To systematically survey the ability of DFDs to supersaturate, we compiled an exhaustive set of 109 structurally independent human DFDs (Supplementary file 1 and Figure 1—figure supplement 1A–B) and then characterized their tendency to spontaneously self-assemble in near-physiological conditions while minimizing interference from other proteins. For this purpose, we used distributed amphifluoric FRET (DAmFRET) in an orthogonal eukaryotic host that completely lacks endogenous DFDs and their associated regulatory machinery – Saccharomyces cerevisiae. DAmFRET produces a snapshot of the population-level distribution of single-cell measurements of ratiometric FRET (‘AmFRET’) between two fluorescent forms of the same protein species (Khan et al., 2018). The data revealed a diversity of behaviors (Figure 1B–C, Figure 1—figure supplement 1C–D, and Supplementary file 1), ranging from no self-association to self-association in all cells at all concentrations. Importantly, the results closely agree with observations of individual DFDs in the literature. Considering the adaptor FADD as an example, its DD was monomeric, whereas its DED assembled robustly, consistent with prior observations that the DD forms exclusively hetero-oligomers while the DED forms homopolymers (Jang et al., 2014; Fosuah et al., 2025; Wang et al., 2010). Twenty-one DFDs transitioned from no to high AmFRET in a discontinuous manner, the signature of a large intrinsic nucleation barrier (Rodriguez Gama et al., 2022; Khan et al., 2018; Kandola et al., 2023).

Nucleation barriers increase with the entropic cost of assembly (Khan et al., 2018; Buell, 2017). Consequently, assemblies with large barriers tend to be more ordered than those without. Ordered assembly by DFDs occurs as a two-dimensional array twisted into a one-dimensional polymer (Lin et al., 2010; Lu et al., 2014; Rodríguez Gama et al., 2021) that can manifest as microscopically visible filaments in cells (Rodriguez Gama et al., 2022; Lu et al., 2014; Shearwin-Whyatt et al., 2000). To evaluate the expected relationship between nucleation barriers and ordered assembly, we used high-throughput confocal microscopy to examine the subcellular distribution of each protein. We observed subcellular assemblies for most of the DFDs that populated AmFRET-positive states, but not those with entirely low AmFRET (Figure 1B and Figure 1—figure supplement 1E). As predicted, the assemblies of DFDs that had transitioned discontinuously had fibrillar morphologies, whereas those that had transitioned continuously (low to high DAmFRET) instead mostly formed spherical or amorphous puncta (Figure 1B, Figure 1—figure supplement 1F–H).

For discontinuous DFDs, the soluble protomers in the low FRET cells are hypothetically poised to assemble; they just lack a structural template to get them started. We tested this hypothesis by co-expressing genetically encoded ‘seeds’ (Figure 1D–E; Rodriguez Gama et al., 2022) and found that the seeds caused cells to switch to the high AmFRET population for most discontinuous DFDs but not continuous DFDs (Figure 1F, Figure 1—figure supplement 1I, Supplementary file 1). We further demonstrated that assembly is more consistent with native DFD subunit interfaces than amyloid-like misfolding (Supplemental Information, Figure 1—figure supplement 2A–B). This distinction is important because amyloid is the only other form of assembly that has been experimentally shown to be capable of forming a discontinuous DAmFRET profile (Khan et al., 2018; Posey et al., 2021). Altogether, these data confirm that a subset of DFDs can supersaturate in a soluble form to power subsequent switch-like assembly.

The magnitude of supersaturation in vivo is determined by the ratio of a protein’s total concentration to its C_sat, which reflects the strength of interactions between subunits in its assembled structure (Kar et al., 2022). Total concentration and C_sat can evolve independently by gene regulation and DFD sequence, respectively. Consequently, if the function of discontinuous DFDs involves supersaturation, we would expect evolution to have lowered their C_sat while raising their expression relative to continuous DFDs. To test these predictions, we first analyzed the relationship of nucleation barriers to C_sat, as determined by each protein’s transition concentration in the presence of seed. As expected, discontinuous DFDs exhibited lower C_sat values than continuous DFDs (Figure 1G, Supplementary file 1), and in the absence of seed achieved soluble concentrations that exceeded them by at least fourfold on average (Figure 1H).

We next evaluated the relationship of nucleation barriers to DFD concentrations in vivo. Using published proteomic datasets (Wang et al., 2015; Karlsson et al., 2021; Jiang et al., 2020), we found that proteins with discontinuous DFDs tend to be expressed more abundantly than those with continuous DFDs, both at the cell type and tissue levels (Figure 2A, Figure 2—figure supplement 1A–B and Supplementary file 1). We then used a high-coverage transcriptomics dataset (Uhlen et al., 2019) to ask how C_sat values relate to DFD gene expression in primary immune cells. Remarkably, transcript abundances significantly anticorrelated with C_sat in 17 of the 18 canonical immune cell populations (Figure 2B and Figure 2—figure supplement 1C). Such a relationship is highly unusual for soluble proteins, whose expression instead tends to strongly correlate with C_sat (Vecchi et al., 2020). These analyses collectively indicate that discontinuous DFDs are likely to be functionally supersaturated in their endogenous physiological contexts.

Given that nucleation barriers are restricted to only a subset of DFDs, we asked if the DFDs with nucleation barriers tend to have different signaling roles than those without. From our prior demonstrations of prion-like activity by ASC, MAVS, and BCL10 (Rodriguez Gama et al., 2022; Cai et al., 2014), we suspected an enrichment among adaptors. However, DFD proteins have not been systematically evaluated with respect to adaptor function. We therefore focused on the defining property of ‘adaptors’ physically connecting other proteins by quantifying their centrality in the DFD protein-protein interaction network. Proteins with high ‘degree centrality’ are hub-like, with many direct interactions. Proteins with high ‘betweenness centrality’ lie on the shortest paths between other nodes and therefore act as bottlenecks in the network. Assessing both measures (see Methods), we observed that discontinuous and/or seedable DFD proteins have significantly greater degree and betweenness centralities (Figure 2C, Figure 2—figure supplement 1D-E). As a point of comparison, we performed the same analysis for the domain family most closely resembling DFDs – Sterile Alpha Motif (SAM). SAM domains have similar size, structure, number (76 in humans), and function as DFDs, often serving as polymeric scaffolds of signaling complexes (Bienz, 2020; Qiao and Bowie, 2005; Knight et al., 2011). Unlike DFDs, however, SAM domains cannot supersaturate (Rodríguez Gama et al., 2021). Consistent with that, we found that they also have much lower betweenness centrality in their physical interaction network as compared to discontinuous DFDs (p=0. 0003; Figure 2—figure supplement 1F).

For a DFD nucleation barrier to enable on-demand assembly, the DFD must remain supersaturated even as a full-length (FL) protein. This means that other parts of the protein must not trigger nucleation prematurely, as through homo-oligomerization. Nor should they disfavor the assembled state, which would raise C_sat. We therefore evaluated the phase behaviors of 21 diverse DFD-containing FL multidomain proteins. Most of the proteins behaved the same way as their respective DFDs (Figure 2D, Figure 2—figure supplement 2A, and Supplementary file 2). Six proteins suppressed the ability of their DFDs to supersaturate, resulting in continuous low or moderate DAmFRET, suggesting they form non-nucleated oligomers. To determine if the DFDs are in an active polymer configuration within these oligomers, we tested the ability of FL NLRP3 to nucleate its cognate adaptor, ASC. It failed to do so (Figure 2—figure supplement 2B), revealing that the oligomers are ‘autoinhibited’ as has been previously demonstrated for multiple DFD-containing receptors and effectors (Huoh and Hur, 2022; Holliday et al., 2019; Ohto et al., 2022; Hochheiser et al., 2022; Sommer et al., 2005; Andreeva et al., 2021; Chuenchor et al., 2014; Green, 2022). This implies that the DFD within FL NLRP3 (and presumably other DFDs that lost supersaturability in their FL context) cannot self-assemble; instead, its joining the polymer structure is driven by the energy released by ligand (e.g. D/PAMP) binding. This stoichiometric requirement limits spurious activation that would otherwise kill or inflame cells unnecessarily (Huoh and Hur, 2022; Holliday et al., 2019; Ohto et al., 2022; Hochheiser et al., 2022; Sommer et al., 2005; Andreeva et al., 2021; Chuenchor et al., 2014). In contrast to these autoinhibited proteins, five of the FL proteins instead retained or enhanced their DFDs’ nucleation barriers: ASC, BCL10, FADD, MAVS, and TRADD (Figure 2D, Figure 2—figure supplement 2A, and Supplementary file 2). Testing a subset of FL proteins further, we found that only this latter group can be self-seeded (Figure 2—figure supplement 3A). These proteins collectively span all four DFD subfamilies, and all of them function as major adaptors in innate immune signalosomes (Figure 2E). ASC, BCL10, MAVS, and TRADD drive inflammation and/or inflammatory cell death downstream of D/PAMP sensing (Martinon et al., 2002; Seth et al., 2005; Kawai et al., 2005; Micheau and Tschopp, 2003; Gross et al., 2006; Bertin et al., 2000; Bertin et al., 2001). FADD signals downstream of death ligands and certain D/PAMPs through various signalosomes that can be either anti- or pro-inflammatory (Henry and Martin, 2017; Tummers et al., 2020; Chen et al., 2005; Gurung et al., 2014; Mouasni and Tourneur, 2018; Balachandran et al., 2004). In contrast, signalosomes with primarily non-immunity functions – the apoptosome, the PIDDosome, and the ectodysplasin (EDA) receptor complex (Li et al., 1997; Tinel and Tschopp, 2004; Headon et al., 2001) – lacked supersaturation (Figure 2G). Consistent with their functioning through supersaturation, only the innate immune set of adaptors is more abundantly expressed than their cognate receptors in the human body (Figure 2F). These data collectively identify the adaptors, specifically, of inflammatory signalosomes as energy reservoirs for signal amplification.

To explore the functionality of supersaturation, we focused on two modes of programmed cell death signaling that differed in this regard: intrinsic apoptosis and pyroptosis. The former occurs downstream of persistent intracellular stresses (Gong et al., 2019; Nano et al., 2023; Häcker and Haimovici, 2023). The latter is instead triggered by minute levels of D/PAMPs and therefore involves greater signal amplification. Our DAmFRET analyses revealed that CARDs of the apoptosome lack nucleation barriers, whereas CARDs of the inflammasome, which drives pyroptosis, have nucleation barriers.

To test if the absence of nucleation barriers limits the sensitivity of the apoptosome, we adapted an optogenetic approach (Rodriguez Gama et al., 2022; Park et al., 2017; Shkarina et al., 2022; Kennedy et al., 2010) to precisely control the initiation of intrinsic apoptosis. HEK293T cells lack the inflammasome constituents NLRC4, ASC, and CASP1, allowing us to repurpose their DFDs in this cell line. Accordingly, we transduced the cells with mScarlet-I fusions of either the non-supersaturable WT apoptosome effector, CASP9, or a chimeric version that harbored the supersaturable DFD of CASP1 in place of its own (CASP9^CASP1CARD). We simultaneously transduced the cells with blue light-inducible seeds of the cognate upstream DFDs – opto-APAF1 or opto-NLRC4, respectively (Figure 3A). We then assessed signal propagation through the reconstituted pathways by measuring activation of CASP9’s downstream target, CASP3/7, after 1 min of blue light stimulation. Both cell lines activated a fluorescent CASP3/7 reporter to the same extent (Figure 3B), confirming that the different modifications do not sterically interfere with the proteins’ activities.

To now assess signal amplification, we evaluated both the persistence of DFD assemblies and their ability to commit cells to apoptosis following an otherwise sublethal stimulus. We anticipated that cells expressing non-supersaturable CASP9 will (1) form transient clusters and consequently (2) survive a short blue light stimulus. In contrast, cells expressing supersaturable CASP9 will (1) form perdurant clusters that signal indefinitely, and consequently (2) die even after a short blue light stimulus.

To test prediction 1, we stimulated the cells with blue light for 1 min and monitored subsequent protein localization. We found that both pairs of proteins formed puncta in essentially all cells within one minute (Figure 3C–D; Figure 3—video 1). The puncta with non-supersaturable CASP9 then dissolved over the course of the next 10 min. In contrast, the puncta with supersaturable CASP9 instead continued to grow at a constant rate for at least the next 20 min (Figure 3C–D; Figure 3—video 2), confirming that the protein’s drive to polymerize exists even without the stimulus.

To test prediction 2, we measured cell death 2 hr after the initiation of blue light stimulation for either 1 min or the entire 2 hr. Neither pair of proteins induced cell death in the absence of blue light (Figure 3E). Non-supersaturable CASP9 induced death in approximately 18% and 55% of cells following the short and long stimulation, respectively (Figure 3E), confirming the expected dose-dependence of signaling. In contrast, supersaturable CASP9-induced death in most cells even upon short stimulation is consistent with the expected amplification of signaling due to supersaturation-mediated DFD polymerization. We note that the increased stability of NLRC4^CARD and CASP1^CARD polymers relative to APAF1^CARD and CASP9^CARD oligomers could increase signal amplification irrespective of polymerization; specifically, due to the slower dissolution of assemblies that formed even during stimulation which could provide CASP9^CASP1CARD more time to activate more molecules of CASP3/7. Consequently, further work will be required to distinguish the contributions of each mechanism to amplification.

We next asked if pyroptosis and extrinsic apoptosis are indeed powered by adaptor supersaturation in vivo. We first induced pyroptosis in human THP-1 monocytes by treating them with poly(dA:dT), a ligand for the inflammasome receptor, AIM2. By 18 hr, approximately 70% of the cells were dead or dying. We then deleted PYCARD to determine if cell death depended on ASC, the adaptor for the inflammasome. Death was delayed, but not eliminated (Figure 4A). We then also deleted FADD, encoding an adaptor that has been shown to co-assemble with AIM2 in pathogen-triggered PANoptosis (Lee et al., 2021). Death was reduced further still (Figure 4A and Figure 3—figure supplement 1B), confirming that AIM2 activation can induce cell death through both ASC and FADD.

To eliminate the potentially confounding effects of orthogonal dsDNA sensors (Maelfait et al., 2020), we next placed AIM2^PYD under blue light control and confirmed that the resulting fusion protein – ‘opto-AIM2’ – grants direct control over seed formation (Figure 4B and C, top). Using this system, a transient blue light exposure induced AIM2 clustering within seconds and cell death within 10 min (Figure 4C–D and Figure 4—videos 1; 2). Next, we evaluated if cell death depended on ASC or FADD. We found that stimulating WT cells with a 1-s blue laser pulse every 15 min killed essentially all of them within 1 hr (Figure 4E). In contrast, only half of cells lacking ASC (PYCARD-KO) died, and they did so with delayed kinetics consistent with FADD-driven signaling (Tummers et al., 2020; Place et al., 2021; Roncaioli et al., 2023).

Finally, we directly assessed inflammasome nucleation by AIM2^PYD seeds. To do so, we reconstituted the PYCARD-KO with mScarlet-I-tagged ASC and titrated its expression to well below endogenous levels so as to circumvent any potential for the fusion tag to enhance the protein’s oligomerization (although this would be unexpected; Bindels et al., 2017; Figure 3—figure supplement 1C–E). We then tracked AIM2^PYD and ASC localization following a 10 laser pulse. Both proteins began to cluster almost immediately. AIM2^PYD cluster growth nearly stopped by 3 min, consistent with excited Cry2 relaxation on a similar time scale (Park et al., 2017; Che et al., 2015). In contrast, ASC clusters continued rapid growth for more than 10 min (Figure 4F), demonstrating that their drive to do so is independent of the triggering stimulus. The plasma membrane concomitantly permeabilized as a result of gasdermin D (GSDMD) activation (He et al., 2015), and to an extent that increased with the level of ASC expression (Figure 4G). These data collectively confirm that endogenous ASC is highly supersaturated prior to stimulation, and that the extent of supersaturation determines the extent of signal amplification.

Given that inflammatory signalosomes frequently initiate programmed cell death, and that a kinetic barrier governs their activity, the susceptibility of cells to aberrant cell death through spontaneous nucleating fluctuations is expected to increase with the level of adaptor supersaturation in vivo. This would constrain a cell’s life expectancy if such nucleation occurs with appreciable frequency. To explore this possibility, we asked how adaptor supersaturation (as approximated by the ratio of mRNA levels to C_sat values) relates to the turnover rates of each cell type in the human body (Sender and Milo, 2021). We found that short-lived cells such as monocytes indeed have greater DFD supersaturation than long-lived cells such as neurons (Figure 4H). Among individual DFD proteins, cell turnover correlated especially strongly with the expression level of ASC (Figure 3—figure supplement 1F–G), suggesting that life expectancy may be limited by the thermodynamic drive for inflammatory signal amplification.

DFDs share the same fold and are co-expressed in many of the same cells. While spontaneous nucleation of supersaturable DFDs poses an inherent risk of inadvertently activating end fate cell signaling, this possibility is multiplied by the risk of cross-activation between DFDs in different pathways. To what extent do they nucleate each other? To determine specificity and systematically map the nucleating interactome of DFDs, we mated our library of seed-expressing yeast strains with a library of strains expressing each mEos-fused DFD (Figure 5—figure supplement 1A) to create over 10,000 arrayed diploid yeast strains that we then screened by DAmFRET (Figure 5—figure supplement 1B and C).

In total, we identified 171 nucleating interactions, representing just ~1.6% of the total library (Figure 5—figure supplement 1D–F and Supplementary file 4). The interactions were largely constrained not just to members of the same subfamily, but to members of the same signaling subnetwork even within subfamilies (Figure 5A–B). For example, CARDs of the CBM signalosome nucleated each other but not CARDs of the inflammasome, and vice versa. As an exception, PYDs of the inflammasome and DEDs of the Death-Inducing Signaling Complex (DISC) nucleated each other (Figure 5C). These interactions are consistent with the residual cell death we observed in the PYCARD FADD double knockout, as well as previously observed crosstalk between these pathways and the close phylogenetic relationship between PYDs and DEDs (Roncaioli et al., 2023; Bedoui et al., 2020; Gullett et al., 2022; Park et al., 2007b).

On the whole, the observed network of nucleating interactions reveals that DFDs from different pathways generally function independently of each other, allowing adaptors to serve as orthogonal energy reservoirs for their respective signalosomes (as schematized in Figure 2E).

Animals acquired DFDs horizontally from bacteria (Aravind et al., 2024; Dalrymple and Jenkins, 1951; Kibby et al., 2023), giving rise to an ancestral DISC (conserved across metazoa) and later inflammasome (conserved across vertebrates; Figure 6A; Yuen et al., 2014; Sakamaki et al., 2014). To investigate conservation of nucleation barriers, we characterized by DAmFRET the phase behaviors of DFD-containing constituents of a basal inflammasome – NLRP3, ASC, and CASP1 from zebrafish (Danio rerio); and a basal DISC – A0A1X7U321, FADD, and CASP8 from the model sponge, Amphimedon queenslandica. As for their human counterparts, the fish and sponge adaptors – but not the receptors and effectors – were supersaturable (Figure 6B and Figure 6—figure supplement 1A–B). These results suggest that the function of DFDs as energy reservoirs preceded the radiation of animals.

Multiple DFDs were found to populate stable ordered polymers in all cells. Although we did not observe a nucleation barrier for these, it is possible that they are supersaturable in vivo at much lower than the micromolar concentrations surveyed by DAmFRET. Similarly, a small number of DFDs that populated only a low AmFRET state did not express to high concentrations in yeast, and it is possible that they can self-assemble under cellular contexts that allow them to reach higher concentrations.

Our study necessarily simplifies and abstracts human innate immune signaling. We considered the cytoplasm of living yeast cells to be a suitable proxy for the physiological context of DFDs as a whole, while acknowledging that the endogenous contexts of DFDs differ from that in our experiments. For example, several of the DFDs normally localize to membranes or the nucleus in human cells, and this could impact their phase behavior. Nevertheless, we find no evidence that such localizations would reduce supersaturation, as none of the DFDs from nuclear proteins exhibited nucleation barriers in our experiments. Likewise, in the case of MAVS, which normally localizes to the mitochondrial membrane, we found that the FL protein (containing its mitochondrial localization signal) was more supersaturable than its DFD alone, suggesting that localizing the protein to the mitochondrial surface increases the nucleation barrier despite restricting its diffusional entropy and increasing its local concentration. Our work also does not explore the many ways that DFD assembly can be regulated transcriptionally and post-translationally. For example, NLRP3 is upregulated and becomes phosphorylated in response to LPS ‘priming’ of the inflammasome (Song et al., 2017), and polyubiquitination of RIG-I promotes its activation of MAVS (Gack et al., 2007). Such modifications imply that some signal amplification can occur upstream of supersaturated adaptors. The level of adaptor supersaturation differs between cell types (as we have shown) and is also likely to be dynamically regulated.

Hygromycin B (Invivogen, ant-hg-1), Penicillin-Streptomycin (Thermo Fisher, 1514014gp), PMA (BioVision, 1544–5), Puromycin (Invivogen, ant-pr-1), Sytox Orange (Thermo Fisher, S11368), CellTox (Promega, G8741), Annexin V Alexa568 (1:200, Thermo Fisher, A13202), Annexin V Alexa488 (1:200, Thermo Fisher, A13201), Incucyte Caspase-3/7 Dye (Sartorius, 4440). Antibodies, anti-ASC (Santa Cruz Biotechnology, sc-514414), anti-FADD (Sigma, 05–486), anti-Actin (Santa Cruz Biotechnology, sc-8432) were obtained from the indicated vendors.

Twelve human proteins contain two DFDs, typically one closely following the other. To determine if the DFDs in such pairs should be evaluated independently or together, we used the predicted aligned error (PAE) matrix generated by AlphaFold3 (Varadi et al., 2022). PAE is the expected positional error at residue x if the predicted and actual structures are aligned on residue y. Seven of the DFD pairs exhibited very low interdomain PAE scores comparable to those of the component DFDs (Figure 1—figure supplement 1B bottom), suggesting a conserved fixed geometric relationship between the domains. We therefore considered these tandem DFDs as single members of their respective subfamilies. Similarly, we excluded the annotated (Wu et al., 2025). PYD-like domain of CENP-N because the PAE matrix and experimental structures show that it is in fact part of a larger non-DFD.

Yeast expression plasmids were made as previously described (Khan et al., 2018). Briefly, we used a high copy episomal vector, V08, which contains inverted BsaI sites to support Golden Gate cloning, followed by a rigid helical linker 4 x(EAAAR) and mEos3.1 (‘mEos’). This vector drives the expression of proteins from a GAL1 promoter and contains the auxotrophic marker URA3. The vector V12 is identical to V08 except that mEos and linker precede rather than follow the BsaI sites, for expressing proteins with an N-terminal fusion. Inserts were ordered as yeast codon-optimized GeneArt Strings (Thermo Fisher) flanked by Type IIs restriction sites for ligation between BsaI sites in V08 and V12. Fusions were made opposite the native N- or C-terminus of each DFD to minimize non-native steric effects. All other inserts were cloned into respective vectors via Gibson assembly between the promoter and respective tag. All plasmids were verified by Sanger sequencing. All expression plasmids are listed in Supplementary file 1.

Lentivirus vectors were as previously described (Rodriguez Gama et al., 2022). Briefly, optogenetic constructs were cloned into pLV-EF1a-IRES-Hygro (Addgene #85134), which encodes a hygromycin B resistance cassette. To create lentiviral vectors expressing the optogenetic constructs fused with miRFP670nano (Oliinyk et al., 2019) and Cry2, the corresponding sequences of AIM2^PYD, APAF1^CARD, and NLRC4^CARD were inserted via Gibson assembly into pLV-EF1a-IRES-Hygro. Finally, the doxycycline-controlled lentiviral vectors were cloned via Gibson assembly with the respective coding sequences from PYCARD, CASP9, CASP1, and mScarlet-I into pCW57.1 (Addgene #41393). All lentivirus vectors are listed in Supplementary file 3.

Unseeded DAmFRET experiments were conducted using S. cerevisiae strains rhy1713, rhy2977, and rhy3078a (Khan et al., 2018; Kandola et al., 2023; Miller et al., 2023). To create strains expressing DFD seeds, we first transformed AseI digests of each DFD plasmid along with a plasmid expressing Cas9 and a guide RNA targeting the URA3 markers into rhy2153 (Rodriguez Gama et al., 2022). This strain contains a genomic landing pad consisting of natMX followed by the tetO7 promoter and counterselectable URA3 ORFs derived from C. albicans and K. lactis, and stop-µNS-mCardinal as described (Rodriguez Gama et al., 2022). Successful integration of the insert replaces the URA3 marker with the gene of interest and fuses to the protein’s C-terminus μNS-mCardinal, under the control of a doxycycline-repressible promoter. Transformants were selected for resistance to 5-FOA and validated for successful seed integration by detection of mCardinal expression using flow cytometry. The arrayed library of resulting strains was then mated to each of the rhy1713 strains expressing separate DFD-mEos fusions, by pinning each pair of strains together onto agar omnitrays containing SD-URA+NAT + dox media. The resulting colonies were then pinned into liquid SD-URA+NAT + dox for continued diploid selection and creation of glycerol stocks. The entire nucleating interaction screening consisted of 384 96-well plates.

We performed DAmFRET as previously described (Khan et al., 2018). Briefly, single transformant yeast colonies were inoculated in 200  μL of SD-URA in a 96-well microplate well and incubated in a Heidolph Titramax platform shaker at 30 °C, 1350 RPM overnight. Cells were washed with sterile water, resuspended in galactose-containing media, and allowed to continue incubating for approximately 20  hr. Microplates were then illuminated for 25 min with 320–500  nm violet light to photoconvert a fraction of mEos molecules from a green (516 nm) form to a red form (581 nm). At this point, cells were either used to collect microscopy data or continue the DAmFRET protocol.

For the nucleating interaction screen, glycerol stock plates were pinned into liquid SD-URA without dox and incubated for 16 hr at 30 °C with 1350 RPM shaking overnight. We then resuspended cells in fresh SD-URA media and continued incubation for an additional 20 hr. After this, we resuspended cells in SGal-URA and continued incubation for 20 hr to induce protein expression. Finally, we resuspended cells in fresh SGal-URA for 4 hr prior to DAmFRET data collection. The library was then consolidated into 96 384-well plates.

DAmFRET data were collected on a ZE5 cell analyzer cytometer. Autofluorescence was detected with 405 nm excitation and 460/22 nm emission; side scatter (SSC) and forward scatter (FSC) were detected with 488 nm excitation and 488/10 nm emission. Donor and FRET fluorescence were detected with 488 nm excitation and 425/35 nm or 593/52 nm emission, respectively. Acceptor fluorescence was detected with 561 nm excitation and 589/15 nm emission. For each well, we collected a volume of 13 μL, resulting in approximately 500,000 events per sample. Data compensation was done in the built-in tool for compensation (Everest software V1.1) on single-color controls: non-photoconverted mEos and dsRed2 (as a proxy for the red form of mEos). For nucleating interactions, we included an additional channel for mCardinal intensity with 561 nm excitation and 670/30 nm emission.

Data were processed on FCS Express Plus 6.04.0015 software (De Novo). Events were gated for single unbudded cells by FSC vs. SSC, followed by gating of live cells with low autofluorescence and positive donor and acceptor fluorescence. With the exception of TNFRSF10A^DD (TRAIL-R1) (rhx2933), which failed to express with either its C- or N-terminus tagged, all expression plasmids were processed. Plots represent the distribution of AmFRET (FRET intensity/acceptor intensity) vs. acceptor intensity (protein expression).

We then analyzed the data as previously described (Kandola et al., 2023). Briefly, FCS files were gated using an automated R script running in flowCore. Before gating, the forward scatter (FS00.A, FS00.W, FS00.H), side scatter (SS02.A), donor fluorescence (FL03.A), and autofluorescence (FL17.A) channels were transformed using a logicle transform in R. Single cells were gated using FS00.A vs SS02.A and FS00.H vs FS00.W. These were gated for expressing cells using FL03.A vs FL17.A. Cells falling within these gates were then exported as FCS3.0 files for further analysis.

DAmFRET histograms were divided into 64 logarithmically spaced bins across a predetermined range large enough to accommodate all data sets. The upper gate values were determined for each bin as the 99th percentile of the DAmFRET distribution in that bin. We used the expression of mEos alone to delineate the region of a DAmFRET plot that corresponds to no assembly. For all samples, cells falling above this region are considered to contain protein assemblies (FRET-positive). The fraction of cells in the assembled population was plotted as a ratio to the total cells in the bin for all 64 bins. The gross fraction of such cells expressing a given protein is reported as fgate.

We initially attempted to classify each DAmFRET dataset as one-state or two-state, and discontinuous or continuous for the latter, using an algorithm previously developed for this purpose (Posey et al., 2021). However, the algorithm invariably misclassified discontinuous datasets as ‘continuous’ when the AmFRET level of the high FRET state changed with concentration, that is exhibited positive or negative slope as for FAS^DD and CASP2^CARD, respectively. Due to this limitation, we adopted a different method.

To determine the continuity of an adequately expressed plasmid, we analyzed the distribution of AmFRET values about the transition region. To do so, we fit a spline to the median AmFRET values across binned concentrations that met a density cutoff defined by a minimum of 20 cells, a minimum cell density of 500, and no more than 25% of cells being defined as an outlier. The bin density is measured by the number of cells divided by the interquartile range (IQR) of AmFRET values within that bin. The number of bins was determined using Scott’s rule for number of bins in a histogram. The median AmFRET values were calculated for each bin and then denoised using the Python SciPy.signal.sosfiltfilt, resulting in the spline fit. This was bootstrapped 100 times with the average of all bootstraps reported as the final spline fit. The resulting spline was used to determine the transition region. The transition point is defined as the concentration with the greatest change in AmFRET. Therefore, this value was calculated as the maximum of the first derivative of the fitted spline, using the numpy.diff package. The transition range is defined as the region between the maximum positive and negative rate of change, indicating where the transition between FRET states is starting and ending, respectively. The transition starting point is calculated as the maximum of the second derivative that lies before the transition point, and the transition endpoint is calculated as the minimum of the second derivative that lies after the transition point. This is done on all 100 bootstrapped splines, with the median of each measure reported. Hartigan & Hartigan’s dip test for unimodality was used to determine the continuity of AmFRET values within the determined transition range. Plots are classified as ‘discontinuous’ or ‘continuous’ for p-values less than or greater than 0.05, respectively.

Plasmids that were determined to have a continuous transition were further classified as uniformly low (‘low’), uniformly high (‘high’), or transitioned from low to high with increasing concentration (‘low to high’). This was done by extracting the minimum and ending AmFRET values from the spline fit. As a reference for the AmFRET value that indicates the start of a high FRET state, the AmFRET value of the transition start point of a control plasmid, rhx0927, was used. For each plasmid, if its minimum and ending AmFRET values fell below the reference, it was classified as low. If both fell above, it was classified as high. Those with a minimum value below and an ending value above were classified as ‘low to high’. The majority classification of replicates for each plasmid is reported.

To determine betweenness and degree of centrality, we extracted interactions involving DFD-containing proteins (listed in Supplementary file 1) from STRING version 12.0 (Szklarczyk et al., 2023), considering only physical interactions with scores of 900 or higher. Using NetworkX v3.1, we analyzed these interactions as an undirected graph to calculate betweenness and degree of centrality. For proteins with multiple DFDs, we classified a protein as discontinuous if any of its DFDs were identified as discontinuous.

We first excluded files with fewer than 2500 events positive for mEos or mean acceptor intensities less than 3.5 p.d.u. From this, any DFD or seed left with less than 25% of their original instances after filtering was removed from the analysis completely. Next, we identified nucleating interactions as DFD pairs that decreased the C50 and increased the fraction assembled (fgate). We standardized all variables for each experimental batch of DFDs to a mean of 0 and variance of 1. We then determined the outlier degree for C50 and fgate based on the number of interquartile ranges below or above the median for these values. This was done directionally on a per-DFD basis. We defined the ‘nucleating interactions’ for a given mEos-fused DFD as those whose mean of these two values (reported as ‘seedability’) is greater than or equal to 3 standard deviations above the mean of all seedability values. We confirmed that the seedability values for most DFD pairs partitioned with that of two negative controls included for each DFD.

To evaluate the reproducibility of our assay, we replicated it for a set of 36 DFDs. This replicate analysis mirrored the original, except it utilized the DFD distributions and cutoff values from the first set. The Pearson correlation (R) between the two sets was 0.91 (p<0.0001). To minimize the impact of random variations in negatives and outliers, we excluded double negative instances, resulting in a slightly altered Pearson correlation (R) of 0.90 (p<0.0001). Of the 3423 DFD +seed combinations reassessed, 16 showed inconsistent hit-calling, indicating an assay consistency rate of 99.53% with a 95% confidence interval ranging from 99.30% to 99.76%.

To generate an average DAmFRET curve, we computed the mean of each histogram bin in the DAmFRET dataset, focusing on bins containing a minimum of 100 cells. The average DAmFRET curve for each DFD in the presence of its self-seed was fit to a Weibull function as follows. We first calculated the average AmFRET value in each concentration bin. The resulting curves resemble the fraction-assembled curves except that the asymptote is the maximum AmFRET value rather than 1. Therefore, we used the following equation for fitting:(1)AmFRET(c)=Amp[1−exp(−ln(2)(cC50)a)]\begin{document}$$\displaystyle AmFRET\left (c\right)=Amp\left [1- exp\left (- ln\left (2\right)\left (\frac{c}{C50}\right)^{a}\right)\right ]$$\end{document}

Here Amp is the AmFRET asymptotic value of the curve, c is the concentration, C50_seeded is the concentration at which the curve, in the presence of its self-seed, has reached 50% of its asymptotic value, and a\begin{document}$a$\end{document} describes the steepness of the stretched exponential. Initial values of the parameters were chosen based on Gaussian smoothed versions of the curves and constrained in the fit to at minimum a twofold change from those initial guesses. The a\begin{document}$a$\end{document} parameter was constrained between 0.1 and 10 based on expected reasonable values of that parameter. At low and high concentration values, the average AmFRET values are clearly unstable and influenced by noise and minor compensation errors. Therefore, we chose the beginning and ending points of each curve by visual inspection, choosing starting points where the curve begins to increase and ending points where the curve levels off. Error values were determined from Monte Carlo simulations as in the fitting of fraction assembly.

To approximate the supersaturability of DFDs, we used the ratio of the average C50 in the presence of the respective DFD’s seed (C50_seeded) to the average C50 in the presence of a null-seed (C50_stochastic). Ratios were only calculated between C50s obtained from the same batch run.

HEK293T cells and THP-1 cells were purchased from ATCC. THP-1 PYCARD-KO (thp-koascz) cells were purchased from InvivoGen. HEK293T cells were grown in Dulbecco’s Modified Eagle’s Medium (DMEM) with L-glutamine, 10% fetal bovine serum (FBS), and PenStrep 100 U/mL. THP-1 cells were grown in Roswell Park Memorial Institute (RPMI) medium 1640 with L-glutamine and 10% FBS. All cells were grown at 37 °C in a 5% CO₂ atmosphere incubator. Cell lines were regularly tested for mycoplasma using the Universal mycoplasma detection kit (ATCC, #30–1012 K).

Stable cell lines were created as described (Rodriguez Gama et al., 2022). Briefly, constructs were packaged into lentivirus in a 10 cm plate 60% confluent of HEK293T cells using the TransIT-LT1 (Mirus Bio, MIR2300) transfection reagent and 7 μg of the vector, 7  μg psPAX2, and 1 μg pVSV-G. Lentivirus was harvested and incubated with 293T with polybrene or infected at 1000 × g for 1 hr for THP-1 cells. For transduction of pCW57.1 derived vectors, HEK293T and THP-1 cells were selected with Puromycin (1 μg/mL) for 7 days. After this time, cells were sorted for positive expression of mScarlet-I and expanded in continuing selection with puromycin. For transduction of plasmids encoding fusions to miRFP670nano-Cryclust, THP-1 and HEK293T cells were selected with hygromycin B (350 μg/mL and 150 μg/mL, respectively) for 7 days. Cells were sorted for positive expression of miRFP670nano and expanded for further experiments with continued selection. To generate THP-1 PYCARD-KO+FADD KO cells, sgRNA targeting FADD exon 1 was cloned into the lentiCRISPR v2-Blast (Addgene #83480). This vector was packaged into lentivirus as described above. THP-1 PYCARD-KO cells were transduced using spinfection and supplemented with polybrene. 24 hr after spinfection, media was replaced. 48 hr after spinfection, cells were selected with blasticidin (1 μg/mL). After 10 days of blasticidin selection, single-cell clonal expansion was done by serial dilution of resistant cells to achieve complete knockouts. Selected wells were analyzed by immunoblot to confirm the absence of FADD protein and sequence-verified.

High-content imaging was performed on the Opera Phenix high-content screening system (PerkinElmer) using a 63 x water immersion objective. Briefly, yeast (rhy2977) transformed with individual plasmids were cultured and induced as for DAmFRET assays. Then, 10 μL were transferred into a well containing 90 μL of SGal-URA of a 96-well optically clear flat-bottom plate (PerkinElmer 6055302). Data analysis of the high content imaging was performed in Fiji. Images of mEos were acquired using 488 nm excitation and a standard GFP filter set. Small z-stacks were acquired over 5 µm total range with 1 µm steps. The image containing the brightest mEos signal was used. The mEos signal was then background-subtracted with a rolling ball radius of 100 pixels, then found and converted to Fiji ROIs using the Fiji Default method of traditional image thresholding. The mean, standard deviation, and aspect ratio (AR) were measured for each object. The coefficient of variation (CV) in pixel intensity was calculated for every object following the formula: CV = Std Dev/Mean*100. The wells were divided into three categories based on their AR and CV. Objects that had a CV >55 and an AR >1.159 were designated ‘fibrillar’. Objects that had a CV >55 and an AR <1.16 were designated ‘punctate’. Finally, objects that had a CV <55 and an AR <1.16 were designated ‘diffuse’. These cutoffs were determined manually from a visual inspection of the data. Plasmids rhx4763 - rhx4767 were acquired in a second data set that used different laser powers and integration times. Hence, for these plasmids specifically, ‘fibrillar’ had a CV >17 and an AR >1.4, ‘punctate’ had a CV >17 and an AR <1.41, and ‘diffuse’ had a CV <18 and an AR <1.41. These cutoffs were determined manually from a visual inspection of the data. The results were then manually verified for all wells. Seven plasmids were inconsistently classified by these cutoffs. Of these, manual inspection confirmed ‘diffuse’ morphology for rhx1033 and rhx1133 and prompted reclassification of rhx2935, rhx2637, rhx1121, rhx1070, and rhx1055 from ‘fibrillar’ to ‘punctate’. Plasmid rhx2232 was incorrectly classified as punctate, but images display fibrillar morphology. Three others (rhx1113, rhx1097, rhx2937) had anomalously high AR due to low expression. Representative microscopy images are included for all plasmids in our repository.

The yeast and mammalian cells were imaged in an LSM 780 microscope with a 63 x Plan-Apochromat (NA = 1.40) objective. T-Sapphire was excited with a 405 nm laser. mEos and mScarlet-I were excited with a 488 nm and 561 nm laser, respectively. For time-lapse imaging, samples were maintained at 37 °C and 5% CO₂ with a stage top incubator. To stimulate Cry2clust, we used the 488 nm laser at a power setting of 50% for a pulse of 10 s, which is the amount of time it took to scan the user-generated region of interest unless indicated otherwise. 561 and 633 nm lasers were used for imaging mScarlet-I and miRFP670nano, respectively. Pyroptosis events were tracked by incorporating the Sytox Orange reagent into the cell. To quantify the CV, images were subjected to an in-house Fiji adapted implementation of Cellpose for cellular segmentation (Stringer et al., 2021). The Cellpose-generated regions of interest (ROIs) were used to measure specified imaging channels.

For quantification of cell death events using IncuCyte (Sartorius), THP-1 cells were plated on a 24- or 96-well plate at a density of 4x10⁸/well or 1x10⁸/well, respectively, with PMA (10 ng/mL) for 16 hr. Media was replaced with fresh media supplemented with Annexin V-Alexa488 and Sytox Orange (1:1000). For AIM2-Cry2clust activation, an initial collection of unexposed measurements was taken for 30 min. Then, the plate was exposed to a 488 nm laser every 5 min. For treatments with poly(dA:dT), cells were treated and immediately subjected to imaging every 30 min for 19 hr. Positive cells for either fluorophore were identified using the integrated software in the IncuCyte instrument.

For optogenetic activation of APAF1^CARD and NCLR4^CARD, HEK293T cells expressing lentivirus constructs were seeded on a 35 mm dish (ibidi) at a density of 4x10⁴/mL with 2 mL of media. The next day, dox was added at a concentration of 1 µg/mL to induce the expression of mScarlet-I tagged proteins. 24 hr after protein induction, media was replaced with fresh media supplemented with Incucyte Caspase-3/7 Dye (1:1000) 2 hr prior to the experiment or Annexin V-Alexa488. Cells were imaged using a spinning-disk confocal microscope (Nikon, CSU-W1) with a ×60 Plan Apochromat objective (NA = 1.40) and a Flash 4 sCMOS camera (Hamamatsu). A region of interest (ROI) was selected to induce optogenetic activation for indicated times using a 488 nm laser at 50% laser power for the indicated time ranging from a fraction of a second to 30 s. ROIs were generated by the Cellpose segmentation algorithm around each cell contour. These ROIs were then used to measure the area, mean, standard deviation, and integrated density of each cell on the 488 nm and 560 nm fluorescence channels.

We performed capillary-based protein immunodetection (Wes, ProteinSimple) as described (Rodriguez Gama et al., 2022). Briefly, protein lysates were prepared as per recommended manufacturer instructions to a final concentration of 1 μg/mL. An assay plate was filled with samples, blocking reagent, primary antibodies (1:50 dilution for anti-Actin, 1:200 dilution for anti-PYCARD), HRP-conjugated secondary antibodies, and chemiluminescent substrate. The plate was subjected to electrophoretic protein separation and immunodetection in the fully automated capillary system. The resulting data was processed using the open-source software Compass (https://www.proteinsimple.com/compass/downloads/) to extract the intensities for the peaks corresponding to the expected molecular weight of proteins of interest. For western blot, cells were centrifuged at 1000 x g for 5 min and resuspended in lysis buffer 50 mM Tris (pH 7.4), 137 mM NaCl, 1 mM EDTA, 1% Triton X-100, 10 mM DTT (dithiothreitol), cOmplete Protease Inhibitor (1 tablet / 10 mL) (Roche, 11697498001). Protein lysates were resolved on a NuPAGE 4 to 12%, Bis-Tris gel and transferred onto a PVDF membrane (IPVH00010, Millipore) using the Pierce Power Blotter (ThermoFisher). The membrane was blocked with 5% skim milk and incubated overnight with the antibodies: anti-Actin (1:1000, sc-8432), anti-FADD (1:500, 05–486) and anti-PYCARD (1:1000, sc-514414). Primary antibody was removed by several washes with TBS +0.1% Tween-20 and then subjected to incubation with secondary antibody (anti-mouse-HRP, 7076 S, Cell Signaling Technology). The detection of protein bands was then carried out using enhanced chemiluminescence (ECL; SuperSignal West Pico Chemiluminescent Substrate, Thermo Fisher, 34577). The chemiluminescent signal was acquired by placing the membrane in a film cassette and exposing it to X-ray film (Kodak) at varying durations in a darkroom. After exposure, films were developed using an automatic film processor.

SDD-AGE was performed as previously described (Khan et al., 2018). Briefly, cells were lysed using a 2010 Geno/Grinder with bead-beating. Samples were prepared with 2% sarkosyl and separated in a 1.5% agarose gel with 0.1% SDS. The distribution of mEos-fused proteins was analyzed directly in the gel with a GE Typhoon Imaging System. Images were processed to remove background using a 250-pixel rolling ball, cropped, and contrast-adjusted.

Two-sided Student’s t-tests were used for significance testing unless stated otherwise for two-sample comparisons. The graphs represent the means  ± SEM of independent biological experiments unless stated otherwise. Statistical analysis was performed using GraphPad Prism 9, Python, and R packages.