We employed the pooled S. cerevisiae deletion library to examine the impact of different cultivation conditions on the physiology of S. cerevisiae and compared potential global population differences between cultivations in shake flasks versus bioreactors. We were able to characterize the impact of process parameters commonly subject to adjustments during fed-batch cultivations on the library. To better simulate industrial processes, we included a two-stage seed train to generate an adequate amount of actively growing cells to inoculate a production bioreactor, as part of our workflow. To avoid any potential impact from amino acid insufficiencies on the fitness of the mutant pools, a prototrophic deletion collection was used in all experiments performed in this study .
Design of Bar-seq scale up fermentation experiments
It has long been recognized that most strains do not perform the same way when cultivated in a bioreactor compared to a shake flask . The main differences between these two general scenarios are dictated by the geometries of vessels and impellers along with enhanced control capabilities available during cultivations in bioreactors. Better gas transfer rates [5, 39, 40] can be achieved through presence of impellers and air spargers, better control of nutrient feed rates through external feeding capabilities as well as robust control over pH through acid or base addition. We examined a selection of these conditions between a shake flask cultivation, a bioreactor cultivation in batch mode as well as in fed-batch format, and individually across different fed-batch strategies. To assess the effect of any given variable on the population dynamics of a pooled S. cerevisiae deletion library, we performed growth competition experiments in two sets, each comprising two seed train stages followed by different final cultivation environments varying in either vessel architecture (set 1: bioreactor versus shake flask) or cultivation parameter (set 2: feeding mode, pH) (Fig. 1, for Additional file 1: Fig S1). For each of the two sets, we used an individual aliquot of the pooled library to inoculate the first seed train stage (seed 1) and measured the relative abundances of mutants within the population at least once every 24 h throughout the course of each cultivation using barcode sequencing .
In contrast to microbial cultivation in shake flasks, where gas transfer and mixing are achieved predominantly by shaking, bioreactors employ oxygen spargers and agitators to improve oxygen transfer in the cultures. To test if the difference in gas transfer impacts the population dynamics of the mutant pool, we performed batch cultivations without DO or pH control. All substrate was added to the culture media at the beginning of the cultivation and the composition of the resulting mutant pool after 72 h was compared. To assess potential differences in oxygen transfer in shake flask cultivations, we tested two different shake flask filling volumes (20% Vf/Vmax and 40% Vf/Vmax) (Additional file 1: Figs. S1, S2).
Additionally, bioreactors allow for pH control to ensure an optimal production environment as well as external feed of additional substrate to extend the process by minimizing nutrient limitations. To assess the impact of a feeding regime and pH on the population dynamics of the mutant pool, we performed four fed-batch cultivations: (i) two pH set points, pH 4 and pH 6, and (ii) two glucose feeding schemes, constant rate feeding and a DO signal-based pulse feeding (Fig. 1, Additional file 1: Figs. S1, S3). In a DO signal feeding regimen, a feeding solution is added upon complete exhaustion of available carbon sources and thus is responsive to the metabolic activity of the cells. Whereas, under constant rate feeding regimen, feed solution is continuously added to the fermentation broth independent of the metabolic activity. If the feed rate is not optimized for the process, then this feeding strategy may result in large accumulation of fermentative by-products, including ethanol, due to potential overfeeding of the culture. In all cases, the fed-batch phase was triggered after initial DO spikes, indicating full consumption of the carbon sources available during the batch phase.
Mutant population succession is dictated by fermentation regime
We interrogated the changes detected in the mutant population structure of each cultivation over time and visualized the data using a correlation matrix showing the diversity of the mutant pool via the total number of observed mutants (Fig. 3). Of the 6002 total barcoded genes present in this deletion collection, 2949 were detected at t0 with at least 10 counts. To minimize statistical noise, we only included genes with a threshold of at least 10 counts in the Seed 2 population in this analysis as described by Payen et al. . It may be noted that the general trend of observed mutants per condition was robust and independent of this threshold (Additional file 1: Fig. S4).
Generally, the population structure of the two individual seed trains were highly similar (r > 0.99 across all seed populations) when compared via Pearson correlation of mutant barcode counts (Additional file 1: Fig. S5, Table S1). In the studies investigating vessel geometry (set 1), microbial growth, measured by optical density (OD), and glucose consumption profiles look similar for all batch cultivations, independent of the cultivation vessel (BR vs. SF1 and SF2 in Additional file 1: Fig. S2a). The overall structure of the populations cultivated in both shake flask conditions (SF1 and SF2) and the batch bioreactor (BR) were highly similar when compared via Pearson correlation (Fig. 3) of mutant barcode counts (r > 0.99 across all time points), with minor changes at 48 h (for additional details see Additional file 1: Fig. S2b and c). As the overall population structure is highly similar in all conditions and time points tested, we conclude that the difference in vessel architecture is not reflected in the population structure of the deletion mutant pool.
For population changes between batch, and a fed-batch environment for the tested conditions, we analyzed the abundance of individual barcode counts in populations under the respective conditions, comparing BR to CF 4, 6 and DF 4, 6. Overall, only 8.6% (212 barcodes) of mutant barcodes initially present in the respective seed 2 culture were retained in all populations cultivated in a bioreactor environment independent of the cultivation mode (batch versus fed-batch), and 0.53% (13 barcodes) of mutant barcodes were lost in all conditions at the final timepoints (Fig. 2b). Upon examination of the mutant populations in the fed-batch conditions specifically, we observed a decrease in the overall diversity of our mutant population over the course of the fermentation for all four conditions (Fig. 2b). As expected from findings from the batch bioreactor study (BR) and the correlation matrix (Fig. 3, Additional file 1: Fig. S3b), over 90% of the mutants in the initial Seed 2 mutant population are still present throughout the batch phase (< 33 h EFT, EFT = elapsed fermentation time) with a highly similar population structure across all conditions (CF 4, 6 and DF 4, 6) before the feed start. Even then, all samples taken up to 48 h EFT show high correlation to each other, indicating little change in the pool of detected barcodes. However, large differences in population composition were observed in the populations grown in a constant rate feeding regime, in CF4 and 6, by the end of the fermentation at 119 h. The constant rate feed resulted in four times higher glucose addition (~ 250 g) but only slightly higher OD values (OD600 ~ 52) than DO signal based feeding (OD600 ~ 40, < 75 g glucose added). Most of the glucose in CF 1 and 2 accumulated as the fermentative by-product ethanol (~ 37 g/L), much higher compared to that in DF4 and 6 (< 1 g/L ethanol produced) (Fig. 2a, Additional file 1: Fig S1). Both populations grown under a DO signal-based feeding regime, DF4 and 6, remained highly correlated to the initial population detected in Seed 2 and each other throughout the cultivation independent of the culture pH (DF4 = pH 4, DF6 = pH 6, r = 0.98). Further, less than 200 mutants (i.e. less than 7.5% of the mutants present in the Seed 2 population) in reactor DF6 were lost over the course of the entire cultivation, indicating that the conditions (pH 6, DO signal based feeding) did not result in sufficient selective pressure for a significant change of the mutant pool population over the course of 119 h. In both reactors cultivated at pH 4, CF4 and DF4, an additional 10% of the mutants in the respective mutant population were lost within a 24 h window in the fed-batch phase between 48 and 72 h compared to DF6. In contrast, the population cultivated in CF6 (pH 6, constant rate feed) exhibited the start of diversity loss earlier, losing 23% of initial mutant barcodes between 33 h EFT and 48 h EFT and additional 55% of these mutants between 48 and 72 h, resulting in a decrease of its diversity to about 35% of its initial mutant pool after 72 h. This observation is also reflected in the correlation matrix: the mutant population of CF4 (pH 4) showed a modest decrease in correlation to DF6 after 48 h (pH 6, r = 0.88), whereas the CF6 population (pH 6) changed dramatically in comparison to DF6 at that time point (r = 0.12), and only showed moderate correlation to CF4 (r = 0.54). Upon comparison of the barcode abundance trends present under the constant feeding conditions in CF4 and 6, we found that the majority of loss in diversity takes place between 48 and 72 h for both of these populations, which is representative of the onset of stationary phase. This time point seems to coincide with significant accumulation of ethanol in these 2 bioreactor environments reaching ethanol concentration of over 12 g/L by 72 h from 3 g/L at 48 h in both reactors (Fig. 2a, red line), which may suggest that the presence of high concentrations of ethanol in CF4 and CF6 contributes to the loss in diversity under these conditions.
Next, we applied standard statistical techniques commonly applied in microbial ecology to better understand how mutant populations diverged from each other over time. These techniques and methods include calculations of Bray–Curtis dissimilarities, which is a statistic to quantify the compositional dissimilarity between two different samples, non-metric multidimensional scaling (NMDS), which attempts to accurately represent the dissimilarity of multidimensional data in lower dimensional space, and beta dispersion, which tests for the homogeneity of dispersion within groups.
An NMDS analysis of Bray–Curtis dissimilarities between mutant populations, revealed that samples diverged from one another as fermentations progressed, indicating that elapsed fermentation time may be an important driver for population divergence (Fig. 4a). To test this hypothesis, we applied common statistical tests on our dataset. Indeed, a one-way ANOVA test revealed that time was a highly significant driver of beta dispersion variance between samples (p = 0.001) and a PERMANOVA test, corrected for repeated measures (which is necessary due to repeated sampling of the same samples over time), showed that time accounted for ~ 36% of beta dispersion variance within the data (p < 0.001), while feeding strategy accounted for ~ 7% of the variance (p = 0.327). Over the course of fermentation, the overall beta dispersion of mutant populations tended to increase (Fig. 4b). When time is not taken into account the beta dispersion in populations from continuously fed reactors was significantly different via ANOVA analysis followed by Tukey HSD, compared to seed cultures (Fig. 4c). These observations indicate that, for this yeast deletion collection, the selection of the feeding scheme which affects the accumulation of the fermentative by-product ethanol may impact the diversity of the mutant pool to a higher degree as compared to feeding regimes that do not accumulate these byproducts.
Deletion mutants of genes associated with mitochondria are commonly lost across all conditions tested
Amongst the deletion mutants that were maintained throughout the entire fermentation, we observed that each condition (shake flask, batch bioreactor as well as fed-batch processes) resulted in overlapping as well as distinct mutant pools with increased or decreased fitness (Fig. 5). Our analysis revealed that 47 mutants were overrepresented specifically in the batch bioreactor, 4 mutants in the shake flask and 30 mutants were uniquely overrepresented in the fed-batch conditions at the final timepoint. While almost 50% of the mutants overrepresented in the shake flask population, were also overrepresented in the batch bioreactor population (3 of 7 barcodes), no overlap was detected for overrepresented mutants between the fed-batch bioreactors and the shake flask. However, out of 47 mutants overrepresented in the fed-batch environment, 17 were also overrepresented in the batch bioreactor but not in shake flasks.
To better understand the biological relevance of these commonly lost and commonly retained mutants, we performed gene ontology enrichment analysis of these gene lists using the Benjamini–Hochberg procedure to decrease the false discovery rate of our analysis. We found that only 3% of the barcodes initially present in Seed 2 (n = 77) were commonly lost across all the fed-batch bioreactors while 31% of the barcodes (n = 796) were commonly retained by the end of the experiment (119 h) (Fig. 2b). Of the 77 deletion mutants commonly lost across all fed-batch bioreactors, almost 50% (28 genes) are annotated to be associated with mitochondria. Of these 28, the majority of genes (20) are annotated to associate to the mitochondrial envelope and 9 genes are associated with mitochondrial organization including the assembly of the mitochondrial respiratory complex. Of the 796 deletion mutants retained in all fed-batch bioreactors with 787 unique gene identifiers, the annotated genes associated with response to salt stress as well as diphosphate activity were observed as significantly enriched.
Amongst the 10 genes annotated to be associated to cellular response to salt stress (GO ID: 0071472) are MSN4, a stress-responsive transcriptional activator involved in the yeast general stress response as well as all four genes MCK1, RIM11, MRK1, and YGK3, that all encode homologues of mammalian glycogen synthase kinase 3 (GSK-3). Yeast GSK-3 homologues are suggested to promote formation of a complex between the stress-responsive transcriptional activator Msn2p, a paralog to Msn4p and DNA, which is required for the proper response to different forms of stress [42, 43]. As these five deletion mutants (msn4−, mck1−, rim11−, mrk1−, ygk3−) were also retained throughout the study performed in the batch bioreactor and SF2, these genes may be promising candidates for gene deletion to improve fitness independent of glucose availability, pH or vessel architecture. However, further experiments need to be performed to confirm this assumption. A summary of identified mutant barcodes and associated annotated gene functions are listed in Additional file 1: Table S2.
Bioreactor-specific stress responses revealed by Bar-seq
In reactors CF4 and 6, which were run under a constant rate feeding scheme, the drastic loss of diversity coincided with the onset of ethanol accumulation. In DF4, the implementation of a DO signal feeding scheme resulted in no substantial accumulation of ethanol but rather a frequent switch from respiratory to fermentative metabolic state triggered by the addition of small glucose boluses that are characteristic of a DO signal feeding scheme (Additional file 1: Fig. S3). To determine whether the groups of mutants that were either lost or selected against during this cultivation period in the respective populations are indicative of the specific stresses imposed on the microbial population, we performed a direct comparison of the mutants specifically reduced or lost in each of the bioreactors followed by gene ontology enrichment analysis of those mutant groups. To allow the determination of statistically significant enrichments of genes in these groups, we excluded the population in bioreactor CF6 from further analysis, as it was reduced to less than 50% of its initial mutant diversity by 72 h.
Overall, 174 mutants were uniquely selected against in bioreactor CF4, 253 in bioreactor DF4 and only 39 mutants were either lost or appreciably selected against in the population cultivated in bioreactor DF6 (Fig. 6). Our analysis revealed that 112 mutants were selected against in both populations cultivated at pH 4 (CF4 and DF4), while only 2 mutants were commonly lost or reduced in all bioreactors at this time period. Glucose addition in DF6 was controlled by measurements of dissolved oxygen within the culture broth (DO signal), which is an indirect indicator for metabolic activity of the cells in suspension. The frequency of DO spikes, and thus the rate at which the cells are consuming all the carbon delivered per feed bolus, significantly decreases over time in case of DF6 compared to DF4, starting at around 50 h EFT. The observed low number of mutants lost in bioreactor DF6 could be attributed to the comparably lower metabolic activity of that culture (Additional file 1: Fig. S3).
The culture in bioreactor CF4 was maintained at pH 4 and was fed at a constant rate with glucose, which resulted in a high accumulation of the fermentative by-product ethanol starting at 48 h. The group of mutants selected against during this period of cultivation in CF4 is enriched with genes associated with cellular stress (5.14-fold enrichment, p = 0.0415), autophagy (5.9-fold enrichment, p = 0.0023), and endosomal transport (6.66-fold enrichment, p = 0.00144). Of those genes associated with cellular stress, the majority is either associated with DNA damage (MSH4, TSA1, GEM1, DPB4, CMR1, RAD57) or protein misfolding (CMR1, SSM4, CNE1). Additionally, we observed an overrepresentation of genes involved in the Cytoplasm-to-vacuole targeting (Cvt) pathway, a constitutive and specific form of autophagy that uses autophagosomal-like vesicles for selective transport of hydrolases to the vacuole, in the pool of mutants that were specifically selected against in bioreactor CF4 between 48 and 72 h (ATG2, ATG3, ATG4, ATG8, ATG9. ATG31, VPS30). Similar to bioreactor CF4, the culture in bioreactor DF4 was also maintained at pH 4 but was fed intermittently based on DO spike signals. Here, the pool of mutants that were significantly selected against was enriched for genes associated with the TCA cycle (6.09-fold enrichment, p = 0.0182) as well as oxidative stress (4.68-fold enrichment, p = 0.0142). The group of mutants selected against predominantly comprised of genes associated with the mitochondria (CIT3, IDH1, COX8, TIM18, UPS3, AIM25, DIC1) as well as general stress response genes (GPH1, GPD1) and oxidative stress responses (MXR1, YBP1, CTA1).
In addition, we found that 220 mutants were underrepresented specifically in the batch bioreactor, 79 in the shake flask and 131 were uniquely underrepresented in the fed-batch conditions at the final timepoint (Fig. 5). Unlike the overrepresented mutant pool analysis, a subset of underrepresented mutants is shared in two conditions and 7 mutants are underrepresented in all three conditions. The majority of these mutants, namely Rps0A−, Rpl8B−, Esa1−, Ric1−, Msc6−, can be associated with ribosomal and translational activity [44,45,46,47]. Similar to the overrepresented mutant pool at the final time point, we observed a large overlap of the underrepresented mutant pool in the shake flask population and batch bioreactor population (47%, 85 of 183 underrepresented in shake flasks). The pool of 131 mutants, specifically underrepresented but still present in fed-batch bioreactors, was enriched with genes associated with mitochondria (p = 0.01, n = 39, GO ID = 0,005,739). Of the 39 mitochondria associated genes, 11 play a role in mitochondrial organization while 7 are associated with cellular respiration.