condensate FLIM-FRET studies targeting the crosstalk between Dcp2, EDC4, and other RNA processing enzymes will help to elucidate the molecular switch controlling Dcp2 and Dcp1A interactions and shed light on the roles of enzyme complexes within P-bodies at different stages of the cell cycle, in different cell types, or under different types of stress.

Lastly, our wavelet-filtered, phasor plot segmentation workflow has been validated for processing TCSPC data from multiple FLIM systems, which will enable the rigorous use of a technology that is increasingly found in imaging core facilities around the world. More broadly, this study provides a blueprint to expand the use of FLIM phasor plots for resolving dilute phase FLIM-FRET, calculating protein partitioning between dilute and condensed phases, and tracking the dynamics of fluorescently labeled RNA and RNA–protein interactions within condensates. Furthermore, organelles perform multiple functions through the enrichment of different types of machinery and their interactions with other organelles (Banani et al., 2016; Alberti and Hyman, 2021; Wu et al., 2018; Lee et al., 2020; Liao et al., 2019; Friedman and Voeltz, 2011). Thus, our automated FLIM-FRET approach provides the potential to resolve functional subdomains within or between organelles to simultaneously evaluate organelle form and function in live cells.

Using Gibson Assembly, we generated LR clonase competent entry vectors (pENTR D-TOPO backbone) with N- or C-termini fluorescence tags (mTQ2, mNG, mScarlet, or HaloTag) with identical multiple cloning sites to pAcGFP1-C1 or -N1 (Clontech/Takara), and second-generation lentiviral destination expression plasmids (pHAGE backbone) driven by a mouse phosphoglycerate kinase promoter lacking mammalian selection genes. G3BP1, Dcp2, Dcp1A, and EDC3 were subcloned by restriction enzyme digest from previously published non-viral mammalian expression plasmids (Lee et al., 2020) and inserted into entry vectors with encoded fluorescent tags. The luminal ER marker, mCh-KDEL, consists of the BiP signal sequence, mCherry, and the KDEL retention sequence (ssBiP-mCh-KDEL), which was PCR amplified from an existing plasmid (Lee et al., 2020) and inserted into a pENTR D-TOPO backbone. DDX6 and Lsm14a were PCR amplified from a HEK293T cDNA library and inserted into entry vectors after XhoI and KpnI, or EcoRI and BamHI digestion, respectively. Genetically encoded fluorescent tags and genes within entry vectors were transferred into lentiviral destination plasmids by Gateway cloning.

DCP1A/1B DKO U-2 OS cells were generated by transient cotransfection of two expression plasmids encoding for Cas9 and DCP1A or DCP1B guide sequences with Lipofectamine 3000 followed by clonal selection. Gene and protein deletion were confirmed by PCR and western blot, respectively. Briefly, primers encoded with two guide sequences targeting a critical exon within DCP1A were used to PCR amplify an insert sequence encoding for constant region 3 (CR3) and the murine U6 promoter. The BsmBI-guide 1-CR3-mU6-guide 2-BsmBI amplicon was inserted into LentiCRISPRv2 plasmid (#52961; Addgene) by Golden Gate Assembly. The same cloning strategy was used to generate a LentiCRISPRv2 plasmid encoding two guide sequences targeting DCP1B.

• DCP1A guides: 5′-CCA​CAA​TTT​GTG​GAA​TGA​GA-3′ and 5′-GGA​CCT​TAT​TCG​TAT​ATC​GA-3′.

• DCP1B guides: 5′-AGG​TCT​GCT​TCT​CCA​AAG​CA-3′ and 5′-GTG​AGT​CTT​TCT​TAC​TAC​AG-3′.

• DDX6 Forward Primer Oligo: 5′-AGC​ACG​GCC​AGA​ACA​GAG​AAC-3′.

• DDX6 Reverse Primer Oligo: 5′-TTA​AGG​TTT​CTC​ATC​TTC​TAC​AGG​CTC​GC-3′.

• Lsm14a Forward Primer Oligo: 5′-AGC​GGG​GGC​ACC​CCT​TAC​ATC-3′.

• Lsm14a Reverse Primer Oligo: 5′-TTA​GGG​TCC​AAA​AGC​TGT​GGT​TTT​CCT​ATA​TTC​AAA​ATC​C-3′.

Human embryonic kidney (HEK293T - CRL-3216) cells were obtained from the American Tissue Culture Collection (ATCC) and maintained in Dulbecco’s Modified Eagle Media (DMEM; Sigma-Aldrich) and 10% fetal bovine serum (FBS; VWR) without antibiotics. Approximately 80% confluent cells in 6-cm dishes were transfected by combining destination plasmid, and the second-generation packaging and envelope plasmids, psPAX2 (12260; Addgene) and pMD2.G (12259; Addgene) with Lipofectamine 3000. After 5–6 h of incubation, cells were rinsed and replenished with 5 ml of fresh media. Lentivirus-containing media was collected after 48 and 72 h, filtered with 0.45-μm polyether sulfone filters, and spun to remove debris. The two harvests were then pooled and spun at 21,000 rcf for 2 h in a microcentrifuge to concentrate lentivirus. After centrifugation, the supernatant was carefully removed and discarded. Lentivirus was resuspended with 500 μl of Lentivirus Stabilizer (OriGene), aliquoted, and stored at −80°C.

Human osteosarcoma (U-2 OS - HTB-96) and alveolar basal epithelial (A549 - CCL-185) cells were obtained from ATCC and maintained in McCoy’s or DMEM-F12 media (both-Sigma-Aldrich) supplemented with 10% FBS and penicillin/streptomycin (Life Technologies). Cell lines were tested for contamination at the start of experiments and every 3 wk after using a PCR-based mycoplasma test (Invivogen). All single- and multigene exogenous expression cell lines used in this manuscript were generated by combining 5 μg/ml polybrene (Sigma-Aldrich) with one or multiple lentiviruses, and transducing ∼80% confluent cells overnight followed by replenishing with fresh media for one recovery day. Cells were split twice over the course of three days prior to experiments on polyclonal cultures, seeded in 96-well plates for clonal selection, or isolated by fluorescence-activated cell sorting (FACS). FACS or clonal selection enabled the enrichment of cell populations with similar fluorescence expression levels, which exhibited quality-controlled organelle standards set by our immunofluorescence characterizations in U-2 OS and A549 cells of endogenous P-body and stress granule abundance and responses to oxidative stress, as well as ER morphology. All cells were maintained at 37°C, 5% CO₂, and 90% humidity.

Live-cell confocal FLIM was performed on a Stellaris 8 FALCON (Leica) confocal microscope equipped with a 78 MHz pulsed white light laser (440–790 nm), hybrid detectors with gallium arsenide phosphide photocathode front plate coupled with an avalanche diode, 1.40-NA plan-apochromat 63× oil immersion objective, and stage-top incubation and objective heater system (Okolab).

mTQ2 was captured with 440-nm excitation and a 460–490-nm emission detection window. mNG was captured with 504-nm excitation and a 520–550-nm emission detection window. mScarlet and JF549 were captured with 570-nm excitation and 585–700-nm emission detection windows. JF549 was captured with a suboptimal excitation wavelength to minimize cross-excitation of mNG. mCh was captured with 587-nm excitation and a 600–700-nm emission detection window. Non-overlapping tile scans with 512 × 512-pixel times were captured with a 150 Hz scan rate, 4× line accumulation, 3.0 zoom, and 0.12-µm pixel size. The incubation chamber was maintained at 37°C, 5% CO₂, and 90% humidity. Cells were imaged in their complete growth medium listed above. NaAsO₂ was dissolved in deionized water (dH₂O) to yield a 0.5 M stock solution just before treatment. NaAsO₂ was spiked into the microscope dish during time-lapse imaging to a final concentration of 0.5 mM. To track cells before and after stress, care was taken to avoid disturbing the dish during NaAsO₂ addition. LAS X Microscope Software (Leica) was used to operate the microscope and capture data.

CellProfiler was used for intensity-based segmentation of ER, stress granules, and P-bodies. The ER was segmented using a modified CellProfiler workflow developed for ER morphology analysis (Garcia-Pardo et al., 2021). Stress granules and P-bodies were segmented with a custom CellProfiler workflow. Briefly, a Gaussian filter was first applied to intensity to increase signal-to-noise ratios (SNRs) between dilute and condensed phases. Primary objects were identified from filtered images using Otsu’s thresholding algorithm. Particle number and area were computed and exported along with binary images.

TCSPC data for experiments and FITC calibration were exported as raw .bin files from LASX FALCON (Leica) and converted to TIFF files in ImageJ. TIFF files for experiments and calibration were imported in the Fluorescence Lifetime Ultimate Explorer (FLUTE) Python GUI (graphical user interface), which then provided either unfiltered or median-filtered G and S coordinates. Unfiltered and median-filtered phasor plots were generated using our custom phasor plot Python script by weighting G and S coordinates with the intensity images. For wavelet-filtered phasor plots, unfiltered G and S coordinates from FLUTE were processed using our custom complex wavelet filter Python script. First, an Anscombe transform is applied to the Fourier coefficients and Intensity to shift G and S coordinate maps with approximate standard Gaussian distributions. A Dual-Tree Complex Wavelet Transform was then applied with nine filtering levels followed by an inverse Anscombe transform to generate wavelet-filtered coefficients and intensity, which were used to calculate wavelet-filtered G and S coordinates.

This section details the methodology behind our automated wavelet-filtered phasor plot segmentation approach to precisely isolate condensate pixels. Our GMM algorithm requires raw photon counts (fluorescence intensity data), wavelet-filtered G and S coordinates, and an accurate prediction of cluster number. However, lifetime clusters from undesired signal sources, such as autofluorescence, might mislead the GMM resulting in the generation of inaccurate cluster positions. Therefore, we focused the GMM on a circular region of the phasor plot with a large defined radius centered around the ideal fluorophore lifetime by removing pixels outside of this circular region. The curated phasor plot enables the GMM to reproducibly identify two Gaussian, condition-blind center coordinates corresponding to the condensed and dilute phases of RNP condensate proteins, and one Gaussian center for membrane-bound organelles, such as the ER. Next, the ROI size for phasor plot segmentation of condensate pixels is determined by the covariance matrix generated by the GMM to locate cluster centers. Our GMM algorithm returns cluster centers, the magnitude and direction of the minimum covariance around the cluster center, and the maximum covariance magnitude and direction around the cluster center. For circular ROIs used to segment unfiltered FLIM phasor data, we use the minimum covariance value for the radius. For elliptical ROIs used to segment wavelet filtered FLIM phasor data, we use the minimum covariance value for the length of the short axis, maximum covariance value for the length of the long axis, and angle of rotation parallel to the direction of the maximum covariance value. Lastly, for elliptical ROIs used to segment dilute and condensed phases, we adjust the ROI size proportional to the partitioning of the condensate protein captured by using a covariance factor. For mNG-G3BP1, which has a low partition coefficient, we used 2.5 times the covariance to determine ROI size, whereas for mNG-DDX6, which has a higher partition coefficient, 3.0 times the covariance was used, respectively. This approach successfully segmented pixels residing in most, if not all, RNP condensates using the first harmonic wavelet-filtered phasor plots. Condensate-specific wavelet G and S coordinates are then used to build fluorescence lifetime histograms or generate condensate masks by performing an inverse Fourier transform.

Reference donor measurements can be obtained from cells expressing only the mNG-Donor or from cells co-expressing mNG-Donor and Halo-Acceptor without JF549. Experimental donor measurements can be obtained from cells co-expressing the mNG-Donor and mScarlet-Acceptor or from the cells co-expressing mNG-Donor and Halo-Acceptor labeled with JF549. Donor FLIM TCSPC data from reference and experimental cells were transformed into unfiltered phasor plots to generate reference and experimental fluorescence lifetime images. To generate reference and experimental condensate masks, donor FLIM TCSPC data from reference and experimental cells were transformed into wavelet-filtered phasor plots, then segmented by the GMM using a covariance factor of 2.0 to evaluate FLIM-FRET in stress granules and P-bodies.

To obtain the mean donor lifetime of reference condensates (T_D), a condensate ROI list was generated from the reference mask in ImageJ using the Analyze particles function with a lower area limit set to 100 nm². The ROI list was overlayed on the reference fluorescence lifetime image to obtain mean fluorescence lifetime measurements for individual condensates, which was used to calculate T_D.

U-2 OS cells were plated on 35-mm dishes with glass bottoms. Cells were fixed using a 37°C solution of 4% paraformaldehyde and 4% sucrose in phosphate-buffered saline (PBS) for 10 min. Cells were permeabilized with 0.1% triton in PBS for 4 min and blocked with 5% normal donkey serum in PBS for 1 h at room temperature. Cells were incubated with primary mouse monoclonal DDX6 antibody (1:100, SAB4200837; MilliporeSigma) in blocking solution for 1 h at room temperature, washed with PBS three times, and then incubated with donkey anti-mouse secondary antibody conjugated to AlexaFluor-647 (1:400, A-31571; Invitrogen) in blocking solution for 1 h at room temperature. Hoechst (Thermo Fisher Scientific) was used for nuclear staining.

Cells were dissociated using Trypsin-EDTA (Thermo Fisher Scientific) and collected by centrifugation at 200 rcf for 2 min. Cell pellets were resuspended with Laemmli’s sample buffer (BioRad) supplemented with fresh β-mercaptoethanol (BioRad) and heated for 5 min at 95°C. Proteins from whole-cell lysates were resolved by size using SDS-PAGE in precast 4–20% Mini-PROTEAN TGX Stain-Free Gradient Gels (BioRad) and transferred to polyvinylidene difluoride membranes (Thermo Fisher Scientific). The following antibodies were used for western blot: DCP1A (1:200, SC-100706; Santa Cruz), DCP1B (1:500, 13233S; Cell Signaling Technologies), and COX IV (1:1,000, 48448; Cell Signaling Technologies). Immunolabeled membranes were then incubated in ECL Prime Western Blotting Detection Reagent (Cytiva) and imaged using an Azure Imaging System (Azure Biosystems).

P values for experiments involving the comparison of FRET efficiencies of individual P-bodies across two conditions (basal versus stressed) were determined using parametric, nested T tests where a normal distribution is assumed, but not formally tested. For experiments involving three conditions (basal, stressed, and recovery), statistical significance was determined using one-way ANOVA, where a normal distribution was assumed. Comparison of intensity versus phasor-based segmentation and of Mander’s colocalization coefficients before and after stress was done using paired Student’s T tests.

Fig. S1 shows that mNG emits photons with mono-exponential decay. Fig. S2 shows that phasor plot filtering is a universal approach to segment stress granule FLIM data, which is independent of the type of FLIM system. Fig. S3 Shows that FLIM phasor plot-guided segmentation outperforms traditional intensity-threshold segmentation approaches for generating ER masks. Fig. S4 shows unfiltered and filtered phasor plots of mNG localized to condensed and dilute phases. Fig. S5 shows stress granule FLIM-FRET detects G3BP1 dimerization in live cells. Fig. S6 shows context-dependent interactions between subunits of the mRNA decapping complex occur within P-bodies. Fig. S7 shows that gene ablation of Dcp1A and Dcp1B in U-2 OS cells resulted in an increase in P-body size and abundance.

All data is available in the main text. Raw and processed data available upon request. Complex wavelet filtering and GMM segmentation scripts are available on GitHub (https://github.com/LeeLabBCM/ComplexWaveletFilter) and ImageJ Macros scripts will be made upon request.

We thank C. Stringari, S.M. Fessehaye, H. Wu, E. Sawyer, S. Gumbin, and S. Hartig for helpful insights and scientific discussions.

J.E. Lee obtained funding for the work, which was supported by the Cancer Prevention and Research Institute of Texas (CPRIT RR200065) and the National Institutes of Health (R35GM151054). J.E. Lee is a CPRIT Scholar in Cancer Research. J.M. Marcus, L.E. Fahim, N.D. Powell, Z.A. Ralston, K. Walgamotte were supported by a grant from CPRIT (RR200065). J.M. Marcus, L.E. Fahim, N.D. Powell were also supported by grants from the NIH, T32ES027801 and R35GM151054.

Author contributions: L.E. Fahim: Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing - review and editing, J.M. Marcus: Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing - review and editing, N.D. Powell: Formal analysis, Investigation, Software, Validation, Visualization, Writing - review and editing, Z.A. Ralston: Resources, Writing - review and editing, K. Walgamotte: Data curation, Formal analysis, Investigation, Writing - review and editing, E. Perego: Investigation, Validation, Writing - review and editing, G. Vicidomini: Conceptualization, Writing - review and editing, A. Rossetta: Conceptualization, Methodology, Software, Validation, Writing - review and editing, J.E. Lee: Conceptualization, Data curation, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Validation, Visualization, Writing - original draft, Writing - review and editing.