The wtf4 meiotic driver used in this work is from S. kambucha, an isolate that is almost identical (99.5% DNA sequence identity) to the commonly studied lab isolate of S. pombe (Rhind et al., 2011; Singh and Klar, 2002). Our previous work demonstrated that the Wtf4^antidote localizes to a region within the spores that inherit the wtf4 gene. The Wtf4^poison protein, however, is found in all four spores and throughout the sac (ascus) that holds them (Nuckolls et al., 2017). Here, we explored the localization of these proteins in greater depth to gain insight into their mechanisms.

We used fluorescently tagged alleles of wtf4 to visualize the proteins. The two Wtf4 proteins have different translational start sites and thus different N-termini (Figure 1A, Figure 1—figure supplement 1A). We took advantage of this feature to visualize the proteins separately. For the Wtf4^antidote, we used an allele with an mCherry tag immediately upstream of the first start codon. This mCherry-wtf4 allele tags only the Wtf4^antidote (mCherry-Wtf4^antidote) but still encodes an untagged Wtf4^poison. We previously demonstrated that this allele is fully functional (Nuckolls et al., 2017). To visualize Wtf4^poison, we used the wtf4^poison-GFP allele. This separation-of-function allele encodes only a C-terminally tagged poison but no Wtf4^antidote protein. We previously demonstrated that this tagged allele is functional but has a slightly weaker phenotype than an untagged wtf4^poison separation-of-function allele (Nuckolls et al., 2017).

We integrated the tagged alleles at the ade6 locus in separate haploid S. pombe strains. We then crossed those two haploid strains to create heterozygous mCherry-wtf4/wtf4^poison-GFP diploids and induced these diploids to undergo meiosis. We imaged the asci using both standard and time-lapse fluorescence microscopy (Figure 1C, Figure 1—figure supplement 1B). We confirmed our previous observations that the mCherry-Wtf4^antidote was enriched in two spores, whereas Wtf4^poison-GFP was found throughout the ascus and often formed puncta of various sizes. In the spores that did not inherit the antidote, Wtf4^poison-GFP also appeared dispersed throughout the spores. In the spores that inherited and thus expressed mCherry-wtf4, however, the localization of Wtf4^poison-GFP was more restricted. Specifically, we observed that the Wtf4^poison-GFP largely colocalized with mCherry-Wtf4^antidote in a limited region of the spore (Figure 1C, Figure 1—figure supplement 1B). In time-lapse microscopy, it was evident that the two Wtf4 proteins consistently colocalized in a defined region of the spore, even as this region changed shape over time. This co-diffusion suggests the two proteins are either physically interacting or are present in the same compartment (Figure 1—figure supplement 1B). It also appeared that the level of Wtf4^poison-GFP protein is reduced in spores containing the antidote. We did not distinguish if this was due to technical reasons (i.e. quenching of the GFP molecules) or biological reasons such as degradation of Wtf4^poison-GFP in spores with mCherry-Wtf4^antidote and/or due to a higher expression of Wtf4^poison-GFP in the spores that inherit it (non-antidote spores) (Figure 1C). We completed Pearson correlation analysis (Adler and Parmryd, 2010) of mCherry-Wtf4^antidote and Wtf4^poison-GFP in the spores (where a result of >0 is positive correlation; 0, no correlation; <0 anti correlation) and obtained a coefficient of 0.61, indicating strong colocalization between the two Wtf4 proteins (Figure 1—figure supplement 1C).

The limited distribution of the Wtf4 poison and antidote proteins within wtf4⁺ spores suggested they may be confined to a specific cellular compartment. To test this idea, we looked for colocalization of Wtf4 proteins with the vacuole, endoplasmic reticulum (ER) and nucleus (see below). For these experiments, we used the fully functional wtf4-GFP allele, which tags both the poison and antidote proteins (Nuckolls et al., 2017).

To assay the localization of the Wtf4 proteins relative to the vacuole, we imaged asci produced by diploids that were heterozygous for both wtf4-GFP and cpy1-mCherry. Cpy1-mCherry localizes to the lumen of the vacuole in vegetative cells (Sun et al., 2013) but has not, to our knowledge, been imaged in spores. We could observe mCherry in two of the four spores – presumably the two that inherited the cpy1-mCherry allele (Figure 1D). This 2:2 spore localization pattern has been previously observed in budding yeast for vacuolar proteins and several other organelles (Neiman, 2011; Roeder and Shaw, 1996; Suda et al., 2007). We found that the Wtf4-GFP and Cpy1-mCherry proteins colocalized within the spores that inherited both tagged alleles, suggesting the Wtf4 proteins are found within the vacuole (Pearson coefficient of 0.89, Figure 1D, Figure 1—figure supplement 2A–B). Interestingly, we also saw colocalization of Wtf4-GFP proteins with an ER marker, pbip1-mCherry-AHDL (Zhang et al., 2012; Figure 1—figure supplement 2C–D). We speculate this colocalization with the ER is due to nitrogen starvation which is required to induce meiosis and promotes organelle autophagy in S. pombe (Kohda et al., 2007; Zhao et al., 2016).

Because we could not distinguish the vacuole and ER within spores, we assayed the localization of the Wtf4 proteins in the absence of nitrogen starvation. To do this, we fluorescently tagged the coding sequence of wtf4^poison (wtf4^poison-GFP) and wtf4^antidote (wtf4^antidote-mCherry) separation-of-function alleles under the control of β-estradiol-inducible promoters (Ohira et al., 2017). We then integrated the wtf4^poison-GFP allele at the ura4 locus and the wtf4^antidote-mCherry allele at the lys4 locus of the same haploid strain. Next, we observed the localization of the Wtf proteins relative to the vacuole (visualized using the CellTracker Blue CMAC lumen stain) or the ER (using Sec63-YFP) following β-estradiol induction. Similar to our observations in spores, we saw that the Wtf4^poison-GFP and Wtf4^antidote-mCherry proteins largely colocalized, with a Pearson coefficient of 0.68 (Figure 1—figure supplement 3D and E). We also found that the Wtf4 proteins colocalized with the CMAC stain (Figure 1E), which suggests that the Wtf4 poison and antidote proteins are largely within the vacuole. However, there were Wtf4^poison-GFP puncta that lined the periphery of the cell and a circle in the middle of the cell, reminiscent of ER localization. These puncta were devoid of Wtf4^antidote-mCherry (Figure 1—figure supplement 3D).

We also attempted to assay the localization of the Wtf4 antidote and poison proteins individually to test if the localization of the Wtf4^poison was altered in the presence of the Wtf4^antidote, as we observed in spores (Figure 1C). We found that the localization of the Wtf4^antidote-mCherry to the vacuole was similar in the absence of the Wtf4^poison (Figure 1F, Figure 1—figure supplement 3B), with a Pearson coefficient of 0.69 (Figure 1—figure supplement 3C). This is analogous to previous observations of the localization of the slightly different Wtf4^antidote protein (82.2% amino acid identity) found in the S. pombe lab strain (Matsuyama et al., 2006). We failed, however, to generate cells carrying the wtf4^poison-GFP allele without the wtf4^antidote-mCherry allele by transformation, or by crossing the strain carrying both wtf4^poison-GFP and wtf4^antidote-mCherry to a wild-type strain (Figure 1—figure supplement 3A). This is likely due to leaky expression of the wtf4^poison-GFP from the inducible promoter even without addition of β-estradiol. Overall, our results suggest that the Wtf4^poison protein is toxic in vegetative cells, but the antidote is still capable of neutralizing the poison, as we could obtain cells carrying both the Wtf4 poison and antidote proteins.

In the process of trying to understand the localization patterns of Wtf4 proteins, we assayed the localization of the Wtf4 proteins relative to the nucleus. For this experiment, we imaged asci produced by wtf4-GFP/ade6⁺ heterozygotes also carrying a tagged histone allele, hht1-RFP (Tomita and Cooper, 2007). Although we did not observe colocalization of Wtf4 proteins and the nucleus, we frequently (24/38 asci) observed that the nuclei in the wtf4^- spores appeared more condensed (Figure 1G, younger ascus). Additionally, in 11 out of 38 asci, one or both of the nuclei in the wtf4^- spores were disrupted and the nuclear contents were dispersed throughout the spores (Figure 1G, older ascus). To address the timing of these nuclear phenotypes, we imaged diploids undergoing gametogenesis using time-lapse microscopy. We saw that all four nuclei tended to look similar shortly after the second meiotic division. As spores matured, however, we observed nuclear condensation sometimes followed by fragmentation in the spores that did not inherit wtf4 (i.e. in spores lacking the enriched GFP expression and antidote function) (Figure 1—figure supplement 4A and B, see Materials and methods). This nuclear condensation and fragmentation are reminiscent of apoptotic cell death (Carmona-Gutierrez et al., 2010; Kerr et al., 1972).

Our experiments in S. pombe suggest that the Wtf4 proteins can act when expressed outside of gametogenesis. However, our inability to induce expression of the Wtf4^poison in the absence of the Wtf4^antidote limited our ability to explore their mechanisms of action in this system. We, therefore, tested if the Wtf4 proteins functioned in the budding yeast Saccharomyces cerevisiae. To do this, we cloned the coding sequences of wtf4^poison-GFP and wtf4^antidote-mCherry under the control of β-estradiol inducible promoters on separate plasmids (Ottoz et al., 2014). We then introduced these plasmids into S. cerevisiae individually and together. We found that cells carrying the wtf4^poison-GFP plasmid were largely inviable when wtf4^poison-GFP expression was induced, indicating the poison is also toxic to S. cerevisiae (Figure 2A). However, cells expressing Wtf4^antidote-mCherry had only a slight growth defect relative to control cells carrying empty plasmids (Figure 2A). Importantly, expression of the Wtf4^antidote-mCherry plasmid largely ameliorated the toxicity of Wtf4^poison-GFP (Figure 2A). Given that S. pombe and S. cerevisiae diverged >350 million years ago (Hoffman et al., 2015), our results suggest that the target(s) of Wtf4^poison toxicity are conserved and the Wtf4^antidote does not require cofactors that are specific to S. pombe or gametogenesis to neutralize Wtf4^poison’s toxicity.

We assayed the localization of the Wtf4 proteins in S. cerevisiae using the inducible wtf4^poison-GFP and wtf4^antidote-mCherry alleles described above. Similar to our observations in S. pombe gametogenesis, we saw that Wtf4^poison-GFP localized as puncta of varying sizes throughout the cytoplasm (Figure 2B). We also observed some of the Wtf4^poison-GFP protein localized to the ER (Figure 2—figure supplement 1A; Friederichs et al., 2011). Analogous to our observation in S. pombe spores, we saw nuclear condensation in cells expressing Wtf4^poison-GFP relative to wild-type cells (Figure 2—figure supplement 1B–1D).

Wtf4^antidote-mCherry, on the other hand, generally localized to one or two large amorphous regions adjacent to the vacuole (Figure 2C). When co-expressed, Wtf4^poison-GFP and Wtf4^antidote-mCherry co-localized to this region next to the vacuole (Figure 2D). In some cells, a faint circle of Wtf4^poison-GFP could also be observed (likely ER localization); however, the majority colocalized with the antidote in the vacuole-associated region (Figure 2—figure supplement 1E). This localization was similar but not identical to our observations in S. pombe cells, where the Wtf4 proteins localize within, rather than adjacent to, the vacuole. To ensure the difference in localization (sequestration to a single puncta) and cell viability of the Wtf4^poison-GFP protein observed in the cells co-expressing Wtf4^antidote-mCherry was not due to the mCherry tag, we also confirmed these results with an untagged Wtf4^antidote (Figure 2—figure supplement 2).

Given that we failed to detect the Wtf proteins enter the vacuole, as we saw in S. pombe, we were interested in testing if the Wtf proteins were entering the vacuole but being rapidly degraded in S. cerevisiae. To test this, we prepared protein samples from cells expressing Wtf4^antidote-mCherry and Wtf4^poison-GFP and performed western blots using α-mCherry and α-GFP antibodies. We analyzed protein from both the pellet and supernatant of our protein preparations because the predicted proteins are likely to be hydrophobic with six predicted transmembrane domains (Figure 1—figure supplement 1A) and their solubility is unknown (TMHMM model, Krogh et al., 2001). Using α-mCherry antibodies, we observed a prominent band at ~65 kDa that showed some smearing consistent with protein degradation. This band was not observed in samples prepared from cells not expressing Wtf proteins, suggesting it is likely full-length Wtf4^antidote-mCherry protein (expected size 65 kDa). The band was more prominent in the pellet than the supernatant consistent with low solubility (Figure 2—figure supplement 3).

Using α-GFP antibodies, we observed multiple bands ranging from ~15–85 kDa, although the expected size of Wtf4^poison-GFP is 60 kDa. The bands were not observed in control samples prepared from cells not expressing Wtf proteins, suggesting the signal is specific to Wtf4^poison-GFP. The large apparent size of some of the Wtf4^poison-GFP bands suggest the protein may have post-translational modifications. In addition, the small size of some of the bands and the overall smeary appearance of the blot is consistent with degradation of the protein. Like the Wtf4^antidote-mCherry, a considerable amount of the Wtf4^poison-GFP was found in the pellet, suggesting both proteins have low solubility (Figure 2—figure supplement 3).

Given that the Wtf4^poison and Wtf4^antidote proteins colocalize and are both found in the pellet fraction of protein preparations, we tested if the proteins physically interact by using acceptor photobleaching Fluorescence Resonance Energy Transfer (FRET, Sekar and Periasamy, 2003) in cells expressing both Wtf4^poison-GFP and Wtf4^antidote-mCherry proteins. This process involves bleaching the fluorescence of a tagged protein (the acceptor) and looking for a corresponding increase in fluorescence of another tagged protein (the donor). If an increase in fluorescence of the donor is observed, the proteins are said to be physically interacting, as they are in close enough proximity (less than 10 nanometers) to transfer energy to each other (Sekar and Periasamy, 2003). When we bleached Wtf4^antidote-mCherry, we saw a corresponding increase in Wtf4^poison-GFP emission, supporting the idea that the two proteins physically interact (Figure 2E and F, Figure 2—figure supplement 1F).

The Wtf4 proteins localize as puncta of varying sizes, so we hypothesized that the proteins assemble into aggregates. To explore the nature of the Wtf4 protein assemblies, we utilized the recently developed Distributed Amphifluoric FRET (DAmFRET) assay (Khan et al., 2018). This approach looks for FRET between red and green versions of the same fluorophore in a partially photoconverted population of mEos3.1-tagged proteins as a measure of the protein’s tendency to self-assemble (Figure 2—figure supplement 4A). We generated wtf4^antidote-mEos3.1 and wtf4^poison-mEos3.1 alleles, both under β-estradiol inducible promoters on ARS CEN plasmids. Both tagged constructs encoded functional proteins in S. cerevisiae, but the mEos3.1-tagged wtf4^poison allele was not as toxic as the GFP-tagged allele (Figure 2—figure supplement 4B). We then carried out DAmFRET analyses on cells producing either Wtf4^antidote-mEos3.1, Wtf4^poison-mEos3.1, or on cells producing both proteins simultaneously. We observed high ratiometric FRET signal (AmFRET) between Wtf4^antidote-mEos3.1 proteins and between Wtf4^poison-mEos3.1 proteins. In fact, all cells expressing Wtf4^poison-mEos3.1 and/or Wtf4^antidote-mEos3.1 proteins exhibited FRET as compared to mEos3.1 negative control, regardless of the expression level of the proteins (Figure 2—figure supplement 4C). The level of AmFRET did not change when both proteins were expressed simultaneously. Collectively, these experiments show that the Wtf4 proteins self-assemble, and that the proteins do not inhibit each other’s assembly.

Interestingly, we did not detect any cells producing purely monomeric Wtf4^poison in our DAmFRET analyses, indicating that self-assembly does not require a rate-limiting nucleation step that is characteristic of prion-forming proteins (Khan et al., 2018). This, combined with the irregular shape of the GFP puncta in our images, suggests that the toxic species of Wtf4^poison is a poorly-ordered assembly of the protein. To explore this idea, we tested if overexpression of chaperones that promote protein homeostasis could neutralize Wtf4^poison aggregates. First, we independently transformed multicopy 2-µm plasmids carrying various galactose-driven chaperones (SIS1, YDJ1, HSP42, JJJ2, HSC82, HSP82, HSP104) into a strain carrying the wtf4^poison-mEos3.1 allele (Figure 2—figure supplement 5; Li et al., 2016b). Overexpression of any of these chaperones failed to ameliorate the toxicity of Wtf4^poison-mEos3.1, even when we reduced the levels of induction of the Wtf4^poison (Figure 2—figure supplement 5). This suggests that neutralization of Wtf4^poison may require the induction of multiple chaperones or stress response pathways at one time.

The Wtf4^poison and Wtf4^antidote proteins share the same 293 C-terminal amino acids (Figure 1A, Figure 1—figure supplement 1A). All the known active Wtf^antidote proteins are highly similar to the Wtf^poison they neutralize (Bravo Núñez et al., 2020a; Bravo Núñez et al., 2020b; Hu et al., 2017). In addition, mutations that disrupt the similarity between a given Wtf^antidote and Wtf^poison can eliminate the ability of the Wtf^antidote to neutralize the Wtf^poison (Bravo Núñez et al., 2018a). Here, we tested the mechanism underlying that requirement using Wtf4 proteins. Given that each Wtf4 protein self-assembles (Figure 2—figure supplement 4), we hypothesized that homotypic interactions between Wtf4^poison and Wtf4^antidote mediate their co-assembly and neutralization of the poison. To test this idea, we mutated sequences at the C-termini of the inducible wtf4^poison-GFP and wtf4^antidote-mCherry alleles in the S. cerevisiae plasmids described above. Specifically, we targeted our mutagenesis to a seven amino acid sequence (IGNAFRG) that is found in varying copy numbers in many members of the wtf gene family (Eickbush et al., 2019). We previously showed that a mismatched number of these repeats between Wtf poison and antidote proteins is enough to disrupt their specificity (Bravo Núñez et al., 2018a). The wild-type S. kambucha wtf4 allele contains ~1.5 repeat units (Figure 3A). To make the mutants, we inserted 18 additional codons into the repeat region of wtf4 to make a total of four repeats. We denote these repeat insertion mutants with an * (Figure 3A).

As expected, the Wtf4^poison*-GFP protein is functional (i.e. toxic) in S. cerevisiae and localizes similarly to the tagged wild-type Wtf4^poison-GFP (Figure 3B, Figure 3—figure supplement 1A). Wtf4^poison*-GFP is neutralized by the matching Wtf4^antidote*-mCherry protein, and the two mutant proteins colocalized as vacuole-associated assemblies, just like the tagged wild-type proteins in S. cerevisiae (Figure 3B and C). The Wtf4^antidote*-mCherry protein on its own also resembled the wild-type Wtf4^antidote-mCherry protein localization (Figure 3—figure supplement 1B). The Wtf4^antidote*-mCherry could not, however, suppress the toxicity of the wild-type Wtf4^poison-GFP (Figure 3B). Similarly, the wild-type Wtf4^antidote-mCherry could not neutralize the Wtf4^poison*-GFP’s toxicity (Figure 3B). The poison and antidote proteins did not colocalize in cells with incompatible poison and antidote proteins, and instead the poison proteins formed distributed aggregates, similar to cells expressing no antidote (Figure 3D and E).

We next used transmission electron microscopy (TEM) to analyze the environment of Wtf proteins within the vacuole-associated aggregates. Similar to our observations made using fluorescence microscopy, we found using immuno-gold labeling that Wtf4^poison-GFP largely clustered near the vacuole in cells also producing untagged Wtf4^antidote (Figure 4A). These images also revealed that the Wtf4 protein aggregates appeared within a cluster of lightly staining organelles resembling lipid droplets (Figure 4A, Figure 4—figure supplement 1A). Very few immunogold particles were found in the cells carrying only empty vectors, suggesting minimal background and high specificity of the GFP antibody used for the immunolabeling (Figure 4—figure supplement 1B).

To look at these Wtf aggregate-associated organelles at higher resolution, we used TEM with a sample preparation method that better maintains cellular morphology (see Materials and methods). We found that the organelles were in fact a mix of lipid droplets and large vesicles with bilayer membranes (Figure 4B arrows, Figure 4—figure supplement 2A–C). We quantified the number of lipid droplets and large vesicles in cells carrying empty vectors and in cells carrying both β-estradiol inducible Wtf4^antidote and β-estradiol inducible Wtf4^poison-GFP. We found that cells expressing Wtf4 proteins had significantly more lipid droplets and large vesicles (Figure 4D, Figure 4—figure supplement 2F). These results indicate that the large aggregates that form in cells expressing Wtf4^antidote are embedded in a cluster of large vesicles and lipid droplets. This phenotype is reminiscent of another aggregation prone protein, α-synuclein, a protein associated with Parkinson’s disease in humans, that when expressed in yeast forms cytoplasmic accumulations in association with clusters of vesicles (Soper et al., 2008). The α-synuclein vesicles, however, appear smaller and more numerous than the Wtf4-associated vesicles. To test if the increase in vesicles and lipid droplets was a common feature of aggregation prone proteins, we expressed (using the β-estradiol system) a different vacuole-associate prion aggregate, Rnq1-mCardinal (Figure 4—figure supplement 2E). We did not observe elevated vesicles or lipid droplets, suggesting that the increase in vesicles is due to the Wtf4 aggregates, not a consequence of the over-expression system or a general feature of aggregation prone proteins.

We also imaged cells expressing only Wtf4^antidote or Wtf4^poison. The morphology of cells expressing only the Wtf4^antidote was indistinguishable from cells expressing both Wtf proteins (Figure 4—figure supplement 2D). It is difficult to interpret the observations from the (dying) cells expressing only Wtf4^poison because the majority of the cells did not maintain cellular integrity during sample preparation (Figure 4—figure supplement 3A). In the few cells we could image, we observed diverse morphologies. We generally did not observe clustering of large vesicles and lipid droplets, as we saw in cells expressing Wtf4^antidote. Instead, organelle integrity often looked disrupted and many cells expressing Wtf4^poison appeared to have undergone extensive autophagy (Figure 4—figure supplement 3B–D).

Our results were reminiscent of other studies in which toxic aggregated proteins were neutralized via sequestration at cellular inclusions (Kaganovich et al., 2008; Liu et al., 2010; Taylor et al., 2003; Chen et al., 2011; Hill et al., 2017; Tyedmers et al., 2010; Kryndushkin et al., 2012; Bagola and Sommer, 2008; Arrasate et al., 2004). In S. cerevisiae, stable, misfolded proteins are generally sequestered to the Insoluble PrOtein Deposit (IPOD), a compartment located near the vacuole and pre-autophagosomal site (PAS) (Kaganovich et al., 2008; Tyedmers et al., 2010; Suzuki and Ohsumi, 2010; Rothe et al., 2018). This compartmentalization of damaged/misfolded proteins mitigates their toxic effects and facilitates their disposal, some of which occurs via autophagy (Marshall et al., 2016).

Given that the Wtf4^antidote and Wtf^poison+Wtf4^antidote aggregates localize adjacent to the vacuole, we hypothesized that they could be at the IPOD in S. cerevisiae. To test this idea, we looked for the localization of the Wtf4 proteins relative to Rnq1-mCardinal and GFP-Atg8. Rnq1 localizes to the IPOD and Atg8 is a component of the pre-autophagosomal structure that is adjacent to the IPOD (Kaganovich et al., 2008; Tyedmers et al., 2010; Rothe et al., 2018). Consistent with our hypothesis, we found that Wtf4^antidote-mCherry either colocalized or was adjacent to Rnq1-mCardinal (Figure 5A, Figure 5—figure supplement 1A–1C). Wtf4^poison-GFP did not colocalize with Rnq1-mCardinal on its own, supporting the idea that Wtf4^antidote recruits the poison to the IPOD (Figure 5—figure supplement 1D).

To visualize the localization of Wtf4^antidote relative Atg8, we expressed Wtf4^antidote-mCherry from a Gal-driven promoter on a plasmid. This is a different expression system than our other experiments described so far, but the protein is functional (Figure 5—figure supplement 2A–B). We found that Wtf4^antidote-mCherry colocalized or was adjacent to GFP-Atg8 (expressed from a plasmid under its endogenous promoter) when both proteins localized outside the vacuole (Figure 5—figure supplement 3D I) Guan et al., 2001. This supports the idea that Wtf4^antidote localizes at the IPOD. However, in the majority of cells expressing both Wtf4^antidote-mCherry and GFP-Atg8, the proteins both localized inside the vacuole (Figure 5—figure supplement 3D, F H). This was different than cells expressing either protein alone, where Atg8-GFP localizes as a focus outside the vacuole (Figure 5—figure supplement 3E) and Wtf4^antidote-mCherry localizes as a large aggregate outside the vacuole (Figure 5—figure supplement 3G). This suggests that the overexpression of GFP-ATG8 enhances the degradation of Wtf4^antidote-mCherry. Additionally, these results could suggest that the increased vacuolar localization of GFP-Atg8 in cells producing Wtf4^antidote-mCherry is due to enhanced recruitment and degradation of autophagosomes. Indeed, we saw increased vesicles in cells expressing Wtf4^antidote (Figure 4).

Proteins in the IPOD tend to be insoluble (Bagola and Sommer, 2008; Kaganovich et al., 2008). To test if the Wtf4^antidote shared this property in S. cerevisiae, we used half punctum-Fluorescence Recovery After Photobleaching (half-FRAP) (Khan et al., 2018; Zhang et al., 2015). This analysis revealed that the Wtf4^antidote-mCherry aggregate has very low internal mobility and is thus more solid-like than liquid-like (Figure 5B). We were curious if the Wtf4^antidote behaved similarly in its native context. To test this, we performed the half-FRAP assay on the Wtf4^antidote-mCherry in S. pombe spores and, consistent with our result in S. cerevisiae, found very low protein mobility (Figure 5—figure supplement 4A B).

To better understand how toxic Wtf4 protein aggregates are neutralized, we screened for genes necessary for survival after induction of wtf4^antidote and wtf4^poison. Briefly, we screened the S. cerevisiae MATa, haploid deletion collection for mutants that failed to survive on galactose media when they carried plasmids containing galactose-inducible wtf4^antidote-mCherry and wtf4^poison-GFP genes (Figure 5—figure supplement 2A B). We found 106 mutants that could grow on galactose when carrying empty vector plasmids, but not when carrying both wtf4 plasmids (Figure 5—figure supplement 2—source data 1). We also provide results of similar screens for comparison (Figure 5—figure supplement 2—source data 1; Willingham et al., 2003 Enyenihi and Saunders, 2003).

Amongst our hits, the only significantly enriched (FDR p<0.05) gene ontology groups were mitochondrial translation and organization (Figure 5—figure supplement 2—source data 1). We speculate this enrichment is due to two known roles of mitochondria in managing protein aggregates. The first is the Mitochondria As Guardian In Cytosol (MAGIC) mechanism in which mitochondria help degrade protein aggregates (Ruan et al., 2017). The second is that mitochondria mitigate the impact of toxic aggregates by promoting asymmetric aggregate segregation in mitosis (Zhou et al., 2014).

We also identified genes involved in cell wall integrity pathways (POP2, MPT5, SLT2 and BCK1) as necessary for survival after induction of Wtf4^antidote and Wtf4^poison (Jin et al., 2015; Li et al., 2016a; Stewart et al., 2007). The cell wall integrity pathway is triggered by diverse stress stimuli (Fuchs and Mylonakis, 2009) and can promote stress-response gene expression and nuclear release of cyclin-C ((Ssn8), also a hit in our screen) (García et al., 2009). Release of cyclin-C into the cytoplasm promotes mitochondrial hyper-fission, stress response gene activation, and either apoptosis or repair of the stress-induced damage (Jin et al., 2015). Consistent with this, we observed separated, significantly smaller mitochondria in cells expressing the Wtf4 proteins (Figure 4—figure supplement 2C and G). Altogether, our screen hits suggest links between the cell wall integrity stress response pathways, mitochondrial fission, and Wtf4^antidote function.

Several other screen hits were genes with known roles in maintaining protein homeostasis and/or aggregate management. For example, we found that several genes involved in vesicle transport, endocytosis, and trafficking to the vacuole (e.g. ATG11, SNF7, and multiple VPS genes) are also required for survival when the Wtf4 proteins are expressed (Figure 5—figure supplement 2—source data 1). These hits suggest vacuolar trafficking pathways contribute to the neutralization of Wtf4 protein aggregates. This is consistent with our EM analyses showing that the Wtf4^antidote inclusion site is enriched with vesicles. Previous work demonstrated these pathways are also important for trafficking other proteins to the IPOD and for neutralizing the toxicity of the aggregation prone TDP-43 (Rothe et al., 2018; He et al., 2006; Liu et al., 2017).

Given our results, which suggest that Wtf4 protein localization is an important factor in mitigating toxicity, we next imaged the localization of Wtf4^poison-GFP and Wtf4^antidote-mCherry in all of the screen hits. We found that the localization of the Wtf4^poison-GFP and Wtf4^antidote-mCherry proteins was disrupted in all 106 hits relative to wild type (where the proteins coalesce to the IPOD). In 81 mutants, the Wtf4^poison-GFP and Wtf4^antidote-mCherry proteins localized as dispersed aggregates throughout the cell. These mutants included deletions of YNL170W, a reported dubious open reading frame, and PHD1, a transcriptional activator (Figure 5—figure supplement 2C D). We noted there were often cells with dispersed Wtf4^antidote-mCherry aggregates or cells with dispersed Wtf4^poison-GFP aggregates, but rarely cells with both. We speculate this is due to toxicity of distributed aggregates and cells expressing both aggregates at the same time being destroyed quickly. Another common feature we observed throughout the screen hits was Wtf4^antidote-mCherry signal in the vacuole. We also observed this vacuolar localization in the C-terminal mutants depicted in Figure 3D and E, so this appears to be a common feature of the Wtf4^antidote-mCherry protein in cells being destroyed by Wtf4^poison. Five mutants appeared to have wild-type looking Wtf4^antidote (single inclusion outside the vacuole) but dispersed Wtf4^poison, suggesting that the mutations may disrupt the interaction of the poison and antidote (Figure 5—figure supplement 2E F). Twenty hits showed very little Wtf4 signal and soluble cytoplasmic localization of both proteins. (Figure 5—figure supplement 2G).

Because the Wtf4^antidote protein is quite similar to the Wtf4^poison and also assembles into aggregates, we were curious if the Wtf4^antidote alone was toxic in the absence of any of our screen hits. We therefore assayed the viability of the 106 deletion mutants when only Wtf4^antidote was expressed. We saw that in approximately half (44/106) of the deletion stains Wtf4^antidote expression reduced viability (Figure 5—figure supplement 2—source data 1). These results are consistent with the idea that active aggregate management pathways are often required for cells to mitigate the toxicity of even the Wtf4^antidote protein in S. cerevisiae.

We also investigated one hit from our screen, VPS1, more thoroughly using our β-estradiol-inducible wtf4 plasmids (described above). Vps1 is a dynamin-like GTPase that is necessary for trafficking of aggregates to the IPOD and/or other inclusion sites (Kumar et al., 2016; Kumar et al., 2017; Hill et al., 2016; Marshall et al., 2016). In the absence of VPS1, we found that the Wtf4^antidote-mCherry and Wtf4^poison-GFP proteins still physically interact (Figure 5C, Figure 5—figure supplement 5A). The Wtf4 protein aggregates did not, however, coalesce to form large inclusions (Figure 5D, Figure 5—figure supplement 5A and B) and Wtf4^antidote-mCherry failed to neutralize the toxicity of Wtf4^poison-GFP in vps1Δ cells (Figure 5E).

Together, these experiments indicate that the physical interaction between the Wtf4^poison and Wtf4^antidote proteins is insufficient to neutralize the toxicity of Wtf4^poison protein aggregates. Trafficking the aggregates to a vacuole-associated inclusion is also required. Interestingly, we also observed enhanced toxicity of the Wtf4^antidote-mCherry protein in the absence of Vps1 and many of our other screen hits (Figure 5E, Figure 5—figure supplement 2—source data 1). These results suggest that the antidote aggregates are more detrimental to cells when they are distributed in the cytoplasm. Importantly, however, even in the vps1Δ mutant, expression of Wtf4^antidote-mCherry is less toxic to cells than Wtf4^poison-GFP. This, and the fact that not all of the 106 hits caused Wtf4^antidote-mCherry to become toxic, suggests there are fundamental differences in the poison and antidote aggregates beyond their propensity to be trafficked to a vacuole-associated inclusion.

Here, we explored how the Wtf4^poison protein kills cells and how the Wtf4^antidote protein neutralizes the toxicity of the Wtf4^poison. We used a combination of genetics and cell biology to study these proteins in three contexts: (1) their endogenous context of S. pombe gametogenesis, (2) vegetatively growing S. pombe cells, and (3) vegetatively growing S. cerevisiae cells. In all three contexts, expression of Wtf4^poison alone kills cells and expression of the Wtf4^antidote rescues the toxicity. The simplest interpretation of these observations is that Wtf4^poison exploits or disrupts a conserved aspect of cellular physiology that is important during both vegetative growth and gametogenesis. Similarly, the Wtf4^antidote neutralizes the Wtf4^poison using conserved cofactors that can act in both vegetative growth and gametogenesis. This conservation suggests that wtf-derived gene drives could be a useful tool for genetically altering populations.

In S. pombe gametogenesis and in vegetative S. cerevisiae cells, we observed the Wtf4^poison-GFP proteins assembled into small foci (aggregates) in the absence of Wtf4^antidote. The aggregates were largely dispersed throughout the cytoplasm, with some ER localization. The assembly of Wtf4 proteins is reminiscent of another meiotic drive element, Het-s, which employs prion-like amyloid polymerization to convert Het-S proteins to a lethal form (Dalstra et al., 2003; Riek and Saupe, 2016). We therefore evaluated whether Wtf4^poison proteins exhibit prion activity in S. cerevisiae using DAmFRET (Khan et al., 2018). We found that Wtf4^poison-mEos proteins assembled with themselves even at very low expression levels (Figure 2—figure supplement 4C). In fact, we were unable to detect cells that lacked self-assemblies, revealing that the toxic form of the protein is not appreciably supersaturated, as would be required for Wtf4^antidote to detoxify it through a simple prion-like mechanism where the antidote templates the deposition of poison monomers. Nevertheless, the sequence-dependent self-assembly of Wtf4 remains consistent with amyloid polymerization. However, given its intimate association with vesicles, extensive testing would be required to further evaluate the structural basis of Wtf4 activity.

The significance of the Wtf4^poison aggregation is not clear. We speculate that the aggregation propensity is intimately tied to the toxicity of Wtf4^poison. We propose that distributed Wtf4 aggregates interact broadly with other proteins and disrupt their folding or localization. Compounding effects of these hypothesized interactions could disrupt protein homeostasis or cellular integrity, leading to cell death. This death may occur via a programmed cell death pathway, as in both S. pombe gametogenesis and in vegetative S. cerevisiae, cells succumbing to the Wtf4^poison exhibit nuclear condensation (followed by nuclear fragmentation in S. pombe). The death may also be related to loss of cell wall integrity, as cell wall integrity pathways are necessary for cell survival upon expression of the Wtf4 proteins. Testing these ideas may be challenging, especially if understanding Wtf4^poison toxicity proves to be as elusive as understanding the intensely studied neurotoxic aggregating proteins TDP-43 and α-Synuclein (Johnson et al., 2009; Cookson and van der Brug, 2008).

Like Wtf4^poison, the Wtf4^antidote also assembles into aggregates in both S. pombe and S. cerevisiae cells. Unlike the Wtf4^poison, however, the Wtf4^antidote aggregates have little effect on the viability of wild-type vegetative cells. This is surprising given the similarity of the two proteins (the Wtf4^poison shares 292 of the Wtf4^antidote’s 337 amino acids). Our data suggest that the localization of the aggregates and/or the exposed aggregate surface area could underlie their differences in toxicity. The Wtf4^poison aggregates (without Wtf4^antidote) remain largely dispersed in the cytoplasm, whereas the Wtf4^antidote proteins are trafficked to a confined region near or within the vacuole.

In S. pombe cells, the Wtf4^antidote aggregates enter the vacuole. In S. cerevisiae cells, the Wtf4^antidote accumulates outside the vacuole in the IPOD, but could also be trafficked into the vacuole at some rate. We observed signs of protein degradation, but this could be due to vacuolar degradation or other degradation mechanisms (Figure 2—figure supplement 3). The different modes of cell division in S. pombe and S. cerevisiae may contribute to the differences between how the species handle Wtf4 aggregates. S. cerevisiae divides asymmetrically by budding and tends to retain aggregates, including Wtf4 proteins, in the mother cells (Spokoini et al., 2012). S. pombe, however, generally divides symmetrically making it difficult for new cells to be born free of aggregates found in the progenitor cell. In some cases, S. pombe can asymmetrically divide damaged proteins to inclusions (Coelho et al., 2014), but the mechanism is not as efficient as S. cerevisiae’s exclusion of aggregates from buds. Due to this, it may be more important for S. pombe to destroy aggregates in the vacuole. The structure of the vacuoles is also fundamentally different in the two species, as S. cerevisiae tends to have one large vacuole per cell, while S. pombe has many small, distributed vacuoles. Additionally, the mechanisms determining the site of inclusions in S. pombe remain to be elucidated. Interestingly, S. pombe wtf4, the homolog of the S. kambucha wtf4 allele used in this study, was previously shown to localize both inside the vacuole and as large cytoplasmic inclusions when overexpressed in S. pombe vegetative growth (Matsuyama et al., 2006). This suggests that some Wtf proteins can be recruited to both vacuoles and inclusions in S. pombe.

When we disrupted the ability of S. cerevisiae cells to transport the Wtf4^antidote aggregates with the vps1∆ mutation, we found that the Wtf4^antidote aggregates were distributed and more toxic than in wild-type cells. This is consistent with the idea that a key feature of Wtf4 protein toxicity relies on the aggregates being widely dispersed in the cytoplasm. When Wtf4^poison and Wtf4^antidote are found together in wild-type cells, the proteins co-assemble into aggregates. The co-assembled aggregates then behave similarly to the Wtf4^antidote aggregates and are trafficked into the vacuole (in S. pombe cells) or to or near the IPOD adjacent to the vacuole (in S. cerevisiae cells) where they cause limited toxicity. Also, like the Wtf4^antidote aggregates, the toxicity of the Wtf4^poison+Wtf4^antidote co-assembled aggregates is greatly enhanced if aggregate transport to the vacuole is disrupted by mutations (e.g. vps1∆).

Together, our observations suggest a mechanistic model for wtf4 function. In this model, wtf4 exploits protein aggregation control pathways to induce selective cell death. The Wtf4^poison forms distributed toxic aggregates and the Wtf4^antidote co-assembles with the Wtf4^poison and neutralizes the aggregate’s toxicity via trafficking to the vacuole (Figure 6). This mechanism is unlike the mechanism of any other meiotic driver described to date (Grognet et al., 2014; Didion et al., 2015; Long et al., 2008; Dawe et al., 2018; Rhoades et al., 2019; Dalstra et al., 2005; Hammond et al., 2012; Vogan et al., 2019; Chen et al., 2008; Akera et al., 2017; Bauer et al., 2012; Pieper et al., 2018; Herrmann et al., 1999; Shen et al., 2017; Yu et al., 2018; Bauer et al., 2007; Wu et al., 1988; Xie et al., 2019; Kruger et al., 2019; Lin et al., 2018; Svedberg et al., 2020), but there are very few mechanistically characterized killer meiotic drive systems (reviewed in Bravo Núñez et al., 2018b).

This study focused on the wtf4 meiotic driver. There is, however, an incredibly diverse array of wtf genes that cause meiotic drive. For example, the poison protein encoded by wtf35 (from the FY29033 isolate) shares less than 23% amino acid identity with Wtf4^poison (Bravo Núñez et al., 2020a). Despite that extreme divergence, both genes cause essentially the same phenotype in S. pombe meiosis: drive of the gene into >90% of the progeny of a heterozygote (Bravo Núñez et al., 2020a). The conserved protein aggregation model offers an explanation for how such a diverse array of proteins can cause the same phenotype. Under our model, the mechanism of the Wtf^poison proteins is dependent upon their aggregation propensity. Presumably, the evolution of a protein that must self-aggregate could be less constrained than the evolution of a protein that must maintain a specific enzymatic activity or interaction partner.

Exon 1 of the wtf4 driver encodes the antidote-specific residues (45 amino acids) that facilitate the recruitment of Wtf4 aggregates to the vacuole in S. pombe (or the IPOD in S. cerevisiae). This antidote-specific function likely relies on unidentified interacting partners (perhaps amongst our screen hits). This specific functional requirement could explain the greater conservation of exon 1-encoded residues amongst bona fide wtf drivers (68–100% amino acid identity) compared to the conservation amongst the remaining exons (30–90% amino acid identity) (Bravo Núñez et al., 2020a).

Importantly, our model also suggests that aggregate management may be a major feature of gametogenesis in S. pombe. The number of wtf genes varies between different isolates, but most have 30 or more wtf genes (Eickbush et al., 2019). Four of these genes are widely diverged from wtf4 and are either not expressed in gametogenesis or their proteins exhibit distinct cellular localization from known Wtf^poison and Wtf^antidote proteins (Bravo Núñez et al., 2020a). The rest of the genes that have been tested are similar to wtf4 in expression and localization (Bravo Núñez et al., 2018a; Bravo Núñez et al., 2020a). It is not clear how many wtf drivers are expressed in a given cell, but they all appear to be transcribed at some level (Eickbush et al., 2019; Kuang et al., 2017). It will be interesting to explore the direct and indirect impacts of these Wtf proteins on S. pombe gametogenesis.

The genetic screen presented in this work identified a number of factors required for cell viability in S. cerevisiae cells expressing Wtf4^poison and Wtf4^antidote. Many of these genes informed our model for Wtf4 protein function and therefore fit nicely within our proposed model. For example, our screen implicated genes involved in the Cytoplasm-to-Vacuole Targeting (CVT) pathway as necessary for survival of the Wtf4 proteins. This pathway has been previously implicated in aggregate management (Kumar et al., 2016; Kumar et al., 2017).

Not all of our screen hits, however, are in genes or pathways with annotated roles that clearly fit our model. Some of the genes have no annotated functions. It is possible that at least some of these genes are not directly involved in aggregate management, but the mutants are especially sensitive to the stresses imposed by Wtf4 aggregates. It is also possible that some of the genes do have roles in mitigating the effects of toxic aggregates. Indeed, in deletions of some genes with unknown functions, we saw distributed Wtf4 aggregates, suggesting these unknown proteins could play a role in sequestration of aggregates. Interestingly, other hits are in well-studied genes, such as multiple acetyltransferases and various kinetochore proteins. Future analysis of these hits will be essential to refine or to potentially reject our current model.

Studying how parasites manipulate their hosts can uncover unexpected insights on the host’s biology. For example, studies of the mouse t-haplotype meiotic driver revealed that gene expression in spermatids can create sperm-autonomous phenotypes, even though spermatids are connected by intercellular bridges (Herrmann et al., 1999). Under our model, a fine line exists between protein aggregates that cells can manage (i.e. Wtf4^antidote) and lethal aggregates that are not effectively managed (i.e. Wtf4^poison). We propose that the wtf4 meiotic driver has exploited this feature for its own selfish advantage. Future studies can now exploit the Wtf4 proteins to learn about protein aggregate toxicity and cellular aggregate management strategies.

Z₃EV promoter system is a titratable inducible promoter system (Ohira et al., 2017). The system requires the Z₃EV transcription factor and a Z₃EV-responsive promoter (Z₃EVpr). β-estradiol induces nuclear import of the Z₃EV protein; therefore, genes placed immediately downstream of Z₃EVpr in a strain expressing Z₃EV become expressed upon β-estradiol addition to the media.

The LexA-ER-AD system (Ottoz et al., 2014) utilizes a heterologous transcription factor containing a LexA DNA-binding protein, the human estrogen receptor (ER) and an activation domain (AD). β-estradiol binds the ER and tightly regulates the activity of the LexA-ER-AD transcription factor. The LexA DNA-binding domain recognizes lexA boxes in the target promoter.

To create the Sec63-YFP strain, we PCR amplified the C-terminus of sec63 (using oligos 939+941) and the sequence downstream of sec63 (using oligos 945+946) using SZY643 as a template. We also amplified a YFP-HIS3 cassette from pYM41 (Janke et al., 2004) using oligos 944+943. We then used overlap PCR (using oligos 939+943) to unite those three PCR products. We then transformed this PCR product into GP1163 with standard lithium acetate protocol (Gietz et al., 1995) (selecting for His+) to integrate the tagged sec63-YFP at its endogenous locus to generate SZY1277. We confirmed the strain via PCR using oligos 2037+2038.

We PCR amplified the vps1Δ::kanMX locus out of strain YKR001C from the haploid yeast knockout MATa collection (Open Biosystems) (using oligos 1850+1851) and transformed the PCR product into SLJ769 using high efficiency lithium acetate protocol (selecting for G418 resistance) to create strain SZY2539. We used PCR (using oligos 1712+1713) and sequenced the locus to confirm the deletion. To add the PACT1-LexA-ER-haB42 transcription factor, we digested FRP718 (Addgene #58431, Ottoz et al., 2014) with NheI and integrated it into SZY2539 at the his3-11,15 locus (using standard lithium acetate protocol [Gietz et al., 1995] and selecting for His+) to create SZY2552.

We induced samples of SZY2072, SZY2070, SZY2159, and SZY2059, with β-estradiol as described above. We then aliquoted these induced samples into a 96-well plate. We then partially photoconverted the mEos3.1 protein by exposing the plate, while shaking at 800 RCF, to 405 nm illumination for 25 min using an OmniCure S1000 fitted with a 320–500 nm (violet) filter and a beam collimator (Exfo), positioned 45 cm above the plate. This exposure yielded a total photo dose of 16.875 J/cm². This photo dose reproducibly achieves the maximum fluorescence of the acceptor (red) form of mEos3.1 while minimizing photobleaching of the green form (Khan et al., 2018). For Figure 2—figure supplement 4C, we assayed the photoconverted samples on a Bio-Rad ZE5 cell analyzer with high-throughput automation. We analyzed 20 µL of each sample to collect approximately 100,000 events per well. We excited the mEos3.1 donor (green form) with a 488 nm laser at 100 mW and collected with 525/35 nm and 593/52 nm bandpass filters, respectively. We excited the acceptor fluorochrome with a 561 nm laser at 50 mW and collected with a 589/15 nm bandpass filter. We performed manual compensation on-instrument at acquisition. We used DeNovo FCS Express for data analysis and visualization and calculated ratiometric FRET as FRET/acceptor signals.

For imaging during S. pombe gametogenesis (Figure 1C, D and G, Figure 1—figure supplement 1B, Figure 1—figure supplement 2A and C), we crossed the two haploid yeast strains to generate heterozygous diploids as previously described (Nuckolls et al., 2017). We placed the diploids on sporulation agar (SPA, 1% glucose, 7.3 mM KH₂PO₄, vitamins, agar) for 2–3 days. We then scraped the cells off of the SPA plates and onto slides coated with 0.2 mg/mL lectin (Sigma) for imaging (Tomita and Cooper, 2007).

For vegetatively growing samples (Figure 1E and F, Figure 1—figure supplement 3B and D), we induced gene expression with β-estradiol as described above. If we used vacuole staining, we took 1 mL of the induced culture, spun to pellet, and resuspended in 1 mL of 10 mM HEPES buffer, pH 7.4, containing 5% glucose with 100 μM CellTracker Blue CMAC (Component B; Invitrogen C2110). We incubated these cells at room temperature for 30 min. We then washed with YEL media and imaged. For imaging, we used the LSM-780 (Zeiss) microscope with a 40x C-Apochromat water-immersion objective (NA 1.2) in photon-counting channel mode. For GFP, we used 488 nm excitation and collected through a 491–552 bandpass filter. For mCherry, we used 561 nm excitation and collected through a 572 longpass filter. For YFP, we used 514 nm excitation and collected through a 500–589 nm bandpass filter. For CMAC, we used 405 nm excitation and collected through a 411–509 nm bandpass filter. Brightness and contrast are not the same for all images. We analyzed at least 20 cells for each strain and chose a representative image. For experiments assaying meiosis/gametogenesis, we used at least two independent progenitor diploids. For cells that were imaged during vegetative growth, we used at least three different starting cultures.

We carried out Pearson correlation analysis (Adler and Parmryd, 2010) as previously described (Slaughter et al., 2013). Briefly, we drew a segmented line (width of two pixels) throughout the spore, randomly covering as much of the spore as we could. We then used an in-house custom written plugin for ImageJ (https://imagej.nih.gov/ij/) to generate a two-color line profile. We calculated the Pearson correlation of the line profile with varying degrees of shifts in at least eight spores or six vegetatively growing cells per sample. We then combined and averaged the trajectories with standard error.

To quantify nuclear size, we calculated the full width at half maximum of the fluorescence intensity of RFP. We quantified 42 spores that inherited wtf4-GFP and 19 that did not, all from a wtf4-GFP/ade6+ heterozygote after 2 days on SPA media. We excluded any nuclei that appeared to have already fragmented.

For the nuclear timelapse (Figure 1—figure supplement 4), we grew diploid cultures to saturation at 32°C overnight in YEL media. We then plated 100 µL of the cultures on a SPA plate, cut a circle punch of agar from the plate, and placed this punch upside down (cells facing down) in a 35 mm glass bottom poly-D-lysine coated dish (MatTek corporation). We placed grease around the edge of the MaTeK dish and a moist kim wipe inside to control for humidity. We then imaged the cells using the Nikon Ti Eclipse coupled to a Yokogawa CSU W1 Spinning Disk, using the 60x oil objective, acquiring images every ten minutes. Here, we excited RFP at 561 nm and collected its emission through a 605–70 nm bandpass filter.

For the gametogenesis timelapse (Figure 1—figure supplement 1B), we grew diploid cultures to saturation at 32°C overnight in YEL media. The next day, we diluted 100 µL of the saturated diploid culture into 5 mLs of PM media (20 mLs of 50x EMM salts, 20 ml 0.4 M Na₂HPO₄, 25 mL 20% NH₄Cl, 1 mL 1000x Vitamins, 100 µL 10,000x mineral stock solution, 3 g potassium hydrogen phthalate, 950 mL ddH₂O, 25 mL of sterile 40% glucose after autoclaving, supplemented with 250 mg/L uracil). We grew the PM culture overnight at 32°C. The next day, we spun to pellet and resuspended the pellet in PM-N media (20 mLs of 50x EMM Salts, 20 mL 0.4 M Na₂HPO₄, 1 mL 1000x Vitamins, 100 µL 10,000x mineral stock solution, 25 mL of sterile 40% glucose after autoclaving, supplemented with 250 mg/L uracil, volume up to 1 L with ddH₂O). We shook the PM-N cultures for 4 hr at 28°C. Then, we took 100 µL of the PM-N culture and mixed it with 100 µL of lectin (Sigma). We took 150 µl of this mixture and added it to a 35 mm glass bottom poly-D-lysine coated dish (MatTek corporation). We waited 5 min to allow the cells to adhere. We then added 3 mL of fresh PM-N to the dish (protocol modified from Klutstein et al., 2015). We imaged using a Zeiss Observer.Z1 wide-field microscope with a 63x (1.4 NA) oil-immersion objective and collected the emission onto a Hamamatsu ORCA Flash 4.0 using μManager software. We acquired the mCherry with BP 530–585 nm excitation and LP 615 emission, using an FT 600 dichroic filter, acquiring images every 10 min.

For all budding yeast images except for the three experiments described below, we induced samples as described above and imaged on an LSM-780 (Zeiss) microscope, with a 40x C-Apochromat water-immersion objective (NA 1.2) in photon-counting channel mode. For GFP and mCherry, we used the same conditions as we did in S. pombe. For mCardinal, we used 633 nm excitation and collected through a 632–696 nm bandpass filter. Brightness and contrast are not the same for all images. We imaged at least 20 cells from at least three starting cultures and chose a representative image for each figure.

For imaging vps1Δ cells (Figure 5), IPOD cytology (Figure 5—figure supplement 1A), and the nuclear timelapse (Figure 2—figure supplement 1B and C), we placed samples in a Millipore Onix 2 Cellasic system to allow for a constant flow of media. We initiated flow of inducing media (SC with 500 nM b-estradiol for vps1Δ and SC galactose for the nuclear timelapse) and took images every 10 min. We used a Perkin Elmer Ultraview Vox spinning disc microscope with a Hamamatsu EMCCD (C9100-23B) with a 40x C-Apochromat water-immersion objective (NA 1.2). We collected GFP and mCherry with 488 and 561 nm excitation as above but collected GFP through a 525–550 nm bandpass filter and mCherry through a 615–670 nm bandpass filter. We had two independent starting cultures for the sample. We chose representative cells and timepoints. Brightness and contrast are not the same for all images.

To quantify nuclear size (Figure 2—figure supplement 1D), we fit each nucleus to a two-dimensional Gaussian function and found the full width at half maximum of the fluorescence intensity of RFP per cell. We quantified at the beginning of the timelapse (early) and 14 hr into the timelapse (late). We quantified 72 Wtf4^poison-GFP expressing cells and 65 wild-type cells at the early timepoint. We quantified 62 Wtf4^poison-GFP expressing cells and 79 wild-type cells at the later timepoint.

We grew overnight cultures of SZY1954 (cells carrying inducible Wtf4^poison-GFP, Wtf4^antidote-mCherry) and SZY1821 (empty vector (EV), untagged control) in SC -His -Ura -Trp (without agar). The next day, we diluted 10 mL of the saturated culture into 90 mL of media of the same type. We then added β-estradiol to a final concentration of 500 nM. We shook the cultures at 30°C to induce. After 2 and 5 hr of induction, we prepared whole-cell lysates as previously described (Gerace and Moazed, 2014), with a few exceptions. First, we used 1% tritonX-100 instead of NP40 in the lysis buffer. For bead beating, we used the FastPrep-24 5G bead beater with Lysing Matrix Y bead tubes (MP Biomedicals) and the S. cerevisiae fast prep program (40 s). We took both the supernatant and pellet protein samples for analysis. Samples were boiled for 1 min at 95°C before loading.

Next, we ran the proteins on NuPAGE 4–12% Bis-Tris Protein Gels (Invitrogen, NP0321) and then transferred to Trans-Blot Turbo Mini PVDF membranes (Bio-RAD #1704156). We stained the membranes with a monoclonal, α-GFP antibody (from cell signaling technology #2956) at 1:1000 and monoclonal, α-mCherry antibody (from EnCor Biotechnology Inc #MCA-1C151) at 1:1000, overnight at 4°C with agitation in Odyssey blocking buffer (TBS, from LI-COR biosciences). A secondary, α-rabbit antibody (800 cW) and secondary, α-mouse antibody (600 cW) was used for fluorescent visualization of the proteins. We imaged the blot on the Odyssey-CLx (LI-COR biosciences). Three independent starting cultures were analyzed to confirm results.

We carried out acceptor photobleaching FRET with β-estradiol induced (described above) SZY1954 (wild type) using a LSM-780 (Zeiss) microscope, with a 40x C-Apochromat water-immersion objective (NA 1.2) in photon-counting channel mode. For vps1Δ cells (SZY2570), we used a Perkin Elmer Ultraview Vox spinning-disc microscope with a Hamamatsu EMCCD (C9100-23B) with 488 and 561 nm excitation. For both samples, we photobleached the acceptor (mCherry) with 561 nm excitation (for bleaching images, see Figure 2—figure supplement 1F for wild type and Figure 5—figure supplement 5A for vps1Δ). We analyzed 22 wild-type and 76 vps1Δ cells.

We made 50 mL saturated overnight cultures of SZY1821, SZY1952, SZY1954, and SZY2731 in SC media lacking histidine, tryptophan, and uracil (to select for retention of the plasmids). The next day, we diluted 10 mL of the saturated cultures into 90 mL of the same media with 500 nM β-estradiol. We shook these cultures for four hours at 30°C, reaching log phase. We then pelleted the yeast cells by filtering and carried out high pressure freezing with the Leica ICE system (Leica Biosystems). We further processed the frozen cell pellets by freeze substitution (FS) using acetone containing 0.2% uranyl acetate (UA) and 2% H₂O was used as FS medium. The FS program was −90° to −80° over 70 hr, −80° to −60° over 6 hr, −60° for 5 hr, −60° to −50° over 6 hr, and −50° to −20°C over 4 hr. After washing extensively with acetone, we then infiltrated, embedded and polymerized the samples into resin.

For Immuno-EM, we used HM-20 resin. We cut 60 nm sections with a Leica Ultra microtome (Leica UC-6) and picked up onto a carbon-coated 150 mesh nickel grid. The grids were labeled with anti-GFP primary antibody (a gift from M. Rout, Rockefeller University, New York, NY) and 12 nm colloidal gold goat anti-rabbit secondary antibody (Jackson Immuno Research Laboratories, Inc). After immuno labeling, we post-stained the samples with 1% UA for 3 min. We acquired images using a FEI Tecnai Biotwin electron microscope. For non-immuno-EM, we used Epon resin to better maintain morphology, but the rest of the procedure was the same. We analyzed the tomographs of at least 10 cells per condition.

For quantification purposes, we also completed array tomography. For array tomography, we cut 60 nm serial sections with a Leica Ultra microtome (UC-6) using an Ultra 35 Jumbo diamond knife (Diatome) and picked up on ITO coated coverslips using the ASH-100 Advanced Substrate Holder (RMC Boeckeler). We post-stained serial sections with Sato’s triple lead stain for two minutes, 4% UA in 70% methanol for two minutes, and Sato’s lead stain again for two minutes. The coverslips were coated with 5 nm of carbon and imaged in a Zeiss Merlin Gemini 2 SEM with 4QBSD detector at 10 kV and 700 pA using Atlas 5 Array Tomography software (Fibics). The obtained dataset was aligned with Midas of the IMOD software package (Kremer et al., 1996) and manually quantified. For better visualization, an image series of an individual yeast cell was cropped and further aligned with registration tools in ImageJ.

For model building, the segmentation was done based on intensity and known organelle structure with Microscopy Image Browser (Belevich et al., 2016) and with IMOD. We used Amira (Thermo Fisher Scientific) software for model rendering and visualization.

Further quantification of mitochondrial volumes was performed on selected cells after training a Unet (Ronneberger et al., 2015). Hand annotation of training data was performed in Fiji. A suite of internally developed Fiji plugins, macros and CherryPy scripts called DeepFiji (see below) sent training data to a pair of in-house NVIDIA Tesla-equipped deep learning machines running Tensorflow. Representative cells were selected, and segmented images inferred using the same macros and deep learning machines before being aligned using a StackReg variant. Mitochondrial volumes were quantified in Fiji using the 3D Segmentation tools.

DeepFiji is a suite of macros and plugins in Fiji, Python, and CherryPy (a Python web framework) that enable end users on any machine with a reasonable amount of RAM to request deep learning training and inference on a remote deep learning box as long as both machines have access to a shared file system.

First, a user selects example sub-images that span the realm of potential objects, background levels and signal levels. Manual annotations are made using Fiji's Region of Interest (ROI) tools and manager, and individual ROI files are saved for each image (in our case the ROIs were each individual cell and the total of the mitochondria inside). A user chooses a small subset of the annotated image/ROI pairs to be used as a validation set, while the remainder becomes the training set. For each image, two binary channels are added from the associated ROIs: a mask channel and an outline channel. The mask channel has all pixels contained within an ROI painted true, while the outline channel only paints true the pixels that were in the ROI's outline. As the training network expects standard image sizes and runs more efficiently with smaller images, the macro next makes image stacks for both annotated training and validation images that break the original images up into 512 × 512 sub-images with 50% overlap between sub-images. For the training set, it also applies a series of random rotations and translations to help the neural network generalize. Both the validation and new training images are saved to a shared file system and the deep learning boxes are notified to begin processing through a call to a webserver running on the box. The sub-image size, file location, and other parameters are configurable by the user at the time of running.

For our in-house system, the deep learning boxes are running ubuntu with NVIDIA Tesla boards and configured with Tensorflow 1.13.1 and CUDA 10.0. CherryPy is configured to listen for web calls on each and, once initiated from a user, begins processing files from the selected directory by calling trainer.py. The trainer first finds the standard deviation (STDEV) and mean (MEAN) of the non-zero pixel intensities and stores those values. The training, validation, and all future inference sets will be processed by (Intensity-MEAN)/STDEV+0.5 first to keep numbers roughly between 0 and 1. The model used is a modified Unet(ref): convolutional layers are alternated with max pool layers, doubling the channel depth at each layer while halving the resolution. The final layer is 32 × 32×512. At each convolutional layer, the image is passed through a leaky rectified linear unit. Convolutional up-sampling brings the image back to its original resolution where it is passed through a tanh() function and the mean square error is calculated with respect to the ground truth image for back-propagation. Every 100th iteration of the training is applied to the validation set and the images are visualized using TensorBoard (which opens automatically on the user's host computer). Training proceeds with the learning rate adjusting over time, until 2200 iterations have passed at which point most networks have either converged or never will.

Once training is completed, and a reasonable iteration point is found in TensorBoard, the user can run Inferer.ijm in Fiji on their host machine to apply their model to a new dataset. Inferer will similarly parse images into sub-images and contact a deep learning box to initiate processing. Once processing is complete the user can run a second script to blank out border regions and de-window their images. The output outline and mask channels are ranged from 0 to 1 and represent probabilities. Typically, thresholding pixel values above 0.5 in the mask channel will suffice for finding objects of interest. However, in cases with frequent object touching, one can subtract the outline probability from the mask.

If consistent mistakes are found in the inferred data, the user can annotate them properly using Fiji and use Retrain.ijm to retrain the network using the new data together with the old to generate a new training set. Retraining starts from the original model so that it does not have to relearn from scratch.

Plugins necessary to run in Fiji are available from the Stowers update site within Fiji. Macros, python code, and CherryPy configurations are available at https://github.com/cwood1967/DeepFiji.