Sourdough starters exhibit similar succession patterns but develop flour-specific climax communities

Flour selection

Five researchers grew a total of 40 starters from 10 different flour types: four replicates each of five flours containing gluten—unbleached all-purpose flour (processed from Triticum aestivum), red turkey wheat (Triticum aestivum), rye (Secale cereale), emmer (Triticum dicoccon), and einkorn (Triticum monococcum)—and five gluten-free flours—teff (Eragostis tef), millet (Eleusine coracana), sorghum (Sorghum spp), buckwheat (Fagopyrum esculentum), and amaranth (Aramanthus caudatus). Cereal grains are small, hard seeds without a fruit layer, grown as crops and harvested to be dried and either fed to livestock; boiled and consumed as a porridge or pilaf; or milled into flour for baking. While most cereals are grasses (i.e., wheat, rye, millet, sorghum, teff), the seeds of other plant families (i.e., buckwheat, amaranth) are also described as “whole grains”. The grains used in this project represent four glutenous and five gluten free cereal lineages that together span more than 527 million years across the angiosperm phylogeny. The all-purpose, red turkey wheat, emmer, einkorn, rye, and millet flours were milled at Boulted Bread Bakery (Raleigh, NC, USA), while the teff, sorghum, buckwheat, and amaranth flours were purchased from commercial vendors online.

Growing and measuring sourdough starters

In studying sourdough starters and other traditional fermentations, there is a tradeoff between using approaches that best match standard laboratory protocol and those that best match actual baking practices. Because we aim to generate results that can be directly applied by bakers, we focused on an approach that is in line with standard baker/bakery practices. We have illustrated our protocol in Fig. 1. On day 0, two level tablespoons of flour and two tablespoons of distilled water were mixed in a half pint wide mouth glass jar with a sterilized spoon. All spoons were washed with soap and water, and wiped with 70% ethanol immediately before mixing. A paper towel was fastened over the mouth of the jar with a rubber band to prevent large particles or insects from settling into the jar while still allowing environmental microbes to colonize the flour-water mixture. The initial height of the flour and water mixture was measured in cm. The jar was labeled with the flour type and replicate number, and left to ferment at room temperature. After 24 h, we measured the maximum height of the starter in cm (i.e., the highest point to which the starter had risen during the preceding 24 h). This point of maximum height can be identified by the residue left on the side of the jar, which resembles a “high tide” mark in cases where an active starter has subsequently deflated by the time of measurement. The height measurements provide a proxy for quantifying carbon dioxide production by each starter community, and changes in community behavior (i.e., function) over time. While sourdough starters grown from different flours may have different gas holding capacities, measuring change in height over time offers insight to gas production for starters grown from the same flour type via baseline comparison.

After measuring height, we removed the paper towel and described the smell(s) of the starter using free association, without the aid of a standardized word bank. Each researcher recorded the aromas produced by every individual starter, each day (1–14). We performed text analysis to classify individual descriptors into super-categories (i.e., earthy, fermented, fruit, grain/cereal, sour, toasted, and other). While sensory assessments may provide subjective interpretation of bread volatiles (Meignen et al., 2001), volatile information still correlates with sensory evaluations (Heenan et al., 2009), making them useful in the initial description of sourdough aromas.

After recording the aromas, we mixed the starter and transferred 1 mL starter to a sterile, labeled 2-mL microcentrifuge tube, which was stored at −20 °C for DNA sequencing (see DNA sequencing). Next, we removed one tablespoon of sourdough (the discard). and measured the pH using short-range (0–6, 0.5 interval) Hydrion Brilliant Paper (MicroEssential Laboratory, Brooklyn, NY, USA). Finally, we added one tablespoon fresh flour and one tablespoon water to the remaining starter and mixed thoroughly with the spoon. These measuring and refreshing steps were repeated once a day (at 24-h intervals) for 14 days. Conventional wisdom holds that a sourdough starter is mature and ready for baking after 14 feedings, although some bakers and instructions for growing a starter from scratch suggest that the starter is mature after only 10 days. After completing data collection, we also baked breads from the starters we grew for this project for an informal tasting, both to celebrate the completion of the project and to verify that all starters successfully leavened bread (Fig. S1).

Characterizing functional stages of microbial succession

We graphed the pH and height of each starter using ggplot2 in R (R Core Team, 2019) to identify distinct stages of succession across time and flour types with regard to starter performance (i.e., production of acid, CO₂, and aromas). We used Kruskal-Wallis and Pairwise Wilcoxon rank-sum tests with Bonferroni correction to identify significant differences in pH and height associated with time (i.e., days elapsed) and flour type. We focus on days 1, 2, and 3 to determine the impact of flour type on early succession dynamics and the potential importance of priority effects on community membership and starter performance at maturity. We include day 6 to represent intermediate succession, and days 10 and 14 to acknowledge two common lengths of time by which bakers assume that starters to have reached maturity. We binned aromas by category (i.e., grain/cereal, toasted, nutty, veggie, fruit, fermented, fermented dairy, sour, earthy, bodily smells, chemical, meat) and quantified (presence/absence) whether each aroma category was detected during each sensory assessment).

DNA sequencing

We collected three 1-mL samples of each flour type (n = 30) and three 1-mL samples of distilled water to represent the resource inputs to each starter on day zero. We also collected a daily 1-mL aliquot from each starter after measuring pH, as part of the discard. All aliquots were stored at −20 °C. After analyzing the pH and height data from all starters to identify functionally significant stages in succession, we elected to sequence starter samples collected on days 1, 2, 3, 6, 10, and 14 (n = 240). Together with the flour and water aliquots, a total of 273 samples were shipped on dry ice for DNA extraction, PCR amplification, and Illumina multiplexed sequencing of the bacterial 16S v4 region on the MiSeq platform, using the 515f (AATGATACGGCGACCACCGAGATCTACACGCT TATGGTAATT GT GTGCCAGCMGCCGCGGTAA) and 806r (CAAGCAGAAGACGGCATACGAGAT XXXXXXXXXXXX AGTCAGTCAG CC GGACTACHVGGGTWTCTAAT) primers for bacteria at the Fierer Microbial Community Sequencing Lab (Boulder, CO, USA) as previously described (Oliverio, Bradford & Fierer, 2017). PCR negative controls were also sequenced to check for contamination. Raw data are available in GenBank (BioProject PRJNA973060, accessions SAMN35108252–SAMN35108524), and the data and code are available in the Dryad database (https://doi.org/10.5061/dryad.bk3j9kdh3).

Bioinformatic analysis

The Fierer Microbial Community Sequencing Lab demultiplexed and processed raw sequences to produce exact sequence variant (ASV) tables with the DADA2 pipeline, as described at (https://github.com/fiererlab/dada2_fiererlab) and performed previously by Landis et al. (2021) to analyze sourdough communities. We used Cutadapt (Martin, 2011) to remove sequences with N’s, then quality filtered sequences using truncLen = 150 for forward reads and 140 for reverse reads, maxEE = 1, and truncQ = 11. Next we inferred ASVs with the DADA2 algorithm, merged paired-end reads, and removed chimeras and contaminants before assigning taxonomy with the Silva database (Quast et al., 2013). We performed all downstream analyses in the R environment (R Core Team, 2019); all associated data and scripts are available in the Supplemental Materials. We filtered out reads assigned to either chloroplast or mitochondria from the bacterial taxa table, along with any reads that were unassigned at the phylum level. We reclassified all ASVs classified as Lactobacillus species using the “lactotax” tool (http://lactotax.embl.de/wuyts/lactotax/) developed by Zheng et al. (2020) to refine taxonomic structure for the clade. For ASVs originally classified as Lactobacillus but without species designations, we used BLAST to determine the new genus designations. First we cross-referenced each of the hits that met the 99% threshold with the new taxonomy. If all of the hits for a specific ASV matched a single genus, we assigned that new genus. If the BLAST results did not agree on a single genus, the ASV was assigned taxonomy only at the family level. We then created two datasets for downstream analyses on alpha and beta diversity:

1. We compared all water, flour, and day 1 starters to assess the influence of inputs on early starter communities. This data set was un-rarefied.

2. We removed all water and flour samples to compare changes in starter community composition from day 1–14. This dataset was rarefied to 1,200 bacterial reads per sample and the ASV table was converted to percent relative abundances for downstream analysis. We retained 232 starter samples for downstream analysis.

We first plotted heatmaps and produced nonmetric multidimensional scaling (NMDS) plots with Bray-Curtis distance calculations using the metaMDS function to compare the bacterial contents of day 1 starters to flour and water inputs. We next turned our attention to focus on the rarefied starter dataset. We calculated alpha diversity as ASV richness and Simpson diversity. We used Kruskal-Wallis tests to measure the effects of days elapsed on alpha diversity and on Lactobacillaceae ASV abundance. As a non-parametric method, the Kruskal-Wallis test is appropriate for comparing microbial membership and other data that do not follow normal distribution (Kruskal & Wallis, 1952). We calculated the Pearson correlation coefficient to quantify the relationship between pH and the relative abundance of Lactobacillaceae to determine the strength of the relationship between microbial membership and metabolic function. We calculated beta diversity between starter samples as Bray-Curtis dissimilarity, using the rarified ASV table. We conducted permutational multivariate analysis of variance (PERMANOVA) using the adonis function in the package vegan to test for relationships between flour type and days elapsed on the composition of bacterial communities. Similarly, we used the envfit function in the vegan package to test for relationships between flour type and aromas on bacterial community composition. We produced nonmetric multidimensional scaling plots with Bray-Curtis distance calculations using the metaMDS function, overlaying the centroids (99% confidence interval for standard error) for each day elapsed. We ran a Kruskal-Wallis test with a Bonferroni correction to determine which bacterial taxa and aromas were significantly different between days (across flour types) and between flour types in mature (day 14) starter communities. We plotted the degree (eigenvalue) and direction (eigenvector) of influence of these significant bacterial taxa and aromas on starter communities using envirofit function from the vegan package. Finally, we created heatmaps using plot_ts_heatmap from the mctoolsr package (https://github.com/leffj/mctoolsr/) to visualize all bacterial genera present at ≥0.1% relative abundance for each time point or flour type, and aromas detected in at least 25% (10/40) of starter samples.

Characterizing functional stages of succession

We identified six functionally distinct succession stages across flour types based on starter pH (Kruskal-Wallis, p < 2.2e−16) and starter height (Kruskal-Wallis, p < 2.2e−16) (Fig. 2 and Table 1). Notably, all starters rose: the height dynamics we observed were robust (Fig. 2C) and similar (Fig. 2D) across flour types. On day 1, starters showed little to no change in either pH or height compared to day 0. On day 2, pH dropped (Pairwise Wilcoxon Rank Sum, p < 3.8e−09) while height increased (Pairwise Wilcoxon Rank Sum, p < 0.00355). pH continued to fall on day 3, (Pairwise Wilcoxon Rank sum test, p < 0.00112), after which starters stabilized at a pH of 3.6; we detected no significant changes in pH throughout the following days of the experiment (Wilcoxon Rank Sum Test, p = 1.0). By contrast, height continued to rise significantly through day 6 (Pairwise Wilcoxon Rank Sum Test comparison between days 2 and 6, p < 0.00112) and day 10 (Pairwise Wilcoxon Rank Sum Test comparison between days 6 and 10, p < 0.00477). By day 14, both pH and height were maintained, suggesting that communities were functionally stable.

Bacterial succession

As predicted, the composition of bacterial taxa in starters on day 1 was driven by taxa associated with flour and inputs (Fig. S2). For example, Pantoea and an unknown genus within Enterobacteriaceae dominated both flour (33.1% and 20.3%, respectively) and day 1 starter samples (23.8% and 26.7%, respectively); and Pseudomonas was present at similar relative abundance in water (13%), flour (10.1%), and day 1 starters (12%) (Fig. S2A). These shared taxa likely drive the overlap between flour and day 1 starters (Fig. S2B), despite flour-specific differences in bacterial membership (Figs. S2C and S2D). All starter communities followed similar succession patterns with regard to a measure of evenness: both Simpson and Shannon diversity initially increased, reaching a maximum between day 3–6 (Figs. 3A and 3B), then decreased after species of Lactobacillaceae became dominant (Fig. 3C) and pH stabilized (Fig. 2A). Bacterial richness followed similar patterns across all flour types except for rye starters, in which bacterial richness increased dramatically from day 3 to day 6 and retained 15–30 more ASVs, on average, compared to other flour types (Fig. S3).

The observed changes in sourdough starter performance correlated significantly with (and are likely driven by) predictable shifts in the composition of bacterial communities. For example, pH decreased significantly as Lactobacillaceae grew more abundant (Pearson correlation = −0.5321391, p < 2.2e−16); Fig. S4). In addition, bacterial community dynamics corresponded with differences in the prevalent aromas produced over time (Fig. 4, Table 1 and Fig. S5). On day 1, when starter communities were dominated by bacterial taxa associated with flour and plants, they smelled predominantly of toasted, grain and cereal aromas. Intermediate stages of bacterial succession produced a mixture of sour, grain/cereal, and earthy aromas. However, by day 14, mature starters smelled like sour, fermented fruit (Fig. 4B). When comparing community succession dynamics among starters from day 1–14, PERMANOVA tests confirmed that starter samples harbor significantly different bacterial taxa on day 1, days 2–3, day 6, and days 10–14. When comparing individual taxa between successional days, we detected significant differences in the relative abundance of eleven genera including Erwinia (Kruskal-Wallis with Bonferroni correction p < 2.15e−15), Pseudomonas (Kruskal-Wallis with Bonferroni correction p < 4.50e−11), Klebsiella (Kruskal-Wallis with Bonferroni correction p < 1.00e−04), Pantoea (Kruskal-Wallis with Bonferroni correction p < 2.88e−11), Enterobacteriaceae unkn (Kruskal-Wallis with Bonferroni correction p < 3.49e−19), Weissella (Kruskal-Wallis with Bonferroni correction p < 1.73e−06), Leuconostoc (Kruskal-Wallis with Bonferroni correction p < 1.03e−04), Latilactobacillus (Kruskal-Wallis with Bonferroni correction p = 2.52e−03), Levilactobacillus (Kruskal-Wallis with Bonferroni correction p < 2.45e−04), Gluconobacter (Kruskal-Wallis with Bonferroni correction p < 4.81e−08), and Pediococcus (Kruskal-Wallis with Bonferroni correction p < 2.72e−14). Erwinia (11.5% on day 1), Pseudomonas (12.8–18.2%), Klebsiella (11.2–13.2%), Pantoea (21.3% on day 1), and Enterobacteriaceae unkn (16.3–28.5%) occurred at higher relative abundance in younger starters (≦3 days old) than in older starters (≧10 days old) (Figs. 4 and S5A). Three genera—Weissella, Leuconostoc, and Latilactobacillus—occurred at higher relative abundance in starters after day 1 but before day 10; and three other genera—Levilactobacillus, Gluconobacter, and Pediococcus—occurred at higher relative abundance in older starters (≧10 days old) (Figs. 4 and S5A).

Differences among mature starters

Whereas the youngest starters (<3 days; Fig. 4) were characterized by variation both among and within flour types, differences among mature starters (day 14) were dominated by differences among flour types (Table 2, Figs. 5 and S6). These flour-specific differences in bacterial membership coincided with differences in specific aromas (Figs. 5A and S5B). In other words, as starters age, they converge toward specific climax communities depending on which flour type they are fed; and different climax communities not only comprise different bacteria but also product different aromas. Two flour types support particularly unique communities: teff starters were dominated by Pediococcus (79.2%) and also contained Leuconostoc, but lacked genera belonging to the family Lactobacillaceae; and amaranth starters were dominated by Latilactobacillus (70.3%, Fig. 5B).

Succession processes

Community membership at each successional stage comprised bacteria with specific functional traits: generalists and plant endophytes gave way to lactic acid producing genera, which were outcompeted by a subset of acid producing genera that are also acid-tolerant. Our results are entirely in line with the three-phase model of succession elaborated through the study of wheat-based sourdough (Ercolini et al., 2013; Minervini et al., 2014; Harth, Van Kerrebroeck & De Vuyst, 2016; Gobbetti et al., 2016). In the first phase (day 1) of succession, sourdough starters are both individually diverse (high alpha diversity; Figs. 3A and 3B) and variable one to the next (high beta diversity), even within flour types. We hypothesize that this initial diversity reflects the effects of source pools and which bacterial taxa happen to arrive in a particular starter, as well as the size of the initial populations of each of these bacterial taxa. Ecologists have pioneered a range of approaches to test how important source pools and the relative abundance of different taxa in source pools are to the initial composition of starters (Miller et al., 2019; Reese et al., 2020). As succession proceeded to phase 2 (days 2, 3, and 6), taxa more characteristic of climax sourdough communities, including LAB, became more abundant and eventually dominated the starters in phase 3 (days 10 and 14). This dominance results in part from the ability of these climax taxa to grow quickly on the available resources and to produce allelopathic compounds, primarily lactic acid, that kill many species (Minervini et al., 2014). It is possible that the production of lactic acid also facilitates the arrival of a subset of acidophilic species that thrive in the final successional stage of the sourdough starters (possible but not yet tested). By including the functional outputs (pH, height, and aromas) associated with dominant bacteria, we offer additional nuance beyond the three-phase model, enabling us to characterize a total of 6 statistically distinct functional stages of sourdough succession (Table 1). The dominant aromas detected at each stage correlate with changes in bacterial community composition (Fig. 4), demonstrating that the succession process has measurable functional consequences for the aesthetic properties of sourdough. Indeed, both the bacterial and aromatic profiles follow classic successional shifts, with intermediate stages exhibiting the most diverse and diffuse representation before converging toward climax membership and functionality (e.g., fruit, sour, and fermented aromas).

Our results support several anecdotal observations shared by bakers and participants in our ongoing series of citizen science sourdough projects (http://robdunnlab.com/projects/science-of-sourdough/). For example, many starter-keepers note a dramatic increase in height on day 2 or 3, followed by a plateau or decrease in activity until roughly day 7. We verify this widely noted behavioral pattern across flour types (Fig. 2). We hypothesize that this early peak in starter height, a proxy for gas production, results from the growth of flour-associated yeasts known to dominate young starters (Ercolini et al., 2013). In addition, bakers often invoke a “10 day rule”, using starters to bake only after they are at least 10 days old. Our results confirm that sourdough starters are functionally mature and stable (i.e., suitable for baking) by day 10 across flour types (Fig. 2), and that this functionality corresponds with dominant bacterial taxa characteristic of sourdough climax communities (Figs. 3 and 4)—specifically, acid-producing genera of Lactobacillaceae (Figs. 3C and S4).

An unexpected result occurred in starters made with rye flour. Starters made with rye flour are very common in some parts of Europe (Weckx et al., 2010). However, they have been less well-studied than starters made with wheat flour (De Vuyst, Van Kerrebroeck & Leroy, 2017). In general, it has been hypothesized that the lactic acid produced by LAB in mature starters limits the diversity of taxa able to survive in mature starters compared to younger starters (e.g., observations in (Minervini et al., 2014; Gobbetti et al., 2016)). The results for all starters supported this prediction except for those made with rye flour, which contained a relatively high diversity of bacterial taxa even after they reached maturity (Fig. S3). Why might rye flour support greater bacterial diversity? Rye contains starches that are easily degraded by amylase compared to more resistant wheat starches (Perten, 1964), making sugars more immediately available for microbial metabolism. We speculate that the diversity of nutrients present in whole grain flours, combined with enhanced starch degradation by amylase in rye flour, create additional niche space to allow the persistence of more species in rye starters compared to other grains—a possibility that deserves more exploration.

Mature starters

Acid producing and acid-tolerant taxa dominate mature starters across all flour types. For example, Gluconobacter, Pediococcus, and other genera of Lactobacillaceae were most abundant in mature starters. The ubiquity of the genus Gluconobacter, which accounted for 12.6–45.8% of the microbial communities present in 7 of 10 flour types studied (Fig. 4), was unexpected. Gluconobacter was recently isolated for the first time from sourdough starters in the Henan Province of China (Xing et al., 2020) and was also detected in a study of sourdough starters collected from four continents (Landis et al., 2021); but the genus has received little attention in studies of starters more generally. Together, these results suggest that acetic acid bacteria may play a greater role in sourdough starters than is currently appreciated.

Climax communities are likely shaped by flour-specific characteristics (Fig. S6), though future research is needed to determine to what extent these climax communities are shaped by flour compounds vs priority effects. When considering acidity, dough height, or the relative abundance of LAB, the starters are far more similar within than between different flour types. We also see key differences among flour types, both in the particular bacterial taxa (ASVs) that are present and the aromas associated with those communities (Table 2 and Fig. 5). Pantoea is much more abundant in mature buckwheat starters (12.2%) than in other starters (≦3.5%). Acid-tolerant Leuconostoc is enriched in teff (18.1%), and an unknown genus of Enterobacteriaceae is enriched in rye starters (7%). Other starters were characterized by the absence of particular taxa: for example, Gluconobacter is present in all starters except teff and amaranth (Fig. 5B). The relative abundance of Pediococcus is negatively correlated with other Lactobacillaceae genera (Companilactobacillus, Latilactobacillus, Latiplantibacillus, Levilactobacillus, and Loigolactobacillus), suggesting that they may compete for resources. This is especially apparent in teff (Latilactobacillus = 0.2%, Pediococcus = 79.2%; fermented aromas) and amaranth (Latilactobacillus = 70.3%, Pediococcus = 2.1%; toasted, meat, and fermented dairy aromas) starters. Landis et al. (2021) previously established that different sourdough communities produce distinct aromatic profiles, and the current project suggests that bakers can use different flour types to select for specific bacteria and aromas (Figs. S5 and S6), which in turn impact bread characteristics. Importantly, while we focused on bacteria in this study, yeast communities also contribute to aromatic profiles through direct formation of volatile compounds (e.g., alcohols, esters, carbonyl compounds (Hansen & Schieberle, 2005)) and via interactions with LAB (Guerzoni et al., 2007).

Many of the bacteria we detected are novel compared to previous studies of the same flour types (Table 2). These differences might indicate that DNA sequencing removes the LAB bias of culture-based methods, enabling us to better appreciate the diversity of sourdough taxa; or they may result from regional differences among grains and/or bacterial species pools. Our study design does not allow us to distinguish why the taxa present in the climax communities differed among flour types. Cereal grains differ in many respects, including enzyme activities, phenolic compounds, and buffering capacity, which have been shown to impact sourdough microbiota in previous studies (Sekwati-Monang, Valcheva & Gänzle, 2012; Dinardo et al., 2019). Another possibility is that nutritional differences among flour types may favor specific taxa and bacterial interactions, yielding predictable community structures that perform particular functions as previously described (Landis et al., 2021). Given that sourdough microbes can reflect grain phenology (Minervini et al., 2015) and cultivation (Rizzello et al., 2015), future efforts to compare starters grown from different flour batches (i.e., the same cereal grown in different fields or years) could yield additional insights to the variation within vs among cereals.

Broader context

The successional patterns characterized here are similar to those previously observed in other lacto-fermented foods. Most spontaneously lacto-fermented foods (e.g., sourdough starter, sauerkraut, and kimchi) take only 2 weeks to develop stable climax communities. Chinese sauerkraut (Xiong et al., 2012) and dongchimi—a traditional Korean kimchi made with radishes (Jeong et al., 2013)—both undergo undisturbed fermentation for 10 days, during which time LAB levels increase and pH drops to ≤4 and taxa shift from opportunistic generalists including Enterococcus faecalis (Xiong et al., 2012), Pseudomonas and Weissella (Jeong et al., 2013), to acid-producing climax communities dominated by Leuconostoc (Jeong et al., 2013), Lactiplantibacillus plantarum and Lacticaseibacillus casei (Xiong et al., 2012). Despite reaching a similar climax community of acid-resistant LAB, these fermentations all differ from sourdough starters in that they don’t involve refreshment. It would be interesting to test how succession might proceed in sourdough starters were they not replenished (with flour). We suspect that failing to feed sourdough starters prevents them from developing their characteristic climate communities. If this were the case, the question is why. One possibility is that cabbage and radish provide a large inoculum of LAB, thereby preventing the establishing of opportunistic bacterial species (Miller et al., 2019). Another possibility is that the diverse complex carbohydrates in cabbage and radish, combined with the salt employed in their fermentation, prevent opportunistic generalists from dominating (through resource diversity effects (Pickett, Collins & Armesto, 1987)) and disfavor species that are not salt tolerant (Yang et al., 2020).