Further support for aneuploidy tolerance in wild yeast and effects of dosage compensation on gene copy-number evolution.

In our prior work by Hose et al., we performed a genome-sequencing survey and reported that aneuploidy was frequently observed in wild strains of S. cerevisiae. We also profiled transcriptome abundance in naturally aneuploid isolates compared to isogenic euploid controls and found that 10-30% of amplified genes, depending on the strain and affected chromosome, show lower-than-expected expression compared to gene copy number. In Hose et al., we argued that this gene group is enriched for genes subject to one or more modes of dosage compensation, where mRNA abundance is decreased in response to higher dosage of that gene. A recent manuscript by Torres et al. refutes our prior work. Here, we provide a response to Torres et al., along with additional analysis and controls to support our original conclusions. We maintain that aneuploidy is well tolerated in the wild strains of S. cerevisiae that we studied and that the group of genes enriched for those subject to dosage compensation show unique evolutionary signatures.

Introduction

In our previous work (Hose et al., 2015), we reported from a genome-sequencing survey that aneuploidy was frequently observed in wild strains of S. cerevisiae, and that 10–30% of amplified genes, depending on the strain and affected chromosome, show lower-than-expected expression in naturally aneuploid isolates compared to isogenic euploid controls. We argued that this group is enriched for genes whose expression is regulated by one or more modes of gene-dosage compensation, which we validated at several genes. Importantly, the group of genes we identified shows non-random associations, including enrichment for specific functional groups, enrichment for genes that are toxic when over-expressed in the lab strain, and a higher rate of gene copy-number variation (CNV) despite higher constraint on expression variation in wild isolates of yeast. A main point of our paper was that dosage compensation at select genes has an important evolutionary effect, by buffering expression against CNV in wild strains.

Torres et al. take issue with the conclusions in our manuscript, arguing that a) wild strains of yeast are not tolerant of aneuploidy and our results emerge as an artifact of culturing wild strains in the lab, and b) errors in our statistical analysis explain the reduced expression of amplified genes and instead there are no genes subject to dosage-responsive control. We disagree with these assertions and provide additional computational analysis that support our original claims.

Results and discussion

Aneuploidy is relatively common and well tolerated in natural isolates of S. cerevisiae

Torres et al. argue that the wild strains analyzed in our study are not tolerant of aneuploidy, since the chromosome-wide average of the relative Illumina read depth measured for each amplified gene is not precisely 1.0, 1.5, or 2.0-fold higher than the euploid control (see Torres et al. Methods). They take this to reflect heterogeneous populations in which cells in the culture have lost the aneuploidy. However, this is not valid for several reasons. Due to technical biases in Illumina sequencing, it is highly unlikely that the mean value of relative gene copy number across whole chromosomes is a precise integer. Indeed, the plots shown by Torres et al. indicate the expected spread in relative read depth across the amplified chromosomes – similar to the spread in read counts of the euploid chromosomes – with mean values very close to the relative DNA abundance we reported. Furthermore, some genes on the chromosomes are not amplified (particularly those near the telomeres [Hose et al., 2015]), which can also slightly reduce the mean value away from a precise integer.

Regardless of the precise mean copy number values, there can be no doubt from the figures presented by Torres et al. that for the strains we analyzed in our original paper, the vast majority of cells in each culture were aneuploid. We point out that several of the chromosomes and strains highlighted by Torres et al. (Figure 1D-F and Figure 2F-H) were not presented in our manuscript or used in any of our analyses (see Hose et al. Figure 1A for strains and chromosomes used in our study). It is true that some chromosome amplifications, namely in sake strains, are variable (appearing or disappearing) across replicates, and thus these chromosomes (Torres et al. Figure 1F and 2F) were not considered as part of this work. The karyotype of these sake strains may indeed be somewhat ‘unstable’ at the culture level; however, the random appearance of extra chromosomes across replicates again suggests a low fitness cost to aneuploidy and an observable rate of mitotic errors (Zhu et al., 2014). Nonetheless, the vast majority of aneuploidies we reported are relatively stable and maintained at high frequencies over many generations and in the absence of any selection.

Instead, our data show that the wild strains we studied are tolerant of aneuploidy, and that it is the laboratory W303 strain that is highly aberrant. 1) By conservative estimation, 30% of the strains we sequenced are aneuploid – these strains were identified in an unbiased sequencing survey in which aneuploidy was not generated or selected for. 2) The aneuploid strains we studied show little growth reduction compared to isogenic euploid strains, both for naturally aneuploid isolates and strains for which we artificially generated aneuploidy (Hose et al., 2015). (In cases where we cited the specific growth rate, we always verified that aneuploidy remained at the end of the experiment by relative qPCR.) Thus, aneuploidy tolerance is not due to unusual adaptation in the lab. In contrast, the tetrasomic W303_Chr12-4n strain – carrying a chromosome reported to be one of the least toxic in this background (Sheltzer et al., 2012) – has a 70% reduction in growth rate compared to its isogenic control (Hose et al. Figure 5B). 3) While extra chromosomes can be lost stochastically, it generally took >200 generations of growth to detect significant chromosome loss in the wild-strain cultures we analyzed. W303_Chr12-4n cultures reproducibly lose the extra chromosome in one culture passaging (~20 generations). The extreme fitness defect incurred by the aneuploid W303 strain explains the rapid emergence of cells that lose the extra chromosome, since the euploid W303 grows nearly twice as fast and rapidly takes over the culture. 4) Aneuploid W303 cells show aberrant gene expression and strong activation of the environmental stress response (ESR, [Gasch et al., 2000]), regardless of the chromosome amplified (Sheltzer et al., 2012; Torres et al., 2007). But as we published, naturally aneuploid strains simply do not activate the ESR (Hose et al. Figure 2A), indicating that they are not experiencing significant stress compared to their euploid controls.

Taken together, these results strongly support our conclusion that aneuploidy is relatively well tolerated in wild isolates of S. cerevisiae, at least for the strains and chromosomes we studied, but not in W303 as previously published (Sheltzer et al., 2012; Torres et al., 2007; 2010). It is certainly possible that some wild strains will not tolerate aneuploidy, and likely that some chromosome amplifications are more problematic than others, regardless of strain background. The prevalence of aneuploidy in our original study is consistent with other reports of aneuploidy in wild isolates (Tan et al., 2013; Strope et al., 2015; Sirr et al., 2015; Muller and McCusker, 2009) and certainly industrial strains (Bakalinsky and Snow, 1990; Bond et al., 2004; Hadfield et al., 1995). Aneuploidy commonly emerges in response to selective pressure (Mulla et al., 2014); interestingly, a recent study by Filteau et al. showed that aneuploidy was relatively frequent in a wild strain, but not a laboratory strain, subjected to the same selection conditions (Filteau et al., 2015). W303 harbors six auxotrophies, several of which affect but do not entirely explain the intolerance of Chr12 amplification (unpublished). Thus, while many nice studies have been done with the W303 laboratory strain, we caution against using it (or any single strain) as the sole representative of S. cerevisiae (Gasch et al., 2016).

A substantial fraction of amplified genes show lower-than-expected expression

To examine expression effects in naturally aneuploid strains, we measured mRNA and DNA abundance in aneuploid isolates compared to closely related or isogenic euploid controls. We did three sets of analyses in Hose et al. We first surveyed expression in six naturally aneuploid strains compared to paired euploid relatives, in biological duplicate. Pooling data across the strains identified 838 out of 2,204 amplified genes (38% over all genes considered in the six strains) whose relative mRNA abundance was reproducibly lower than the relative DNA abundance measured in the strain pairs (Hose et al. Figure 4A). The reduced expression could be because of a response to the aneuploidy, a response to the increased gene copy number, or due to heritable polymorphisms that reduce expression – we emphasize the presentation in Hose et al. that these genes should not be taken as dosage compensated from this analysis alone.

To distinguish the above possibilities, we performed two analyses on isogenic strain pairs, in which the only difference between strains was chromosome copy number. First, we measured mRNA abundance (in biological duplicate) for three aneuploid strains and their isogenic euploids and compared relative mRNA abundance to relative DNA abundance measured in those strain pairs (outlined in more detail below). This identified 163 of 882 genes (or 18% of the total set of genes assessed across these three strains) with lower-than-expected expression in aneuploid cells (Hose et al. Figure 4B and Supplementary File 3). Second, we generated isogenic strain panels for two other naturally aneuploid strains, in which diploid cells carried two, three, or four copies of the amplified chromosome. Using a mixture-of-linear regressions (MLR) model, we defined genes whose relative mRNA abundance did not increase proportionately to relative DNA abundance measured in the aneuploid versus isogenic euploid strains. The MLR analysis identified 172 genes in this class out of 773 genes (or 22% of the genes assessed in these two strain panels, Hose et al., Table 1, Class 3a). We then combined the gene groups from the two analyses of isogenic strain sets for downstream analysis. Because the paired strain analysis was less stringent (in part because only duplicates were analyzed), we required that genes be identified in both Figure 4B (isogenic strain pairs) and Figure 4A (non-isogenic strain pairs). This left 73 genes whose expression was lower-than-expected in four biological measurements from isogenic strain pairs plus the 172 genes identified by the more sensitive MLR analysis from the isogenic strain panels, for a combined total of 245 genes out of 1655 (15%) total genes assessed in the two analyses of isogenic strains. In Hose et al., we cited that 10–30% of genes, depending on strain and chromosome, met our criteria. There was an error in the abstract of the published article that we request to be changed, where '>30%' should have read 'up to 30%' of genes may be subjected to dosage compensation. We regret the error, but point out that the correct values are cited clearly throughout the manuscript.

Torres et al. disagree with our analysis methods, stating that i) the data were misnormalized, ii) the thresholds used were not valid, iii) the mixture of linear regressions (MLR) model was inappropriate, and iv) we did not correct for false discovery. They further propose that expression differences in aneuploid strains fit a normal distribution across the transcriptome, which they take as a null model for no dosage compensation. Below, we briefly address each of these points in turn.

First, as published in our original manuscript, several normalization methods were compared to ensure accuracy, including RPKM, RPKM excluding the amplified chromosomes, and the most accurate method of normalizing based on the number of collected cells. The latter is done by doping a fixed number of Schizosaccharomyces pombe cells relative to carefully counted S. cerevisiae cells in the collection, such that the S. cerevisiae sequencing reads can be scaled according to the known distribution of Sz. pombe reads across the samples (see Hose et al. for specifics). This is the most appropriate method of normalization, since it makes no assumptions about the data; however, it is particularly challenging for wild yeast that are often flocculent and difficult to count. Nonetheless, for several of the strains we analyzed the data normalized by Sz. pombe ‘spike-in’ agreed as well with the same data normalized by RPKM as did biological replicates normalized by RPKM (Table 1). (The exception was YJM428_Chr16 for which the spike-in normalization was clearly off based on comparing calculated Chr16 relative abundance to qPCR-measured Chr16 relative abundance from the same culture pairs, not shown). Data values were otherwise similar for different normalization methods, indicating that RPKM is a valid approach (Hose et al., 2015). Importantly, for our final analysis the DNA and RNA samples were normalized with the identical RPKM procedure – this produced global data centers (i.e. mean log2 values across all measurements) that were very similar for both mRNA and DNA measurements, indicating that the data were normalized in a comparable manner. Thus, there is no evidence that misnormalization of the data dramatically skewed our results.

Table 1.

The center (mean) of log2 distributions for mRNA or DNA ratios measured across all unamplified genes in isogenic aneuploid-euploid strain comparisons are shown. mRNA data from each strain was normalized by either ‘spike-in’ of Sz. pombe cells to the S. cerevisiae cell collections or by reads-per-kb per million mapped reads (RPKM) of S. cerevisiae genes only. Normalized data were then compared across strain pairs to provide a log2 ratio of relative mRNA or DNA abundance for each gene.

DOI: http://dx.doi.org/10.7554/eLife.14409.003

	Mean log2 mRNA ratios (spike-in)	Mean log2 mRNA ratios (RPKM)	Mean log2 DNA ratios (RPKM)	1 SD used for threshold
T73_Chr8-4n vs -2n rep1	0.08	−0.019	−0.072	0.197
T73_Chr8-4n vs -2n rep2	0.102	−0.020	n.a.
YJM428_Chr16-4n vs -2n rep1	−0.401	−0.037	−0.121	0.168
YJM428_Chr16-4n vs -2n rep2	−0.457	0.004	n.a.
YPS163_Chr8-2n vs -1n rep1	n.a.*	-0.146	0.045	0.242
YPS163_Chr8-2n vs -1n rep2	n.a.*	−0.135	n.a.
NCYC110_Chr8-3n vs -2n rep5	0.014	−0.005	n.a.
NCYC110_Chr8-4n vs -2n rep5	0.273	0.186	n.a.
NCYC110_Chr8-4n vs -2n rep1	n.a.	0.058	−0.076
NCYC110_Chr8-4n vs -2n rep2	n.a.	0.011	−0.108
NCYC110_Chr8-4n vs -2n rep3	n.a.	0.078	n.a.
YPS1009_Chr12-4n vs -2n rep1	n.a.	0.014	0.003
YPS1009_Chr12-4n vs -2n rep2	n.a.	-0.094	−0.229
YPS1009_Chr12-4n vs -2n rep3	n.a.	−0.184	n.a.

*We attempted spike-in normalization for haploid YPS163-disomic (‘2n’) and -monosomic (‘1n’) strains but were unable to accurately count cells due to differences in flocculation across aneuploid-euploid strains. As described in Hose et al. (2015), RPKM normalization produced data that agreed as well or better across replicates compared to spike-in normalization and in the case of YJM428_Chr16-4n agreed better with qPCR measurements of Chr16 abundance in the culture (not shown). Note data from NCYC110 replicate (rep) 5 were not used in the analysis but were generated during the Hose et al. manuscript revision stage for normalization controls. The SD of DNA ratios on the affected chromosome that were subtracted from gene-level measurements of relative DNA abundance (i.e. to generate the gene-specific thresholds, see text) are shown for reference where relevant.

Second, the description of our analysis methods presented by Torres et al. does not recapitulate what was done in our manuscript. For isogenic strain pairs shown in Hose et al. Figure 4B, we generated biological duplicate RNA-seq replicates and single DNA-seq samples, where DNA and one RNA sample were taken from the same culture, for both the aneuploid and euploid strains. For each replicate, we measured the relative mRNA abundance for each gene in the aneuploid versus euploid strain, as well as the relative DNA abundance for that gene in the aneuploid versus euploid strain. We then defined a gene-specific threshold to identify genes with lower-than-expected mRNA abundance: we took the relative DNA abundance measured for a given gene, minus one SD of the chromosome-wide mean of relative DNA abundances (Table 1). We then identified genes whose relative mRNA abundance in the aneuploid versus euploid strain was lower than the gene-specific cutoff in both biological mRNA replicates. Comparing measured mRNA ratios to measured DNA ratios (as opposed to theoretical DNA ratios) is critical to capture systematic and stochastic variation in the measurements. Our thresholding method allowed us to account for gene-specific biases in sequencing counts while incorporating measurement noise (and minimizing sequencing costs). Importantly, to meet our criteria for downstream analysis, a gene must have also met these criteria in the comparable analysis of non-isogenic strain pairs (Hose et al. Figure 4A). Thus, each gene had to be expressed below the threshold in four biological replicates. We note that it would be inappropriate to analyze the mean gene-level values from small numbers of replicates, which was not done in our analyses. In attempt to estimate the false discovery rate (FDR), we performed random permutations on the data (see Methods). We estimate the FDR for this analysis to be below 15%.

Torres et al. argue that our approach for selecting thresholds is not valid, citing the high standard deviation (SD) for mRNA abundance levels (RPKM) across the transcriptome (Torres et al. Figure 8B). However, we did not use abundance values, but rather relative abundance across isogenic strains, and thus the SD of RPKM values across all transcripts is not relevant to our analysis. What is relevant is the SD of replicate mRNA abundance ratios compared to the SD of the relative DNA values used to define the threshold applied to replicates. The average SD of the replicate mRNA ratios for each gene ranged from 0.12–0.3, for amplified genes and for unamplified genes. This was the same range as the SDs of the relative DNA abundance ratios (0.17–0.24) used to define the gene-specific thresholds (Table 1), which were applied to four biological measurements of relative mRNA abundance. Torres et al. attempt to estimate the false discovery rate of our method on permuted data; while their methods are not entirely clear, they identify hundreds more genes on randomized data than we did on real data, and the SD values cited in their Table 2 are not close to our values, suggesting critical differences in methods.

In a second analysis, we generated two sets of isogenic strain panels in which diploid cells carried two, three, or four copies of a chromosome. Triplicate mRNA and duplicate DNA measurements were generated for each strain in each panel, and the data were fit using a sensitive mixture of linear regressions (MLR) model, which did not use the same cutoff approach to call genes. In fitting the MLR model, we did not average the replicates to reduce data to three points in a scatterplot per gene – such an approach could be adversely affected by outliers. Instead, we used all replicate measurements on all genes simultaneously, contrary to the assertion in Torres et al. The model takes into account variation in the replicate mRNA measurements (see Hose et al. for details). For simplicity, we classified genes based on their maximum posterior probability, and no cutoff on the magnitude of the expression effect was used – indeed, we showed in the paper that many genes have only partially reduced expression (Hose et al. Figure 4 and 6). Though it was not a focus of the original report, the MLR mixture model allows for a conditional FDR assessment, following the approach in Newton et al. (2004). The 142 genes called Class 3A in the YPS1009 panel correspond to 17% FDR; the 30 genes called Class 3A in the NCYC110 panel correspond to 5% FDR by this method.

An independent check on MLR computations comes by using ordinary linear regression on the same input data, and using the regression to test the null hypothesis (per gene) of no deviation from proportional expression (i.e. the null is intercept=0 and slope=1). Under normal theory, the likelihood ratio test statistic has a chi-squared distribution on 2 degrees of freedom. An estimate of the proportion of null genes comes by processing the collection of chi-square p-values through Storey's q-value calculator (Storey, 2003). For both strain panels, these null proportion estimates are remarkably close to the Class 1 proportion estimates from MLR (in the YPS1009 panel: 14% by q-value and 15% by MLR; in the NCYC110 panel: 7% by q-value and 8% by MLR, both among the set of genes showing linear relationships on the log scale mRNA versus DNA). The likelihood ratio approach does not offer a direct classification of non-null genes, in contrast to the favored MLR approach used in Hose et al., but does provide a check on inferences about the breath of putative-dosage compensation effects. In summary, the two methods we used to identify genes expressed lower-than-expected in isogenic aneuploid versus euploid strains identified 245 genes for downstream analysis. This amounts to 10 to 30% of genes depending on the strain and affected chromosome, or 15% of all genes assessed, at an FDR between 5 and 17%.

Torres et al. perform their own analysis to identify potentially dosage compensated genes, focusing on global averages, correlations, and distributions across the chromosomes. They argue that the distribution of expression effects across each aneuploid strain versus its isogenic euploid fits a normal distribution, in which there is a similar number of genes with higher expression on the affected chromosome as genes with lower expression. The authors cite that this is the null expectation in the absence of any genes subject to dosage compensation. While such a global approach would capture chromosome-wide dosage compensation as seen in sex chromosomes, we argue that this approach will miss many individual genes potentially subject to dosage control.

As published in Hose et al., there are genes on and off the amplified chromosome that have higher expression in the aneuploid strains. The MLR analysis reported nearly the same number of amplified genes with higher expression as with reduced expression (Hose et al. Table 1). We regret the magenta coloring to highlight amplified genes in Figure 4, which was added during the revisions to make a separate point and thus the genes were not selected with the same criteria as genes with lower expression. Nonetheless, as we describe in Hose et al., the two tails of the distribution are enriched for different functional groups: whereas amplified genes with lower-than-expected expression are enriched for particular functional groups (see below) and for genes that are toxic when over-expressed in the lab strain, the set of amplified genes with higher-than-expected expression is enriched for genes encoding membrane and cell-surface proteins. As we cited, the higher-than-expected expression of amplified genes (and of genes on unamplified chromosomes) is likely an indirect response to the known influence that ploidy has on cell size/shape and flocculence (Wu et al., 2010). Therefore, we do not believe that the distribution across all genes can be used to select individual genes subject to dosage compensation, nor did we focus on global averages or correlations. Instead, we identified individual genes that have the lower-than-expected expression phenotype and analyzed the pooled set to explore enriched features, as described below.

Unique features of selected genes suggest the group is enriched for dosage-compensated genes

A major point of our original manuscript that was not addressed by Torres et al. is that the genes we identified through the above analyses are strikingly enriched for unique patterns. The set of genes with lower-than-expected expression in aneuploid strains is strongly enriched for certain functional groups, including genes known to feedback on their own expression (see more below), genes that are toxic when over-expressed, and genes that are under higher expression constraint but display elevated copy number variation (CNV) in wild populations (Hose et al., 2015). Such non-random associations are unlikely to occur if the gene selection was truly random and driven by noisy data. We ruled out the possibility that these genes reflect a common, indirect response to aneuploidy. However, it is true that a subset of the genes we identified could be responding transcriptionally to the particular chromosome amplified, rather than the gene’s copy number per se. We attempted to minimize these effects in our original downstream analysis by pooling amplified genes from five aneuploid strains and across three amplified chromosomes. Here, we provide an additional series of controls to show that the trends seen for the gene group we identified cannot be explained as indirect transcriptional responses to aneuploidy.

As presented by Torres et al., chromosome-specific responses to aneuploidy should affect groups of functionally related genes, whether or not the genes are amplified. In an attempt to remove amplified genes that may be a part of such indirect responses, we first identified unamplified genes in each aneuploid strain with significantly reduced expression compared to its isogenic euploid (FDR 0.01) and identified enriched functional groups (see Methods). We then removed from consideration all amplified genes in that strain belonging to any of those strain-specific functional groups. (Note: we consider cytosolic and mitochondrial translation factors as separate groups). This analysis removed ~30% of the genes classified as potentially dosage compensated in the original analysis; the majority of these were linked to mitochondrial function and is consistent with the common mitochondrial response we reported (even though, interestingly, the removed genes were not affected in multiple aneuploid strains). We note that our procedure may remove genes that are legitimately dosage compensated, especially those that show only partial compensation, which could cause indirect effects on unamplified genes in the same functional category (e.g. partial compensation of a transcriptional repressor).

The revised set of amplified genes with lower-than-expected expression remains enriched for functional groups that cannot be explained by common or chromosome-specific responses, including cytosolic RPs (p=6e-7), other proteins involved in ribosome biogenesis (p=5e-4), and an overlapping gene set linked to translation (p=3e-6). RPs are known to feedback on their own expression (Warner et al., 1985; Vilardell and Warner, 1997; Dabeva and Warner, 1987; Dean et al., 1981; Tsay et al., 1988), and we validated the trend at two genes (Hose et al., 2015). The revised group is also weakly enriched for sequence-specific DNA binding proteins (p=0.006), including several transcription factors that bind their own promoters and/or regulate their own expression (including, Hap1, Stb5, Bdf1, and Rsc30 [Venters et al., 2011; Deckert et al., 1995; Hon et al., 2005; Denby et al., 2012]). Importantly, these enrichments remained significant when genes from each chromosome were held out (p<0.007, except for one case where DNA binding protein enrichment p=0.04), showing that the result is not skewed by one particular aneuploidy. As cited in Hose et al., these categories are not enriched among unamplified genes whose expression is affected by the common aneuploidy response (Hose et al., 2015). Thus, our analysis is clearly enriching for genes subject to dosage compensation.

All of the evolutionary signatures we reported in the original manuscript remain statistically significant on this reduced gene group (Figure 1). The revised gene set enriched for dosage-compensated genes displays significantly higher levels of CNV in S. cerevisiae populations, compared to all genes and compared to amplified genes with proportionately elevated expression (Figure 1A). They also have significantly higher CNV buffering scores (Figure 1B, see Methods and (Hose et al., 2015)), indicating a higher level of CNV despite increased expression constraint. If these trends are driven by genes subject to repression or indirect effects of aneuploidy, we would expect to see the same patterns for unamplified genes that show reduced expression compared to euploid controls. But this is not the case: unamplified genes with reduced expression (identified either by edgeR or by the identical methods used to call dosage-compensated genes) show no increase in CNV propensity or buffering scores compared to the control group. The trends for the affected amplified genes remain if we define genes using a more stringent threshold to call affected genes, if we remove from the analysis all genes belonging to any of the enriched functional groups, or if we separately hold out each amplified chromosome from the analysis (see Methods). A final prediction is that genes with larger effect sizes in terms of the level of compensation should have more striking evolutionary patterns, and this is indeed the case (Figure 1). Therefore, the results we report cannot be explained as a secondary transcriptional response to the amplified chromosome – we believe that the group is legitimately enriched for genes subject to dosage control and that dosage compensation has an important effect on genome evolution.

Figure 1.

Gene sets enriched for dosage-compensated genes show unique signatures.

Gene sets and the numbers contained within them are indicated by the key. The revised list of amplifed genes with lower-than-expected expression were partitioned into thirds, based on genes with the greatest (top third) or smallest (bottom third) reduction in expresison compared to expectation. (A) The fraction of genes in each group for which at least three of 103 strains showed gene amplification. (B) The distribution of Buffering Scores for genes with CNV. Here, Buffering Score represents the number of 103 strains with a gene duplication divided by expression constraint (Vg/Vm, see Hose et al. for details). Higher values indicate a higher propensity for CNV despite expression constraint. The orange line indicates the median value for amplified genes with proportionate expression as a reference point. Asterisks indicate statistical significance (p<0.035) compared to the amplified genes with proportionate expression. In some cases, the trends were consistent but not significant (likely owing to small sample sets).

DOI: http://dx.doi.org/10.7554/eLife.14409.004

Our original intention in Hose et al. was not to claim that dosage compensation is ‘widespread’ in S. cerevisiae, but rather that it exists at a significant number of genes and has an important role in evolution. We certainly agree with Torres et al. that dosage compensation does not function at most amplified genes, as seen for sex chromosomes in other organisms. Our goal was not to define individual genes subject to dosage compensation, but rather to investigate the group – we caution that no single gene from our lists should be taken as dosage compensated without orthogonal evidence. In the end, the revised gene set we identified here at the highest confidence encompasses 157 (13%) genes out of 1,243 interrogated across all the isogenic strain sets. It is true that some of the genes we identified, particularly those with only subtly reduced expression patterns, are false positive calls from our analysis. However, the number of genes we identified is inline with the 14% of dosage compensated genes reported by Springer and colleagues (Springer et al., 2010) (using GFP reporters for 730 genes interrogated) and is consistent with (but on the lower end) our original report of 10–30%. Future work will undoubtedly clarify the precise number and identity of genes subject to dosage compensation, as well as the genetic basis for phenotypic differences in aneuploidy tolerance in laboratory versus wild strains.

Methods

In the process of this work, we verified our original analyses to define gene groups. We estimated the FDR of our thresholding method as follows: 1) We calculated the log2 ratio of mRNA in aneuploid versus euploid cells, minus the corresponding log2 ratio of relative DNA abundance for that gene minus 1SD of the chromosome-wide mean of relative DNA values. A negative value indicates that the relative mRNA expression was below the relative DNA copy number minus 1SD. 2) These difference values were randomly permuted with regard to the gene labels, within each of two replicates in both the isogenic and non-isogenic strain pairs. 3) We calculated the number of genes for which all four of the randomized data values were negative, over 10,000 iterations. The FDR (average number of genes identified from 10,000 iterations on random data divided by the number of genes identified in real data) was calculated to be below 15.0%. We note that some fraction of true positives will meet the threshold in randomized trials.

We also generated a revised list of putatively dosage compensated genes as follows: Unamplified genes with significant expression differences in each strain were identified using edgeR (Robinson et al., 2010) comparing RPKM values in aneuploid versus isogenic diploid strains. Biological duplicate RNA-seq data were analyzed for T73_Chr8-4n versus T73_Chr8-2n, YJM428_Chr16-4n versus YJM428_Chr16-2n, and YPS163_Chr8-disomic versus YPS163_Chr8-monosomic; biological triplicates were analyzed for YPS1009_Chr12-4n versus YPS1009_Chr12-2n and for NCYC110_Chr8-4n versus NCYC110_Chr8-2n (see Hose et al. for all other details). Unamplified genes with an FDR <0.01 (Robinson et al., 2010) and a negative mean log2(fold-change) in aneuploid versus euploid expression were taken as 'lower expressed'. Functional enrichment was done on the set of lower-expressed unamplified genes for each strain separately, using the program FunSpec (Robinson et al., 2002) and taking p<0.0004 as significant. Amplified genes from that strain belonging to any of the strain-specific enriched functional categories (excluding those based on cellular localization) were removed from the original list of putatively dosage compensated genes. Note that we considered cytosolic and mitochondrial RPs and translation factors as separate groups. One of the strains (YJM428_Chr16-4n) had a substantial secondary response at unamplified genes, amounting to two thirds of the 1,456 lower-than-expressed unamplified genes pooled across the isogenic strain pairs – the effected genes from YJM428_Chr16-4n were admittedly enriched for translation factors and rRNA biogenesis genes, causing the removal of amplified Chr16 genes in those categories from the list of affected genes. Nonetheless, the functional categories described in the text remained significant for the remaining 157 genes with reduced expression in aneuploid strains. For analysis in Figure 1, the genes were ranked and partitioned into thirds based on the magnitude of expression reduction (‘effect size’, where ‘top third’ represents amplified genes with the strongest reduction in expression), taken as the average relative mRNA ratio for the isogenic aneuploid and euploid strains compared to the relative DNA ratio between those strains.

As a separate approach to call putatively dosage compensated genes, we used the same methods described for Figure 4B of Hose et al., except that we selected genes whose relative mRNA values were less than the relative DNA ratio measured for that gene minus 3 SD of the chromosome-wide mean in that strain. A gene had to be below that gene-specific threshold in all (two or three where relevant) biological replicates of isogenic strain pairs. This identified 56 genes across the five strains. We applied an identical method to identify non-amplified genes with reduced expression; for this, we used the same SD as applied for amplified genes (although this was generally very similar to the SD for all genes on unamplified chromosomes). This identified a total of 454 unamplified genes with lower expression (‘3SD’ in Figure 1A).

We plotted the fraction of genes for which at least three of 103 strains previously analyzed had evidence of gene amplification (see Hose et al. for details). Trends were similar when plotting the fraction of genes in which at least two strains displayed CNV. For simplicity, the buffering score plotted in Figure 1 represents the number of 103 strains with CNV divided by the expression constraint score Vg/Vm (see Hose et al. for details). We performed a variety of other controls, ensuring that trends were the same when each chromosome was held out separately, when genes belonging to any functional group were held out, and using different cutoffs for CNV. The trends were consistent in all cases (although not always statistically significant owing to small datasets in some cases).

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your work entitled "Further support for aneuploidy tolerance in wild yeast and effects of dosage compensation on gene copy-number evolution" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by Duncan Odom as the Reviewing Editor and Randy Schekman as the Senior Editor.

The reviewers have discussed the reviews with one another and Dr Odom, the Reviewing Editor, has drafted this decision to help you prepare a revised submission. Despite the criticisms below, all reviewers in this eLife discussion board unanimously shared a commitment to allowing you a fair and open response in print to the scientific challenge offered by Torres et al.

Because your submission is an author-response to Torres et al., a manuscript that is currently in press at eLife challenging your prior work, I have appended the complete reviews from three independent reviewers, as well as one notable (anonymized) comment from our discussion. Note that lightly edited versions of these reviews will be available for the community to see. There were a number of criticisms of your lines of reasoning and the clarity of your presentation, as well as suggested experiments by Reviewer 1.

For re-submission of this Short Report, the only required changes are:

a) To revise the clarity and argument structure, both of which were considered at times weak by all the reviewers

b) Slowly cross-check all your numbers, as well as your references to numbers in your original work, and explain any discrepancies clearly.

c) Rewrite the Abstract to clarify, per Reviewer 3.

d) Revise your Discussion to admit possible weaknesses in your analyses. Do please consider including some or all of the following points: that the number of replicates used may be too low to reliably use means, that some of the strains may well be unstable, and all things considered, the number of genes that are dosage-compensated could be quite small – and are likely to be smaller than claimed in the original Hose paper.

Dr Odom, the Reviewing Editor, suggests that careful (and extensive) experimentation that begins from the types of experiments suggested by Reviewer 1 could serve as the nucleus of a more comprehensive story for a later submission that is entirely independent of the current dosage compensation discussion.

Reviewer #1:

“Further support for aneuploidy tolerance in wild yeast and effects of dosage compensation on gene copy-number evolution” is a rebuttal of a previous manuscript of Torres, Springer and Amon, which was itself a rebuttal of a paper published by Gasch, Hose et al. in 2015. It is a sad reality that the two groups were not able to find a common ground for a professional discussion that would result in a collaborative manuscript addressing the key issues. The reviewers and editors have to help to ensure that the rebuttal remains factual and professional, supported by clear arguments and genuine attempts to find a solution.

The rebuttal focuses on two main points. First, they argue that, oppositely to the claim of Amon and Torres, aneuploidy is indeed more tolerated in wild budding yeast strains. This claim is irrelevant, as Amon and Torres did not question the claim that wild yeast strains are more tolerant to aneuploidy (they questioned that this tolerance is due to the ability to compensate the gene dosage). In the second part, they argue that their analysis of dosage compensation of mRNA levels of amplified genes is correct and that a significant proportion of genes is compensated in wild yeast strains. This part is confusing and flawed. Several points need to be addressed:

1) Subheading “A substantial fraction of amplified genes show lower-than-expected expression”: It is questionable that the normalization strategies used for mRNA and genomics NGS data should be identical. RPKM is used exclusively for mRNA normalization, as far as we know. It is not clear how "careful cell counting and doping with Schicosaccharomyces pombe cells at the time of collection" should have helped to improve the normalization. This requires further explanation. Finally, log2 difference of paired RNA-DNA sample at -0.04 feels dangerously low – this similarity can rarely be found even between technical replicates. Using this precise number as a comparison between two very different sequencing experiments seems to be in stark discrepancy to the argument in paragraph one, subheading “Aneuploidy is relatively common and well tolerated in natural isolates of S. cerevisiae”, where the authors state that due to technical biases in Illumina sequencing the mean value isn´t a precise measurement and can´t be used to conclude on a unbalanced karyotype.

2) Paragraph four, subheading “A substantial fraction of amplified genes show lower-than-expected expression”: Definition of the cut off is the critical point. Why the cut off for mRNA levels should be based on DNA abundance is beyond our understanding – the variance of DNA abundance and variance of RNA abundance are usually very different. The rationale is not explained.

3) The next part of the paragraph brings a flurry of numbers that is difficult to follow. They first state that they identified 163 of 882 genes tested to be dosage compensated – and not 838, as stated by Torres. But in the Hose et al. manuscript, there is indeed the number 838, and the number 163 or 882 is nowhere to be found. They also claim that their subsequent analysis defined stringent criteria for dosage compensation to be met by 88 genes out of 738. They cite Figure 4A in Hose et al., but again, this number is not there. This means dosage compensation in 12% of amplified genes, but at 15% FDR. None of these numbers was mentioned in the previous manuscript.

4) Paragraph five, subheading “A substantial fraction of amplified genes show lower-than-expected expression”: Why the SD across all transcripts is not relevant is beyond our understanding. Both SDs across replicates and SD across all transcripts within one sample are relevant, each of them provides different information.

5) Paragraph six, subheading “A substantial fraction of amplified genes show lower-than-expected expression”: The criticism of MLR by Torres and Amon was based on the fact that only 3 samples were used for the analysis, which is not sufficient; here it is claimed that 12 samples were used – however, this number includes the biological replicates of 3 samples only, which is indeed not sufficient. All the additional analysis does not help the fact that the model is based on 3 samples.

6) They agree with Torres and Amon that the distribution of the gene expression of amplified genes is normal (not skewed), yet argue that the downregulated genes represent dosage compensation, whereas the upregulated genes are a part of the general response to aneuploidy – it is very unclear why each of these two outliers' groups should be interpreted differently. Their argumentation in the chapter "Unique features of selected genes...." is a prime example of scientific tautology: genes are dosage compensated because it is disadvantageous for the cells to overexpress them. Accordingly, for these genes a propensity to CNV is observed. They do not observe propensity to CNV for genes from these categories when they are present in normal copy numbers – but this is most likely because they are not overexpressed and therefore there is no need to select against them! They also never say what gene categories are downregulated among unamplified genes (likely the same categories as on the amplified chromosomes).

7) There are also genes that are up-regulated – the identified categories broadly agree with the previous gene categories identified to be upregulated in response to aneuploidy (they cite the work from Fink lab, which relates to tetraploidy, but see more relevant Amon, Torres, and Storchova lab results). Remarkably, the dosage compensated genes closely match the categories identified previously to be downregulated in response to aneuploidy – ribosome biogenesis, translation, transcription – in several different species (see Amon, Torres, Storchova, Foijer laboratories). This option should be considered.

8) In paragraph five, subheading “Unique features of selected genes suggest the group is enriched for dosage compensated genes”, they cite a paper by Springer et al. (Mol Sys Biol 2010) in support of the presented work. But Springer's manuscript is focusing entirely on dosage compensation on protein level, whereas Hose considers mRNA; of individual proteins only, measured one-by-one by FACS, whereas Hose et al. considers whole chromosomes only, and Springer analyzes loss of copies, whereas Hose analysis gain of gene copies.

In conclusion, the arguments do not seem strong enough to justify their analysis approach. Secondly, even if we would agree with the analysis, they found that the mRNA of 12 – 13% of genes might be dosage compensated at a FDR of almost 15% – a result that in fact says that there is no general dosage compensation.

Reviewer #2:

This short paper by Gasch, Hose and colleagues is a reply to a paper by Torres and colleagues, which is in turn a reply to a first paper by Hose, Gasch et al. in which they report that several wild yeast strains show high levels of aneuploidy and that the expression of some genes does not reflect their copy number changes, suggesting that these wild yeasts have some mechanisms of gene dosage compensation that reduces the expression effect of copy number variation (which may explain why such high levels of aneuploidy are tolerated).

The paper by Torres and coworkers notes two key issues in the work by Gash et al:

1) The copy number variation is based on changes in read depth of illumina genome sequencing. However, the read depth does not always change in integer numbers, which may indicate that only part of the population shows the CNV and that the populations are heterogeneous.

2) The statistical methods used by Hose, Gasch et al. are not correct – most notably, Torres and coworkers argue that Gasch et al. do not account for multiple testing and do not use appropriate error calculations, since the expression levels were calculated with very noisy mRNA abundance levels (RPKM).

Gasch, Hose et al. now argue that:

1) The non-integer changes in read depth are due to sequencing bias

2) The error calculations were OK since everything was based on (less noisy) relative abundance levels.

Major comments:

1) As I already indicated in my review of Amon et al., I am also puzzled by the non-integer changes in sequencing depth. I do think that this suggests that the strains could indeed be unstable. However, I also noted that we need to be sure that sequencing biases are not involved. I suggested a few relatively easy experiments to verify whether the populations are homogeneous and whether they show the expected aneuploidies. It seems easy enough to measure the copy number of a few key regions in a few key strains using qPCR, and to do this on a few different (small) colonies of the same culture. This would close the discussion.

2) I follow the arguments about the error calculations. Although I am not an expert in this area, I think that, Gasch, Hose and coworkers indeed used correct calculations, but their sample size (number of replicates) are very low. Again, a few extra experiments using qPCR would help to close the discussion.

3) The number of genes that may show dosage compensation (245) seems quite low, especially since many of these might be false (false discover rate…). Even if some GO enrichment is found in this set, it is not clear to me that the number is high enough to call this "widespread". My best guess is that strains become aneuploid and the relatively quickly acquire mutations that normalize the expression of a few key genes for which changes in expression are not tolerated. Whether this is special for wild yeasts and should be called "dosage compensation" is a difficult question (obviously, the dosage is compensated, but using the name also suggests a general, dedicated mechanism, and this may very well be lacking…)

Reviewer #3:

This manuscript by Gasch et al. is a rebuttal to the Torres et al. criticism of Hose et al. Both Hose et al. and Torres et al. were published in eLife. I will preface by saying that I firmly believe Gasch should have the opportunity to present a response (this manuscript).

At the heart of this dispute is to two points: namely (1) is aneuploidy relatively common in wild strains and well tolerated there? and (2) to what extent (fraction) do genes show a lower-than-expected expression?

Here Gasch provides a strong rebuttal of the Torres concerns regarding whether aneuploidy is common and tolerated in wild strains. In my opinion the Torres arguments on this point were weak and Gasch clearly articulates the relevant counter argument.

On the second point, namely to what extent do genes show a lower than expected expression, Gasch clearly outlines the details of the methods used in Hose and why the criticisms of Torres were off the mark at times. Gasch also correctly points out the Torres methods for estimating FDR were poorly described. Gasch provides additional details on the methods of Hose and additional, more stringent analysis that reaches many of the same conclusions of Hose.

One notable (anonymized and lightly edited) comment from the reviewers’ discussion:

On the second issue, the extent of dosage compensation, I think that there is a fundamental disconnect between the groups in how to assess dosage compensation. Gasch (in Hose and Gasch) seek to take a gene specific view that accounts for gene-to-gene variability. This is noble and philosophically sound, but fraught with problems they don't address – namely that they have way too small a sample size to really do this (as also noted by both other reviewers). The Amon approach is to take a more distribution-based approach – which is, in principle, better in low sample sizes – but as Gasch correctly points out (and was a major criticism of the Torres manuscript) – the Torres methods are so poorly described as to be impossible to assess. Even if you think the Gasch method has merit, in the end they do back peddle on the number of dosage compensated genes (in Gasch compared to Hose) leading one to suspect we are talking about something that is either a small class or even non-existent.

For re-submission of this Short Report, the only required changes are:

a) To revise the clarity and argument structure, both of which were considered at times weak by all the reviewers

b) Slowly cross-check all your numbers, as well as your references to numbers in your original work, and explain any discrepancies clearly.

c) Rewrite the Abstract to clarify, per Reviewer 3.

d) Revise your Discussion to admit possible weaknesses in your analyses. Do please consider including some or all of the following points: that the number of replicates used may be too low to reliably use means, that some of the strains may well be unstable, and all things considered, the number of genes that are dosage-compensated could be quite small – and are likely to be smaller than claimed in the original Hose paper.

We have revised the text in several places, providing additional detail as well as a new Table 1 to clarify our explanations and reasoning. We have also provided a more detailed summary of Hose et al. including the genes selected at each step and references to where those gene lists/data were reported in Hose et al. We double-checked the values from the original paper and clarified the Abstract as suggested by Reviewer 3. We hope that these changes have clarified our arguments in these places.

We have also revised the Discussion in several places as requested: We added a statement in paragraph two, subheading “Aneuploidy is relatively common and well tolerated in natural isolates of S. cerevisiae” that some of the strain karyotypes may be unstable; we added a sentence to the last paragraph that states that some of the genes on our list will be false positives, and another statement that our revised gene set is on the lower end of what we originally reported when considering all genes as a single group. These changes are in addition to a statement that individual genes from our list should not be taken as dosage compensated without orthogonal evidence. We also added a statement to paragraph five, subheading “A substantial fraction of amplified genes show lower-than-expected expression” stating that it is inappropriate to use mean gene-level expression values from small numbers of replicates, which was not done for any of our analysis.

Reviewer #1:

“Further support for aneuploidy tolerance in wild yeast and effects of dosage compensation on gene copy-number evolution” is a rebuttal of a previous manuscript of Torres, Springer and Amon, which was itself a rebuttal of a paper published by Gasch, Hose et al. in 2015. It is a sad reality that the two groups were not able to find a common ground for a professional discussion that would result in a collaborative manuscript addressing the key issues. The reviewers and editors have to help to ensure that the rebuttal remains factual and professional, supported by clear arguments and genuine attempts to find a solution.

We wholeheartedly agree with the reviewer on these points.

The rebuttal focuses on two main points. First, they argue that, oppositely to the claim of Amon and Torres, aneuploidy is indeed more tolerated in wild budding yeast strains. This claim is irrelevant, as Amon and Torres did not question the claim that wild yeast strains are more tolerant to aneuploidy (they questioned that this tolerance is due to the ability to compensate the gene dosage).

In fact, Torres et al. spend considerable effort to argue that the aneuploid strains are ‘unstable’ and therefore not tolerant to aneuploidy. As indicated by reviewer #3 below, our response to these points is strong and convincing.

1) Subheading “A substantial fraction of amplified genes show lower-than-expected expression”: It is questionable that the normalization strategies used for mRNA and genomics NGS data should be identical. RPKM is used exclusively for mRNA normalization, as far as we know. It is not clear how "careful cell counting and doping with Schicosaccharomyces pombe cells at the time of collection" should have helped to improve the normalization. This requires further explanation. Finally, log2 difference of paired RNA-DNA sample at -0.04 feels dangerously low – this similarity can rarely be found even between technical replicates. Using this precise number as a comparison between two very different sequencing experiments seems to be in stark discrepancy to the argument in paragraph one, subheading “Aneuploidy is relatively common and well tolerated in natural isolates of S. cerevisiae”, where the authors state that due to technical biases in Illumina sequencing the mean value isn´t a precise measurement and can´t be used to conclude on a unbalanced karyotype.

We apologize for the confusion, as the description of the doping normalization was included in the original Hose et al. manuscript. We have added a clearer description and referencing for this normalization procedure to paragraphs two and three, subheading “A substantial fraction of amplified genes show lower-than-expected expression”. We also added a table to present the centers (means) of log2 distributions for normalized data, which this reviewer may have misunderstood. The important points from this section are: 1) the most accurate method of Sz. pombe spike-in normalization gives comparable normalization to the more robust RPKM (aside of one strain, for which the spike-in normalization was clearly off, for reasons we describe in the text); 2) the center of the distributions for RPKM-normalized mRNA ratios was very similar to the center of distributions for RPKM-normalized DNA ratios; 3) the SDs used for thresholding are significantly higher than any difference in data mean centers. Thus, there is no evidence that misnormalization of the data has dramatically affected our results. The point raised by this reviewer about capturing technical biases in Illumina data is the precise reason why we compare relative mRNA abundance measured in each strain pair to relative DNA abundance measured in the same way (see more below). We again highlight that the significant enrichment for functionally related genes among the selected group strongly suggests that we have not selected random genes, but rather enriched for functionally related groups.

2) Paragraph four, subheading “A substantial fraction of amplified genes show lower-than-expected expression”: Definition of the cut off is the critical point. Why the cut off for mRNA levels should be based on DNA abundance is beyond our understanding – the variance of DNA abundance and variance of RNA abundance are usually very different. The rationale is not explained.

There is simply no other way to do this analysis – it would be wholly inappropriate to compare a measured value for mRNA (subject to both systematic and stochastic technical variation) to an expected value for DNA abundance. In fact, we would have identified many more genes if we had used an expected ratio of DNA abundance that does not capture measurement biases that can compress measured ratios. We added a statement to more clearly outline our rationale: “Comparing measured mRNA ratios to measured DNA ratios (as opposed to theoretical DNA ratios) is critical to capture systematic and stochastic variation in the technical measurements. Our thresholding method allowed us to account for gene-specific biases in sequencing counts while incorporating measurement noise (and minimizing sequencing costs).”

3) The next part of the paragraph brings a flurry of numbers that is difficult to follow. They first state that they identified 163 of 882 genes tested to be dosage compensated – and not 838, as stated by Torres. But in the Hose et al. manuscript, there is indeed the number 838, and the number 163 or 882 is nowhere to be found. They also claim that their subsequent analysis defined stringent criteria for dosage compensation to be met by 88 genes out of 738. They cite Figure 4A in Hose et al., but again, this number is not there. This means dosage compensation in 12% of amplified genes, but at 15% FDR. None of these numbers was mentioned in the previous manuscript.

We have attempted to clarify this section by outlining in paragraph one, subheading “A substantial fraction of amplified genes show lower-than-expected expression” the three sets of expression analyses done in Hose et al. The first compares aneuploid to non-isogenic euploid strains and identified 838 out of 2,204 (38%) genes with lower-than-expected expression in aneuploid strains, based on biological duplicates. However, these 838 genes include those responding to the aneuploidy and genes whose expression is affected by heritable polymorphisms across the non-isogenic strain pairs. We were careful to explain in Hose et al. that this group does not represent dosage compensated genes. Torres et al. imply this and apparently perform their analysis on this gene set.

We then outline the two analyses we did to call dosage compensated genes: the first compared biological duplicate measurements across three aneuploidy-euploid strain pairs (179 of 882 genes (20%) met these criteria, Hose et al. Figure 4B). The second compared biological triplicate measurements across two strain panels using the MLR model (172 out of 773 genes (22%) met the relevant classification, Hose et al. Table 1). We then cite how the genes from the two analyses were combined: because the paired-strain analysis was less stringent (owing to duplicated instead of triplicated replicates), we required that the genes also be identified from the non-isogenic strain comparisons in Figure 4A, in other words the genes had to pass the threshold in four biological replicates. This left 73 of the 179 genes that were added to the 172 genes from the MLR analysis, for a total of 245 genes for downstream analysis (as outlined in Hose et al. main text). We hope this has clarified our analysis without muddying the text with more numbers.

4) Paragraph five, subheading “A substantial fraction of amplified genes show lower-than-expected expression”: Why the SD across all transcripts is not relevant is beyond our understanding. Both SDs across replicates and SD across all transcripts within one sample are relevant, each of them provides different information.

Torres et al. cite the SD of RPKM values across all transcripts in the cells – transcript abundance across the transcriptome varies several orders of magnitude, and thus the SD of all RPKM values is very large (as it should be). The reason this SD is not relevant is that we never used raw RPKM values for any of our work, but rather the relative RPKM in each aneuploid versus the isogenic euploid. As we presented, the SD of the replicate mRNA ratios and the SD across all mRNA ratios on each affected chromosomes are in the same range as the SD of the DNA ratios. The SDs are certainly important – but only the SDs of the data types being studied.

5) Paragraph six, subheading “A substantial fraction of amplified genes show lower-than-expected expression”: The criticism of MLR by Torres and Amon was based on the fact that only 3 samples were used for the analysis, which is not sufficient; here it is claimed that 12 samples were used – however, this number includes the biological replicates of 3 samples only, which is indeed not sufficient. All the additional analysis does not help the fact that the model is based on 3 samples.

The important point is that the MLR model is fit to all of the data for a given panel at once, which includes three ratios for each gene in each of three strain comparisons plotted against the measured DNA. We argue that fitting a linear model across nine relative mRNA measurements and comparable DNA measurements per gene is as accurate a method as we can think of to define the genes we’re interested in.

6) They agree with Torres and Amon that the distribution of the gene expression of amplified genes is normal (not skewed), yet argue that the downregulated genes represent dosage compensation, whereas the upregulated genes are a part of the general response to aneuploidy – it is very unclear why each of these two outliers' groups should be interpreted differently. Their argumentation in the chapter "Unique features of selected genes...." is a prime example of scientific tautology: genes are dosage compensated because it is disadvantageous for the cells to overexpress them.

Our argument for this was based on data: the group of lower-than-expressed genes we identified is enriched with statistical significance for genes that are toxic when over-expressed in the lab strain (see Hose et al. for references). These genes are also enriched for genes known to be dosage compensated (see paragraph two, subheading “Unique features of selected genes suggest the group is enriched for dosage compensated genes”). The genes with amplified expression are explained by known effects of ploidy on cell size (Wu et al. 2010).

Accordingly, for these genes a propensity to CNV is observed. They do not observe propensity to CNV for genes from these categories when they are present in normal copy numbers – but this is most likely because they are not overexpressed and therefore there is no need to select against them! They also never say what gene categories are downregulated among unamplified genes (likely the same categories as on the amplified chromosomes).

It is not entirely clear what this reviewer is referring to, but perhaps he or she is misunderstanding the analysis. We quantified CNV based on publicly available datasets of S. cerevisiae strains, then simply analyzed trends across the different gene groups we defined here and in Hose et al. The genes we identified with lower-than-expected expression show, as a group, a higher propensity for gene duplication in publicly available datasets measuring CNV. With regard to the second point, in this work we removed from consideration all amplified genes belonging to functional groups enriched among the unamplified genes with a significant expression difference in that strain (see subheading “Unique features of selected genes suggest the group is enriched for dosage compensated genes” and Methods). Aside of mitochondrial genes, most of the other functional categories reported in Hose et al. remain enriched in our gene group. The enrichments remain significant when each chromosome is held out of the enrichment analysis, showing that the effect is not driven by a particular strain/chromosome. The enrichments are not significant for the set of unamplified genes that are repressed in multiple strains as part of a common aneuploidy response (Hose et al.).

7) There are also genes that are up-regulated – the identified categories broadly agree with the previous gene categories identified to be upregulated in response to aneuploidy (they cite the work from Fink lab, which relates to tetraploidy, but see more relevant Amon, Torres, and Storchova lab results). Remarkably, the dosage compensated genes closely match the categories identified previously to be downregulated in response to aneuploidy – ribosome biogenesis, translation, transcription – in several different species (see Amon, Torres, Storchova, Foijer laboratories). This option should be considered.

Respectfully, this reviewer is incorrect. The work by the Fink lab showed that fully tetraploid cells are bigger and have higher expression of genes encoding cell-surface proteins – there is no mention of anything related to ribosome biogenesis or translation factors in that work. To clarify one of our main points in this work: genes encoding ribosomal proteins (RPs) and translation factors can be repressed as a coherent group as part of the environmental stress response – indeed, the strains studied by Amon and Torres are very stressed and show robust activation of the ESR, including down-regulation of the entire group of RP genes and translation factors defined in the ESR, whether those genes are amplified or not. Our results are distinct in that a) the ESR as defined is not activated in our strains and b) the reduced expression is largely specific to the amplified genes. The more likely explanation that we favor is that at least some of these genes, especially RPs that are known to be dosage compensated by feedback, are regulated in response to their own gene copy number.

8) In paragraph five, subheading “Unique features of selected genes suggest the group is enriched for dosage compensated genes”, they cite a paper by Springer et al. (Mol Sys Biol 2010) in support of the presented work. But Springer's manuscript is focusing entirely on dosage compensation on protein level, whereas Hose considers mRNA; of individual proteins only, measured one-by-one by FACS, whereas Hose et al. considers whole chromosomes only, and Springer analyzes loss of copies, whereas Hose analysis gain of gene copies.

We have added a clarification to the Discussion that Springer used a GFP reporter with a non-native 3’UTR for that work.

In conclusion, the arguments do not seem strong enough to justify their analysis approach. Secondly, even if we would agree with the analysis, they found that the mRNA of 12 – 13% of genes might be dosage compensated at a FDR of almost 15% – a result that in fact says that there is no general dosage compensation.

We respectfully submit that the numbers cited in this work are on the same order as the numbers claimed in the original Hose et al. manuscript. As outlined in the last paragraph of this manuscript, we did not intend to claim that dosage compensation was widespread or functioned at most genes, and we never used the term. One of our key points in the original work and in this paper is that dosage compensation likely plays an important role in evolution, particularly in facilitating CNV. All of our work remains consistent with this notion.

Reviewer #2:

1) As I already indicated in my review of Amon et al., I am also puzzled by the non-integer changes in sequencing depth. I do think that this suggests that the strains could indeed be unstable. However, I also noted that we need to be sure that sequencing biases are not involved. I suggested a few relatively easy experiments to verify whether the populations are homogeneous and whether they show the expected aneuploidies. It seems easy enough to measure the copy number of a few key regions in a few key strains using qPCR, and to do this on a few different (small) colonies of the same culture. This would close the discussion.

We thank the reviewer for this suggestion. For five of the strains we worked with, we have done a similar analysis, first at the culture level and finally on individual colonies from a passaged culture, that shows that in most cases the aneuploidies are quite stable (but can be lost stochastically, see text for details). Therefore, we have opted not to add more of these comparisons at this time.

2) I follow the arguments about the error calculations. Although I am not an expert in this area, I think that, Gasch, Hose and coworkers indeed used correct calculations, but their sample size (number of replicates) are very low. Again, a few extra experiments using qPCR would help to close the discussion.

We provided qPCR and DNA microarray analysis of several genes from Chr12 in Hose et al. It is true that to measure very small, but real, differences in expression would require more than biological triplicates of mRNA data.

3) The number of genes that may show dosage compensation (245) seems quite low, especially since many of these might be false (false discover rate…). Even if some GO enrichment is found in this set, it is not clear to me that the number is high enough to call this "widespread". My best guess is that strains become aneuploid and the relatively quickly acquire mutations that normalize the expression of a few key genes for which changes in expression are not tolerated. Whether this is special for wild yeasts and should be called "dosage compensation" is a difficult question (obviously, the dosage is compensated, but using the name also suggests a general, dedicated mechanism, and this may very well be lacking…)

Respectfully, we never used the term widespread (that was used extensively by Torres et al.). Indeed, we do not believe that dosage compensation functions at most yeast genes. Our main interest in this manuscript was to show that genes potentially subject to dosage compensation display unique evolutionary patterns. We agree that there is unlikely a single or universal mode of dosage compensation analogous to the silencing of sex chromosomes, for example, and we did not indent to claim that in our original manuscript. With regard to the acquisition of rapid mutations that down-regulate genes: this was the main motivation for deriving the isogenic strain panels, since heritable polymorphisms that down-regulate genes as part of an adaptive response would be detected in our analysis. It is certainly possible that this occurs at some of the genes with heritably reduced expression that we identified in Hose et al., but it cannot explain the reduced expression per gene copy in our isogenic strain groups.

One notable (anonymized and lightly edited) comment from the reviewers’ discussion:

On the second issue, the extent of dosage compensation, I think that there is a fundamental disconnect between the groups in how to assess dosage compensation. Gasch (in Hose and Gasch) seek to take a gene specific view that accounts for gene-to-gene variability. This is noble and philosophically sound, but fraught with problems they don't address – namely that they have way too small a sample size to *really* do this (as also noted by both other reviewers). The Amon approach is to take a more distribution-based approach – which is, in principle, better in low sample sizes – but as Gasch correctly points out (and was a major criticism of the Torres manuscript) – the Torres methods are so poorly described as to be impossible to assess. Even if you think the Gasch method has merit, in the end they do back peddle on the number of dosage compensated genes (in Gasch compared to Hose) leading one to suspect we are talking about something that is either a small class or even non-existent.

It is true that our group versus Torres et al. have a philosophical difference in describing dosage compensation: we have focused on gene-level analysis without implying mechanism; they have focused on chromosome-wide effects. In this regard, both our groups agree that there is no evidence for a chromosome-wide mechanism of dosage compensation. We do believe that we have statistical power to identify genes subject to dosage compensation, and the fact that the gene groups we identified are enriched for myriad features of interest reveals that our results simply cannot be explained away as noise. Our more stringent analysis in this work presents a restricted gene set, which while on the lower end of the range we cited in Hose et al. remains consistent with our original claims. Fundamentally important to us are the evolutionary signatures and functional enrichments among the gene set we identified – while the numbers may appear small depending on perspective, it is the impact on evolution that we center on.