Volatility as a Concept to Understand the Impact of Stress on the Microbiome

The microbiome-gut-brain-axis is a complex phenomenon spanning several dynamic systems in the body which can be parsed at a molecular, cellular, physiological and ecological level. A growing body of evidence indicates that this axis is particularly sensitive to the effects of stress and that it may be relevant to stress resilience and susceptibility. Although stress-induced changes in the composition of the microbiome have been reported, the degree of compositional change over time, which we define as volatility, has not been the subject of in-depth scrutiny. Using a chronic psychosocial stress paradigm in male mice, we report that the volatility of the microbiome significantly correlated with several readouts of the stress response, including behaviour and corticosterone response. We then validated these findings in a second independent group of stressed mice. Additionally, we assessed the relationship between volatility and stress parameters in a cohort of health volunteers who were undergoing academic exams and report similar observations. Finally, we found inter-species similarities in the microbiome stress response on a functional level. Our research highlights the effects of stress on the dynamic microbiome and underscores the informative value of volatility as a parameter that should be considered in all future analyses of the microbiome.


Introduction
The mammalian gut plays host to approximately 1 trillion microbial organisms collectively known as the gut microbiome (Sommer and Bäckhed, 2013). The microbiome is highly sensitive and reactive to the effects of stress to the extent that it is now accepted that the stress response is not solely the domain of brain function, but rather that it results from a synergy of mechanisms that constitute the gut-brain axis Bastiaanssen et al., 2020;Cruz-Pereira et al., 2020;Cryan et al., 2019;Dinan and Cryan, 2012;Foster et al., 2017). In particular, studies in rodents have correlatively linked alterations in microbiota composition to the effects of stress on behaviour (Bharwani et al., 2016;Bharwani et al., 2017;Burokas et al., 2017;Marin et al., 2017;Szyszkowicz et al., 2017;Xu et al., 2020) and the central/peripheral inflammatory milieu (Bailey et al., 2011;Bharwani et al., 2016;Bharwani et al., 2017;Burokas et al., 2017;Szyszkowicz et al., 2017). Going further, manipulation and perturbation of the microbiome have been shown to alter the reaction to stress, further solidifying the regulatory role of the microbiome in the stress response (Donoso et al., 2020;Jašarević et al., 2017;Kuti et al., 2020;Langgartner et al., 2018;Morais et al., 2020;Pearson-Leary et al., 2019;Provensi et al., 2019;Stothart et al., 2019;Wang et al., 2020). Moreover, in humans there have been a number of studies confirming a relationship between stress and microbiome composition across the lifespan (Allen et al., 2016;Hemmings et al., 2017;Messaoudi et al., 2011;Papalini et al., 2019;Zijlmans et al., 2015).
These observations notwithstanding, a common feature of the above cited studies is the fact that they are a 'snapshot-in-time' analyses of samples and therefore do not consider the compositional and temporal instability of the microbiome (Caporaso et al., 2011;Claesson et al., 2011). Temporal variance as a feature of the microbiome has been discussed before. Initially, the volatile nature of the microbiome was pointed out (Weinstock, 2011) and relatively soon thereafter, the term volatility was used to refer to the degree of change between timepoints (Goodrich et al., 2014). Aside from these studies, the term volatility has been scarcely used in the context of the microbiome. One example of its rare use is in the context of the rapidly shifting microbial communities on shoes and cell phones (Lax et al., 2015). In the context of stress, recent studies have shown differential effects on microbiota beta-diversity dependent on when the samples were takenshortly after commencement of social stress or towards the end of the protocol (Gautam et al., 2018), though the degree of change was not pursued, nor linked to the severity of the stress response. More broadly, there have been reports of microbiome beta-diversity being altered after stress compared to control, though often these reports do not relate severity of stress to difference in beta diversity, nor do they take the baseline microbiome into account (Langgartner et al., 2018;Xu et al., 2020). Recently, in a large human cohort, patients with inflammatory bowel disease (IBD), a condition that is associated with an increased level of stress, were shown to have a more volatile microbiome than healthy volunteers Ryan et al., 2020). While these studies provide some evidence that volatility may be related to stress, the impact of stress on microbial volatility measures remains unknown. Furthermore, deciphering what makes a microbial community volatile can provide new insights into its role in mediating the effects of stress. To further investigate volatility in relation to stress, we exposed mice to chronic psychosocial stress and assessed volatility. Based on the findings of this discovery cohort, we then set out to validate our findings in an independent cohort of a different strain of mice under the same experimental conditions. Finally, we also took advantage of an ongoing study involving healthy volunteers undergoing academic examination stress and sought to examine whether the interaction between volatility and stress occurred across species.

Animals
For this study we used two cohorts of adult male mice. The first cohort will be referred to as the Discovery cohort and consisted of (B6;129-Gt(ROSA)26Sor tm1(CAG-cas9,-EGFP)Fezh /J; https://www.jax. org/strain/024857). The second cohort will be referred to as the Validation cohort and consisted of adult male C57BL/6 (Envigo, UK). There were no differences in terms of experimental treatment and handling between the two cohorts. There were three weeks between the arrival of the animals and the start of singly housing. Approximately one week before commencement of social defeat sessions, all mice were singly housed and weighed daily over the course of the experimental protocol ( Fig. 1). For the chronic social defeat stress procedure, non-experimental singly housed adult male CD1 were used as aggressors (Envigo, UK). Mice were kept under a 12 hr light/dark cycle (ON 7:30AM, OFF 7:30PM) in a temperature/humidity controlled environment (21 • C, 55.5%) with food and water ad libitum. The main behavioural and physiological responses to chronic stress of the Discovery cohort have been initially reported elsewhere (Gururajan et al., 2019) and are used here in a correlative capacity with the microbiome analysis.

Chronic social defeat stress
Mice were randomly assigned to either the social defeat stress or control groups. Defeat sessions were performed as previously described (Gururajan et al., 2019). Briefly, for each defeat session, test mice were subjected to a pre-screened aggressor CD1 mouse until the first attack, expression of submissive posturing or until 5 min had passed. The mice were then separated by a perforated plexiglass wall that allowed only non-physical contact for 2 hrs. Subsequently, the separator was removed and, after another defeat, mice were transferred back to their home-cage. This was repeated with a different CD1 aggressor each day for 10 days. Control mice remained in their home-cages over the course of the stress protocol but were handled to an equal extent as the stressed mice in the process of measuring daily body weight and collecting tail-blood samples. Across the duration of the defeat protocol, to prevent contamination during defeat procedures, the experimenter removed any traces of faecal boli produced by the aggressor or the stressed mice. Based on the findings in the first cohort of mice, we repeated the experiment in a larger cohort. We refer to these cohorts as the Discovery cohort and the Validation cohort, respectively. Further details can be found in Supplementary Methods.

Social interaction test
The social interaction testing of mice was used to assess avoidance of the CD1 aggressors the day after the last defeat session and was carried out as previously described (Gururajan et al., 2019). Briefly, the test was carried out in a plastic box (41 × 32 × 24 cm) containing a wire mesh cage (9.5 × 7.5 × 7.0 cm) against one wall. In the first 2.5 min trial, the test mouse was allowed to explore freely. The mouse was then returned to its home cage for 1 min. During this time, an unfamiliar aggressor CD1 mouse was placed inside the wire mesh cage. This was followed by a second 2.5 min trial in which the test mouse was allowed to explore the area freely in the presence of the caged CD1. Social interaction (SIT) ratios were generated based on social investigation time in an arbitrarily defined interaction zone around the wire mesh cage. Further details can be found in Supplementary Methods.

Plasma sampling for corticosterone
Collection and analysis of plasma samples for corticosterone was carried out as previously described (Gururajan et al., 2019;van de Wouw et al., 2018). Briefly, tail bleeds were carried out within 1 hour of the lights turning off . Whole blood was collected in sterile eppendorfs. Whole blood was centrifuged (3500 g, 10 minutes, 4 degrees C) and plasma was collected. Plasma samples were analyzed in duplicate using the Enzo® Corticosterone ELISA kit plate (ADI-900-097, Enzo, Exeter, United Kingdom) according to the manufacturer's instructions. ELISA plates were read using a Multiskan® microplate photometer (Thermofisher Scientific®, Waltham, MA, USA) at 405 nm. See Supplementary Methods for further detail. From Day 1 to 10, mice were randomly assigned to either the control condition (not shown) or chronic social defeat stress. On Day 11, plasma and faecal boli samples were again collected and social behaviour was assessed. The following day, all mice were culled. Trunk blood was collected for flow cytometry and brain tissue was processed for gene expression analyses.

Gene Expression Analysis
Gene expression analysis was carried out as previously described (Gururajan et al., 2019) Briefly, total RNA was extracted using the miRVana™ miRNA Isolation kit (Ambion/Life Technologies, Paisley, UK) according to the manufacturer's instructions. RNA was reverse-transcribed to complementary DNA using the Applied Bio-system® High Capacity cDNA Reverse Transcription Kit (10X RT Buffer, 25X dNTP mix (100 mM), 10X RT Random Primers, Multiscribe® Reverse Transcriptase) on the Applied Biosystem® GeneAmp PCR System 9700 (Thermofisher®, Waltham, MA, USA). qRT-PCR was carried out on the StepOnePlus® PCR machine (Thermofisher®, Waltham, MA, USA) using primer assays designed by Integrated DNA Technologies (Skokie, Illinois, USA). See supplementary information for primer IDs. Experimental samples were run in duplicates and to check for amplicon contamination, each run also contained template free controls for each probe used. PCR data were normalised using β-actin, with an average ct value of 20.9 (+/-0.5SD) across all samples, and transformed using the ΔΔCt method as previously described (Stilling et al., 2018).

Human Healthy Volunteer Study
Briefly, healthy volunteer study participants were recruited via advertisement and direct contact to the student population of University College Cork (UCC). A total of 84 volunteers responded to advertisement and direct contact; 54 were pre-screened by telephone call (64%); 36 were invited to a screening visit (43%); and thirty were enrolled in the study and randomised to treatment (36%). Inclusion criteria: participant must be able to give written informed consent; be between 18 and 30 years of age; be male; be in generally good health as determined by the investigator. Prior to testing days, participants were asked to refrain from strenuous exercise and alcohol 24 hours before the session, and from caffeine three hours prior to the session. At the screening visit, two weeks before baseline measurement, study participants were asked about their demographics, general medical history, medication record, and other metadata. Furthermore, the participants were screened using the MINI International Psychiatric Interview (to exclude subjects with a significant DSM-V psychiatric diagnosis). Participants attended for study visits during 2 semesters in UCCboth 8 weeks prior to an exam period and during an exam period. The exam visit took place during the participant's exams, but not on the day of an exam. The measures taken during the visit included Cohen's Perceived Stress Scale (PSS). Faecal samples from the morning of the visit were collected into plastic containers containing an Anaerogen sachet. Participants were instructed to keep the sample in a cool place until delivery at the study visit. Samples were immediately frozen at − 80 • C at the visit. 200 mg was taken from the core after thawing and subsequently used for DNA extractions and 16 s sequencing. Salivary Cortisol Awakening Response was assessed using the Cortisol ELISA kit plate ADI-900-071, Enzo, Exeter, United Kingdom) according to the manufacturer's instructions. ELISA plates were read using a Multiskan® microplate photometer (Thermofisher Scientific®, Waltham, MA, USA) at 405 nm. These data were derived from control participants from a larger healthy volunteer study examining the effects of an intervention on the microbiota-gut-brain-axis .

16S rRNA Gene Sequencing
Faecal samples were collected during the weighing of mice before the first day and after the last day of social defeat. Boli were transferred to eppendorfs and stored at − 80 • C. DNA was extracted from faecal samples and prepared for sequencing using an Illumina 16S Metagenomic Sequencing Library Protocol. See supplementary methods for further details.

Bioinformatics analysis
Three hundred base pair paired-end reads were pre-filtered based on a quality score threshold of >28 and trimmed, filtered for quality and chimaeras using the DADA2 library in R (version 3.6.3). Only samples with >10.000 reads after QC were used in analysis. Taxonomy was assigned with DADA2 against the SILVA SSURef database release v138. Parameters as recommended in the DADA2 manual were adhered to unless mentioned otherwise. ASVs were aggregated at genus level; those that were unknown on the genus level were not considered in downstream analysis, as were genera that were only detected as non-zero in 10% or few of total samples. As ratios are invariant to subsetting and this study employs compositional data analysis techniques (Aitchison et al., 2000;Gloor et al., 2017).

Statistical analysis
Further data-handling was done in R (version 3.6.3) with the Rstudio GUI (version 1.1.453). Custom R scripts are available at at https://gith ub.com/thomazbastiaanssen/Tjazi (Bastiaanssen, 2018). Stacked barplots were generated by normalizing counts to 1, generating proportions. Genera that were never detected at a 1% relative abundance or higher were aggregated and defined as rare taxa for the purposes of the stacked barplots. Principal component analysis was performed on centred log-ratio transformed (clr) values using the ALDEx2 library (Fernandes et al., 2014). Number of permutations was always set to 1000. Volatility was measured as distance between before and after the experiment or treatment and was calculated as the Aitchison distance between the two timepoints. Unlike other distance metrics such as Bray-Curtis and Jensen-Shannon divergence, the Aitchison distance takes into account the compositional nature of microbiome datasets (Aitchison et al., 2000). Piphillin was used for functional inference from 16S rRNA gene seqeunce of mouse stool samples in the form of KEGG orthologues (Iwai et al., 2016). Gut-Brain Modules (GBMs) and Gut-Metabolic Modules (GMMs) were calculated using the R version of the Gomixer tool (Valles-Colomer et al., 2019). Differential abundance of both microbes and functional modules were calculated using implementations of the ALDEx2 library. As part of testing for correlations between volatility and metadata, skadi, an implementation of jackknifing and Grubb's test, was used to assess reliability of the data and detect outliers (Bastiaanssen, 2018). Correlation was assessed using Spearmans's rank correlation coefficient in the case of low N or heteroskedacity. In all other cases the Pearson correlation coefficient was used. Normality was assessed using the Shapiro-Wilk test. For normally distributed data, between-group differences were analysed using ANOVA or unpaired two-tailed t-test and Tukey's test for post-hoc analysis. For datasets in which the condition of normality was violated the non-parametric Kruskal-Wallis test was used and post-hoc analysis was done using the Wilcoxon test. A p-value of <0.05 was deemed significant in all cases. To correct for multiple testing in tests involving microbiota or Functional Modules, the Benjamini-Hochberg (BH) post-hoc was performed with a q-value of 0.1 as a cut-off. R scripts are available online on GitHub (https://github.com/thomazbastiaanssen/V olatility).

The gut microbiome is differentially volatile in response to chronic social defeat stress
After filtering, 137 different genera were detected in the mouse microbiome samples ( Fig. 2A, B). Differential abundance analysis can be found in the supplementary files at genus (Supplementary Figure S1) and at higher taxonomic levels (Supplementary Figure S2). In the Validation cohort, stressed animals showed more changes in their microbiome compared to control animals. To quantify volatility, which we define as the degree of compositional change of the microbial ecosystem over time, we calculated the intra-subject Aitchison distance between the genus-level count tables from the same subject taken before and after the experiment using the clr-transformation (Fig. 2C). Stressed mice showed a trend toward a significantly higher degree of volatility when compared to controls in the Discovery cohort (Fig. 2D). In the Discovery cohort, we found that volatility correlated with several measurements associated with the stress response, including social avoidance behaviour (Fig. 3). In the first cohort, we randomly selected samples for 16S rRNA gene sequencing from the original 27 stressed and 29 control mice (Gururajan et al., 2019). All data regarding the Discovery cohort in this manuscript only refers to the subset of animals of which the faecal microbiome was sequenced. We ran a second larger independent cohort under the same conditions, the Validation cohort, with the intent to verify our results from the smaller Discovery cohort. In the Validation cohort, we again found a higher degree of volatility in stressed animals compared to controls, this time significantly so (Fig. 2E-F).

Volatility of the gut microbiome is correlated with aspects of the stress response
Pursuing the elevated volatility in stressed animals compared to controls, we found a significant correlation between social avoidance behaviour and volatility. This finding was again replicated in the validation cohort (Fig. 4).

Volatility is correlated with absolute change in measures of alphadiversity
As alpha-diversity and beta-diversity are related metrics, we asked whether changes in beta-diversity, volatility, would be related to changes in alpha diversity. We computed alpha-diversity based on the first three hill-numbers; Chao1, Simpson and Shannon and found correlations between these metrics and volatility in both cohorts in the stressed mice, but never in controls (Fig. 5).

Perceived Stress in Humans
To investigate whether the relation between volatility and stress was observed in humans, we tested for correlation between volatility and stress as measured by the Perceived Stress Scale (PSS) in a cohort of students undergoing academic exams. We found a significant correlation between volatility and PSS during stress, but not under non-stress conditions i.e., before the exam period (Fig. 6). On their own, we detected no differential abundance in terms of genera, Gut-Brain Modules or Gut Metabolic Modules in the human cohort after FDR (Supplementary Tables 6-8).
Additionally, we set out to compare the relationship between blood cortisol and corticosterone and volatility in our human and mouse cohorts, respectively. In humans, we found a correlation between the cortisol awakening response (AUC) and volatility. Analogously, we found a significant positive correlation between evening corticosterone levels in the discovery mouse cohort as well as a trend in the same direction in the validation mouse cohort (Fig. 7) using the Pearson correlation coefficient, even though we did not find significance in the validation cohort using the Spearman correlation coefficient, likely due to low N (Fig. 3C).

Comparing Microbiome response to stress across cohorts
Finally, we set out to investigate whether other types of changes could be replicated between cohorts. We assessed differential abundance on the genus level in both discovery and validation cohorts as well as differential abundance on the level of Gut-Brain Modules (GBMs) and Gut-Metabolic Modules (GMMs). These modules represent functional pathways curated from literature that have been reported to take place in the microbiome and are involved in either gut-brain communication or in microbiome metabolism, respectively (Valles-Colomer et al., 2019). In order to compare the responses between the two cohorts, the effect sizes representing the change per microbiome feature were extracted and those modules that were present in both cohorts were tested for correlation. In the theoretical case of perfect agreement between cohorts, the rho for the correlation of the effect sizes would be 1. We found no correlation on the genus level in any animals, but in the stressed animals we found a strong positive correlation in the effect sizes on both the GBM and GMM level (Fig. 8). The same procedure was then carried out comparing the functional changes in the mouse cohorts to those in the human cohort. The mouse cohorts were aggregated in order to promote interpretability (Fig. 9). We observed significant correlations between the responses to chronic stress for both GBM and GMM across human and mouse host species. In particular, in terms of effect sizes, GABA synthesis and isovaleric acid synthesis seemed to increase the most in both host species in terms of GBMs, while mucin degradation increased the most in terms of GMMs. Complete differential abundance features results can be found in the supplementary files in terms of genera (Supplementary Table 3 Table 4) and Gut-Metabolic Modules (Supplementary Table 5). A heatmap visual representation of these differentially abundant features can be found as well (Supplementary Fig. 1).

Discussion
Microbiome volatility is a relatively underutilized concept in microbiome ecology. With regard to stress it has not been explored previously. One exception is in the context of irritable bowel syndrome (IBS) (Halfvarson et al., 2017), a condition which has been linked to physical and psychosocial stress exposure (Mayer et al., 2001) or Inflammatory Bowel Disease (IBD), which has also been linked to stress and anxiety (Mawdsley and Rampton, 2005), both of which found display more volatility (though it was not referred to as such) in patients compared with healthy controls Ryan et al., 2020). In this study, we further investigated the concept of volatility and, for what is to our knowledge the first time, report its potential influence on stress-related central and peripheral phenotypes.
We firstly showed that mice which had higher values in biological measures commonly associated with stress, such as changes in corticosterone levels also showed an increased volatility. Secondly, we observed a significant negative correlation between volatility and social behaviour. Notably, this correlation was found in both the discovery and validation cohorts. The implication is that severity of the stressor is related to degree of volatility, indicating volatility is related to stress susceptibility and resilience. Clearly, some stressed animals showed a higher degree of volatility than others. There are two possible explanations for this observation. The first is that volatility is determined by the microbiome, which would imply that a more volatile microbiome is a marker of stress susceptibility. Conversely, a more stable microbiome would then be a marker of stress resilience. Second, an elevated volatility after stress could be the result of a more severe reaction to stress. Indeed, exposure to stress has often been associated with changes in the microbiome (Bharwani et al., 2016;Langgartner et al., 2018). Stress is known to change host physiology and behaviour including diet, both of which are known to impact the microbiome (David et al., 2014;Lucking et al., 2018;O'Connor et al., 2020;O'Connor et al., 2019). This makes it seem likely that volatility is determined at least to a degree by stress. However, recently more attention has been brought upon the temporal dynamics including stability and drift in the microbiome, viewing it as T.F. S Bastiaanssen et al. an ecological system (Kenney et al., 2020). Previous research reported that both external and microbiome composition itself influence microbiome stability (Gibbons et al., 2017). Determining causality in non-linear systems like the microbiome brings its own set of challenges. For instance, different factors often cannot be separated as they together represent the system (Sugihara et al., 2012). Taken together, it is entirely possible that stress and volatility could be in a positive feedback loop with each influencing and potentially exacerbating the other. Research incorporating more frequent and numerous timepoints could shed light on this question. A similar approach was recently employed to great effect to characterise the interaction between diet and the microbiome (Johnson et al., 2019). There, diet was shown to partially explain changes in microbiome composition over very short timeframes.
While inconclusive, we initially sought to identify features in the baseline microbiome that could explain the degree of volatility after stress (data not shown), however, such differences in baseline did not hold up in the validation cohort or in the human study. Here, we were unable to find predictor features in the baseline microbiome that were generalizable over all three cohorts. In future studies, the volatility at baseline as opposed to single point measures in the current study could be measured to address this question. Specifically, one may hypothesize that hosts with the most volatile microbiomes during neutral conditions could be the most susceptible to stress and that low volatility is a predictor of stress resilience. Additionally, future studies are needed to examine the correlation between volatility and other phenotypes of relevance following chronic stress.
The findings of this study have potential translational implications in understanding volatility in the context of human health. Indeed, the fact that we observed a correlation between self-reported stress during academic exams and volatility strengthens the notion that volatility is closely associated to stress and stress resilience. For example, one could consider volatility in the context of microbial-based interventions to treat stress-induced psychopathologies, formulations designed to stabilise the microbiome could be administered over a period of time to improve response. Alternatively, given that psychotropics are themselves known to influence the microbiome , we speculate that pre-treatment with psychobiotics which introduce specific keystone species into the microbiome may make it more receptive to the therapeutic effects of antidepressants or anxiolytics. This latter approach could be relevant especially for patients who are resistant to treatment using conventional approaches. This also opens the door to keystone species, species that when absent will destabilize the gut ecosystem, in psychobiotic formulations. On its own, it is unclear what the impact a more volatile microbiome could be on host health, if any. One could hypothesise that volatility destabilizes the microbiome resulting in an increased susceptibility for bacterial taxa to colonize. We did not find evidence of this in this study, but this might be due to the sanitary housing conditions of the animals and that mice are    T.F. S Bastiaanssen et al. coprophagic. Indeed, fecal microbiota transplantation, representing a high-alpha-diversity pool, has been shown to expedite colonization rate over natural recovery (Suez et al., 2018). Moreover, the microbiome has been shown to regulate microbiome ecosystem stability (Gibbons et al., 2017). More research is warranted to test this hypothesis.
We also found consistent changes across cohorts and even between mice and humans in the microbiome after chronic stress. We did not, however, find such agreement at a taxonomic level. This is likely due to the differences in baseline microbiome between the two cohorts and the humans. Indeed, in humans, it is well-known that interpersonal variability is much lower on the functional level than on the taxonomical level (Human Microbiome Project Consortium, 2012;Mehta et al., 2018). From our findings we extrapolate that in the context of stress, while the taxonomical changes of the microbiome seem to be cohort-dependent and ultimately baseline microbiome dependent, there is a strong agreement in how the functional microbiome changes after a stressor. A stress response in the microbiome that seems invariant of the baseline condition could indicate some sort of adaptive stress response, either on the level of the microbiome or on the host level. Per definition, the GMMs and especially the GBMs have functional implications for host health. In a recent study, GBMs were shown to be influenced by diet (Butler et al., 2020;Valles-Colomer et al., 2019). Together with the finding that stress influences these modules in a specific manner, this opens up the door for psychobiotics that specifically aim to control the levels of specific modules that are known to be altered by stress. Notably, GABA synthesis was altered in both our human and mouse cohorts. The GABAergic system has been previously shown to be modulated by the microbiome in the context of the stress response (Bravo et al., 2011) and GABA-modulating bacteria have been implicated in stress-related disorders such as depression in humans (Strandwitz et al., 2019). In addition, the synthesis pathway of the short-chain fatty acids propionate and isovalerate were inferred to be upregulated after stress. These metabolites have been shown to influence stress-resilience (van de Wouw et al., 2018). Similarly, tryptophan metabolism and quinolinic acid synthesis was found to be increased after stress and have been implicated in mental health (Cervenka et al., 2017;Schwarcz et al., 2012). Alterations in menaquinone (vitamin K2) synthesis after stress were also found. Menaquinone has anti-oxidant properties and has been reported to confer neuroprotective effects (Farhadi Moghadam and Fereidoni, 2020). Notably, most of the significant changes after stress had a positive effect size, indicating consistent gains of functions occurring. This together with the correlations we reported in the changes in GBMs and GMMs following stress supports the notion that there is a directed response to stress in the microbiome. Indeed, there has been speculation that some alterations in the microbiome due to a stressor could actually be adaptations to protect the host (Walter et al., 2020). These alterations should be pursued in future research.
To conclude, we propose that an analysis of volatility should be considered in all future longitudinal microbiome research projects. Given the novelty of this concept, we make some basic recommendations as to how examine this variable. The approach to calculating volatility presented here relies on Aitchison distance. This metric was selected because it was specifically designed to deal with compositional data, such as the microbiome (Aitchison et al., 2000;Gloor et al., 2017). While other metrics for beta-diversity do exist (Bokulich et al., 2018), Aitchison distance has the added benefit of satisfying the criteria for being a Euclidean distance, making comparisons between two distances within the same analysis possible. Other popular metrics like Bray-Curtis or Unifrac do not have this property, but rather give relative distance on a scale from zero to one, making them less suitable for the purpose of assessing volatility. We speculate that further convergence of high-dimensional mathematics, microbiology and genetics will lead to newer algorithms which prove to be more useful and easier to use. For instance, the phILR beta-diversity metric combines the taxonomic framework of Unifrac with the compositionally appropriate Ilr-transformation and should be explored in further volatility research (Silverman et al., 2017). Lastly, in this study, volatility was calculated by assessing the distance 'travelled' between two points over time. Future studies should consider collecting samples over multiple time points (e. g. during stress exposure) to produce higher-dimensional geometric shapes in microbiome-space which could lead to more nuanced insights into the role of the microbiome as a mediator of the stress response.

Funding
This research was funded by a Marie-Sklowdowska-Curie Actions Individual Fellowship to AG from the European Research Council (Grant Number 704995), MyNewGut (human study; Grant number 613979) and conducted in the APC Microbiome Institute, which is funded by Science Foundation Ireland (SFI; grant nos. SFI/12/RC/2273). TGD and JFC are also supported by the Irish Health Research Board, the Dept. of Agriculture, Food and the Marine and Enterprise Ireland. TGD and JFC are the principal investigators in the APC Microbiome Institute, University College Cork. The authors declare that they have no competing interests. The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. The APC Microbiome Institute has conducted research funded by many pharmaceutical and food companies. TGD has been an invited speaker at meetings organized by Servier, Lundbeck, Janssen and AstraZeneca and has received research funding from Mead Johnson, Cremo, Wellness, Nutricia and 4D Pharma. JFC has been an invited speaker at meetings organized by Mead Johnson, Yakult, Alkermes and Janssen and has received research funding from Mead Johnson, Cremo, Suntory Wellness, Nutricia and 4D Pharma.

Availability of data and materials
All data generated or analysed during this study are included in this published article and its supplementary information files. In particular, mouse microbiome data (Supplementary Table 2) and stress-associated measurements (Supplementary Table 1) from the validation cohort.

Author contributions
TFSB, AG and JFC designed the study. AG and MVW carried out all stress procedures and behavioural experiments. CML carried out the perceived stress experiments in human subjects. AG, MVW, NLR, GM, CML and JL carried out the behavioural and physiological analyses. NCW, ABM and FF performed 16S-RNA sequencing. TFSB carried out the microbiome and bioinformatics analyses. TFSB and AG analysed the data and co-wrote the manuscript. CS, MC, TGD and JFC provided supervision. All authors contributed to the interpretation of the data, critically revised the manuscript and read and approved the final version before submission.

Ethics approval and consent to participate
All animal experiments were conducted conducted in accordance with European Directive 86/609/EEC, Recommendation 2007/526/65/ EC, and approved by the Health Products Regulatory Authority (AE19130-P031) as well as the Animal Experimentation Ethics Committee of University College Cork. All efforts were made to minimize animal suffering and to reduce the number of animals used.
The human study was approved by the Clinical Research Ethics Committee of the Cork Teaching Hospitals (study number APC080) and conducted in accordance with the ICH Guidelines on Good Clinical Practice and the Declaration of Helsinki. Written informed consent was obtained from all participants at the screening visit, before any study procedures were conducted. Participants were free to withdraw from the study at any time.

Declaration of Competing Interest
The authors report no declarations of interest.