Unravelling metabolic cross‐feeding in a yeast–bacteria community using 13C‐based proteomics

Abstract Cross‐feeding is fundamental to the diversity and function of microbial communities. However, identification of cross‐fed metabolites is often challenging due to the universality of metabolic and biosynthetic intermediates. Here, we use 13C isotope tracing in peptides to elucidate cross‐fed metabolites in co‐cultures of Saccharomyces cerevisiae and Lactococcus lactis. The community was grown on lactose as the main carbon source with either glucose or galactose fraction of the molecule labelled with 13C. Data analysis allowing for the possible mass‐shifts yielded hundreds of peptides for which we could assign both species identity and labelling degree. The labelling pattern showed that the yeast utilized galactose and, to a lesser extent, lactic acid shared by L. lactis as carbon sources. While the yeast provided essential amino acids to the bacterium as expected, the data also uncovered a complex pattern of amino acid exchange. The identity of the cross‐fed metabolites was further supported by metabolite labelling in the co‐culture supernatant, and by diminished fitness of a galactose‐negative yeast mutant in the community. Together, our results demonstrate the utility of 13C‐based proteomics for uncovering microbial interactions.


Introduction
Prevalence of metabolite exchange in natural microbial communities is widely supported by bioinformatic as well as experimental evidence (Henriques et al, 2020;Blasche et al, 2021;Giri et al, 2021;Machado et al, 2021;Estrela et al, 2022;Pontrelli et al, 2022;Yu et al, 2022). The latter, however, is much more limited, both in terms of the coverage of the communities and the depth of the evidence. Most experimental evidence stems from associations between nutritional perturbations and community compositional dynamics, with few studies pinpointing the exchanged metabolites and the corresponding donors and receivers. A major challenge is the universality of central metabolites, for example, amino acids, nucleotides and organic acids, across the tree of life. Unlike DNA and RNA, these metabolites cannot be assigned to a species once outside the cell. While genomic auxotrophies can suggest exchanged metabolites in natural communities (Zelezniak et al, 2015;Giri et al, 2021) and provide a basis for engineering synthetic communities (Chang et al, 2021;Guillen et al, 2021;Konstantinidis et al, 2021), even then the identification of the exchanged metabolites can be difficult due to the complexity of the metabolic networks. Cellular regulation can switch off certain biosynthetic pathways in an environment-dependent manner, and many different metabolites can compensate for a given auxotrophy. While some of the obstacles can be alleviated by (meta-) transcriptomics/proteomics measurements, their utility in predicting metabolite secretion is limited even in single species studies. Direct metabolic measurements are surprisingly uncommon, even for synthetic communities, and thus, the list of experimentally validated inter-species metabolic interactions is small. Thus, mechanistic understanding of community dynamics in response to, for example, environmental change, as well as rational modulation of microbial communities, is currently limited.
Metabolic flux analysis using stable isotope tracing has been successfully used for resolving the fate of nutrients in single species (Wiechert, 2001;Schwechheimer et al, 2018;Xiao et al, 2022). However, the analysis methods, which are based on proteogenic amino acids or direct metabolite measurements, cannot be directly applied to communities as most metabolites cannot be ascribed to the producer/consumer member species. An alternative approach is to separate DNA extracted from the community by mass, which is indicative of the degree of mass-isotopomer incorporation; subsequent sequencing can then identify the community members that incorporate a labelled substrate (Radajewski et al, 2000;Bernard et al, 2007). Similarly, the protein stable isotope probing (Protein-SIP) technique (Jehmlich et al, 2010(Jehmlich et al, , 2016von Bergen et al, 2013;Sachsenberg et al, 2015) has been used to unambiguously identify the peptides and the metabolically active phylogenetic groups in the community (Jehmlich et al, 2008). This requires tailored mass spectrometry data analysis methods that simultaneously deconvolute potential mass shifts and species identities (Sachsenberg et al, 2015, Jehmlich et al, 2016. Both the DNA and protein-based methods, as well as similar approach to lipid labelling have been key to tracing mass-flow and identifying key players in hydrocarbon-degrading communities Herbst et al, 2013;Vogt et al, 2016). Yet, while these methods can identify direct utilizers of the labelled substrates and potential interactors, intra-cellular flux distributions and cross-fed metabolites cannot be easily resolved. Towards resolving intra-cellular fluxes, the 13 C labelling combined with meta-proteomics can in principle be used to infer labelling at the level of amino acids within community members if enough peptides can be uniquely assigned (Ghosh et al, 2014;Gebreselassie & Antoniewicz, 2015;Sachsenberg et al, 2015). However, this concept has, to our knowledge, remained experimentally untested beyond engineered conspecific strains (Gebreselassie & Antoniewicz, 2015).
Here we used a yeast-lactic acid bacterial community to evaluate the peptide labelling strategy for elucidating inter-species metabolic exchange. While of much lower complexity than naturally occurring microbial communities, this two-member community is an excellent model to develop approaches to tackling this highly complex experimental problem. Yeast Saccharomyces cerevisiae and lactic acid bacteria such as Lactococcus lactis co-exist in many natural and fermented food environments such as sourdough, kefir and wine (Ferreira & Mendes-Faia, 2020;Blasche et al, 2021;Pino et al, 2022). These species are thus important for ecology as well as widely used as model organisms. We have previously shown the critical role of metabolic cross-feeding in this symbiosis and identified nitrogenous compound overflow (esp. amino-acid) as a key mechanism of lactic acid bacteria (LAB) support by yeast (Ponomarova et al, 2017). A notable feature of this community is that it can be readily assembled in a defined medium and the one-way (commensal) relation can be turned into mutualism by changing a single nutrient, namely glucose to lactose. In the mutualistic scenario, yeast, being unable to breakdown lactose, depends on LAB for carbon, while the LAB relies on the yeast for nitrogen. While the latter is known to be mediated by amino acids, a key open question is which metabolites are shared by the lactic acid bacteria with the yeast (Fig 1A). We here answer this question by using 13 C label-tracing in peptides and extracellular medium, providing first proof of concept for using proteomics for inferring inter-species metabolic interactions.

Results
Assembly of a mutualistic community of S. cerevisiae and L. lactis We assessed the capacity of several lactic acid bacteria to form a mutualistic community with S. cerevisiae on a defined medium containing lactose and 8 different amino acids. Two L. lactis strains isolated from the fermented milk drink Kefir (Blasche et al, 2021), but not the L. lactis-type strain IL1403, stained positive for betagalactosidase activity and formed characteristic 'crescent moon' shapes when pinned next to S. cerevisiae (Fig 1B). Strain SB261 was chosen for all further experiments and showed robust growth on lactose, glucose, a glucose-galactose mix, but not galactose alone ( Fig 1C). The media used in all following experiments was chosen so that the two species grow together but neither L. lactis nor S. cerevisiae could grow alone due to their respective requirements for amino acids and carbon from the other species (Fig 1D).
13 C incorporation analysis in co-cultures by proteomics Some LAB strains harbour galactose-lactose antiporters and thus export galactose in the presence of lactose (Hutkins & Ponne, 1991;Benthin et al, 1994). To probe whether galactose is the carbon source shared by L. lactis with yeast, we performed co-culture proteomics analysis using 13 C-labelled lactose (Fig EV1). Yeast and L. lactis were grown together with either non-labelled lactose, lactose labelled in the carbons corresponding to the glucose backbone lactose-[ 13 C-glucose], or lactose labelled in the carbons corresponding to the galactose backbone lactose-[ 13 C-galactose].
The mass spectrometry data from the co-cultures were analysed using MetaProSIP (Sachsenberg et al, 2015) tool within the OpenMS framework (Rost et al, 2016;Fig 2A). The assignment of labelled and unlabelled peptides to bacteria or yeast was based on the reference peptides from non-labelled lactose sample corresponding to 2,693 and 6,721 proteins respectively (Dataset EV1). A target decoy database including peptides from other organisms was used to detect any incorrect assignments of isotope-labelled peptides. Among the 49,365 peptides identified in the samples, only 279 were deemed false positives, indicating the incorporation false positive rate (iFPR; Griss et al, 2014;Sachsenberg et al, 2015) to be circa 0.005 (Dataset EV1). We estimated the degree of incorporated 13 C atoms within each peptide in terms of Relative Isotope Abundance (RIA, Materials and Methods, Dataset EV2). Both yeast and bacterial proteomes showed good reproducibility in RIA values across peptides ( Fig 2B) indicating overall uniformity in isotope incorporation.
Peptide-labelling patterns reveal galactose as a metabolite shared by L. lactis When the lactose was labelled in the galactose subunit, the labelling degree in the yeast peptides was considerably higher (~30%, we note that the growth medium contains unlabelled amino acids) than when the glucose subunit was labelled (~5%; Fig 2B). This difference indicates that the bacteria, upon lactose breakdown, preferentially use glucose, while galactose is secreted and taken up by yeast. In bacterial proteome, the pattern was expectedly reversed with degree of labelling being higher when lactose-[ 13 C-glucose] was fed (~19%) than in the case of lactose-[ 13 C-galactose] (~15%; Fig 2A). This suggests that lactose is metabolized intracellularly by L. lactis and galactose is released, as with other dairy lactococci. (Leenhouts et al, 1995). The relatively small difference in the labelling degree between two substrates, as well as the overall smaller labelling than yeast, is consistent with the bacterial reliance on amino acids in the growth medium and those secreted by the yeast. The yeast cells, on the other hand, channelled more carbon from the galactose to protein synthesis. Our data do not exclude that a small fraction of the galactose is metabolized by L. lactis, although its poor growth on this substrate ( Fig 1C) and preference for glucose utilization makes this unlikely to be at an appreciable level.
To trace the flow of labelled carbon from lactose to the yeast and bacterial peptides, we estimated the contribution of individual amino acids to peptide RIA values using multiple linear regression (Materials and Methods). This allowed us to estimate the average, relative, contribution of individual amino acids to the mass shift/labelling state of peptides, specific to each species and labelling regime (galactose or glucose moiety labelled). (Dataset EV2). Principal component analysis (PCA) showed clear distinction of the yeast-lactose-[ 13 C-galactose] in accord with the highest labelling incorporation seen from overall RIA distribution (Fig 2C). At the level of individual amino acids, the contribution of amino acids provided in the growth medium to the labelling was minimal as expected, while those secreted by yeast (glycine, alanine, glutamate, glutamine, serine, threonine and aspartate; Ponomarova et al, 2017) had the highest contribution (Fig 2C and D). Three amino acids with a close link to the central carbon metabolism, viz., alanine, glycine and cysteine, had stronger labelling contribution under lactose-[ 13 C-galactose] growth, attesting the entrance of 13 C in yeast via glycolysis. The dependency of L. lactis on the yeast-supplied amino acids was also well reflected in the bacterial peptide labelling ( Fig 2D). The overall distribution of individual amino acid contributions in bacterial peptides was thus similar to that of yeast (Fig 2C and D). The activity of the galactose utilizing Leloir pathway in yeast was evident in the expression of Gal1, Gal2, Gal3, Gal7, Pgm1/2 and Gal10 in the unlabelled proteomics data set (Dataset EV2). Together, the labelling patterns reveal galactose produced by L. lactis as  A Schematic hypothesis of metabolite exchanges in the microbial community. B Screen of LAB strains and E. coli for beta-galactosidase activity (top) and growth in close proximity to S. cerevisiae colonies on CDM35-lactose plates (bottom). Strains SB261 and SB17 are L. lactis strains isolated from Kefir in a previous study (Blasche et al, 2021); IL1403 and WCSF1 are L. lactis and Lactiplantibacillus plantarum type strains respectively. The "crescent moon" shape observed in spots with SB261, SB17 and WCSF1 indicates that both yeast and LAB benefit from proximity to each other, that is, indicate the potential to form a cooperative co-culture. Strain SB261 was selected to be used in all following experiments. C Bacterial growth (OD 600 ) of L. lactis strain SB261 in different carbon sources after 26 h. Bar heights indicate the mean. Values for each carbon source were compared to growth in galactose and the statistical significance is indicated (n = 3 biological replicates, two-sided Student's t-test, P-values: glucose: 0.00059; glucose + galactose: 0,0.0027; lactose 0.018). D Dynamics of L. lactis SB261 growth in co-culture with S. cerevisiae WT S90 strain (blue line) and both species in monoculture (green and orange lines) in CDM35lactose media. The lines show the mean of three biological replicates.
the main carbon source of yeast and the reciprocity of yeast through amino acids that enter bacterial biomass.
Exo-metabolome reveals a multi-factorial cross-feeding The labelling enrichment in yeast peptides under lactose-[ 13 Cglucose] compared to unlabelled control indicates that yeast receives bacterial metabolites other than galactose. To identify these metabolites, and to affirm the peptide-based inference of cross-fed metabolites, we used GC-MS to analyse 13 C labelling in the metabolites in the co-culture supernatants (Materials and Methods, Dataset EV3). Circa 70% of pyruvate and lactate were found to be labelled in all carbon atoms during growth on lactose-[ 13 C-glucose], but only 40% in the case of lactose-[ 13 C-galactose] supplementation    Figure 2. Analysis of labelling patterns in species-specific peptides by proteomics identifies cross-fed amino acids.
A Proteome data were acquired for yeast-LAB communities grown on lactose with either the glucose or galactose part labelled. The MetaProSIP workflow was used to identify peptides, their species identity, and relative isotope abundance (RIA, fraction of carbons labelled). The contribution of individual amino acids to peptide labelling was estimated using multiple linear regression. The resulting coefficient vectors for the four species-substrate pairs were then compared using Principle Component Analysis (PCA). B Box plots of RIA distributions in the co-culture assay. All peptides identified for each species were grouped based on their biological replicates and discriminated by the labelling conditions. Numbers above the boxes indicate the number of peptides in this group. As is standard, this boxplot shows the median of the data as central line, the quartiles as box and the extend of the rest of the distribution as whiskers. Points which are 1.5 times the inter-quartile range beyond the high and low quartiles are considered outliers and shown individually. C PCA discriminates the yeast-lactose-[ 13 C-galactose] from the other three groups. D Predicted estimated coefficients relationship between the lactose-[ 13 C-galactose] and lactose-[ 13 C-glucose] in yeast (left) and bacteria (right) based on 3 different AA groups. Higher values indicate more 13 C labelling. (Fig 3). This is consistent with the bacteria being the main utilizers of glucose and lactic acid being a main fermentation product of L. lactis metabolism. Pyruvate, which is only a single enzymatic step away from lactic acid, had a singular labelling pattern as lactic acid. Furthermore, co-culture proteomics data showed presence of lactate dehydrogenase (Dld1p, detected in 3/3 unlabelled samples) consistent with bacterial secretion of both lactate and pyruvate into the medium. Amino acids that were added to the medium showed, as expected, little 13 C labelling ( Fig EV2). Among the other amino acids, glycine and aspartate had approximately 15-20% fractions labelled but did not show appreciable difference whether the lactose was labelled in galactose or glucose. Both amino acids thus appear to have been synthetized and secreted by yeast as well as L. lactis. Alanine had 20% higher labelling when the galactose part of lactose was labelled, attesting the proteomic results wherein the contribution of alanine to peptide labelling was high in lactose-[ 13 Cgalactose] samples for both organisms.
Circa 60% of glutamate in the supernatant was labelled in at least one carbon atom in the case of lactose-[ 13 C-galactose] and 40% in the case of lactose-[ 13 C-glucose] indicating that carbon flow from both species and that from other, unlabelled, media components contributed to glutamate synthesis. Yet, the labelling pattern in terms of the number of labelled carbons was different between lactose-[ 13 Cgalactose] and lactose-[ 13 C-glucose]. When lactose-[ 13 C-galactose] was fed, glutamate molecules had 4 or 5 carbons labelled; while when fed with lactose-[ 13 C-glucose], only 1, 2 or 3 carbon atoms were labelled. Succinate followed the same patterns as glutamate since it is a downstream metabolite. Glutamate thus clearly originated from different initial carbon sources or from different biosynthetic pathways between the two cases, that is, lactose-[ 13 C-galactose] versus lactose-[ 13 C-glucose]. L. lactis lacks a complete TCA cycle (Morishita & Yajima, 1995) and thus the difference can be, in part, ascribed to the glutamate originating from yeast (4-5 carbons labelled) or L. lactis (1-3 carbons labelled). Overall, the metabolite labelling in the co-culture supernatant corroborates the cross-feeding interactions revealed by proteomics. Together, these show that bacteria share, beyond galactose, pyruvate, lactic acid, glycine, aspartate and glutamate into the medium and contribute to carbon into yeast proteome, while yeast cells primarily share amino acids.  Metabolic modelling, spent medium assay and gene knockout mutants complete the cross-feeding scenario We next used genome-scale metabolic models to link the observed metabolic cross-feeding to genomic features, esp. auxotrophies, of the yeast and L. lactis strains (Materials and Methods). At the level of individual species, flux balance simulations confirmed the amino acid auxotrophies of L. lactis and the inability of yeast to utilize lactose. Community simulations confirmed the inter-dependencies of two species, particularly that of yeast on galactose and that of bacteria on amino acids secreted by yeast. The modelling also suggested other possible interactions mediated by aromatic amino acids or lactate (Materials and Methods, Dataset EV4).
To independently confirm the carbon flow between L. lactis and yeast, we cultivated yeast in bacterial spent medium (Fig 4A and B) and followed the metabolites in the supernatant using GC-MS (Dataset EV3). Galactose and lactate were the main components in the condition medium of which galactose had been consumed after 20 h while only a small reduction was seen in lactate after the galactose was exhausted (Fig 4C and D). Other metabolites in the bacterial secretome, such as fumarate, succinate and pyruvate, were not consumed by the yeast, while pyruvate was secreted (Fig EV3). The spent medium assay thus further confirmed bacteria-supplied galactose as the primary carbon source of yeast, yet also suggested that other metabolites like lactate and pyruvate could potentially support yeast growth. To genetically test this hypothesis, we replaced the wild-type yeast by gal1Δgal3Δ mutant, which is not able to utilize galactose (Fig 5B). The co-culture of the mutant yeast and L. lactis had a reduced fitness, yet the growth was not abolished, supporting our hypothesis (Fig 5C). Together, the peptide and supernatant A Conditioned medium assay design. B Bar graph representing yeast growth in different percentages of bacterial spent medium. Bar heights represent the mean of n = 3 biological replicates. The statistical significance of the difference in growth between 0 and 80% bacterial conditioned is indicated for days 2 and 7 (two-sided Student's t-test, P-values: day 2: 1.27 × 10 −6 ; day 7: 0.0025). C Targeted metabolomics using GC-MS applied to the bacterial conditioned medium, metabolites produced by LAB (relative to the mean concentration in bacteriaconditioned media). Data represent three biological replicates. D Exo-metabolome dynamics revealed by GC-MS targeted metabolomics, metabolites that potentially cross-fed from bacteria to yeast, produced by LAB and consumed by yeast (see Dataset EV3 for the complete list of metabolites measured). Data represent three biological replicates, lines represent the mean.
metabolite labelling patterns, modelling, spent medium assays, and gal1Δgal3Δ mutant community affirm galactose as a preferred but not the only carbon flow from bacteria to yeast (Fig 5A). A Qualitative scheme of the metabolic interchange in the co-culture growing in CDM35-Lactose. Dashed lines indicate processes not conclusively supported by experimental evidence. Arrows with an overlaid circle represent processes supported by modelling. B The yeast gal3Δgal1Δ deleted mutant is not able to grow in CDM35-galactose. Data represent three biological replicates, lines represent the mean. C Co-culture growth is compromised but not completely impeded if the LAB co-cultured with yeast is unable to metabolize galactose. Co-cultures and mono-cultures were grown in CDM35-lactose medium. Data represent three biological replicates, lines represent the mean.

Discussion
Ó 2023 The Authors Molecular Systems Biology 19: e11501 | 2023 some contribution towards understanding the set of expressed metabolic genes, but ultimately fall short as enzyme levels are not easily relatable to metabolic fluxes, and even less so for the import/export of metabolites from the environment (Hackett et al, 2016;Sauro, 2017;Zelezniak et al, 2018). Direct experimental evidence of metabolite sharing is therefore required to establish the molecular interaction network, but so far is only sparsely available. One approach has extended classical metabolic flux analysis (MFA) by linking two metabolic models with exchange fluxes and fitting intra-and inter-cellular flux distributions using mass isotopomer distributions (MIDs) obtained from peptide labelling data. But so far, these studies were either within a species (Gebreselassie & Antoniewicz, 2015) or theoretical (Ghosh et al, 2014). Other approaches to disentangling communal metabolism have used the incorporation of stable isotopes into proteins, albeit with the a limited aim of identifying species (groups) consuming/degrading a specific isotope-labelled substrate (Jehmlich et al, 2016) rather than metabolite exchange between cells.
We here built on the protein-SIP concept to capture cross-feeding interactions, providing a first experimental proof-of-concept for inferring interspecies metabolic exchange using isotope tracing in coculture proteomics. The peptides provided the species identity while the isotope enrichment indicated the flux channelling; the method thus circumvented the limitations of classic metabolite-based flux analysis wherein the metabolites cannot be ascribed to the species of origin. Labelling analysis of metabolites in the co-culture supernatant agreed well with the peptide-based inference and helped to identify additional cross-fed metabolites. Metabolic modelling accounting for the total biosynthetic capabilities (or lack thereof) were also consistent with the findings. Importantly, few of the modelling predictions (e.g. tryptophan and phenylalanine) were not supported by the data. This is, however, expected, as while the models outline possible exchanges, enzyme expression and other operational or regulatory constraints are not accounted for, and are difficult to do so in a generalized fashion. Metabolite sharing is not a consequence of nutrient release by dead cells as label incorporation into protein depends on live, metabolically active cells. Furthermore, the community is stable over long timescales and can be evolved to enhance metabolite sharing (Konstantinidis et al, 2021), and the fraction of dead cells observed in yeast-LAB co-cultures is minimal (Ponomarova et al, 2017).
Yeast and lactic acid bacteria are quite distinct in their protein sequence space; in communities involving phylogenetically closely related species and/or those with larger membership, disentangling species identity and label incorporation will become difficult. Application of the peptide-based flux analysis to more complex communities would, therefore, require both theoretical and experimental development. Furthermore, the approach requires differential preference of carbon substrates. Yet, as our results exemplify, proteomics, metabolomics and metabolic modelling could be jointly used to disentangle complex metabolic interactions.

Strains, media and growth conditions
All cultures were grown statically at 30°C. Chemically defined medium, CDM42-lactose for bacteria-conditioned medium and CDM35-lactose for bacteria and yeast co-culture growth assays, were prepared based on the medium designed in our previous work (Fig EV4; Ponomarova et al, 2017). The lactose concentration in the medium was 4%. Labelled lactose was purchased from Omicron including [UL-13 C 6 glc ]-lactose and [UL-13 C 6 gal ]-lactose monohydrate (LAC-006, LAC-007). All solutions were sterilized by filtration. Liquid medium experiments were started at 0.1 OD 600 for bacteria and 0.05 OD 600 for yeast cultures, using pre-cultures washed twice with phosphate-buffered saline (PBS). We used CDM35-galactose liquid media (2% and 0.05%) and SD-galactose 2% agar plates to test the phenotype of the gal1Δgal3Δ double mutant. For assays on solid media, 5 μl of the washed LAB cultures (0.1 OD 600 ) and yeast culture (0.2 OD 600 ) were spotted on CDM35-lactose plates (2% agar). LAB cultures were placed in proximity to the yeast inoculum and further away on the same plate as a control. CDM35-lactose agar plates were incubated for 9 days. Cell growth was monitored by measuring the optical density at 600 nm (OD 600 ) using a spectrophotometer (Biochrom Ultrospec 2100 pro, or BioTek Synergy for growth assays in 96-well plates). Growth experiments were conducted in biological triplicates.

Yeast strain construction
The double knockout mutant gal1Δgal3Δ was constructed for this study in the prototrophic haploid laboratory S90 genetic background. Genomic DNA of a gal1Δ::kanMX4 mutant in BY4741 background (Mulleder et al, 2012) was extracted and a fragment containing the kanMX4 cassette with 25 base pairs of flanking genomic regions was PCR amplified using previously described A and D primer pairs (Winzeler et al, 1999; here called PNG45_A and PNG47_D). Two deoxyoligonucleotides were designed based on the gal3 nucleotide sequence and the multiple cloning site present in the pCfB2194 vector (Stovicek et al, 2015). To generate gal3Δ::hphMX4 cassette the deoxyoligonucleotide PNG53 comprised 25 bp of the 5 0 region of the GAL3 5UTR sequence (+603 to +628), and 22 bp of the pCfB2194 multiple cloning site. The deoxyoligonucleotide PNG57 contained 25 bp corresponding to the 3UTR GAL3 gene (+2,836 to +2,861) and 20 bp from the pCfB2194 multiple cloning site. Qiagen-purified pCfB2194 DNA was used as template for PCR amplification. These PCR products were used to transform wild-type S. cerevisiae S90 as described previously (Gietz & Schiestl, 2007) with some modifications. A mid-log culture (OD 600 = 0.7, 50 ml) was centrifuged, washed, and re-suspended in 1 ml of sterile water. Then 100 μl of cell suspension were mixed with transformation mix (240 μl PEG 3500, 50% w/v; 36 μl lithium acetate, 1.0 M; 50 μl boiled single-stranded carrier DNA, 2 mg/ml; 34 μl of PCR amplification product described above). After incubation for 40 min (42°C water bath), cells were re-suspended in YPAD and incubated for 3-4 h to allow for the expression of the integrated antibiotic marker. Clones were selected on YPAD plates with G418 (300 μg/ml) and subsequently on hygromycin B (Formedium Ref No: HYG1000) plates (200 μg/ml). Success of homologous recombination was verified by colony PCR using PNG46_C-PNG47_D, KanB-PNG45_A, PNG48_A-KanB and PNG49_C-PNG50_D primer pairs.
Conditioned medium assay L. lactis SB261 was cultured in CDM42-lactose medium for 4 days (approx. 0.5 OD 600 ) at 30°C without shaking to create bacteria-conditioned medium. Yeast was grown for 7 days in 80% (supplemented with 20% of fresh medium) or 100% of bacteriaconditioned medium at 30°C without shaking. At indicated time points, 1 ml of cell suspension was sampled and centrifuged at 4,000 g at 4°C for 1 min. Supernatants were passed through 0.2 μm syringe filter PES (Millipore). Metabolites of the bacteriaconditioned medium and the subsequent medium after yeast growth were measured by GC-MS.
Targeted metabolomics analysis of extracellular spent growth media with GC-MS Supernatants of bacterial monocultures in CDM42-lactose, supernatants of yeast in CDM42-lactose bacteria-conditioned medium and yeast-L. lactis co-cultures in CDM35 with labelled lactose were filtered with 0.2-μm PVDF syringe filters and 3-kDa MWCO filters (Millipore). Ribitol (Adonitol; Alfa Aesar, UK) was added as an internal standard at a final concentration of 50 μg/ml. Polar metabolites were extracted from 100 μl of the sample by addition of 200 μl of methanol (Biosolve Chimie, France), followed by incubation at 72°C for 15 min and addition of 200 μl of cold MilliQ water. Following centrifugation (4,000 g, 10 min, 4°C), the supernatants were collected and dried using a GeneVac EZ-2 plus evaporating system (mode: hplc fraction, temperature: 30°C). The dried metabolite extracts were stored at −80°C until metabolomics analyses.
The dried metabolite extracts from the bacteria-conditioned media were derivatized to (MeOx)TMS-derivatives through reaction with 100 μl of 20 mg/ml methoxyamine hydrochloride (Alfa Aesar, A19188) solution in pyridine (Sigma-Aldrich, 437611) for 90 min at 40°C, followed by reaction with 200 μl N-methyl-trimethylsilyltrifluoroacetamide (MSTFA; Alfa Aesar, A13141) for 12 h at room temperature (Kanani & Klapa, 2007;Kanani et al, 2008). GC-MS analysis was performed as previously described (Radic Shechter et al, 2021). Briefly, a Shimadzu TQ8040 GC-(triple quadrupole) MS system (Shimadzu Corp.) equipped with a 30 m A~0.25 mm A~0.25 μm ZB-50 capillary column (7HG-G004-11; Phenomenex) was used. One microlitre of sample was injected in split mode at 250°C using helium as a carrier gas with a flow rate of 1 ml/min. GC oven temperature was held at 100°C for 4 min followed by an increase to 320°C with a rate of 10°C/min, and a final constant temperature period at 320°C for 11 min. The interface and the ion source were held at 280 and 230°C respectively. The detector was operated both in scanning mode, recording in the range of 50-600 m/z, and in MRM mode for specified metabolites. For peak annotation, the GC-MS solution software (Shimadzu Corp.) was utilized. The metabolite identification was based on an in-house database with analytical standards being utilized to define the retention time, the mass spectrum and marker ion fragments for all the quantified metabolites. The metabolite quantification was carried out by integrating the area under the curve of the MRM transition of each metabolite. The data were further normalized to the area under the curve of the MRM transition of ribitol. All peaks were manually checked.
For 13 C stable isotope tracing analysis, the dried extracts from the co-cultures were derivatized with 50 μl of 20 mg/ml methoxyamine hydrochloride (Alfa Aesar, A19188) solution in pyridine (Sigma-Aldrich, 437611) for 90 min at 40°C, followed by reaction with 100 μl N-tert-Butyldimethylsilyl-N-methyltrifluoroacetamide +1% tert-Butyldimethylchlorosilane (Sigma-Aldrich, 00942) for 1 h at 60°C. The samples remained at room temperature for 10 h and then analysed by GC-MS. The GC-MS was operated as described above with the following difference: GC oven temperature was held at 100°C for 3 min followed by an increase to 300°C with a rate of 3.5°C/min and a final constant temperature period at 300°C for 10 min. The detector was operated in single ion monitoring (SIM) mode for the ion fragments corresponding to all possible mass isotopologues of the specified metabolites. Mass isotopologue distributions were determined by integrating the area under the curve of the ion fragments. The data were corrected for natural isotope abundance using the Isotope Correction Toolbox (ICT; Jungreuthmayer et al, 2016).

Protein extraction
Cell pellets were washed first with water and then with ammonium bicarbonate 50 mM. Pellets were resuspended in RapiGest SF (Waters, 186001860) lysis buffer composed of 1 mg RapiGest ST 0.1%, 4 M urea, 100 mM NH 4 HCO 3 , 5 mM DTT and protease inhibitor cocktail. Cells were lysed by 3 rounds of bead beating (45 s beating, 1 min rest at 4.5 m/s and 4°C) with zirconia/silica beads. Cell lysates were centrifuged for 1 min at 800 rpm and 4°C, supernatants were recovered and sonicated for 90 s at 4°C (30 s, 3 times at 80% amplitude in 0.5 s on-off cycles with 1 min cooling intervals on ice). Lysate protein concentration was measured by Bradford assay. Protein samples were processed in-solution and all reagents were prepared in 50 mM HEPES (pH 8.5). Cysteines were reduced using dithiothreitol (56°C, 30 min, 10 mM). The sample was then cooled to 24°C and alkylated with iodoacetamide (room temperature, in the dark, 30 min, 10 mM). A novel protocol using paramagnetic beads, termed Single-Pot Solid-Phase-enhanced Sample Preparation (SP3; Hughes et al, 2014) was used to prepare the samples for LC-MS/MS. The proteins were digested at 37°C overnight using trypsin (Promega) with an enzyme to protein ratio of 1:50. Peptides were cleaned up using OASIS ® HLB μElution Plate (Waters).

Proteomics measurements by LC-MS
Peptides were separated using an ultra-performance liquid chromatography (UPLC) system (nanoAcquity, Waters) fitted with a trapping column (nanoAcquity Symmetry C18, 5 μm, 180 μm × 20 mm) and an analytical column (nanoAcquity BEH C18, 1.7 μm, 75 μm × 200 mm). The outlet of the analytical column was coupled directly to a linear trap quadrupole (LTQ) Orbitrap Velos Pro (Thermo Fisher Scientific) using a Proxeon nanospray source. Solvent A was water with 0.1% formic acid and solvent B was acetonitrile with 0.1% formic acid. The samples were loaded with a constant flow of solvent A at 5 μl/min onto the trapping column. Trapping time was 6 min. Peptides were eluted via the analytical column with a constant flow of 0.3 μl/min. During the elution step, the percentage of solvent B increased in a linear fashion from 3 to 7% in 10 min, then increased to 25% in 100 min and finally to 40% in a further 10 min. The peptides were introduced into the mass spectrometer (Orbitrap Velos, Thermo) via a Pico-Tip Emitter 360 μm OD × 20 μm ID; 10 μm tip (New Objective) and a spray voltage of 2.2 kV was applied. The capillary temperature was set at 300°C. Full scan MS spectra with mass range 300-1,700 m/z were acquired in profile mode in the FT (Fourier transform) with resolution of 30,000. The filling time was set at a maximum of 500 ms with a limitation of 1.0 × 106 ions. The most intense ions (up to 15) from the full scan MS were selected for sequencing in the LTQ. A normalized collision energy of 40% was used and the fragmentation was performed after accumulation of 3.0 × 10 4 ions or after a filling time of 100 ms for each precursor ion (whichever occurred first). MS/MS data were acquired in centroid mode. Only multiply charged (2+, 3+, 4+) precursor ions were selected for MS/MS. The dynamic exclusion list was restricted to 500 entries with a maximum retention period of 30 s and a relative mass window of 10 ppm. Lock mass correction using the ion 445.12003 m/z was applied.

Proteomic data analysis
Rstudio v1.2.5033 and R v3.5.1 were used for data analysis. For protein-SIP analyses, the automated and freely available bioinformatics solution for protein-SIP-based metaproteomic experiments, MetaProSIP (Sachsenberg et al, 2015) was used, as well as the workflow engine Konstanz Information Miner (KNIME, v3.6.2) for the analysis of large data sets and production of high-quality visualizations (Berthold et al, 2009) and the OpenMS framework (Rost et al, 2016). A total of 9 LC-MS/MS raw files were analysed, corresponding to three different labelling conditions: lactose nonlabelled, lactose-[ 13 C-galactose] and lactose-[ 13 C-glucose], in triplicates. For peptide and protein identification with the MetaProSip, raw files were converted to mzML using MSConvertGUI (ProteoWizard Tools, v3.0.18199-78f1f2280;Kessner et al, 2008). Subsequently, mzML files were analysed through a 2-step relative KNIME pipeline, where first unlabelled peptides were identified, followed by identification of their labelled isoforms. Data were searched against the yeast (Uniprot proteome UP000002311) and bacteria proteome (Blasche et al, 2021). The decoy database contained peptides from human, pig, Podospermum laciniatum, Escherichia coli and bovine. The unlabelled peptides were identified by the search engine OpenMS MSGFPlusAdapter. The following options were included into the search parameters: a mass error tolerance of 10 ppm for full scan (MS1); trypsin as protease with a maximum of two missed cleavages; carbamidomethylation of cysteines as a fixed modification, oxidation of methionines as a variable modification; a minimal peptide length of 7. After that, through the PeptideIndexer the protein references for all the peptide hits from the idXML files were refreshed and target/decoy information was added to them. The FDR was set to 0.01 on the peptide (PSM) level based on q-value and the maximal number of ambiguous amino acids (AAAs) for matching to the protein database was 3. Furthermore, we used the feature finder tool to detect quantitative features in MS1 data based on information from peptide identifications (derived from MS2 spectra).
Data were then transferred as featureXML-files to the second part of the workflow where the labelled peptides were detected, the related peaks were extracted and eventually the Relative Isotope Abundance (RIA) was calculated through MetaProSip. Proteins were inferred from the peptide identification results by requiring a minimum of two unique peptides per protein. The stability and the data variability were measured by the coefficient of variation (CV) for peptides identified in at least two replicates. Peptides that were identified in only two out of the three replicates and which also had a high CV (> 40%) were filtered out. The median RIA value across the three replicates was computed for each peptide. MLR analyses were performed to elucidate the impact of the AA carbons on each peptide. The standardized values of the amino acid A in peptide P, stvalues P A , is defined as: where o P A is the number of occurrences of amino acid A in peptide P, c A is the number of carbons in amino acid A and len p is the length of peptide P. We then implemented a linear regression model of the normalized values against the RIA to estimate the relative coefficients.

Construction and models simulations
To develop a mechanistic understanding of the interactions between yeast and bacteria we performed in silico Flux Balance Analysis (FBA). We initially constructed single-species models from genome assemblies and the two species were then combined to model the syntrophic growth and metabolite accumulation of the co-culture in a given media. The single-species and community models were built with CarveMe (Machado et al, 2018) and the Reframed Python package. For the bacterial model, Lactococcus lactis subsp. Cremoris MG1363 from the Bigg database (King et al, 2016) was used as a reference model. Minor modifications were made to the CDM35lactose and CDM42-lactose media composition used in simulations to include vitamins and minerals that were essential in silico. We first explored the relevance of the presence of lactose and bacterial beta-galactosidase activity by comparing the different FBA growth rate predictions. The same procedure was followed for the coculture models. In this case, the pre-built CarveMe models were used along with the modified CDM35-lactose medium. All the applied modifications regarding the exchange reactions, metabolites and CDM35-lactose composition were implemented with Python version 3.6. The Smetana tool (Zelezniak et al, 2015) was used to map all possible inter-species metabolic exchanges, while the simulation results for each community were summarized as SMETANA score, which estimates the strength of metabolic coupling. For the Smetana community simulation, the same medium composition was used as with Reframed, while water was excluded from calculations.

Data availability
The data sets and code produced in this study are available in the following databases: • Code for metabolic models, simulation results and scripts: GitHub (https://github.com/ChMaga/13C-proteomics).
Expanded View for this article is available online.

Acknowledgements
We thank S. Blasche for kefir L. lactis isolates, proteomic core facility (EMBL, Heidelberg) for the protein mass spectrometry, Timo Sachsenberg