Mitotic checkpoint gene expression is tuned by coding sequences

The mitotic checkpoint (also called spindle assembly checkpoint, SAC) is a signaling pathway that safeguards proper chromosome segregation. Proper functioning of the SAC depends on adequate protein concentrations and appropriate stoichiometries between SAC proteins. Yet very little is known about SAC gene expression. Here, we show in fission yeast (S. pombe) that a combination of short mRNA half-lives and long protein half-lives supports stable SAC protein levels. For the SAC genes mad2+ and mad3+, their short mRNA half-lives are supported by a high frequency of non-optimal codons. In contrast, mad1+ mRNA has a short half-life despite a low frequency of non-optimal codons and despite the lack of known destabilizing motifs. Hence, different SAC genes employ different strategies of expression. We further show that Mad1 homodimers form co-translationally, which may necessitate a certain codon usage pattern. Taken together, we propose that the codon usage of SAC genes is fine-tuned for proper SAC function. Our work shines light on gene expression features that promote spindle assembly checkpoint function and suggests that synonymous mutations may weaken the checkpoint.


26
The spindle assembly checkpoint (SAC; also called mitotic checkpoint) is a eukaryotic signalling 27 pathway that delays cell cycle progression when chromosomes have not yet become properly 28 attached to microtubules during mitosis (Kops et  Yet, expression of these genes has not been studied in any detail. 34 The protein network of the SAC, on the other hand, is well understood. While the SAC is active, 35 it forms the mitotic checkpoint complex (MCC), which prevents the anaphase-promoting complex 36 (APC/C) from initiating anaphase (Pines, 2011). A key effector of the SAC is the Mad1/Mad2 37 complex, a tetramer of two Mad1 and two Mad2 molecules (Chen et al, 1999;Sironi et al, 2002) 38 ( Fig 1A). Mad1 homodimerizes through a long, parallel inter-molecular coiled-coil at its N-terminus, the process, forming C-Mad2/Cdc20 through the same seat belt type of binding (Luo et al., 2002). 52 Mad3-GFP were in the range of 0.2, whereas that for Nmt1-GFP was around 0.6 (Fig 2A). 138 This raised the question how the protein concentrations of Mad1, Mad2 and Mad3 can be 139 homogeneous across the population when the mRNA numbers are highly variable. We considered 140 a simple gene expression model with a constitutively active promoter and different mRNA and 141 protein synthesis and degradation rates (see Methods for details) that would all yield mRNA and 142 protein numbers similar to those that we observe for mad1 + , mad2 + and mad3 + . The longer the 143 mRNA half-life, the longer a state of low or high mRNA numbers persists; and the shorter the protein 144 half-life, the more closely protein concentrations follow the mRNA numbers ( Fig 2B). Hence long 145 mRNA half-lives and short protein half-lives favour noise, whereas short mRNA-half lives and long 146 protein-half-lives suppress noise (Fig 2B,C; S2B). In the latter case, the long persistence time of 147 proteins buffers fast fluctuations at the mRNA level (Fig 2B). 148 To ascertain whether this prediction is met by SAC genes, we measured mRNA and protein 149 half-lives. We determined mRNA half-life by metabolic labelling followed by depletion of the labelled 150 pool and quantification of the remaining pool by quantitative PCR. The mRNA half-lives for mad1 + , 151 mad2 + and mad3 + were all in the range of a few minutes (mad1 + : 5.6 min, mad2 + : 7.7 min, mad3 + : 152 5.2 min) ( Fig 2D). This was consistent with the half-lives determined for these genes in a large-153 scale study using metabolic labelling ( Fig S2D)  but even in this study SAC genes were at the lower end of mRNA half-lives ( Fig S2D). As controls, 157 we measured two unrelated genes with reportedly long and short half-life (Eser et al., 2016), act1 + 158 and ecm33 + , which behaved as expected (Fig 2D). We determined protein half-lives by translation 159 shut-off using cycloheximide, followed by immunoblotting. The half-lives of Mad1, Mad2 and Mad3 160 were in the range of many hours, considerably longer than the typical S. pombe cell cycle of 161 2.5 hours (Fig 2E; S2E of codon optimality such as the percentage of optimal codons or the tRNA adaptation index (tAI) 178 ( Fig S3A). Since the SAC genes had short mRNA half-lives, we expected them to have low CSCg 179 values. Indeed, mad2 + and mad3 + were among the 20 % of protein-coding genes with the lowest 180 CSCg values (Fig 3A,B). This result was independent of which large-scale mRNA half-life data or 181 which correlation parameter was used (Fig S3C,D). These results raise the interesting possibility 182 that codon usage in mad2 + and mad3 + contributes to their short mRNA half-life. The mad1 + gene 183 showed different characteristics, which we will discuss below. 184

Codon-optimization increases the mRNA concentration of mad2 + and mad3 + 185
To test if codon usage contributes to the short mRNA half-lives, we codon-optimized mad2 and 186 mad3 and inserted the codon-optimized sequence at the respective endogenous locus (Fig 3C;  187 S3B,F). The GFP tag, which remained unchanged, mitigated but did not abolish the effect of the 188 codon optimization on the CSCg value of the fusion genes ( Fig S3B). An increase in mRNA half-189 life should result in an increased steady-state mRNA number if synthesis was unchanged. Indeed,190 we found an increased mRNA number for codon-optimized mad2 and mad3 compared to the wild-191 type gene (Fig 3D). Cytoplasmic mRNAs showed a 27 % increase (Fig EV3). For mad2 + , the 192 increase was restricted to the cytoplasm and not observed in the nucleus, strongly suggesting 193 stabilization of the mRNA (Fig EV3). deletion of ste13 + significantly increased mad2 + and mad3 + mRNA half-lives-from about 8 min to 198 14 min for mad2 + , and 5 min to 10 min for mad3 + (Fig 3E, EV2C). This indicates that mad2 + and 199 mad3 + mRNA are subject to Ste13-mediated degradation. The steady-state mRNA numbers were 200 not greatly affected by ste13 + deletion (Fig 3D, EV2B, EV3). This is consistent with a global 201 'buffering' of mRNA  ). This may explain why mRNA numbers increased after codon-optimization, but not after 206 ste13 + deletion. Overall, our results support the hypothesis that non-optimal codons in mad2 + and 207 mad3 + contribute to the short mRNA half-life of these genes. 208

Codon-optimization, but not ste13 + deletion, increases the protein concentration of Mad2 209
and Mad3 210 To ask whether the consequences of codon-optimization propagate to the protein level, we 211 quantified Mad2-and Mad3-GFP protein expressed from the wild-type or codon-optimized genes. 212 Both immunoblotting and fluorescence microscopy showed an increase in protein concentration 213 after codon-optimization (Fig 4), which can partly be explained by the increase in mRNA (Fig 3) 214 and might be enhanced by an increased translation efficiency. In contrast, the Mad2 and Mad3 215 protein concentrations in ste13D cells remained largely stable when analyzed by immunoblotting 216 (Fig 4B,C), consistent with the RNA results ( Fig 3D). Altogether, these data support that codon 217 usage bias towards non-optimal codons in mad2 + and mad3 + lowers their protein concentration but 218 supports a short mRNA half-life, thereby establishing a gene expression pattern that lowers cell-to-219 cell variability. 220 Mad1 + expression regulation differs from that of mad2 + and mad3 + 221 The mad1 + gene shares a short mRNA half-life with mad2 + and mad3 + (Fig 2D). Different from 222 mad2 + and mad3 + , though, mad1 + has a higher fraction of optimal codons and a CSCg value above 223 the median of all protein-coding S. pombe genes (Fig 3A,B; S3A,B). This was surprising because 224 we expected similar features within the SAC network. Unlike for mad2 + and mad3 + , the mad1 225 mRNA number did not increase after codon-optimization, but rather decreased slightly (Fig 5A,B; 226 EV5). A second codon-optimized mad1 whose sequence was considerably different from the first 227 (77% nucleotide identity; Fig S3F; Table S3) showed the same trend (Fig EV4A, EV5). Similar to 228 mad2 + and mad3 + , mad1 + mRNA half-life was still prolonged in ste13D cells (from 6 min to 10 min; 229 Fig 5C), but unlike for mad2 + and mad3 + not reaching statistical significance ( Fig EV4E). Thus, the 230 short mad1 + mRNA half-life is less dependent on codon usage bias and Ste13, and hence, different 231 modes of regulation bring about the short mRNA half-life of these SAC genes. 232 The ecm33 + control mRNA was strongly stabilized in ste13-deleted cells (Fig 5C, EV4E), despite 233 a high fraction of optimal codons in ecm33 + (Fig 5D). This highlights that-despite some overall 234 correlation-the relationships between codon-optimality, mRNA half-life, and susceptibility to 235 ste13 + deletion are far from predictable ( Fig EV4F) (He et al, 2018). It is worth noting that Ecm33 236 is a membrane-binding protein, which, extrapolating from work in other organisms (Jungfleisch et 237 al, 2017;Weber et al, 2020), may explain the Ste13-mediated mRNA destabilization. 238

Codon-optimization of mad1 + decreases its protein concentration 239
Unlike Mad2-and Mad3-GFP, whose protein concentration increased after codon optimization, 240 that of Mad1-GFP decreased, both by immunoblotting and fluorescence microscopy (Fig 6). Mad1 241 protein formed from the codon-optimized mRNA had a similar stability to that formed from wild-type 242 mRNA ( Fig S4A,B), and still bound Mad2 (Fig S4C). The reduction, rather than increase, in protein 243 concentration after codon-optimization of mad1 + corroborates that the codon usage pattern of 244 mad1 + serves a different purpose than that of mad2 + and mad3 + . Deletion of ste13 + had hardly any 245 influence on the Mad1 protein concentration (Fig 6B,C), consistent with the largely unchanged 246 mRNA concentration (Fig 5B). 247 We previously found that SAC function was well preserved when Mad1 levels were lowered to 248 30 % (Heinrich et al., 2013). Consistently, we did not observe an obvious growth defect when cells 249 expressing codon-optimized mad1 were grown in presence of the microtubule-drug benomyl (Fig  250   EV1B), and we did not observe a SAC defect in a live-cell imaging assay where microtubules were 251 depolymerized (Fig S4D,E). To test SAC function in a more sensitive assay, we deleted the gene 252 for the microtubule-interacting protein Alp7 (Sato et al, 2003). This also activates the SAC, but less 253 robustly than microtubule-depolymerization. Using this assay, cells expressing codon-optimized 254 mad1 tended to exit mitosis more quickly than cells expressing wild-type mad1 + (Fig 6F, S4F). The 255 difference did not reach the level of statistical significance but was reproducible with independent 256 strains. This suggests that synonymous codon changes, without any change in the protein 257 sequence, can impair SAC function. 258

Upstream and downstream sequences of mad1 + are insufficient for proper expression 259
The lower mRNA concentrationa after mad1 codon-optimization (Fig 5B, EV4A) suggested that 260 the concentration of mad1 + mRNA is not purely determined by regulatory sequences upstream and 261 downstream of the coding sequence. This is supported by our observation that merely fusing GFP 262 to mad1 + , without altering surrounding sequences, increases its mRNA number (Fig 1E). Further 263 supporting this notion, but rather surprisingly, we found that replacing the mad1 + coding sequence 264 with GFP produced neither significant amounts of mRNA nor protein (Fig S5A,B). This again 265 contrasted with the mad2 + and mad3 + genes, which produced comparable amounts of mRNA and 266 protein when the original coding sequence was replaced with GFP ( Fig S5C,D). Hence, the 267 sequences surrounding the mad1 + coding sequence are insufficient to establish mad1 + -like 268 expression, and contributions from the coding sequence are required. Preserving the first 66 or 108 269 base pairs of mad1 + partly rescued both mRNA and protein levels but not completely (Fig S5A,B). 270 While this suggests that the 5' region of the mad1 + coding sequence carries signals that are 271 important for mRNA synthesis or stabilization, some other genes contain sequences that can 272 compensate. Introducing an nmt1 + -GFP fusion gene or fusions between S. cerevisiae GCN4 and 273 N-terminally truncated versions of S. pombe mad1 + (Heinrich et al, 2014) allowed for expression 274 from the mad1 + locus (Fig S5A,B). What these genes share, that GFP does not, remains unclear. 275 Altogether, these results indicate that mad1 + expression has some unique aspects: mad1 + uses 276 a different mode for reducing mRNA half-life than mad2 + or mad3 + , and its coding sequence carries 277 elements that help transcribe, stabilize or translate RNA. 278

Mad1 homodimers assemble co-translationally 279
We considered whether mad1 may have a certain codon usage pattern to facilitate protein 280 production or complex formation (Liu et al, 2021). Mad1 forms a homodimer through a long N-281 terminal coiled-coil (Piano et al., 2021;Sironi et al., 2002), but-except in a very recent genome-282 wide study (Bertolini et al, 2021)-how this homodimer forms has not been examined. If formation 283 was co-translational rather than post-translational, this may require a certain pattern of codon 284 usage for proper complex formation. To assess dimer formation, we examined cells expressing 285 both tagged and untagged Mad1. If Mad1 dimer formation was post-translational, it should be 286 possible to observe interactions between tagged and untagged Mad1. However, in haploid strains 287 expressing a C-terminally GFP-tagged and an untagged mad1 + gene, a GFP immunoprecipitation 288 almost exclusively precipitated Mad1-GFP, but not untagged Mad1 (Fig 7A). In contrast, a Mad1 289 immunoprecipitation precipitated Mad1-GFP and Mad1 in approximately the same ratio in which 290 they were present in the extract. These experiments used a monomeric version of GFP. Thus, it is 291 unlikely that this pattern is driven by dimerization of GFP. With two versions of Mad1 being 292 expressed, a slight bias towards the form that is being pulled down would be expected even when 293 heterodimers between these forms were generated with equal likelihood as homodimers 294 (Fig EV6A). At a 1:1 ratio of the isoforms in the extract, a 2:1 ratio would be expected in an 295 immunoprecipitation or pull-down. However, the bias that we observed always exceeded the 296 expected bias, usually vastly (Fig 7, EV6). Hence, we propose that Mad1 forms homodimers 297 between isoforms more efficiently than heterodimers. This is most easily explained by co-298 translational assembly of Mad1 dimers from the nascent chains of two ribosomes translating mad1 + 299 from the same mRNA molecule (Fig 7B). 300 We further corroborated this finding by using diploid strains expressing Mad1-GFP and Mad1-301 Strep from the two endogenous loci. Again, a GFP-immunoprecipitation isolated Mad1-GFP but 302 very little Mad1-Strep, whereas a Strep pull-down isolated Mad1-Strep but very little Mad1-GFP 303 ( Fig 7C, EV6B). We obtained similar results after in vitro translation of Mad1 ( Fig EV6C): when 304 Mad1-GFP and Mad1-flag-His were co-translated in a rabbit reticulocyte lysate, a subsequent GFP 305 immunoprecipitation isolated very little Mad1-flag-His, and a His-pull-down isolated very little Mad1-306 GFP. Heterodimerization between C-terminal Mad1 fragments has previously been reported in an 307 in vitro translation (Kim et al., 2012). However, in our experiments, even C-terminal fragments 308 showed a strong bias towards the form that was being precipitated, both in yeast extracts and after 309 in vitro translation (Fig S6). To exclude that heterodimer formation between Mad1-GFP and 310 untagged Mad1 was unphysiologically prevented by the large GFP tag, we tested a combination of 311 Mad1-flag-His and untagged Mad1 in an in vitro translation. Again, His pull-down almost exclusively 312 isolated Mad1-flag-His, whereas a Mad1 immunoprecipitation isolated both forms in approximately 313 the same ratio in which they were present in the extract (Fig 7D). 314 To further test the idea that Mad1 dimer assembly occurs on a single mRNA molecule (Fig 7B), 315 we examined mad1 mRNA. Consistent with few heterodimers on the protein level, we did not 316 observe colocalization between two different mad1 isoform mRNAs present in the same cell (Fig  317   7E). Intensity measurements of mRNA FISH spots suggested the presence of single mRNAs, not 318 mRNA doublets, when both untagged mad1 + and mad1 + -GFP were expressed (Fig 7E, left; EV6D). 319 Further supporting this finding, the number of mRNA spots for a given isoform was identical in the 320 absence or presence of another isoform (Fig 7E, right), indicating that the isoforms do not co-321 localize. We additionally tested the possibility that mRNAs of the same isoform may co-localize by 322 comparing FISH spot intensities with probes against GFP between mad1 + -GFP mRNA and mad3 + -323 GFP mRNA (the latter coding for Mad3 monomers). We did not find any difference in spot intensity 324 ( Fig 7F). Hence, we conclude that mad1 + mRNAs rarely, if ever, co-localize, and we favour the idea 325 that Mad1 homodimers emerge from two ribosomes co-translating a single mRNA (Fig 7B). 326 The fact that Mad1 homodimers form co-translationally is consistent with the idea that 327 synonymous codon changes may subtly impair complex formation and therefore translation 328 efficiency and mRNA stability. Overall, these results suggest that codon usage bias within mad1 + 329 contributes to maintaining proper mRNA and protein levels, possibly by supporting Mad1 folding 330 and dimerization. 331

332
Proteins are the workhorses of cells. The deployment of this workhorse army is controlled by 333 regulatory elements encoded in DNA that are still incompletely understood. The spindle assembly 334 checkpoint is sensitive to expression changes, and we therefore asked which features of gene 335 expression may be important for its proper function. Our results suggest that a combination of short 336 mRNA half-lives and long protein half-lives is important to keep protein variability low. We also find 337 that-despite their closely shared function-mad1 + differs in its expression features from mad2 + 338 and mad3 + . The coding sequences of mad2 + and mad3 + contribute to the short mRNA half-life of 339 these genes, whereas that of mad1 + contributes to maintaining mRNA ( Fig S5) and protein levels 340 (Fig 6). We propose that the choice of synonymous codons in mad1 + is optimized for the formation 341 of the Mad1 homodimer and, ultimately, the Mad1/Mad2 complex. 342

Short mRNA half-life of constitutively expressed SAC genes favours low noise 343
The short mRNA half-lives of mad1 + , mad2 + and mad3 + , along with their long protein half-lives, 344 can explain the low protein noise of SAC genes despite low and variable mRNA numbers (Fig 1,2 . At least two elements seem to play a role for 363 mad2 + and mad3 + : our data suggest that the mRNA half-lives are shortened by a high fraction of 364 non-optimal codons (Fig 3); in addition, the mad2 + and mad3 + 3' UTRs contain sequence motifs 365 that are associated with a short mRNA half-life (Eser et al., 2016). We previously found higher 366 mRNA numbers after traditional tagging, which changed the 3' UTR to that of a highly expressed 367 gene (Heinrich et al., 2013), suggesting that the predicted motifs in the 3' UTR may indeed be 368 functional. For mad1 + , in contrast, overall codon usage bias seems to play a lesser role (Fig 5), and 369 the mad1 + 3' UTR does not contain reported motifs implicated in half-life shortening (Eser et al., 370 2016). We suspect that other elements that influence translation efficiency may be important. 371 Generally, less efficiently translated mRNAs are less stable (Hanson & Coller, 2018), and mad1 + 372 seems to be translated less efficiently than mad2 + or mad3 + (Rubio et al, 2020). 373 However, we suggest that the capacity of an N-terminal Mad1 fragment to dimerize would need to 385 be based on assessing self-association rather than assessing association with Mad1 expressed 386 from a different locus. Of note, C-terminal Mad1 fragments also dimerize, possibly post-387 translationally (Kim et al., 2012), although our own experiments still suggest a preference of 388 homodimerization ( Fig S6). 389 While we propose that assembly of the Mad1 homodimer occurs co-translationally, the 390 assembly of the Mad1/Mad2 tetramer does not occur in synchronous co-translational fashion, since 391 the mRNAs for mad1 + and mad2 + do not co-localize in the cytoplasm (Fig 1). This leaves open the 392 possibility of post-translational assembly of the tetramer or of asynchronous co-translational 393 assembly, where one protein is already fully formed and binds the other that is being translated 394 the idea that the tetramer assembles while one of the proteins is being translated, and it will be 398 interesting to test whether the mad1 + mRNA binds Mad2 protein or vice versa to facilitate such an 399 assembly. It will also be interesting to examine whether different eukaryotes use the same 400 assembly pathway for the highly conserved Mad1/Mad2 complex. 401

Potential SAC malfunction from synonymous mutations 402
Overall, our data suggest that the coding sequences of mad1 + , mad2 + and mad3 + modulate 403 gene expression. Hence, even synonymous mutations carry some risk of impairing the SAC. We 404 suspect that mad1 is most susceptible to single synonymous substitutions, given the need for co-405 translational homodimer assembly (Fig 7), which may be facilitated by controlling the speed of 406 ribosome movement (Liu et al., 2021). In S. pombe, a cluster of non-optimal codons follows the 407 coiled-coil region of mad1 + (Fig S3E, S7), which may ensure that the N-terminal coiled-coil is fully 408 formed before the remainder of Mad1 is translated. 409 It will be interesting to test whether synonymous mutations found in cancer samples can

Experimental Models
Schizosaccharomyces pombe strains This study Table S1 Saccharomyces cerevisiae strain Nick Buchler, NC State University, USA Table S1 Recombinant DNA Indicate species for genes and proteins when appropriate sgRNA sequences This study Table S2 Codon-optimized mad1, mad2, and mad3 This study Table S3 PCR fragments for in vitro transcription This study  Alanine 206 to Arginine (A206R), which is expected to reduce dimerization (Zacharias et al, 2002). 434 Codon-optimization used proprietary algorithms by two different companies, and sequences are 435 listed in Table S3. The strain with two differently tagged versions of mad1 + has mad1 + -ymEGFP 436 along with 110 bp upstream and 164 bp downstream of the coding sequence integrated between 437 the leu1 + and apc10 + gene. 438

Cycloheximide treatment for determination of protein half-lives 450
Cells were grown in EMM (plus supplements required for auxotrophic mutations) to a final 451 concentration of around 1 x 10 7 cells/mL. Cultures were diluted to 8 x 10 6 cells/mL, transferred to 452 a 30°C water bath for 30 minutes and a sample was taken prior to addition of cycloheximide (CHX) 453 to a final concentration of 1 mg/mL. Cells were collected at specified timepoints, spun down at 454 980 rcf and frozen in liquid nitrogen before processing. 455

In vitro transcription and translation 456
The T7 promoter was appended 5' of the mad1 transcription start site by PCR. Precise 457 sequences are available in Table S6. Full-length mad1 + was amplified from cDNA generated using 458 the SuperScript IV First Strand Synthesis System (ThermoFisher). Mad1 fragments 3' of the intron 459 were amplified from genomic DNA. PCR fragments were purified using the Wizard SV Gel and 460 PCR Clean-Up System (Promega). In vitro transcription was carried out with the HiScribe T7 ARCA 461 mRNA Kit (with tailing) (New England Biolabs) using between 25 and 70 ng/µL template DNA. 462 Reactions were run at 32ºC or 37ºC for 2 hours. RNA was purified using the Monarch RNA Cleanup 463 Kit (New England Biolabs). RNAs were mixed and diluted as required before adding them to rabbit or anti-rabbit conjugated to HRP (Dianova) and quantified by chemiluminescence using 526 SuperSignal West Dura ECL (ThermoFisher) and imaged on a Bio-Rad Gel Doc system. 527 Chemiluminescence signals were quantified on non-saturated images using Image Lab software 528 (Bio-Rad). Measurements from a reference dilution series were used to create a standard curve, 529 which was used to determine the concentration of sample relative to the reference. Membranes 530 with radioactive proteins were dried and exposed to a phosphorscreen (GE Healthcare), which was 531 read-out on a Typhoon phosphorimager (GE Healthcare/Cytiva). 532

Quantification of GFP fusion proteins in single cells (3D segmentation) 533
To quantify GFP fusion proteins in single cells, cells were grown in EMM (plus supplements that 534 were required for auxotrophic mutations) at 30ºC to a final concentration of 6-9 x 10 6 cells/mL. 535 Cultures of GFP-positive and GFP-negative cells were mixed at a 1:1 ratio to a final concentration 536 of 2.5-6.0 x 10 6 cells/mL and incubated for 30 minutes at 30ºC. To ensure a uniform and flat 537 imaging plane, cells were loaded into a Y04C microfluidics trapping plate (Millipore Sigma) and 538 incubated inside a climate-controlled microscope chamber for 2 hours at 30ºC with constant flow 539 of fresh media. Imaging was performed on a DeltaVision Elite system equipped with a PCO edge 540 sCMOS camera and an Olympus 60x/1.42 Plan APO oil objective. Images were acquired for 541 ymEGFP, tdTomato and brightfield as 7.2 µm or 10 µm stacks with images separated by 0.1 µm. 542 The acquired image area was 1024 x 1024 pixels with 1 x 1 binning. All images were deconvolved 543 using SoftWoRx software. To correct for uneven illumination, deconvolved fluorescence images 544 were flatfielded individually for each channel using a custom FIJI script (Baybay et al., 2020). 545 The Pomegranate image analysis pipeline (Baybay et al., 2020) was used to segment nuclei 546 (using TetR-tdTomato-NLS) and whole cells (using brightfield signal and spherical extrusion of the 547 midplane segmentation) (Fig S1A). We corrected for chromatic aberration and for stretching of 548 distances in the Z direction (Baybay et al., 2020). Further analysis was conducted in R (R-Core-549 Team, 2020) and figures were produced using the package ggplot2 (Wickham, 2016). 550 Only information from mono-nucleated cells for which both the whole cell and the nucleus had 551 been segmented was retained. Cells were excluded if one or more of the following conditions were 552 met: the nuclear segmentation protruded beyond the three-dimensional bounds of the cell; whole-553 cell segmentation was cut-off by more than two slices because insufficient slices in Z had been 554 recorded; cell was at the image edge and incompletely recorded; the nucleus had an aspect ratio 555 (diameter in Z to diameter in XY) of less than 0.8 or more than 1.2; cell volume was in the 0.1 st or 556 99.9 th percentile. Cells with or without GFP signal were distinguished by k-means (k = 2) clustering 557 ( Fig S1D-F), except for Nmt1-GFP, where the threshold for each image was set manually. One 558 image, where the autofluorescence of GFP-negative cells deviated by more than three standard 559 deviations from that of other images, was excluded. One additional image, where the cells had 560 visibly moved during acquisition, was also excluded. 561 To subtract autofluorescence and other background, we averaged the fluorescence intensity 562 per cell or nuclear volume for GFP-negative cells in an image and subtracted that value from the 563 fluorescence intensity per cell or nuclear volume of each GFP-positive cell in the image. For a rough 564 estimate of absolute concentration in nanomolar, we used our previous estimate of about 70 nM 565 Mad3-GFP in the cell nucleus (Heinrich et al., 2013) and normalized all background-subtracted 566 data to this value. 567 Even after background subtraction, we observed some variation of mean intensities between 568 single images ( Fig S1F) and we could not distinguish whether these differences were a 569 consequence of sampling or came from conditions on the microscope stage while recording the 570 image. We therefore opted to determine the coefficient of variation (CV = standard deviation / 571 mean) for each protein not across all images, but instead for each image separately; Fig

Quantification of GFP in single cells (2D segmentation and projection) 584
For experiments evaluating fluorescence signals after replacing the coding sequences of mad1 + , 585 mad2 + , and mad3 + (Fig S5), quantification was performed on projections, using 2D segmentation 586 of cells. Cells were grown in minimal medium, collected by centrifugation from liquid cultures, 587 mounted in medium on a slide, and brightfield and fluorescence images were collected immediately 588 at room temperature. At least two slides were prepared and imaged for each strain. Images were 589 recorded on a Zeiss AxioImager M1, using Xcite Fire LED illumination (Excelitas), a Zeiss Plan-590 Apochromat 63x/1.40 Oil DIC objective and an ORCA-Flash4.0LT sCMOS camera (Hamamatsu) 591 with Z sections spaced by 0.2 µm. 592 Cells were segmented based on an in-focus brightfield image using YeaZ (Dietler et al, 2020). For analyses with cytoplasmic or nuclear RNA counts (except the mad1/mad2 colocalization 634 experiment; Fig 1F), nuclei were re-segmented in three dimensions using a FIJI macro adapted 635 from https://github.com/haesleinhuepf/cca_benchmarking (Robert Haase, MPI-CBG, Dresden). 636 Analysis was limited to cells whose nuclei were entirely contained within the image stack. To measure co-localization of mad1 and mad2 mRNA, a two-color FISH experiment was 642 performed targeting mad1 with gene-specific probes and mad2 + -ymEGFP with ymEGFP probes. 643 The three-dimensional coordinates of each spot were recorded and corrected for relative chromatic 644 aberration in Z. Distances were then calculated from each mRNA to its nearest neighbor of the 645 other species within the same cell. In order to determine a distance cut-off for classifying RNA 646 molecules as either co-localized or unpaired, the same two probe sets were used in another two-647 color FISH experiment in which both probes targeted mad1 + -ymEGFP. Nearest-neighbor distances 648 were calculated in the same way, and the distribution of these distances was used to determine the 649 co-localization distance cut-off value. This cut-off was applied to the distances in the original 650 experiment to classify each mad1 or mad2 mRNA molecule as co-localized or unpaired. 651 To test if mad1 mRNA forms dimers, we used RNA FISH experiments to measure spot 652 intensities and counts of RNA in the cytoplasm in strains with the following genotypes: (1) untagged 653 mad1 + expressed from the endogenous locus, (2) untagged mad1 + expressed from the 654 endogenous locus and mad1 + -ymEGFP expressed from the exogenous leu1 + locus, (3) 655 endogenous mad1 + deleted and mad1 + -ymEGFP expressed from the exogenous leu1 + locus, and 656 (4) mad1 + -ymEGFP and mad3 + -ymEGFP expressed from the endogenous loci. All samples were 657 hybridized with a combination of mad1-and ymEGFP-targeting probes in two-color FISH 658 experiments. FISH probe spots were quantified separately for each imaging channel. Colocalized 659 spots of different colors were then paired using the same colocalization method as described for 660 mad1/mad2 colocalization above. Intensity analysis used two measurements of spot intensity 661 To test for differences in mean RNA levels between genotypes, the categorical fixed effect 677 variable genotype was added to the model. The interaction between cell length and genotype was 678 also included if a likelihood ratio test comparing models with and without the interaction term 679 showed that it improved the model's ability to explain the data significantly (p < 0.05). P-values for 680 the likelihood ratio test were obtained both by comparing the test statistic to a chi-square distribution 681 and generating a null distribution by bootstrapping (1000 replicates) using the 'PBmodcomp' 682 function from the package pbkrtest. In all cases, the results were consistent between the two 683 methods. Only a few models required the interaction term: comparison of wild-type with codon-684 optimized mad2 (whole cell, cytoplasmic and nuclear RNA counts), comparison of wild-type with 685 codon-optimized mad2 + ste13D (cytoplasmic RNA only) and comparison of untagged and GFP-686 tagged mad1. 687 Genotype coefficients and corresponding 95 % confidence intervals presented in the paper 688 were exponentiated (e^) and represent the ratio of expected RNA levels between the two genotypes 689 in each comparison. RNA levels were considered significantly different between genotypes if the 690 exponentiated confidence interval excluded 1. 691

Assay for spindle assembly checkpoint function using nda3-KM311 692
Strains expressing the tubulin mutant nda3-KM311 were grown in EMM (plus supplements 693 required for auxotrophic mutations) at 30ºC to a concentration of 0.5-1.0 x 10 7 cells/mL. Cells were 694 diluted with EMM to a final concentration of 7.5 x 10 5 or 1.5 x 10 6 cells/mL. 300 µL of each strain 695 were loaded into a lectin-coated Ibidi µ-Slide glass-bottom chamber and incubated about one hour 696 at 16ºC on the microscope stage prior to imaging. Cells were imaged at 16ºC on a DeltaVision Elite 697 system with a PCO edge sCMOS camera (PCO) and an Olympus 60x/1.42 Plan APO oil objective 698 and EMBL environmental chamber. Images were acquired every 5 minutes for GFP and mCherry 699 over an 18-hour period using an 'optical axis integration' (sum projection) over a 3.2 µm Z-distance. and dot-like GFP signals were therefore measured in the direct vicinity to Plo1-mCherry. An area 703 of the same size for each cell was used to capture the kinetochore signal and was also used to 704 measure the intensity in the nucleoplasm for background subtraction. GFP intensities from multiple 705 cells were aligned to the time point of Plo1-mCherry appearance and averaged for each timepoint. 706 Assay for spindle assembly checkpoint function using alp7D 707 Cells were grown in EMM at 25ºC to a concentration of 0.5-1.0 x 10 7 cells/mL, diluted to 1.5 x 708 10 6 cells/mL, and 300 µL of this dilution were loaded into a lectin-coated ibidi µ-Slide glass-bottom 709 chamber. Cells were incubated on the microscope stage at 30ºC for 35 minutes before imaging. 710 Images were acquired at 30ºC every 55 seconds to 1.5 minutes for 2-3.5 hours using an 'optical 711 axis integration' (sum projection) over a 3.6 µm Z-distance. Cells were segmented based on the 712 brightfield image using YeaZ (Dietler et al., 2020). All pixels within the cell were quantified and the 713 0.1 st percentile value was subtracted from the 99.9 th percentile value to obtain the "maximal 714 intensity". The localization of Plo1-tdTomato to spindle-pole bodies or Mad1-GFP to kinetochores 715 ( Fig S4F) is reflected in higher maximal intensities. Time in mitosis was determined from a custom 716 Matlab script that detects strong increases and decreases in signal. Some cells could not be 717 analyzed in an automated fashion (e.g. due to overlapping other cells) and were analyzed manually. 718 The analysis mode is reported in the source data. 719 where "target" is the mRNA of interest, "reference" is the reference gene, "sample" is the sample 745 of interest and "control" is the control sample being normalized to. The denominator is the geometric 746 mean of the reference genes (act1 and cdc2), and efficiencies were estimated from the slopes of 747 four-step, serial 1:5 dilution standard curves. by randomizing the codon order 10,000 times. Observed values that deviated by more than 2 856 standard deviations from the null mean were marked with filled circles. 857

Gene expression model -simulations and theoretical predictions 858
Protein noise predictions (Fig 2C; S2B,C) were made by assuming a constitutively active 859 promoter, and only considering stochastic mRNA and protein synthesis and degradation and 860 ignoring cell growth and division. The coefficient of variation (CV = standard deviation / mean) for 861 protein is calculated as: 862 where P is the protein number per cell, M the mRNA number per cell, kdegM the mRNA degradation 864 rate and kdegP the protein degradation rate (Swain, 2004). For the predictions in S2C, we assumed 865 a protein number of 6,000 per cell, mRNA numbers of 1 to 1,000, and we varied RNA degradation 866 rate in a range corresponding to half-lives of 1 to 60 minutes, and protein degradation rate in a 867 range corresponding to half-lives of 15 to 600 minutes, which we consider a physiologically 868 plausible range. Predictions were excluded when mRNA synthesis or protein synthesis rates 869 became unrealistically high. We assumed this to be the case when mRNA synthesis rate was higher 870 than 25 minute -1 or protein synthesis rate higher than 20 mRNA -1 minute -1 . Assuming a gene with 871 characteristics similar to a SAC gene (protein number = 6,000, mRNA number = 3.5, protein half-872 life = 360 minutes, mRNA half-life = 4 minutes) yields a CV prediction of 0.0575. In the figure, we 873 labelled CV predictions less than 0.06 in light grey (low noise) and those equal or higher than 0.06 874 in dark grey (high noise). 875 The stochastic simulation of mRNA and protein numbers (Fig 2B)  General linear mixed models and generalized linear mixed models were fit using the functions 886 'lmer' and 'glmer,' respectively, from the lme4 package. Default function settings were used except 887 for the optimizer in 'glmer,' which was set to 'bobyqa.' Bootstrapping using the function 'bootMer' 888 (10,000 replicates; lme4 package) was used to obtain 95 % confidence intervals for fixed effects 889 model coefficients and 95 % confidence bands for predicted regression curves. Nonlinear least 890 squares regression models, Wilcoxon rank sum tests and t-tests were performed using 'nls', 891 'wilcox.test' and 't.test,' respectively, from the package stats. Poisson distributions were fit to data 892 frequency distributions using 'fitdistr' from the package MASS.  numbers per cell using FISH probes against the endogenous genes and using either strains 1197 expressing the GFP-tagged gene or the endogenous, untagged gene. The difference for mad1 is 1198 statistically significant, that for mad2 is not (Fig EV1E). A lower mRNA number for untagged mad1 1199 was also observed in a different strain. (F) Co-staining by smFISH using probes against mad1 and 1200 GFP either in a strain expressing mad1-GFP as a positive control or in a strain expressing wild-1201 type mad1 and mad2-GFP. Cytoplasmic mad1 (green) or GFP mRNA spots (magenta) were 1202 quantified as co-localizing or not with the respective other. For the mad1-GFP strain, 544 cells and 1203 a total of 1,641 mad1 spots and 1,839 GFP spots were analyzed; 48 cells were not considered 1204 since they did not contain at least one spot of each type in the cytoplasm. For the mad1 mad2-GFP 1205 strain, 571 cells and a total of 1,107 mad1 spots and 1,537 GFP spots were analyzed; 158 cells 1206 were not considered since they did not contain at least one spot of each type in the cytoplasm. 1207   Frequency distribution of mRNA numbers per cell using FISH probes against the endogenous genes and using either strains expressing the GFP-tagged gene or the endogenous, untagged gene. The difference for mad1 is statistically significant, that for mad2 is not (Fig EV1E). A lower mRNA number for untagged mad1 was also observed in a different strain. (F) Co-staining by smFISH using probes against mad1 and GFP either in a strain expressing mad1-GFP as a positive control or in a strain expressing wild-type mad1 and mad2-GFP. Cytoplasmic mad1 (green) or GFP mRNA spots (magenta) were quantified as co-localizing or not with the respective other. For the mad1-GFP strain, 544 cells and a total of 1,641 mad1 spots and 1,839 GFP spots were analyzed; 48 cells were not considered since they did not contain at least one spot of each type in the cytoplasm. For the mad1 mad2-GFP strain, 571 cells and a total of 1,107 mad1 spots and 1,537 GFP spots were analyzed; 158 cells were not considered since they did not contain at least one spot of each type in the cytoplasm.      The alp7 + gene was deleted to increase the likelihood of spindle assembly checkpoint activation. Localization of Plo1-tdTomato to spindle pole bodies was used to judge entry into and exit from mitosis (also see Figure S4). Exp1: n = 73 (WT) and 94 cells (co); Exp2: n = 126 (WT) and 152 cells (co). Difference between WT and co: p = 0.14 (Exp1) and 0.15 (Exp2) by Kolmogorov-Smirnov test.