SalmoSim: the development of a three-compartment in vitro simulator of the Atlantic salmon GI tract and associated microbial communities

Background The aquaculture sector now accounts for almost 50% of all fish for human consumption and is anticipated to provide 62% by 2030. Innovative strategies are being sought to improve fish feeds and feed additives to enhance fish performance, welfare, and the environmental sustainability of the aquaculture industry. There is still a lack of knowledge surrounding the importance and functionality of the teleost gut microbiome in fish nutrition. In vitro gut model systems might prove a valuable tool to study the effect of feed, and additives, on the host’s microbial communities. Several in vitro gut models targeted at monogastric vertebrates are now in operation. Here, we report the development of an Atlantic salmon gut model, SalmoSim, to simulate three gut compartments (stomach, pyloric caecum, and midgut) and associated microbial communities. Results The gut model was established in a series of linked bioreactors seeded with biological material derived from farmed adult marine-phase salmon. We first aimed to achieve a stable microbiome composition representative of founding microbial communities derived from Atlantic salmon. Then, in biological triplicate, the response of the in vitro system to two distinct dietary formulations (fishmeal and fishmeal free) was compared to a parallel in vivo trial over 40 days. Metabarcoding based on 16S rDNA sequencing qPCR, ammoniacal nitrogen, and volatile fatty acid measurements were undertaken to survey the microbial community dynamics and function. SalmoSim microbiomes were indistinguishable (p = 0.230) from their founding inocula at 20 days and the most abundant genera (e.g., Psycrobacter, Staphylococcus, Pseudomonas) proliferated within SalmoSim (OTUs accounting for 98% of all reads shared with founding communities). Real salmon and SalmoSim responded similarly to the introduction of novel feed, with majority of the taxa (96% Salmon, 97% SalmoSim) unaffected, while a subset of taxa (e.g., a small fraction of Psychrobacter) was differentially affected across both systems. Consistent with a low impact of the novel feed on microbial fermentative activity, volatile fatty acid profiles were not significantly different in SalmoSim pre- and post-feed switch. Conclusion By establishing stable and representative salmon gut communities, this study represents an important step in the development of an in vitro gut system as a tool for the improvement of fish nutrition and welfare. The steps of the system development described in this paper can be used as guidelines to develop various other systems representing other fish species. These systems, including SalmoSim, aim to be utilised as a prescreening tool for new feed ingredients and additives, as well as being used to study antimicrobial resistance and transfer and fundamental ecological processes that underpin microbiome dynamics and assembly. Video abstract Supplementary Information The online version contains supplementary material available at 10.1186/s40168-021-01134-6.

metagenomics and amplicon sequencing approaches can provide a detailed taxonomic assessment of the microbial composition of the gut, they may be less useful for day-to-day monitoring of speci c taxa [18].
In view of the above criteria, the aim of the current study is to develop a synthetic, continuous salmon gut microbial fermentation simulator, representative of generalised marine lifecycle stages of the Atlantic salmon. Salmonids are gastric shes [19], with their guts characterised by a clearly de ned stomach followed by a pylorus with attached blind vesicles called pyloric caeca, as well as a relatively short and non-convoluted posterior (mid and distal) intestine leading to the anus [20]. Our experimental gut system simulates the stomach, the pyloric caeca, and the midgut regions of the gastrointestinal tract of farmed Atlantic salmon. To validate the system, we compare the performance of the simulator with in vivo modulation of the gut microbial community during a feed trial that ran in parallel. We aimed to rst assess microbial community dynamics within SalmoSim over time; then compare the SalmoSim microbiome to that found in real salmon; and nally, to evaluate the response to the two systems (SalmoSim and real salmon) to a change in diet.

Methods
Experimental set-up and sample collection in an aquaculture setting The Atlantic salmon (Salmo salar) In vivo feed trial was performed by MOWI ASA at their research site in Averøy, Norway. Prior to commencement of the feed trial, salmon were fed on a sh meal diet (FMD) until they reached circa 750 grams in mass. Fish were separated into 5x5 meter marine pens (150 randomly distributed sh per pen) in 4x4 modular design. Four pens were randomly assigned to each of the trial diets. This study focused on eight pens housing sh fed on FMD and sh meal free diet (FMF) (Supplementary Table 1, Figure 1D). The feed trial was conducted over ve months (November 2017-March 2018. For in vivo samples recovered from actual salmon, three randomly selected sh were collected at the end of the feed trial for two different feeds (N=3 sh/feed) and sacri ced by MOWI employees. After, samples from three salmon gut compartments were collected (stomach (N=3/feed), pyloric caeca (N=3/feed) and mid gut (N=3/feed) (approximately 20 cm from the vent)), placed into 1.5 ml cryovials and kept on ice before long term storage in -80 o C conditions. For in vitro initial inoculum samples (the founding community for SalmoSim runs), a further three sh fed on FMD were sacri ced and samples from three distinct gut compartments were collected (Stomach (N=3), pyloric caecum (N=3) and midgut (N=3)), transferred to 15 ml Falcon tubes containing 30% glycerol and kept on ice before long term storage in -80 o C conditions. Details of samples collected from farmed Atlantic salmon have been described previously [21].
In vitro feed trial within SalmoSim system Physicochemical conditions within Atlantic salmon gastrointestinal tract and microbiome sampling Physicochemical conditions (temperature, pH, dissolved oxygen) were directly measured in adult Atlantic salmon from Mowi salmon farm in Loch Linnhe, Scotland (Supplementary Figure 1A-C). Bacterial inoculums were prepared for the in vitro trial from the different gut compartments sampled from individual sh (three biological replicates, three gut compartments per sh -stomach, pyloric caecum and midgut) collected at the start of the in vivo feed trial in Averøy, Norway. Prior to SalmoSim inoculation, inoculums that were stored in 15 ml falcon tubes containing 30% glycerol solution in -80 o C freezer were dissolved in 1 ml of autoclaved 35 g/L Instant Ocean® Sea Salt solution. Distinct individual sh collected in Averøy formed the founder community for each distinct replicate of the in vitro trial (i.e., a true biological replicate (Figure 1)).
In vitro system 'feed' preparation In vitro system feed media was prepared by combining the following for a total of 2 litres: 35 g/L of Instant Ocean® Sea Salt, 10 g/L of the FMD or FMF used in the MOWI feed trial (Supplementary Table 1), 1 g/L freeze-dried mucous collected from the pyloric caecum and 2 litres of deionised water. This feed was then autoclave-sterilised, followed by sieving of the bulky occulate, and nally subjected to the second round of autoclaving.
In vitro system preparation Three 500 ml Applikon Mini Bioreactors were lled with four 1cm 3 cubes made from sterile aquarium sponge lters used as a surface for bio lm formation, assembled by attaching appropriate tubing and probes (redox, temperature, and dissolved oxygen, Figure 1A), and autoclaved. Bioreactor preparation was followed by attachment of reactor vessels to the Applikon electronic control module, connection of feed and acid and base bottles (0.01M hydrochloric acid and 0.01M sodium hydroxide solutions ltered through 0.22µm polyethersulfone membrane lter unit (Millipore, USA)). Nitrogen gas was periodically bubbled through each vessel to maintain anaerobic conditions and dissolved oxygen continually monitored. The bioreactors were then allowed to ll with 400 ml of feed media. Once the system was set up, media transfer, gas ow and acid/base addition occurred for 24 hours in sterile conditions (without microorganisms present in the system) in order to stabilise the temperature, pH, and oxygen concentration with respect to levels measured from adult salmon.
Initial pre-growth period during in vitro trial In order to allow bacterial communities to proliferate in the in vitro environment without washing through the system, the microbial populations within the inoculum from real salmon were pre-grown inside the SalmoSim system for four days. During this phase, the system was lled with FMD media preparation and inoculum, and no media transfer occurred.
Performing feed trial within SalmoSim system After the initial pre-growth period, each validation experiment was run for 20 days while supplying SalmoSim system with FMD. After 20 days, SalmoSim was run for 20 additional days while supplying FMF food. During the full 44-day experiment (4-day pre-growth period, 20-day system fed on FMD, and 20day system fed on FMF) physiochemical conditions within three SalmoSim gut compartments were kept similar to the values measured in real salmon: temperature inside the reactor vessels were maintained at 12 o C, dissolved oxygen contents were kept at 0% by daily ushing with N 2 gas for 20 minutes, and pH was kept stable in each bioreactor by the addition of 0.01M NaOH and 0.01M HCl (stomach pH 4.0, pyloric caecum pH 7.0, and mid intestine pH 7.6). During this experiment (apart from the pre-growth period) the transfer rate of slurry between reactor vessels was 238 ml per day (5% of the total bioreactor volume -the maximum volume of sampling without disturbing microbial community structure). Finally, 1 ml of ltered salmon bile and 0.5 ml of sterile 5% mucous solution were added daily to the reactor simulating the pyloric caecum compartment. The schematic representation of SalmoSim system is visualised in Figure 1A and full feed trial within SalmoSim is visually summarised in Figure 1B and C.
Sampling was performed in several steps. First, samples from initial inoculums were collected for each gut compartment. Once SalmoSim main experiment started, the sampling from each bioreactor vessel was performed every second day throughout the 40-day run period (20 samplings in total). The SalmoSim samplings entailed collecting 30 ml of the bioreactor contents into 50 ml falcon tubes, centrifuging them for 10 minutes at 5000 rpm speed, and freezing the pellets at -20 o C for storage.
Measuring nitrogen metabolism within the SalmoSim system At each sampling point, the protein concentration in each chamber of the system was measured using Thermo Scienti c™ Pierce™ BCA Protein Assay Kit (Thermo Fisher Scienti c, USA) and the ammonia concentration using Sigma-Aldrich® Ammonia Assay Kit (Sigma-Aldrich, USA) to assay the bacterial community activity. Both methods were performed according to manufacturer protocol by using The Jenway 6305 UV/Visible Spectrophotometer (Jenway, USA).

Measuring Volatile Fatty Acid (VFA) production in SalmoSim
The last two time points for each diet were selected from the SalmoSim system (for all three gut compartments) for VFA analysis: 18 and 20 for FMD; and time points 38 and 40 for FMF, respectively, to ensure that the bacterial communities had enough time to adapt to SalmoSim system (for FMD time points) and the feed change (for FMF time points). During runs, 1ml of supernatant from SalmoSim bioreactors was frozen in -80°C which, was then used for VFA extraction. The protocol involved combining 1ml of supernatant with 400µl of sterile Phosphate-buffered saline (PBS) solution (Sigma Aldrich, USA) and vortexing the mixture for 1 minute. The sample was then centrifuged at 16,000 g for 30 minutes, followed by two rounds of supernatant removal, before centrifuging at 16,000 g for 30 minutes. Finally, the supernatant was then ltered through 0.2µm Costar SpinX centrifuge tube lters (Corning, USA) at 15,000 g for 2 minutes until clear. The extracted VFAs were sent for gas chromatographic analysis at MS-Omics (Denmark). In order to determine if the VFA concentrations were statistically different between SalmoSim fed on FMD and FMF diets, measured VFA values dataset were subjected to statistical analysis using linear mixed effect models (See Supplementary methods 2). Results are shown in supplementary Figure 6.
In vivo phenotypic sh performance fed on two different feeds Phenotypic performance data ( sh length, weight, gutted weight, carcass yield, gonad, and liver weights) was collected and provided at the end of the in vivo feed experiment by MOWI. The differences between each feed group (n=32 sh per feed) for each phenotype were visualised and statistical analysis was applied (independent two-sample t-test) to identify statistically signi cant differences between the two feed groups.

Measuring bacterial population dynamics in SalmoSim
Genomic DNA extraction The DNA extraction protocol followed was previously described by [21]. In short, samples were subjected to a bead-beating step for 60 seconds by combining samples with MP Biomedicals™ 1/4" CERAMIC SPHERE (Thermo Fisher Scienti c, USA) and Lysing Matrix A Bulk (MP Biomedicals, USA). Later, DNA was extracted by using the QIAamp® DNA Stool kit (Qiagen, Valencia, CA, USA) according to the manufacturer's protocol [22].

NGS library preparation and sequencing
In the rst instance, microbial population dynamics in SalmoSim were measured in near real-time using a set of qPCR primers including published and custom sequences to enable the stability of the system to be monitored (See supplementary Methods 1 and data Supplementary Figure 3). Subsequently 16S rRNA sequencing was deployed to provide a fuller picture community dynamics. The commonly used 16S ribosomal hypervariable region 4 primers were shown to cross-amplify Salmo salar 12S ribosomal gene sequences [21,23] and hence were not used in this study. Rather, ampli cation of the 16s V1 hypervariable region was adopted as an alternate approach [24]. Ampli cation of 16S V1 hypervariable region from diluted DNA samples was achieved using redundant tagged barcode 27F and 338R at nal concentration of 1 pM of each primer. Primer sequences are summarised in Supplementary Table 3. Firstround PCR was performed in triplicate on each sample and reaction conditions were 95°C for ten minutes, followed by 25 cycles at 95°C for 30 seconds, 55°C for 30 seconds and 72°C for 30 seconds, followed by a nal elongation step of 72°C for 10 minutes. After the triplicate reactions were pulled together into one, their concentration was measured using a Qubit® uorometer (Thermo Fisher Scienti c, USA), and all of them were diluted to 5ng/µl using Microbial DNA-Free Water (Qiagen, Valencia, CA, USA). The secondround PCR, which enabled the addition of the external multiplex identi ers (barcodes), involved only six cycles with otherwise identical reaction conditions to the rst. The detailed composition of second-round PCR primers is summarised in Supplementary Table 4. This was followed by the DNA clean-up using Agencourt AMPure XP beads (Beckman Coulter, USA) according to the manufacturers' protocol. The cleaned DNA was then gel-puri ed by using the QIAquick Gel Extraction Kit (Qiagen, Valencia, CA, USA) and then quanti ed by using Qubit® (Thermo Fisher Scienti c, USA). All the PCR products were pulled together at 10nM concentration and sent for sequencing using HiSeq 2500.
Bioinformatic analysis of 16S rRNA gene amplicon sequencing data Sequence analysis was performed with our bioinformatic pipeline as described previously with slight modi cations [21]. First, quality ltering and trimming (>Q30 quality score) was performed on all the reads of the 16s rRNA V1 hypervariable region using Sickle version V1.2 software [25]. Second, read error correction was performed using the BayesHammer module within SPAdes V2.5.0 software to obtain highquality assemblies [26]. Third, paired-end reads were merged (overlap length 50bp) using PANDAseq v2.11 software with simple_bayesian read merging algorithm [27,28]. After overlapping, paired end reads merged reads were dereplicated, sorted, and ltred chimaeras using GOLD SILVA reference dataset [29] and singletons were removed by using VSEARCH version 2.3.4 tool [30]. Merged pair-end ltered reads were clustered in operational taxonomic units (OTUs) using VSEARCH software at 97% identity followed by a decontamination step by mapping OTUs against the host (Salmo salar) reference genome (available on NCBI) DNA using bwa aligner implemented in DeconSeq v0.4.3 tool [31]. Taxonomic assignment of OTUs was achieved using the Naïve Bayesian Classi er [32] implemented in the QIIME 2 platform using SILVA 132 database [33,34]. Phylogenetic trees of OTUs were generated using FastTree software after using MAFFT for multiple sequence alignment [35]. The resultant OTU table was converted to a biological observation matrix (BIOM) format for the post-OTUs statistical analysis [36].
Post-OTUs statistical analysis: diversity metrics and community structure and composition analysis All statistical analysis of the OTU table was performed by using R v 4.0.1 and RStudio v 1.3.959 [37].
Alpha diversity analysis was performed using Rhea pipeline [38], supplemented by microbiomeSeq [39], and PhyloSeq [40] for ANOVA and visualisation steps. Two alpha diversity metrics were calculated: microbial richness (estimated number of observed OTUs) and Shannon diversity (an estimate of balance of the community using effective Shannon index [41,42]. Before calculating effective microbial richness, proportional ltering was performed at a relative abundance of 0.25% in each community to minimise the in ation in microbial richness caused by the very low abundant OTUs. Afterwards, a one-way ANOVA of diversity between groups was calculated with the p-value threshold for signi cance (p-value <0.05) represented using boxplots.
To investigate the effect of time on the bacterial community stability, beta diversity analysis was performed using different phylogenetical distances metrics to assay phylogenetic similarities between samples (weighted, generalised, and unweighted UniFrac). To compare communities isolated from various sources (SalmoSim, inoculum and real salmon), only samples fed on FMD were included as initial inoculum were collected from sh fed on FMD alone. Furthermore, only SalmoSim samples from the last 3 time points fed on FDM were selected as they are considered stable time points (once bacterial communities had over two weeks to adapt and grow within SalmoSim system). In short, the resulting dataset contained: real salmon samples fed on FMD, all inoculum samples and stable SalmoSim time points fed on FMD (days 16, 18, and 20). This dataset was then subdivided into several smaller datasets that included OTUs, shared by various percentage of samples (60%, 50%, 40% and 30% of samples), with the aim of minimising the impact of rare OTUs (low prevalence) on comparisons and only focusing on prevalent OTUs between samples (see details in Supplementary Table 4).
To analyse the response of microbes to the diet change (see supplementary Table 1 for feed formulation) in real salmon and SalmoSim, in addition to the full dataset (in vivo and in vitro samples); three different full dataset subsets were used to perform beta diversity analysis: samples from in vivo study, all samples from SalmoSim (all data points), and samples only from SalmoSim once it had achieved stability (the last 3 time points fed on FMD (days: 16, 18, and 20) and FMFD (days 36, 38, and 40). These datasets were used to compute ecological (Bray-Curtis and Jaccard) and phylogenetic (unweighted, weighted, and generalised UniFrac) distances with vegdist() function from the vegan v2.4-2 package and GUniFrac() function (generalised UniFrac) from the Rhea package [38,43] Both ecological and phylogenetical distances were then visualised in two dimensions by Multi-Dimensional Scaling (MDS) and non-metric MDS (NMDS) [44]. Finally, a permutational multivariate analysis of variance (PERMANOVA) by using calculated both distances was performed to determine if the separation of selected groups is signi cant as a whole and in pairs [44].
To provide an overall visualisation of microbial composition across all samples, a principal Coordinates Analysis (PCoA) was performed using the microbiomeSeq [39] package based on phyloseq package [45] with Bray-Curtis dissimilarity measures calculated and visualised for four different subset-datasets: the full dataset (real salmon, inoculum and all SalmoSim samples), and, three different subsets each containing only one of the free biological replicate samples from SalmoSim (Fish 1, 2, or 3), along with all real salmon and inoculum samples. Differential abundance was calculated by using microbiomeSeq based on DESeq2 package [39,45]. BIOM generated OTU table was used as an input to calculate differentially abundant OTUs between selected groups based on the Negative Binomial (Gamma-Poisson) distribution.

Results
Stabilisation of microbial communities within the SalmoSim system Effective richness (Figure 2 A) indicates that within the stomach and midgut compartments the initial inoculum contained the highest number of OTUs compared to later sampling time points from SalmoSim system: in the stomach compartment, effective richness was statistically different between time point 0 Our results reveal that within the stomach compartment over time (including initial inoculum), the effective Shannon diversity was stable with a downwards but non-signi cant trend over the course of the experiment (Figure 2 B). A similar downwards trend was observed in the pyloric caecum, with signi cant differences between later time points, but no signi cant differences were noticed between the inoculum and SalmoSim. Within the midgut compartment Shannon diversity was statistically lower than the inoculum (time point 0) over most time points (sampling days 2-40).
Taken together, diversity and richness metrics suggest some loss of microbial taxa as a result of transfer into SalmoSim in the pyloric caecum and midgut, but not in the stomach. Subsequently, richness and evenness were then stable over the time course of the experiment in stomach and mid gut compartments (some instabilities seen only between initial inoculum and later time points), whilst much more instability within alpha diversity metrics were detected in the pyloric caecum compartment.
To assess the compositional stability of the system, comparisons over time were undertaken with reference to pairwise beta-diversity metrics. Signi cant differences in composition between time points represent instability in the system. To compare different samples types (inoculum, real salmon, SalmoSim) sample sizes were rst balanced by examining a reduced dataset that contained: real salmon samples, all inoculum samples and stable SalmoSim time points fed on FMF (days 36, 38, and 40). Alpha diversity comparisons between inoculum, real salmon and SalmoSim are shown in Figure 2.
Compositional comparisons between different samples types (inoculum, real salmon and SalmoSim) were made using several pairwise beta-diversity metrics (phylogenetic (unweighted, generalised, and weighted UniFrac and ecological distances (Bray Curtis and Jaccard)) ( Table 1). Ecological metrics could not distinguish between SalmoSim and inoculum. In contrast, metrics that incorporate phylogenetic differences between taxa (i.e., Unifrac) did identify signi cant differences, indicating that there is variability between the inoculum and SalmoSim, but the taxa involved are closely related. To explore the impact of rare OTUs when accounting for observed differences between sample types, the dataset was partitioned and analysed. Partitioned datasets indicated that progressive removal of rare OTUs increased the compositional similarity of the inoculum and SalmoSim via all metrics (Table 1). Between real salmon (in vivo) and SalmoSim, and between real salmon and inoculum, however, statistically signi cant differences were found in the majority of cases using all metrics regardless of inclusion of rare OTUs. These observations are consistent with the fact that SalmoSim and inoculum samples originated from the same individuals, while real salmon samples were, by necessity, collected from different individuals during the in vivo trial.

Effect of changing diet on the microbiome of real salmon (in vivo) and SalmoSim (in vitro).
The impact of diet on the abundance of individual taxa: In response to the change of diet, the relative abundances of individual taxa in salmon vs SalmoSim also revealed some differences, as well as multiple similarities in response of the two systems ( Figure 4). In this respect, the abundance of the vast majority of OTUs (SalmoSim -97%; Salmon -95%; Figure 4 C) were unaffected by the change in feed; these included 161 OTUs shared by SalmoSim and the real salmon assayed. For OTUs whose individual abundance was impacted by feed across the two systems, only a single common OTU changed in the same way in both Salmon and SalmoSim (Figure 4 A). qPCR-based estimates of taxon abundance variation in response to diet (Supplementary Table 6), and corresponding data for the same taxa from 16S OTU pro les ( Figure 4D) show several similarities and differences between SalmoSim and real salmon. Again, however, the overall pattern is that of limited change in both in vivo and in vitro systems in response to the change in diet. Invariance observed in the microbiome in response to feed were re ected in estimates of physical attributes of sh in response the change in feed formulation. As such, no statistically signi cant differences in various phenotypic measurements ( sh length, weight, gutted weight, carcass yield, gonad, and liver weights) were noted in salmon fed on the two different diets used in the experiment (see Supplementary Figure 5). Invariance was also observed in VFA production data, in which no signi cant differences were observed in SalmoSim between the sh meal and sh meal free diets (See Supplementary Figure 6).
Microbial composition in SalmoSim and real salmon fed different feeds: Most gut compartments for real salmon, SalmoSim, and the salmon used to inoculate SalmoSim were dominated by Pseudomonas, Psychrobacter and Staphylococcus genera, suggesting that genera present in the marine phase salmon are generally maintained in SalmoSim ( Figure 5). In terms of change in alpha diversity, the only statistically signi cant difference in response to the switch in feed was observed in the pyloric caeca compartment of the SalmoSim compartment based on the Shannon diversity metric ( Supplementary  Figure 4), where a slight decrease alongside the FMF occurred. Otherwise, the change in feed formulation did not impact alpha diversity in any gut compartment, either in real salmon, or in SalmoSim.
To provide an overview of microbial composition and variation in the experiment, a PCoA (Principal coordinates Analysis) based on Bray-Curtis distance was performed and plotted ( Figure 6 A-D). Biological replicate (the sh providing the founding inoculum of each SalmoSim run) appears to be a major driver of community composition in the experiment (Figure 6 A). Taxonomic composition represented in stacked bar plots in Figure 5 also supports this observation. Once individual SalmoSim runs (biological replicates) are visualised separately changes to microbial communities in response to the feed become apparent (Figures 6 B-D). Statistical comparisons based PERMANOVA show there is a clear effect of feed on microbial composition in both salmon and SalmoSim (Table 2), however, based of OTU differential abundance data (above) the effect seems to be small. Samples from real salmon fed on the different diets also diverge from one and other (supported by Table 2, Figure 5), however, not necessarily along the same axes as each SalmoSim replicate. This divergence is potentially indicative of an effect of the biological replicate (i.e., inter-individual variation). Consistent with Figure 5, inoculum for the respective SalmoSim replicates cluster among SalmoSim samples for the sh meal diet in each case.

Discussion
Our ndings suggest a loss of microbial taxa diversity and richness as a result of transferring initial inoculums from real salmon into the SalmoSim system in the pyloric caeca and mid gut compartments. Several lines of evidence from our core OTU analysis suggest that low prevalence (rare) OTUs make up most of the taxa lost, and progressive removal of rare OTUs increased the compositional similarity between inoculum and SalmoSim samples using both phylogenetic and ecological distances. A general trend was observed, in which all gut compartments became increasingly stable throughout the 40-day of experiment, with little-observed impact of introducing the different feed at day 20. Comparison of real salmon and SalmoSim samples at the microbial level showed signi cant differences using both ecological and phylogenetic metrics. These differences may be explained by the fact that samples used for real salmon and SalmoSim originated from different individuals, whereas initial inoculum and SalmoSim samples for a given run originated from the same sh. Correspondingly, we observed that the biological replicate (the founding inoculum of each SalmoSim run) was the major driver of community composition in the experiment. This could be explained by the fact that feed used in the in vitro study was sterile, thus the only bacterial communities within the SalmoSim system originated only from real salmon inoculums. Once the individual runs were separated, phylogenetic and ecological distances suggested that changing feed was the driver of community composition in both real salmon and SalmoSim. However, the vast majority of OTUs remained unchanged by the switch in feed in both systems and no changes were noticeable in the bacterial activity (VFA production) within the system after the introduction of plant-based feed, nor in phenotypic performance of Atlantic salmon fed on two different feeds ( sh length, weight, gutted weight, carcass yield, gonad, and liver weights) in in vivo trial.
Many of the microbes we detected, and cultured, from the salmon gut microbiome have been reported previously in this species. For example, gram-negative Pseudomonas and Psychrobacter, the most abundant genera we observed, are among the core bacterial taxa known to reside within the real salmon gut [24,47,48]. Staphylococcus genera have also been reported widely in fresh-water and marine farmed salmon [49]. SalmoSim was able to maintain these species in culture throughout the experimental run, and although some diversity was lost, no statistical differences could be detected between the composition of SalmoSim and that of the sh gut communities used to found the different biological replicates via ecological metrics. Notable by their scarcity were mycoplasma OTUs, which occurred at relatively low abundance in both the in vivo and in vitro systems in this study. Mycoplasma OTUs were recovered from most SalmoSim gut compartments at low abundances (see supplementary Table 7), suggesting that these fastidious microbes can survive in the bioreactors. Our group and several others have widely reported Mycoplasma species from marine and freshwater stage of wild and farmed Atlantic salmon (higher abundance in farmed salmon), where many proliferate intracellularly in the gut epithelial lining [21,[50][51][52]. Establishing whether mycoplasma can actively proliferate in SalmoSim would require the use of founding communities rich in these organisms in a future experiment, and we found, as has been reported in a number of other microbiome studies (e.g [53], that inter-individual variability (in our case affecting the initial inoculum) was a main driver for gut microbial composition divergence.
We identi ed that a change in feed resulted in an overall shift in microbial community structure in both real salmon and SalmoSim system, as has been found to be the case in many previous studies [7,54,55].
The direction of this shift, and the microbial taxa involved, were not equivalent in SalmoSim and real salmon, although no overall trend was observed at higher taxonomic levels in either system. Importantly, it is also the case that the vast majority of OTUs within both real salmon and SalmoSim were not affected by the switch in feed. Furthermore, it was found that change in feed did not affect VFA production in the SalmoSim system. As such, it is not clear whether any relevant functional shifts occurred in the microbiome of SalmoSim or real salmon as a result of the treatment. Furthermore, we did not identify any phenotypic changes ( sh length, weight, gutted weight, carcass yield, gonad, and liver weights) within in vivo trial of Atlantic salmon fed on two different feeds. This lack of change is not unexpected, considering the plant-based feed was developed to have similar macronutrient composition to a Fish meal-based feed. One difference is a slightly higher crude bre (fermentable substrate) proportion in Fish meal free diet, which could explain higher microbial diversity in in vivo samples fed on plant-based feed [56].
The use of in vitro systems to study and model the microbial communities of monogastric vertebrates is becoming increasingly widespread, with systems simulating: Sus scrofa (pig) [57], Gallus gallus (chicken) [58], Canis lupus (dog) [59] and other vertebrate guts. Using in vitro gut simulators is also a widely accepted approach to study the human gut microbiome [15,60,61]. One of the most established systems is the Simulator of the Human Intestinal Microbial Ecosystem (SHIIME) that mimics the entire gastrointestinal tract incorporating the stomach, small intestine and different colon regions [16]. This system was used to study the effects of many different dietary additives on human microbiome [62,63].
The value of in vitro simulators in providing genuine insights is limited only by the research question and the corresponding level of sophistication required. The host component of the system, for example, is often poorly modelled, although cell lines, arti cial mucosae and digestion / absorbance systems can be included, which can provide speci c physiological and metabolic insights [60,64]. As we found, interindividual variability may be an important consideration, and adequate biological replication is necessary to enable reliable interpretation of results, a consideration that can be overlooked by even the most sophisticated systems. Prior to the current study, only one other attempt was made to study the effect of diet on Atlantic salmon gut microbial composition in vitro [65]. In this preliminary study a simple in vitro system was used to assess the impact of different feed formulations on the microbial communities of faecal slurries prepared from live salmon. However, no direct comparison was made with a true in vivo trial; nor were the different gut compartments present in salmon modelled in any detail and the predictive value for such simple in vitro systems in not immediately clear. Nonetheless, the work provided an important catalyst for the development of more sophisticated systems.

Conclusions
Our results indicate that SalmoSim can not only maintain stable microbial communities from real salmon, but also demonstrates similar responses in experimental treatments as those seen in real salmon. These results are encouraging, however, the nature of the treatment applied in this study: a switch between two similar feeds that had little effect on the gut microbiota in vivo, suggests that further experimentation with SalmoSim would be bene cial. For example, the survival and in uence of probiotics within the system or the in uence of known prebiotics could also be assessed, as has been previously studied in other in vitro gut systems [59]. Gut models such as SalmoSim could provide a powerful role in aquaculture, where there is considerable interest associated with the development of feed and feed additives [66][67][68], but where the capacity for in vivo trials is limited. The aim of such systems could be to provide pre-screening tool for new feed ingredients and additives with the aim of reducing the cost and scale of in vivo testing. In parallel, an in vitro gut model for salmon could also be exploited to understand questions of public health importance (e.g. antimicrobial resistance and transfer [58]), as well as the fundamental ecological processes that underpin microbiome dynamics and assembly.