Free-living cells are frequently challenged by changes in their environment, which results in a need to allocate their resources accordingly. One of the most important cellular processes to manage is the production of ribosomes10,11. Ribosome synthesis, although essential for growth and division, consumes a substantial fraction of cellular resources12. To assemble into functional complexes, the coordinated expression of tens of ribosomal RNA and ribosomal protein (RP) genes is necessary13. In the budding yeast, which underwent a whole-genome duplication followed by the loss or divergence of most duplicated genes14, RPs are largely conserved as duplicates (ohnologues) with high sequence identity2. Why evolution may have favoured the retention of RP ohnologues remains an open question; possible reasons include increased gene dosage, genetic robustness towards mutations, and distinct biological roles or differential regulation of the duplicated genes2,14,15,16,17,18,19,20,21,22. Recently, it has been suggested that duplicated transcription factors, which are also preferentially retained, might be an evolutionary solution to a trade-off between tight regulation of expression level and responsiveness to environmental changes, such as stress9. In this scenario, one ohnologue provides a precise level of protein expression that is required irrespective of the environment, whereas the other one generates population heterogeneity, allowing for a flexible response when the environment changes. While the phenotypic effects of deleting the less-expressed copy of duplicated ribosomal genes are usually observed only under stress17, it is unclear whether this paradigm applies to duplicated RPs.

The duplicated RP genes in yeast are highly enriched for introns. Although fewer than 5% of all yeast genes contain an intron23, 94 out of 118 (80%) RP genes with an ohnologue do3. Differential expression of RPs through intronic regulation could expand their functional repertoire, even within the low sequence divergence; hence, introns may be involved in the evolutionary conservation of RP ohnologues5,18. Yeast introns in RP genes affect the expression level of the corresponding gene and, in some instances, that of its ohnologue6,18,24,25,26. Concerted intron retention in transcripts occurs in response to stress, suggesting that splicing regulation has a functional role in yeast27. Such a role is supported by the observation that intron deletion in the budding yeast results in growth alterations under conditions including drug treatment, starvation or population saturation4,5,6. Introns are thus clearly relevant for adapting to environmental changes; nevertheless, the determinants and functional outcomes of intron-mediated responses remain unclear.

As ribosome synthesis and protein translation are tightly coupled to growth rate28, it is crucial to dissect cellular responses that are specific to particular stressors from responses to non-specific growth-inhibition when studying the effects of stress on ribosomal function. To this end, we previously established a method termed isogrowth profiling7. It is based on exposing cells to a combination of two drugs in an antiparallel concentration gradient discretized into separate liquid cultures, while keeping the overall inhibition constant. Analysing the transcriptome along the growth isobole then enables responses that are specific to each drug, or to their combination, to be distinguished from the general stress and growth inhibition responses.

Here, to investigate the role of introns and RPs in stress response, we extend isogrowth profiling from RNA sequencing (RNA-seq) to single-cell protein-level measurements. We found that lithium chloride (LiCl) induces extensive retention of introns in RP transcripts and an intron-dependent bimodal expression of Rps22B, a component of the small ribosomal subunit. The two subpopulations exhibit differential fitness under conditions of starvation and recovery. We show that whereas yeast in standard rich laboratory growth medium do not exhibit Rps22B bimodality, cells in medium with a high concentration of glucose do as they approach stationarity. Together, these results suggest that yeast has evolved an intron-mediated regulation mechanism of Rps22B to cope with uncertainty regarding the possible replenishment of nutrients at the end of exponential growth in a high-glucose environment.

LiCl inhibits the splicing of RP transcripts

To study the effect of stress on the regulation of ribosomes, we started by investigating the transcriptional response to two growth inhibitors—LiCl, a pleiotropic drug that induces cationic and osmotic stress29; and cycloheximide, an inhibitor of the large ribosomal subunit30. As regulation of ribosomes is sensitive to changes in growth rate and medium composition28, we used an antiparallel concentration gradient of the two drugs applied in separate liquid cultures7, which consistently lowered the relative growth rate to near 50% (Fig. 1a, b). We inoculated the drug-free control at a lower cell density so that all samples reach a comparable cell density at the time of collection (Fig. 1b). To determine intron retention, RNA was extracted from the samples, ribo-depleted and sequenced (Methods). This procedure on the series of drug mixtures enabled us to observe the specific effects of stressors on intron retention in a way that is not confounded by growth-rate or cell-density effects.

Fig. 1: LiCl-induced stress inhibits the mRNA splicing of ribosomal proteins.
figure 1

a, Schematic of the isogrowth profiling that was used to characterize changes of splicing in mRNA. The coloured dots indicate the points in the two-drug gradient used to extract total RNA that was subsequently ribo-depleted and sequenced. IC50, half-maximum inhibitory concentration. b, Growth curves of samples used for RNA-seq. The no-drug control was inoculated at a lower density to reach a comparable optical density at the time of extraction. The shaded area denotes measurements used to determine the exponential growth rates (right). Colour code as in a. OD600 nm, optical density at 600 nm. c, Intron retention rate (intron read density/CDS read density; Methods) for all nuclear introns when treated with cycloheximide or LiCl, normalized to no-drug control (y axis) versus non-normalized (x axis). The most significantly enriched Gene Ontology (GO) cellular component gene set for the LiCl sample (‘ribosome’) is shown in green with the corresponding P value of a hypergeometric test. CDS introns are represented by dots and 5′ UTR introns by plus signs. See Extended Data Fig. 1 for data at other points of the isobole. d, RNA-seq counts for the RPS22B locus, colour-coded as in a. Top, schematic of the RPS22B introns and exons and snR44 locus according to the Saccharomyces Genome Database.

We found that LiCl induces extensive intron retention compared to the drug-free control (Fig. 1c). Intron retention was not due to changes in growth rate or general stress, as no such increase was observed when cells were treated with cycloheximide (Extended Data Fig. 1a), or with two other drugs with different targets (Extended Data Fig. 2a). RP transcripts were the functional gene set most strongly affected by this increase in intron retention (Benjamini–Hochberg-corrected hypergeometric test, Padj= 3 × 10−25, Fig. 1c, Extended Data Fig. 1b, c, Methods). The increase in intron retention was correlated with a decrease in their host transcript level (Extended Data Fig. 1d, e). Analysing RNA-seq reads that span the exon–intron junctions confirmed that introns are retained in mature transcripts, rather than being spliced out but not degraded (Extended Data Fig. 1f); this observation was further corroborated by observing the increase in intron retention also in the polyadenylated fraction of RNA (Extended Data Fig. 2b). It has previously been proposed that splicing of RP genes is downregulated by the accumulation of stable excised introns in the stationary phase4. However, these introns were not retained differently to other introns under LiCl stress (Extended Data Fig. 1g), suggesting that the high retention of introns in RP transcripts observed with LiCl treatment is not mediated by the stable linear excised introns previously reported.

Osmotic stress induces Rps22B bimodality

RPS22B was a clear outlier with respect to intron retention under LiCl treatment. RPS22B encodes a component of the small ribosomal subunit and is one of only nine Saccharomyces cerevisiae genes containing two introns31. The intron in the 5′ untranslated region (5′ UTR) of RPS22B showed the highest increase in intron retention among 5′ UTR introns under osmotic stress (Fig. 1c), whereas there appeared to be numerous copies of the coding sequence (CDS) intron of RPS22B, presumably owing to a small RNA encoded within the intron that is under separate regulation (Fig. 1d; snR44). The prevalence of stop codons in the introns of RPS22B (Extended Data Fig. 3a, b), typical of yeast introns32,33, suggested that intron retention in the transcript could manifest itself in altered protein levels rather than protein isoforms. To observe the single-cell behaviour of Rps22B at the protein level, we coupled the rigorously controlled combinatorial drug treatment to flow cytometry using a yeast strain with a GFP-tagged Rps22B8. Notably, Rps22B exhibited two clearly distinct levels of expression at intermediate LiCl concentrations (Fig. 2a). Other osmotic stressors similarly induced bimodal expression of Rps22B (Extended Data Fig. 3c).

Fig. 2: The 5′ UTR intron mediates bimodal Rps22B protein expression under LiCl, consistent with a bistable regulatory loop.
figure 2

a, Histograms of flow cytometry measurement of strains with GFP-tagged Rps22B or Rps22A. The RP gene RPS22B, which contains the 5′ UTR intron with the largest increase in retention due to LiCl, exhibits bimodal protein expression at intermediate LiCl concentrations, whereas its ohnologue, RPS22A that contains no introns, does not. See Extended Data Fig. 3 for Rps22B expression in other osmotic stresses. AU, arbitrary units. b, As in a, but with seamless deletion of either of the RPS22B introns (Δ5′ UTRi, deletion of 5′ UTR intron; ΔCDSi, deletion of CDS intron). See Extended Data Fig. 3d for a strain with both introns deleted. c, As in a, but for strains with the 5′ UTR of RPS22B fused to GFP, with or without the 5′ UTR intron. d, Ad hoc definition of the extent of Rps22B bimodality (Methods). e, The extent of Rps22B bimodality (coloured lines) is overlaid on the median protein level of Rps22B (greyscale) in a two-drug gradient of LiCl and cycloheximide. f, Extent of Rps22B bimodality as a function of median Rps22B protein level is shown for all wells in the 2-drug gradient (dots) alongside the running average with a window of 30 data points (line). Rps22B bimodality peaks at a certain level of median Rps22B expression, rather than at a certain growth rate or LiCl concentration. See Extended Data Figs. 4, 5 for comparison with LiCl–myriocin drug pair.

The 5′ UTR intron mediates Rps22B bimodality

Although LiCl induced bimodal expression of Rps22B, no such expression pattern was observed for Rps22A, the intronless ohnologue of Rps22B (Fig. 2a), the protein sequence of which is identical to that of Rps22B except for a single amino acid (Extended Data Fig. 3d). As the promoters of RPS22A and RPS22B behave similarly under different conditions34, we hypothesized that the bimodal expression of Rps22B that is observed under osmotic stress depends on the presence of one or both introns; the 5′ UTR intron in particular is highly conserved both within Saccharomycetaceae23 and in S. cerevisiae itself (Extended Data Fig. 3e). Deletion of the 5′ UTR intron, but not the CDS intron, abrogated the bimodal expression of Rps22B under LiCl stress (Fig. 2b, Extended Data Fig. 2f). In addition, it resulted in higher levels of RPS22B transcripts (Extended Data Fig. 3g) and protein (Fig. 2b, Extended Data Fig. 3f) under LiCl stress compared to the parental strain. No growth defects that could explain the differences in protein expression between the strains were observed (Extended Data Fig. 3h). This suggests that, in the absence of the 5′ UTR intron, cells under LiCl stress can no longer downregulate the production of Rps22B, consistent with previous reports that the double-stranded RNA structure contained in the intron is necessary for the regulated degradation of the transcript35. Fusing the 5′ UTR of RPS22B to GFP conferred bimodal GFP expression under LiCl stress, but only if the 5′ UTR intron was left intact (Fig. 2c). These results show that the 5′ UTR of RPS22B is a cis-regulatory element sufficient to confer bimodal protein expression to an unrelated gene, and corroborate that the evolutionarily conserved 5′ UTR intron is necessary for this effect.

To further elucidate the role of the 5′ UTR intron in the protein heterogeneity of Rps22B, we quantified intron retention in different situations. Fluorescence-activated cell sorting (FACS) of the LiCl-induced Rps22B-bimodal cells showed that Rps22B-high cells have lower intron retention and higher CDS transcript level; however, these differences alone appear insufficient to account for the pronounced bimodality in protein expression (Extended Data Fig. 3g). Moreover, sequencing the polyadenylated RNA fraction revealed that although LiCl induces a strong increase in 5′ UTR intron retention in the Rps22B transcript, this increase is not necessary for the induction of Rps22B protein bimodality by other osmotic stressors (Extended Data Fig. 3g). This suggests that the bimodal expression of Rps22B under osmotic stress is mediated by a post-transcriptional mechanism that requires the presence of the 5′ UTR intron and amplifies cell-to-cell differences in intron retention at the protein level.

Rps22B expression is potentially bistable

A bimodal pattern of expression suggests—but does not necessarily imply—an underlying bistable regulatory circuit, such as a positive feedback loop36. A general hallmark of such a regulatory circuit is the existence of an unstable fixed point. Here, this fixed point would correspond to a protein level at which a small increase or decrease causes the cell to go to one of the two stable fixed points that correspond to high or low expression states37, respectively. To address whether such an unstable fixed point exists for Rps22B, we examined the Rps22B bimodality in the presence of LiCl while using another drug to perturb the growth rate and, as measured, the overall level of Rps22B. We used an automated liquid handling set-up that enables continuous monitoring of culture over a fine two-drug gradient distributed over six 96-well microtitre plates7 (Extended Data Fig. 4a), and at the end of the incubation measured the single-cell protein expression by flow cytometry (Methods). We used LiCl in combination with two drugs with disparate mechanisms—the translation inhibitor cycloheximide and the sphingolipid synthesis inhibitor myriocin38. We found that the extent of how clearly the expression levels of the two subpopulations are separated into two peaks depends primarily on the median protein level of Rps22B (Fig. 2d, e, Extended Data Fig. 4c, d), with a maximum at an intermediate level (Fig. 2f, Extended Data Fig. 4e). This behaviour is consistent with the existence of an unstable fixed point at this protein level, causing what would be an otherwise log-normally distributed population centred around this point to divide equally into the two presumably stable subpopulations.

Single-cell isogrowth profiling

To investigate the extent to which osmotic or translation stress induces distinct cellular states, as observed for Rps22B, we took advantage of the yeast protein–GFP library8, which contains 4,156 strains with single protein–GFP fusions, and performed genome-wide single-cell isogrowth profiling (Extended Data Fig. 6a–c). We profiled the entire library in four of the conditions used previously (Supplementary Table 1) and selected strains that, on visual inspection, did not show a clearly unimodal expression pattern for further characterization using a more detailed antiparallel gradient (Supplementary Table 2). We found several instances of protein expression heterogeneity resulting from non-specific growth inhibition (Extended Data Fig. 6d) or specifically from either of the stresses (Extended Data Fig. 6e). Cycloheximide induced bimodal expression of Hsp12, the budding yeast persistence marker39, and of Aro9, an enzyme involved in production of the yeast quorum-sensing molecule tryptophol40, a trigger for invasive growth in low-nitrogen environments40. LiCl induced heterogeneity in Rps9A, another small RP subunit that contains an intron in its gene. Similar to Rps22B, the stress-induced heterogeneity of Rps9A levels was intron-mediated (Extended Data Fig. 6f). In addition, there was a general trend in which an increase in intron retention in LiCl correlated with a decrease in the level of the respective protein (Extended Data Fig. 6g). Overall, these observations indicate that intron retention is used by the yeast cell to control the protein level and, in the case of Rps9A and Rps22B, also the protein level heterogeneity.

Differential fitness during starvation

Introns have a key role in preparing the yeast population for starvation5,27. Therefore, we hypothesized that the two subpopulations, defined by distinct levels of Rps22B expression, have differential fitness under starvation stress. To test this idea, we subjected an exponentially growing culture of the Rps22B–GFP strain to LiCl stress using a microfluidic system (Methods). After the establishment of bimodality, we replaced the LiCl-containing growth medium with spent medium. In this way, we induced starvation of varying duration, followed by a sudden switch to rich medium (Fig. 3a). At first, the spent medium triggered a rapid disappearance of the strong nucleolar signal of Rps22B–GFP (Supplementary Videos 1, 2), which was presumably caused by the redistribution of nucleolar proteins across the cytoplasm upon stress, as occurs in other model systems41. After the switch to rich medium, the nucleolar signal of Rps22B re-emerged in cells that started growing again. Notably, the fitness of cells depended on their levels of Rps22B expression immediately before the starvation stress. Longer starvation stress favoured the survival of cells that expressed Rps22B at low levels, in that they lysed less frequently than the highly expressing cells. By contrast, shorter starvation favoured the cells that expressed Rps22B at high levels, in that they budded more in the period following nutrient replenishment (Fig. 3b–f). FACS corroborated this phenotypic difference, as the Rps22B–GFP-high subpopulation exhibited poorer survival under starvation (Fig. 3h), but a higher growth rate in rich medium (Fig. 3i, j) compared to the Rps22B–GFP-low subpopulation. This marked diversification of the population observed for Rps22B, with clear fitness effects under a subsequent stress, may reflect a bet-hedging strategy42,43,44.

Fig. 3: Rps22B expression level confers a selective advantage under starvation stress.
figure 3

a, Schematic of the temporal sequence of media used in microfluidic microscopy experiments. b, c, Example micrographs of Rps22B–GFP yeast cells after the application of LiCl, showing LiCl-induced bimodality of Rps22B expression (b). Rps22B-high cells (dark red arrows) lyse after prolonged starvation and nutrient replenishment (c). d, e, As in b, c, but for a shorter period of starvation. After a shorter period of starvation and medium replenishment, the Rps22B-high cells marked with dark red arrows (d) budded more than other cells (e) (Supplementary Videos 1, 2). Scale bars, 10 μm. f, Quantification of the time-lapse micrographs. Top, cumulative distribution of Rps22B expression measured before starvation in cells that have either lysed (green) or not (grey) within 6 h after medium replenishment following a long starvation. Bottom, cumulative distribution of Rps22B expression before starvation in cells that have either budded (green) or not (grey) within 6 h after medium replenishment following a shorter starvation. Rps22B expression values between the two panels are not comparable. A two-sided Mann–Whitney U test was used to determine significance. g, Histograms of Rps22B–GFP expression in LiCl before and after FACS. h, Survival curves (colony-forming units) of the sorted populations as a function of time spent in phosphate-buffered saline (PBS) after the sorting and before plating on rich medium. The two-sided bootstrapped P value is shown (Methods). i, Growth curves of cultures in rich medium after sorting. The shaded rectangle denotes the optical density range that was used to quantify growth rates. j, Quantification of growth rates from i. Significance was determined using a two-sided t-test. k, Fraction of cells that were dead after starvation in time-lapse microscopy experiments comparing wild-type (WT) Rps22B and the RPS22B 5′ UTR intron deletion mutant. For experimental set-up and further quantification, see Extended Data Fig. 7. Significance was determined using a one-sided permutation test.

To test whether the link between intron-mediated Rps22B protein heterogeneity and phenotypic heterogeneity is causal, we performed time-lapse microscopy with the RPS22B 5′ UTR intron deletion mutant (Extended Data Fig. 7a, Supplementary Videos 3, 4). Deletion of the 5′ UTR intron in RPS22B not only increased the fraction of Rps22B–GFP-high cells, but also the fraction of cells that died under starvation (Fig. 3k, Extended Data Fig. 7b, d). Notably, it also partially abolished the coupling between phenotypic heterogeneity and Rps22B protein heterogeneity (Extended Data Fig. 7e), suggesting that downregulation of Rps22B specifically through the intron-mediated mechanism is important for phenotypic benefit under starvation. Population-level assays with the intron-deletion mutant confirmed that its survival times under starvation are less heterogeneous and lower on average (Extended Data Fig. 8a), and that its growth rate is increased under LiCl stress (Extended Data Fig. 8b). These observations confirm that the RPS22B 5′ UTR intron can mediate not only Rps22B expression heterogeneity, but also phenotypic heterogeneity.

Bimodal Rps22B expression in high-glucose conditions

Bet-hedging strategies can evolve if the eliciting signal is probabilistically followed by stress in which the subpopulations exhibit differential fitness in the natural environment42,45. However, yeasts nearing saturation, and hence starvation, in standard rich laboratory medium do not exhibit Rps22B bimodality (Fig. 4a). Therefore, we wondered whether there is a plausible natural setting in which Rps22B bimodality is triggered just before the onset of starvation. We reasoned that yeasts may often be exposed to hyperosmotic sugar concentrations, such as those in ripe fruits. High glucose elicits osmotic responses that are different from those to salt46. In addition, a high concentration of glucose, despite being an osmotic stressor, leads to comparatively higher expression levels of RPs than LiCl stress (Extended Data Fig. 9). As Rps22B exhibits the maximum extent of bimodality at a certain median expression level (Fig. 2d–f), we hypothesized that a high-glucose environment might specifically induce Rps22B bimodality as yeast nears saturation, when RP levels decrease owing to growth slowdown.

Fig. 4: A high-glucose environment induces bimodal Rps22B expression as cells near saturation.
figure 4

a, b, Histograms of Rps22–GFP expression as measured by flow cytometry after 10 h of incubation in standard rich medium (2% (w/v) glucose; a), or in a yeast peptone medium supplemented with 20% (w/v) glucose (b). The red hue indicates increasing cell density in the initial inoculum. Examples of exponential and near-saturated cultures are highlighted by shading.

To test this idea, we inoculated Rps22B–GFP cultures into a yeast peptone medium with a high concentration of glucose (20% w/v) that emulates the total hexose concentration in ripe grapes47, at various initial cell densities. Exponentially growing cells in high-glucose medium generally resulted in less clear bimodal expression of Rps22B–GFP compared to cells treated with LiCl; however, cultures in high-glucose medium that were nearing saturation exhibited a pattern of Rps22B expression that was clearly bimodal (Fig. 4b). This observation suggests that part of the yeast population is attuned to a probabilistic event of nutrient replenishment that can follow growth in a high-glucose environment, while the rest of the population is preparing for starvation.


In the budding yeast, the overall expression level of RPs results from two sets of RP genes with high amino-acid sequence similarity, which often differ in the presence and identity of intronic sequences6,18. Here, we have uncovered an intron-mediated regulation of protein expression heterogeneity. We showed that LiCl leads to a widespread retention of introns48 that is independent of growth rate perturbations and predominantly affects RP transcripts. The small ribosomal subunit protein gene RPS22B manifested bimodal protein expression under conditions of osmotic stress, and this effect was mediated by its 5′ UTR intron, which is conserved throughout the Saccharomycetaceae23. By contrast, its intronless ohnologue RPS22A exhibited a unimodal pattern of protein expression irrespective of stress. This behaviour of the RPS22 gene pair offers a paradigm to explain the function of introns in duplicated RP genes, in that they enable differential and versatile regulation for duplicated genes as an evolutionary trade-off between environmental responsiveness and precise regulation9.

We observed population diversification with respect to starvation stress and subsequent recovery, which was marked by Rps22B expression levels and partially dependent on the RPS22B 5′ UTR intron. Reminiscent of bet-hedging strategies42, we found that in high-glucose medium, yeast populations diversify in their expression of Rps22B as they enter the stationary phase, as if anticipating a probabilistic replenishment of nutrients. Why such a behaviour should have evolved for a high-glucose environment, but does not manifest at glucose levels present in the standard laboratory ‘rich’ medium, is unclear. One plausible scenario is that for yeast living on the skin of grapes, a high concentration of glucose is an environmental signal that the fruits are ripening and bursting. Osmotic bursting of fruits is known to follow a probabilistic trajectory over time49 and is known to happen predominantly during rainfall50; hence, in a cluster of grapes, vigorous yeast growth due to bursting of one of the berries might be followed by nutrient replenishment due to bursting of a neighbouring berry, determined by the probabilistic aspects of rain duration and intensity.

The fact that we found osmotic-stress-induced, intron-mediated phenotypic heterogeneity only for Rps22B and Rps9A, although most other pairs of duplicated RP genes also contain at least one intron, is notable, especially as the effect of introns on protein expression level has been reported for other RPs in the budding yeast6. It is thus tempting to speculate that for other RPs, there might exist other levels of osmotic stress—or other stressors altogether—that would trigger population diversification with respect to the expression level of a given RP. Such a set of diversification mechanisms could then present a versatile stress toolkit for the yeast population, enabling yeast to brace against continued stress and at the same time maintaining a small, stress-sensitive subpopulation that is poised to rapidly exploit a short window of fitness advantage should the stress suddenly disappear. Our study thus highlights the need for further study of the intronic regulation of population diversification, and demonstrates the utility of using graded, growth-rate controlled perturbations.


Transcriptional isogrowth profiling

For isogrowth RNA-seq, S. cerevisiae strain BY4741 was grown in seven conditions with varying ratios of LiCl (Sigma-Aldrich, L9650) and cycloheximide (Sigma-Aldrich, 37094), ensuring 50% growth inhibition, and in YPD (yeast extract (Sigma-Aldrich Y1625) 1% w/v, peptone (Sigma-Aldrich 91249) 2% w/v, dextrose (Sigma-Aldrich D9434) 2% w/v) containing no drug (Extended Data Table 1). The frozen stock was diluted 130 times into drug-containing wells and 100 times more into wells without drug. The cells were incubated for a total of around 17 h to a final absorbance of around 0.1 on the 96-well plate, corresponding to an OD600 nm of approximately 0.5. Cells were then collected and the RNA extraction was performed using the RiboPure RNA Purification Kit for yeast (Thermo Fisher Scientific, AM1926). The extracted RNA was sent to the Next Generation Sequencing Facility of the Vienna Biocenter Core Facilities, where it was rRNA depleted using the Illumina Ribozero Yeast Kit, multiplexed using Illumina True Seq adapters, single-end 50-bp sequenced using Illumina HiSeqV4 and demultiplexed.

RNA samples from treatments with single stressors were obtained from cultures with automated re-inoculation every 8 h (ref. 7), with a total incubation time of 24 h, to keep cultures in exponential phase and ensure the drug effects have taken place. Strains were grown in YPD with a drug, either in a concentration gradient to select samples closest to 50% of growth inhibition (LiCl, fenpropimorph and myriocin), or in a single concentration (0.4 M NaCl and glucose 20%). Fenpropimorph (Sigma-Aldrich 36772) was tested in a 0.5–1.5 µM gradient; sequenced samples were treated with 0.85 µM fenpropimorph. The myriocin (Sigma-Aldrich 476300) gradient encompassed 0.25–0.75 µg ml−1; 0.32 µg ml−1 myriocin samples were sequenced. For each sample of extracted RNA, both mRNA and ribo-depleted total RNA sequencing libraries were prepared at the Cologne Center for Genomics (CCG) and sequenced using 2 × 100-bp paired-end reads.

Intron retention analysis

The reads resulting from sequencing were aligned to the annotated reference S. cerevisiae genome R64-2 using TopHat51 or HISAT252. The reads mapping to introns or CDS of intron-containing genes were quantified using featureCounts53 using custom produced .saf files extracted from the reference .gff annotation file. The intron retention rate was calculated as the ratio of intron counts normalized to the intron length over the CDS counts normalized to the length of the CDS region. For genes containing multiple introns, the introns were treated jointly for CDS introns and separately for 5′ UTR introns. Alignments were visualized using Integrative Genomics Viewer54.

For the analysis shown in Extended Data Fig. 1f, first the number of contiguous reads (no mismatches allowed) overlapping the individual intron ends was determined. This count was then averaged between the 5′ and the 3′ end of the intron ends and divided by the read length (50 bp). The intron retention rate was then calculated as the ratio of the resulting value over the CDS counts normalized to the length of the CDS region.

For GO enrichment analysis, intron-containing genes were ordered into a ranked list on the basis of the fold change increase in intron retention in LiCl compared to the no-drug control, either in decreasing (Fig. 1c) or increasing (Extended Data Fig. 1b) order. The ranked list was then analysed using GOrilla55, for either cellular component or biological process, respectively. The adjusted P values reported here correspond to the false discovery rate reported by GOrilla. Gene enrichments were visualized using REVIGO56.

Rps22B intron conservation analysis

To gauge the evolutionary conservation of the Rps22B 5′ UTR intron within the natural population of S. cerevisiae (Extended Data Fig. 3e), previously published sequencing data for 1,011 isolates57 were multiple-sequence aligned using Clustal Omega58 and visualized using MSA-BIOJS59.

Screening the entire yeast protein–GFP library for drug-induced bimodality

The whole yeast protein–GFP library8, contained in 43 96-well microplates, was screened for drug-induced bimodality by culturing in four conditions: YPD medium containing no drug, cycloheximide, LiCl, or a combination of LiCl and cycloheximide. The screen was divided into six experimental batches. In every batch, up to nine plates of the library were processed, with the same strain being cultured in four conditions on the same day and on the same microtitre plate, giving rise to up to 36 microtitre plates per experiment. Before the experiment, empty microtitre plates and YPD medium were pre-incubated at 30 °C to ensure reproducibility of gene expression measurements. LiCl, cycloheximide or a mixture of the two were diluted with YPD to a final concentration needed to achieve 40% inhibition and ensuring that in the mixed condition, the LiCl and cycloheximide were mixed in an equipotent manner, meaning that on their own they both elicited approximately the same (smaller) growth inhibition (Extended Data Table 1). The drug solutions were pipetted row-wise into a 96-well microtitre plate, including a drug-free YPD control, enabling the cultivation of 24 strains in 4 conditions per microtitre plate. The library plates stored at −80 °C were thawed and, using an electronic multi-step multi-channel pipette, 1.5 μl saturated glycerol stock of the corresponding strain was inoculated into each drug-containing well and 0.2 μl of the saturated glycerol stock was inoculated into drug-free YPD as a control. The smaller inoculum size for the drug-free control was designed to ensure that the final cell density at the end of the experiment was comparable to that of the drug-containing wells. Plates were incubated for around 14 h in an automated incubator (Liconic StoreX) kept at 30 °C and greater than 95% humidity, vigorously shaken at more than 1,000 rpm. During the incubation, OD600 nm was measured every approximately 45 min in a Tecan Infinite F500 plate reader. In addition to shaking during incubation, directly before each measurement, plates were shaken on a magnetic shaker (Teleshake; Thermo Fisher Scientific) at 1,100 rpm for 20 s. The automated set-up was programmed using the Tecan FreedomEvo software. The growth rates were inferred by fitting a line to the log-linear part of OD600 nm measurements between 0.01 and 1.

After incubation, the yeast cells in 96-well microtitre plates were twice centrifuged at 1,050g for 3.5 min and resuspended in ice-cold Tris-EDTA buffer by vigorous shaking at 1,000 rpm on a Titramax shaker for 30 s. After another centrifugation at 1,050g for 3.5 min, the cells were resuspended in 80 µl of Tris-EDTA and immediately stored at −80 °C. On the day of the flow cytometry measurement, the plates were thawed on ice for around 3 h and kept on ice until the measurement. The fluorescence was measured for 10,000 cells using BD FACS Canto II equipped with a high-throughput sampler, using FACS Diva software. The green fluorescence measured in FITC-H channel was normalized to the forward scatter FSC-H channel.

Strains exhibiting bimodality, minor peaks or prominent changes in gene expression at visual inspection of the expression histograms were manually selected for a more detailed screen. The selected strains were re-streaked from the yeast protein–GFP library microtitre plates onto YPD-agar plates and single clones for each strain were picked and cultured in a 96-well microtitre plate to saturation. Glycerol was added to a final concentration of 15% and the plates were frozen. The plates were screened in eight conditions (Extended Data Table 1) and analysed using flow cytometry in a way similar to that described above. The identity of the relevant strains—Rps22A, Rps22B, Aro9, Hsp12, Cit1 and Tdh1—was confirmed by PCR and gel electrophoresis as reported before8, using a common F2CHK reverse primer and strain specific oligos taken from ( The strain Smi1 appeared bimodal in the screen; however, its identity could not be confirmed as above. The identity of other strains was not tested.

Construction of GFP-labelled intron deletion strains and flow cytometry experiments

The seamless intron deletion strains for Rps22B (Rps22B-∆i1-∆i2, rps22bDi_JPY138F4; Rps22B-∆i1, rps22bD1i_MDY125A4; Rps22B-∆i2, rps22bD2i_MDY125A9) and Rps9A (Rps9A-∆i, rps9aDi_MDY133H8), as well as the parental haploid strain WT_JPY10H3 (MATa ura3∆0 lys2∆0 leu2∆0 his3∆200) were a gift from J. Parenteau and S. Abou Elela5. The parental and intron deletion strains were labelled with GFP fused to the protein of interest following homologous recombination of the PCR-amplified fluorescent marker from the matching yeast protein–GFP library strains. To achieve this, the DNA of the Rps22B–GFP and Rps9A–GFP strains from the library was extracted using the Yeast DNA Extraction Kit (Thermo Fisher Scientific, 78870). PCR amplification of the GFP label with 40 base pairs flanking sequences was done using the corresponding F2 and R1 pair of primers from a previous study8 ( in the following reaction conditions: 2 ng ml−1 DNA in a final mix of 20 μl, with 0.4 units of Phusion High-Fidelity DNA polymerase (Thermo Fisher Scientific, F530SPM), 4 μl of 5× Phusion HF buffer, 0.2 μM dNTPs and 0.5 μM primers. PCRs were incubated for 3 min at 98 °C, followed by 34 cycles of 98 °C for 30 s and 72 °C for 2.5 min, and a final incubation at 72 °C for 5 min. The resulting PCR products were transformed into the parental WT_JPY10H3 and ∆i strains as in a previous report60, selecting on SD -HIS agar plates (Takara Bio, 630411; Diagonal GmbH&CoKG, Y1751). The correct genomic context insertion of GFP of single colonies was confirmed by PCR using a combination of primers that also allowed the confirmation of the intron deletion strain. To confirm the Rps22B–GFP fusions in the Rps22B-∆i and WT_JPY10H3 strains, we used primers rps22b-899F 5′-CCGTTATTCTTCTCGCAACC-3′ binding upstream of the 5′ UTR intron, and RPS22B-CHK (ref. 60) 5′- ACTAGATGGTGTGATCGGGC-3′ binding in the CDS intron sequence in combination with the reverse primer F2CHK (ref. 60) 5′- AACCCGGGGATCCGTCGACC-3′, complementary to the GFP sequence. To confirm the Rps9A–GFP fusions in the Rps9A-∆i and WT_JPY10H3 strains, we used rps9a-800F 5′- GTTCGATTTCTTGGTCGGACGC-3′ upstream of the open reading frame and F2CHK. Once the successful construction of the GFP reporter strains was confirmed, 20 μl of saturated overnight cultures were inoculated into 96-well microtitre plates containing 180 μl of YPD with the respective concentration of LiCl. The microtitre plates were incubated at 30 °C and continuous shaking at 900 rpm on a Titramax shaker for 7 h, collected and measured on a flow cytometer as described above. Strains with genome-integrated GFP with different 5′ UTR fusions were a gift from the laboratory of G. Stormo26.

Measurement of Rps22B expression in detailed two-dimensional drug gradients

The re-inoculation set-up, as reported previously7, in conjunction with flow cytometry measurements was used to measure Rps22B expression in detailed two-dimensional (2D) gradients of LiCl and cycloheximide, and LiCl and myriocin (Sigma-Aldrich, M1177). In brief, a S. cerevisiae strain with the RPS22B gene fused to GFP protein from the ORF-GFP library8 was grown in YPD broth in a conical flask overnight and then distributed into a 96-well plate. A customized robotic setup (Tecan Freedom Evo 150) with eight liquid handling channels and a robotic manipulator was used to produce a 2D discretized two-drug 24 × 24-well gradient in YPD spread over 6 96-well plates and to inoculate the yeast overnight culture to a final liquid volume in the well of 200 μl and a final absorbance 0.03 on the 96-well plate, corresponding to a standardized OD600 nm of 0.15. Working drug solutions were prepared either by adding the respective amounts of concentrated DMSO drug stocks thawed from −20 °C storage (no refreezing) previously prepared from stock chemicals (cycloheximide and myriocin), or by dissolving directly in YPD and sterile-filtering (LiCl). The six plates were incubated for three iterations, each lasting around 8 h. After the incubation, the cells were collected and measured using flow cytometry as described above. Rps22B protein bimodality was quantified as the depth of the trough on the Rps22B protein single-cell expression histogram; that is, the y-axis distance between the trough and the lower of the two peaks (Fig. 2d). To this end, the MATLAB function findpeaks was used to determine the prominence of the peak that was created from the trough through inversion of the histogram values.

FACS of Rps22B–GFP and experiments with GFP-sorted cells

Constitutive cytosolic mCherry expression was added to the parental and intron deletion GFP-fusion strains. To do so, a TDH3::mCherry construct was inserted in the HO locus by homologous recombination after transformation of a NotI-digested SLVA06 plasmid61, provided by the laboratory of M. Springer. The Rps22B–GFP strain with constitutive mCherry expression was inoculated into YPD and incubated at 30 °C for 16 h. The overnight culture was diluted 20-fold into 4.5 mg ml−1 LiCl in YPD, incubated for 6 h and washed twice with PBS. Positive mCherry cells were sorted into low-GFP and high-GFP populations with a Becton Dickinson INFLUX cell sorter (for gating strategy, see Supplementary Fig. 1). Around 8 × 106 sorted cells of each condition were used for RNA extraction and sequencing (see ‘Transcriptional isogrowth profiling’). Samples of sorted cells at a final concentration of 4,000 cells per ml were prepared with PBS for the following experiments. Immediately after sorting, 15 µl of the samples were inoculated into fresh YPD medium in 24 replicates each. Optical density was measured every 20 min for 48 h in a plate reader incubated at 30 °C with constant shaking. OD was corrected by subtracting the background, and growth rates were inferred by a linear fit to the log-linear part of the corrected OD measurements between 0.02 and 0.2. Growth rates higher than 0.85 h−1, which resulted from bacterial contaminations, were discarded. A two-tailed t-test was performed to compare the growth-rates from the two sorted populations (using the function ttest2 in MATLAB R2019a). For scoring survival after starvation, sorted cells were incubated in PBS at 30 °C with shaking at 200 rpm. At each sampling time point, 50 µl of starving cultures were spread onto an OmniTray plate (Thermo Fisher Scientific) filled with 35 ml of YPD-agar, in two replicates for each sorted population. Agar plates were cultivated at 30 °C, pictures were taken after around 36 h, and colonies were counted using ImageJ. To quantify the difference between low-GFP and high-GFP populations (Fig. 3h), we used the difference Δ between the total low-GFP and high-GFP colony counts along all timepoints. To test the statistical significance of Δ, we used bootstrapping. In brief, we created 104 surrogates (artificial datasets) complying with the null hypothesis that there is no difference between the populations. To create one such surrogate, we sampled—for each replicate and time point—the colony count from the Poisson distribution with a rate λ corresponding to the mean colony count for that time point. We determined the P value as the fraction of surrogates that exhibit a higher Δ than the original experimental dataset.

Time-lapse imaging of the Rps22B–GFP strain during starvation

For micrographs shown in Fig. 3, the ORF-GFP library Rps22B–GFP strain was used; for experiments with the intron deletion mutants (Extended Data Fig. 7), we used the 5′ UTR intron-deletion strain and its parental Rps22B–GFP strain, both transformed with a plasmid constitutively expressing mCherry (see ‘FACS of Rps22B–GFP and experiments with GFP-sorted cells’) to allow for easy tracking of cellular integrity, because cells lose the mCherry signal after lysis; this approach has been used previously for tracking the lysis of bacterial cells62. Strains were inoculated into YPD at 30 °C with shaking overnight. The resulting culture was diluted 20-fold and loaded to the CellASIC ONIX2 haploid yeast microfluidic plates (MerckMillipore) at 55 kPa pressure, using the CellASIC ONIX2 system software. A sequence of YPD, LiCl-YPD, starvation medium (spent medium or PBS) and YPD were flushed through the microfluidic chamber in a time-controlled manner at 10-kPa pressure. Spent medium was produced by incubation of the Rps22B–GFP strain in YPD medium for seven days with subsequent sterile filtration. Starvation times were empirically selected such that an intermediate number of cells died in the long starvation condition or budded after the short starvation. The specimen was imaged every 5 min at multiple locations with a Nikon Eclipse Ti inverted microscope through a 100× oil objective in a 30-°C cage incubator, using NIS Elements software. The micrographs of yeast from just before the starvation stress (Fig. 3) were segmented and fluorescence-quantified using the CellStar MATLAB plug-in63 and the fluorescence readout was log10-transformed. The time-lapse images were then visually scored to determine which cells budded or lysed after the replenishment of nutrients. The time-lapse videos from the experiments with the intron-deletion mutant were segmented and tracked using YeaZ64.

RPS22B 5′ UTR intron-deletion mutant fitness experiments

The 5′ UTR intron-deletion strain and its parental strain, with Rps22B–GFP fusion and constitutive mCherry expression, were grown overnight in YPD. For growth rate determination experiments in liquid media, 5 ml of each saturated overnight culture was standardized to OD600 nm = 0.1 and inoculated in a 1/10 dilution into YPD containing 0, 2.25, 4.5, or 9 mg ml−1 LiCl in 12 replicates of each strain and condition. OD was measured every 5 min for 24 h in a plate reader during incubation at 30 °C with constant shaking. OD and growth rates were processed as above (see ‘FACS of Rps22B–GFP and experiments with GFP-sorted cells’). For scoring survival under starvation, 1 ml of each overnight culture was inoculated into 19 ml of YPD with 9 mg ml−1 LiCl. Cultures were incubated at 30 °C with 200 rpm shaking for 6 h. Cultures were standardized to OD600 nm = 1 and washed twice by centrifuging at 1,050 g for 5 min and resuspending pellets in 20 ml of PBS. Dilutions of 5,000 cells per ml were incubated at 30 °C with constant shaking. Plating onto YPD-agar, incubation and analysing colony counts were done as explained above (see ‘FACS of Rps22B–GFP and experiments with GFP-sorted cells’).

Rps22 bimodality in other osmotic stresses and in high glucose

The ORF-GFP library Rps22B–GFP cultures were inoculated into YPD medium, YPD containing 0.6 M NaCl, 2 M KCl (Sigma Aldrich, S3014 and P9541, respectively) or into a yeast peptone medium with 5%, 10% or 20% (w/v) glucose (Sigma Aldrich, G8270). To measure the fluorescence intensity of cells from different cell densities and growth stages, six microtitre plates were prepared, each with the same conditions but with a gradient of initial inoculum sizes; in this way, each plate accounted for the stress array with different cell density and could be incubated and analysed at different time points. To achieve this, saturated overnight cultures were diluted with YPD in a 4/5 serial dilution (that is, 4 ml was transferred to the final volume of 5 ml) to obtain (4/5)0 to (4/5)23 dilutions. Subsequently, 15 μl of cell dilutions were inoculated in 185 μl of medium. Plates were incubated in the Liconic StoreX and kept at 30 °C and greater than 95% humidity, with constant shaking. Plates were processed at 8, 12, 16, 20 and 24 h after inoculation with an automated robotic system consisting of a ACell System (HighRes Biosolutions) integrated with a Lynx liquid handling system (Dynamic Devices), a plate reader (Synergy H1, BioTek) and a CytoFLEX flow cytometer (Beckman Coulter). First, plates were shaken and measured as described above to obtain the OD600 nm values. Cultures were then transferred to a new plate and, when necessary, diluted to an absorbance of around 0.3 with Tris-EDTA buffer to avoid high cell densities, and measured by flow cytometry.


Rps22B protein bimodality in LiCl was replicated independently by two researchers, at least five times, using two different yeast strains; all attempts at replication were successful. Rps22B protein bimodality in NaCl and KCl was determined twice; all attempts at replication were successful. The loss of bimodality in the RPS22B 5′ UTR intron deletion strain was replicated three times; all replication attempts were successful. The bimodal expression of the RPS22B 5′ UTR intron–GFP fusion in LiCl was observed twice; all attempts were successful. The Rps22B bimodality on the entry to the stationary phase after growth in high glucose was reproduced twice; the exact timing of observing maximum bimodality varied, so multiple time points were measured as described in ‘Rps22 bimodality in other osmotic stresses and in high glucose’. All attempts at replication using the incubation method described were successful; one attempt at replication using a different microtitre plate shaker (Titramax 1000, Heidolph) resulted in less pronounced heterogeneity, which is likely to be due to a different degree of aeration or shaking. Intron retention determined by sequencing was confirmed independently by two researchers, once by each; all attempts at replication were successful. Time-lapse microscopy phenotype of Rps22B-high and Rps22B-low cells was replicated twice in slightly modified set-ups and with independent strains as detailed in Fig. 3, Extended Data Fig. 7; all attempts at replication were successful. For the experiment shown in Extended Data Fig. 7, two starvation conditions, in which the length of the starvation period was varied, were tested to achieve an intermediate number of cell deaths within the time frame of the experiment as assessed visually and the chosen condition was then quantified. The whole yeast–GFP library isogrowth scan was performed once owing to the scale of the experimental effort; about 10% of strains that did not show clear unimodal expression on visual inspection of the expression histograms were restreaked to exclude the possibility of contaminations and the measurement was repeated once on a more detailed antiparallel gradient as detailed in the manuscript; the bimodality of Rps9A and Aro9 was additionally replicated once in a detailed discretized 2D drug concentration gradient similar to the one shown in the manuscript for Rps22B. To determine the phenotypic effect of the Rps22B expression level, cells were sorted once and survival and growth assays performed in 2 and 24 replicates, respectively; the phenotypic effect was independently confirmed by other methods (time-lapse microscopy and phenotypic assays using the intron-deletion mutant in Extended Data Fig. 8).

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this paper.