An in vitro model maintaining taxon-specific functional activities of the gut microbiome

In vitro gut microbiome models could provide timely and cost-efficient solutions to study microbiome responses to drugs. For this purpose, in vitro models that maintain the functional and compositional profiles of in vivo gut microbiomes would be extremely valuable. Here, we present a 96-deep well plate-based culturing model (MiPro) that maintains the functional and compositional profiles of individual gut microbiomes, as assessed by metaproteomics, while allowing a four-fold increase in viable bacteria counts. Comparison of taxon-specific functions between pre- and post-culture microbiomes shows a Pearson’s correlation coefficient r of 0.83 ± 0.03. In addition, we show a high degree of correlation between gut microbiome responses to metformin in the MiPro model and those in mice fed a high-fat diet. We propose MiPro as an in vitro gut microbiome model for scalable investigation of drug-microbiome interactions such as during high-throughput drug screening.

alter the composition and function of the gut microbiota [5][6][7][8][9] , which in turn impacts the host. Therefore, the gut microbial ecosystem, specific microbes, and microbial pathways are novel targets in drug discovery.
In vitro culture models could provide time-and-cost saving solutions to discover microbiome responses to drugs. However, current culture models do not maintain the functional and compositional profiles of the initial gut microbiome. To mimic an in vivo microbial ecosystem, it is key to develop culturing models and culturing media capable of conserving the composition and functional activities of an individual's microbiome. Current in vitro culturing methods have been reported to sustain microbial diversity and achieve proper cultured bacterial coverage [10][11][12][13] . However, we should not simply equal a microbiome's diversity and coverage to its structure and functionality. Profound shifts in proportions of different taxa have been frequently described compared to the inoculum, in both batch culturing 10,14 and continuous flow culturing models, such as the Chemostat 13 and the SHIME 12 . Importantly, shift of in microbiome composition can result in alternations of functional properties and ecological processes 15 , subsequently would result in different microbial responses to a drug stimulus. To the best of our knowledge, there has been no report on model maintaining microbiome properties and activities similar to the inoculum. In particular, preservation of functional activities hasn't been described elsewhere.
Moreover, comprehensive study of drug effects will require high-throughput screenings 16 . Continuous flow models (e.g. Chemostat 13 , SHIME 12 and M-SHIME 17 ) as well as microfluidic models (e.g. HuMiX 18 , and gut-on-a-chip 19 ) cannot be readily adapted for high-throughput approaches, partially due to model size and the long period of time required for setting up and stabilization with these bioreactors.
For gaining a deep insight into drug responses of a microbiome, a technique that can precisely quantify microbial functional activities is required. Among the various meta-omics tools, metaproteomics directly quantifies the microbial functional responses at the end-product level, i.e. expressed proteins 20 .
The development and application of mass spectrometry (MS)-based metaproteomics technology in gut microbiome research has thrived in recent years. The identification coverage and sensitivity of MS-based metaproteomics has increased dramatically, enabling in-depth analysis of microbiome functional activities 21,22 . Moreover, compared with sequencing-based techniques, metaproteomics has been validated to be more accurate for biomass estimates on a species level 23 , and it also enables the study of taxon-specific functions through annotation of unique peptides 24 . It is important that taxon-specific functional profiles in the microbiome is compared for revealing cultured stability and drug responses. This is because overall functions can maintain relatively stable due to the redundancy of functional genes among species in a gut microbiome 25 . Therefore, metaproteomics is an adequate tool for insights into in vitro drug responses of gut microbiome.
Here, we report the development and validation of a high-throughput in vitro model for the Maintenance of gut microbiome Profiles (MiPro). Briefly, the MiPro model adopts an optimized culture medium and a 96-deep well plate-based format for microbiome culture ( Figure 1A). The culture medium and culturing conditions were improved from our previous medium composition study 26 . The model was first evaluated for its ability to maintain gut microbiome profiles in vitro, followed by testing and evaluation of the model's in vitro correlation with in vivo drug response. This culture model enables, in combination with metaproteomic analysis, the assessment of drug effects on the microbiome at the compositional and functional levels. This model maintains high microbiome compositional stability and > 83% taxon-function similarity over 24 hr. In addition, we demonstrate that our MiPro model recapitulates the in vivo effects of metformin observed on individual mouse gut microbiomes.

Establishment of the Mipro model
The first objective of this work was to develop a high-throughput in vitro model for the maintenance of gut microbiome profiles. We have recently evaluated the composition of the culture medium for optimal culturing of ex vivo microbiotas 26 . Here, we improved upon the composition of our previously reported medium by assessing the effect of bile salts formula on the gut microbiome. Two formulae were compared: (1) a mixture of primary bile salts, i.e. 1:1 (w/w) sodium salt forms of cholic acid (CA) and chenodeoxycholic acid (CDCA), and (2) a commercialized 1:1 (w/w) mixture of sodium salt forms of CA and deoxycholic acid (DCA). The commercialized bile salts mixture has been adopted in a majority of gut microbiome culture media 13,14,[27][28][29][30][31][32][33] . The ratio of Firmicutes to Bacteroideteshas been extensively used to characterize microbiomes 34 . Globally, the inoculum ratio of Firmicutes to Bacteroidetes better approximated to the ratio from the gut microbiota grown for 24 hr in the presence of CDCA and CA than that from the microbiome grown in presence of commercialized bile salts ( Figure 1B).
Additionally, we assessed whether the gut microbiome culture was affected by the culturing conditions, namely (1) tube-based and (2) 96-deep well based culturing, while keeping all other conditions, including medium, inoculum, temperature, and container material (of polypropylene) constant.
Culture tubes are the most frequently used containers in batch culture experiments 14,26,33 . Notably, the 96deep well was covered with a silicone-gel cover, which was perforated at the top of each well. This cover could prevent gas-exchange with the outer environment in the chamber, so as to preserve the partial pressure of gases and volatile metabolites in each well, which could subsequently preserve certain levels of dissolved gas molecules in the culture medium. In contrast to the use of 96-deep well plates, employing culture tubes resulted in a remarkable change in the proportions of Firmicutes and Bacteroidetes ( Figure 1B).
From the above results, we established our MiPro model ( Figure 1A): a 96-deep well based culturing in combination with the optimized medium which contains 1:1 (w/w) of CA and CDCA (hereafter called MiPro medium).

MiPro increased viable bacteria ratio and quintupled the count
We next evaluated the ability of the MiPro model to sustain a viable microbiome, which is vital for an effective in vitro microbiome response study. Temporal changes in parameters including microbial biomass, viability and diversity were compared with the 0 hr baseline sample. The commonly used basal culture medium (BCM) was included for comparison. At 24 hr the OD 595 (bacterial intensity) was 0.8 ± 0.1 fold higher in the optimized medium as compared to that cultured in BCM ( Figure 1C). Also, our MiPro medium achieved high ratio of viable bacterial cell throughout the culture period (95.77% at 24 hr compared with 71.24% at 0 hr, Figure 1D). We observed a 4.4-fold increase of viable bacterial count in MiPro medium after 24 hr of culturing, whereas a 3.0-fold maximum increase was detected in the BCM medium at 9 hr post-culturing.

Mipro model maintains taxon-specific functional profiles of gut microbiome
In order to evaluate the effectiveness of MiPro to simulate the in vivo features of the gut microbiome inoculum, we used a metaproteomic approach 35,36 to characterize the taxonomic and functional stability of three individuals' gut microbiomes over 9, 24, 34 and 48 hr of growth in either MiPro or BCM media.
A minimum of three cultured, technical replicates were analyzed at each time point by LC-MS/MS.
Ninety high-quality MS raw files were obtained with a total of 2,066,069 MS/MS spectra. A total of 58,848 peptides and 16,326 protein groups were identified with a false discovery rate (FDR) threshold of 1% (Supplementary Figure S2A). A high concordance (Pearson's correlation coefficient r = 0.97 ± 0.02) was observed between the technical replicates of each group, indicating robust experimental reproducibility (Supplementary Figure S2B). Using an LCA approach on the MetaLab 36 , a total of 21,839 peptides were assigned with a taxonomic lineage, resulting in 788 assigned species. The quantitative information (summed peptide intensities) was used to assess the species-level biomass contributions 35,36 . 121 species that were quantified with ≥ 3 peptides were included in the comparison of species biomass contributions.
We applied a Bray-Curtis dissimilarity-based approach 37 Figure S4). In addition, both MiPro and BCM media achieved well-maintained alpha-diversity (Shannon-Wiener index) overtime across the three volunteers (Supplementary Figure S5).
To assess the stability of functional activities, the identified proteins were annotated with clusters of orthologous group (COG) categories and the abundance of each COG category was calculated by summing the LFQ intensities of all the proteins belonging to the same COG category 38 . Principal component analysis (PCA) was used to assess the relatedness of the samples based on the functional makeup of the metaproteomes. The first two components, PC1 and PC2, explained 71.7% and 21.4% of the total variance, respectively ( Figure 2C). The largest functional variability was found in response to the culture condition using the BCM medium, indicating a better maintenance of the inoculum's microbial functional profile by using the MiPro medium ( Figure 2C and Supplementary Figure S6). In order to assess the maintenance of taxon-specific functional traits, we carried out a taxon-function-coupled analysis using the iMetaLab platform (http://shiny.imetalab.ca/) 39 . In total we identified 1,066 unique COGs of proteins corresponding to 419 taxa, generating a three-dimensional dataset (sample-taxonfunction) for between-sample comparisons. Pearson's correlation coefficient r of the taxon-specific functional profiles was calculated between the inoculum and the cultured microbiome for each time point

Evaluation of in vitro-in vivo correlation of microbial response to metformin treatment
We then evaluated the in vitro-in vivo correlation (IVIVC) of microbiome drug response in the MiPro model using high-fat diet (HFD) -fed C57/BL6 mice. Metformin is a widely prescribed drug for treating type 2 diabetes and it has been reported that in human 30% of the oral dose can be recovered in ers he for in feces 40 . Several studies have focused on the effect of metformin on gut microbiota composition and functions 5,6,41 . For these reasons, we employed metformin to validate our MiPro model against in vivo studies by investigating the impact of metformin exposure on gut microbial communities in mice and in our model. Briefly, the MiPro model was inoculated with the stool microbiome from each mouse and cultured for 24 hr in presence or absence of metformin. Mice were then treated daily for 28 days with 300 mg/kg of metformin through gavage, and stool samples were collected at days 0, 14 and 28. All across all comparisons, and a 75% agreement in these changes was observed at the species level.
Metformin-treatment increased the abundance of A. municiphila, which was in agreement with several previous studies [42][43][44][45] . analysis (PLS-DA) was performed on shared proteins, and proteins with a VIP score > 1 were regarded as differential protein groups, which were mainly involved in 12 KEGG pathways ( Figure 4A). A total of 11,222 (out of 17,646) proteins that correspond to these KEGG pathways were extracted from the original 1 protein group file. All LFQ intensities of protein groups that were assigned to each of the 12 pathways were summed in each sample for a KEGG pathway level evaluation. Figure 4A shows that the relative abundances of selected pathways were uniformly altered in both in vitro and in vivo metformin treated samples in comparison to untreated microbiomes. Subsequently, IVIVC were visualized by comparing the changed value of these KEGG pathways ( Figure 4B). In most cases, changed proteins appeared in Quadrants I and III suggesting agreement of in vitro-in vivo responses. Among these, glycolysis/glucogenesis, ABC transporters and two-component system showed high levels of variation.
Here we took a deeper look into the shift of functional balance by normalizing the intensities of each enzyme against the summed protein intensity of the glycolysis/gluconeogenesis pathway ( Figure 4C). In In order to build an in vitro model to assess microbiome function or to address microbiome's response to xenobiotics, it is critical to preserve both the compositional and functional profiles of an individual's microbiome. Although some studies have shown the culturability of the gut bacterial community in vitro 10,13,14,47 , these studies have not demonstrated the maintenance of the microbiota's functional activities. The majority of gut microbiome culture media (including BCM) 13,14,27-33 contain a mixture of commercialized bile salts constituted of the sodium salt forms of CA and DCA at a 1:1 ratio (w/w). DCA (secondary bile acid) is known to act as an antimicrobial agent due to its high-hydrophobic and detergent properties on bacterial membranes 49 . Studies have noted a decrease in Firmicutes abundance in response to DCA 50 , whereas an increased level of primary bile acids drove Firmicutes' enrichment 51 . Hence, we replaced DCA with CDCA, a major primary bile acid produced in human 52 .
This replacement was effective in maintaining the in vitro microbiome composition (Figure 2A). In terms of culturing condition, a silicon-gel cover was perforated at the top of each well in order to create tiny vent holes permitting the escape of the gas produced by the gut microbiota while minimizing inward gas diffusion from the anaerobic chamber. We found that culture tubes with loosen caps resulted in marked microbiome composition changes, the 96-well format maintained the microbiome composition ( Figure   2B). The gut microbiome produces gases such as H 2 , CH 4 , H 2 S and NO x , etc. 53 and volatile organic compounds, such as SCFAs 54 . As opposed to culturing in a tube, the silicon-gel cover increased the partial pressure of gases and volatile metabolites in each well, which subsequently preserved certain levels of dissolved gas molecules in the culture medium. Some dissolved metabolites are important factors for bacteria cross-feeding 54 and the control of pathogens 55 , and thus, are presumably essential for the maintenance of an in vitro microbiome. Besides the application in our high-throughput model, these optimizations may also be adoptable in previously-reported fluidic-based models [17][18][19]56 , in which the optimized medium would help to maintain microbial functional activities, and a similar action that preserves a proper gas metabolite pressure should be considered.
As an outcome of these optimizations, our MiPro model quintupled viable bacteria count, and had similar alpha-diversity and microbial composition at 24 hr post-inoculation to the inoculum at time 0.
These findings are in agreement with the premise that high biodiversity may enhance the temporal stability of microbial communities 57 . In terms of functions, a general view of functional stability using summed protein abundances in each functional category is not sufficient due to the complexity of functional constitution in a gut microbiome ecosystem. Study has shown that at the microbiome level, functional pathways are stable within a health human population 58 . This is stability owes to the functional redundancy among species in a gut microbiome 25 . Functional compensation can happen among different taxa, which preserves long-term average ecosystem performance, as one species can increase a function for loss or decline in another 59 . This mechanism shifts the functional balance between two taxa in a gut ecosystem. Therefore, for maintenance of an in vitro gut microbiome culturing system, it is very important to preserve the stability of taxon-specific functional profiles in the microbiome. And results showed that this has been well achieved in our model. Together, our MiPro model comprises two phases, a growth phase with significant viable biomass accumulation and a lag phase where > 83% of the taxonspecific functional activities of the inoculum are maintained. Hence, this model can be effectively used to amplify a microbial community and to investigate the effect of perturbations on the microbial community composition and functional activities.
The IVIVC study estimated the power of the MiPro to recapticulate in vivo drug responses on different levels of taxonomic biomass contributions and functional activities. Changes of major taxonomic responders on phylum, genus and species levels as reported in several studies 5,[41][42][43][44][45] , were captured in both our in vitro and in vivo treatments. Additionally, our functional profiling analysis showed agreements of in vitro -in vivo responses, and as well as with previous studies 5 . significant responses were found in pathways of glycolysis/gluconegenesis, pyruvate metabolism, fatty acid biosynthesis, fatty acid degradation, nitrogen metabolism, ABC transporters, and the two-component system. A recent study 5 indicated that metformin alters multiple microbial pathways including ABC transporters, twocomponent system, fructose and mannose metabolism and pyruvate metabolism. All of these responses were found in our model. Two microbial ABC transporters and the two-component systems were significantly decreased, while the glycolysis/ gluconeogenesis pathway was similarly increased following drug treatments in both the MiPro and in vivo models ( Figure 5C). Shin et al. has reported the role of metformin in improving glucose homeostasis in HFD-fed mice 42 . Changes in glycolysis/gluconeogenesis pathway inside the gut microbial environment could play a key role in glucose absorption across the intestinal mucosal layer 60 . Notably, metaproteomic responses within the glycolysis/gluconeogenesis pathway revealed high IVIVC of the Mipro on the enzyme level ( Figure 4C). This highlighted the depth of our model to recapitulate in vivo microbiome functional activities in response to drug treatment.

Conclusions
In this work, we evaluated the performance of the MiPro model to maintain microbial taxon-function stability and the utility of this model for drug-microbiome interactions studies. We optimized the medium and culture model for high-throughput drug-microbiome co-culturing. The optimized model showed improved performance in sustaining viability, diversity, compositional and functional profiles of the inoculum microbiome. A high degree of in vitro -in vivo correlation of compositional and functional responses were observed with metformin treatment. Our work provides an effective experimental platform for drug-microbiome interaction studies.

Medium preparation and gut microbiome culturing
The medium composition for MiPro was based on our previously suggested medium composition 26 , which comprises: 2.0 g L -1 peptone water, 2.0 g L -1 yeast extract, 0.5 g L -1 L-cysteine hydrochloride,

Microbiome growth and viability tests
At 0, 3, 6, 9, 12, 24, 34 and 48 hr, two 100 μ l aliquots were removed from each sample for OD 595nm measurements. One of the aliquots was centrifuged at 16,000 × g and the supernatant was used as the medium blank for the OD measurement. Viability of the microbiomes were tested at 0, 9,24

MiPro model to in vivo comparison and validation
In vitro and in vivo effects of metformin on the murine gut microbiome associated with a high-fat diet (HFD) were tested. Briefly, 7-week old male litter-mates from inbred C57/BL6 mice were singlehoused and fed a 42% fat calories diet (ENVIGO, TD.09682) for 6 weeks to allow stabilization of their microbiomes on diet. Fresh stool pellets from each mouse were collected for in vitro culturing on day-0.
The microbiomes were cultured in the absence or presence of metformin (cat# PHR1084, Sigma-Aldrich) using the MiPro model. The in vitro concentration of metformin was set to 6 mg/ml, emulating the 30% fecal recovery ratio of metformin previously reported from in vivo experiments 40

Trypsin digestion, desalting and LC-MS/MS analysis
Protein extraction, trypsin digestion, and desalting steps were carried out as previously described 38 , with a minor modification of the cell washing step. Briefly, for general analysis, the samples were washed two times with PBS at 14,000 × g, 4°C for 20 min. Then large debris were removed with 300 × g centrifugation at 4°C for 5 min, followed by pelleting cells at 14,000 × g, 4°C for 20 min. The microbial Protein concentrations were determined by DC (detergent compatible) protein assay before an overnight trypsin digestion at 37 o C following protein reduction and alkylation, as previously described [25]. to 1800 m/z, followed by data-dependent MS/MS scan of the 12 most intense ions, a dynamic exclusion repeat count of two, and repeat exclusion duration of 30 s. All samples were run on LC-MS/MS in a random order. In addition, for the pre-experiment that evaluated bile salts composition and culture conditions, samples were analyzed on an Orbitrap XL following a 6 hr gradient using previously described parameters 26 . All raw data from LC-MS/MS have been deposited to the ProteomeXchange Consortium (http://www.proteomexchange.org) via the PRIDE partner repository.

Metaproteomics data processing
Protein/peptide identification and quantification were carried out using the MetaLab software (version 1.0) 36 , which automates the MetaPro-IQ approach 38 . An iterative database search strategy using gut microbial gene catalogs (for cultured human microbiomes, human gut microbial gene catalog with 9,878,647 sequences from http://meta.genomics.cn/ and for mice microbiomes, mouse gut microbial gene catalog database comprising 2,572,074 genes, obtained from http://gigadb.org) and a spectral clustering strategy were used for database construction, the peptide and protein lists were generated by applying strict filtering based on a FDR of 0.01, and quantitative information of proteins were obtained with the maxLFQ algorithm; quantitative taxonomic analyses were achieved by assigning identified peptides with For human gut microbiomes, protein groups were filtered with the criteria that the protein should be present in ≥ 25% of the samples (Q25). Bray-Curtis dissimilarity and analysis of similarities (ANOSIM) were performed using the R package "vegan". Principal coordinates analysis (PCoA), principle component analysis (PCA), and hierarchical clustering were visualized with R (version 3.4.3). Taxonomic composition visualization and taxon-function coupled analysis were performed and visualized using iMetaLab (http://imetalab.ca/). The database of clusters of orthologous groups (COG) of proteins was used for functional annotation. For each sample, taxon-specific functional proteins with protein intensity was then generated from iMetaLab. With these set of tables, the Pearson's correlation coefficient r of the taxon-function coupled profile between any two samples was calculated using R to generate a correlation matrix, and the correlation between 0 hr and the cultured microbiome at each subsequent time point was obtained. For visualizing the taxon-function coupled enrichment, p < 0.05 was set as the threshold for both taxonomic and functional enrichment, and the top 30 connections were selected from enriched taxon-function matches.
For the in vitro-in vivo mice microbiome experiments, protein groups were filtered with the criteria that the protein groups should be present in ≥ 50% (Q50) in each of the listed subgroups (including in vitro untreated vs. in vitro 24 hr treated, in vivo untreated vs. in vivo day 14 or day 28 treated samples), partial least squares -discriminant analyses (PLS-DA) test was performed on shared proteins among listed subgroups using the online tool MetaboAnalyst (www.metaboanalyst.ca/). Protein groups with a VIP score > 1 were annotated to the corresponding KEGG categories. Then all protein groups annotated with these KEGG categories were extracted from the original protein group file. In vitro -in vivo correlation of the microbiome drug response was evaluated using taxonomic and pathway change of metformin-treated microbiome. Pearson's correlation coefficient r was calculated using R. Taxonomic composition analysis was done using MetaboAnalyst. Figure S1. Gating of live, dead and unstained bacteria according to stained gut microbiome cells, stained and heattreated microbiome cells, and unstained microbiome. Figure S2. Metaproteomic data quality.  Figure S4. Comparison of Faecalibacterium Praunsnitzii biomass change in the MiProand BCM-cultured microbiomes. Figure S5. Shannon-Weiner index suggesting well-maintained alpha-diversity of the microbiomes cultured from volunteers V1-3 over 48 hrs. Figure S6. PCA scores plot with hierarchical clustering based on COG functional categories of microbiome proteins from volunteers V2 and V3. Figure S7. Taxonfunction-coupled profile in comparison with 0 hr baseline samples.