Ultra-sensitive isotope probing to quantify activity and substrate assimilation in microbiomes

Stable isotope probing (SIP) approaches are a critical tool in microbiome research to determine associations between species and substrates. The application of these approaches ranges from studying microbial communities important for global biogeochemical cycling to host-microbiota interactions in the intestinal tract. Current SIP approaches, such as DNA-SIP or nanoSIMS, are limited in terms of sensitivity, resolution or throughput. Here we present an ultra-sensitive, high-throughput protein-based stable isotope probing approach (Protein-SIP), which cuts cost for labeled substrates by ∼90% as compared to other SIP and Protein-SIP approaches and thus enables isotope labeling experiments on much larger scales and with higher replication. It allows for the determination of isotope incorporation into microbiome members with species level resolution using standard metaproteomics LC-MS/MS measurements. The analysis has been implemented as an open-source application (https://sourceforge.net/projects/calis-p/). We demonstrate sensitivity, precision and accuracy using bacterial cultures and mock communities with different labeling schemes. Furthermore, we benchmark our approach against two existing Protein-SIP approaches and show that in the low labeling range used our approach is the most sensitive and accurate. Finally, we measure translational activity using 18O heavy water labeling in a 63-species community derived from human fecal samples grown on media simulating two different diets. Activity could be quantified on average for 27 species per sample, with 9 species showing significantly higher activity on a high protein diet, as compared to a high fiber diet. Surprisingly, among the species with increased activity on high protein were several Bacteroides species known as fiber consumers. Apparently, protein supply is a critical consideration when assessing growth of intestinal microbes on fiber, including fiber based prebiotics. In summary, we demonstrate that our Protein-SIP approach allows for the ultra-sensitive (0.01% to 10% label) detection of stable isotopes of elements found in proteins, using standard metaproteomics data.


26
Stable isotope probing (SIP) approaches are a critical tool in microbiome research to determine 27 associations between species and substrates. The application of these approaches ranges from 28 studying microbial communities important for global biogeochemical cycling to host-microbiota 29 interactions in the intestinal tract. Current SIP approaches, such as DNA-SIP or nanoSIMS, are 30 limited in terms of sensitivity, resolution or throughput. Here we present an ultra-sensitive, high-31 throughput protein-based stable isotope probing approach (Protein-SIP), which cuts cost for 32 labeled substrates by ~90% as compared to other SIP and Protein-SIP approaches and thus 33 enables isotope labeling experiments on much larger scales and with higher replication. It allows 34 for the determination of isotope incorporation into microbiome members with species level 35 resolution using standard metaproteomics LC-MS/MS measurements. The analysis has been 36 implemented as an open-source application (https://sourceforge.net/projects/calis-p/). We 37 demonstrate sensitivity, precision and accuracy using bacterial cultures and mock communities 38 with different labeling schemes. Furthermore, we benchmark our approach against two existing 39 Protein-SIP approaches and show that in the low labeling range used our approach is the most 40 sensitive and accurate. Finally, we measure translational activity using 18O heavy water 41 labeling in a 63-species community derived from human fecal samples grown on media 42 simulating two different diets. Activity could be quantified on average for 27 species per sample, 43 with 9 species showing significantly higher activity on a high protein diet, as compared to a high 44 fiber diet. Surprisingly, among the species with increased activity on high protein were several 45 Bacteroides species known as fiber consumers. Apparently, protein supply is a critical 46 consideration when assessing growth of intestinal microbes on fiber, including fiber based 47 Introduction 52 Microbial communities drive chemical transformations from global element cycling to human 53 nutrition. Unfortunately, the overwhelming complexity of these communities is often a barrier to 54 unraveling their functionality. Use of isotopic or chemical labeling is a powerful solution to that 55 problem. Even in the context of complex microbial communities, labeling enables assigning 56 activities and functions to taxa, tracking metabolic pathways and resolving trophic relationships 57 among species [1][2][3][4][5] . Current labeling approaches include use of click-chemistry (BONCAT) [6] , 58 nanoscale secondary ion mass spectrometry (nanoSIMS) [2] , Raman microscopy [7] , genomic 59 sequencing of isotope labeled DNA/RNA (DNA/RNA-SIP) [8] , separated from unlabeled 60 DNA/RNA with density gradient centrifugation, and protein-based stable isotope probing 61 metaproteomics (Protein-SIP) [9] . Some of these approaches use labels with defined chemistry 62 such as non-canonical amino acids in BONCAT [6] , which are directly assimilated into biomass. 63 Others use more generic labels, such as substrate molecules labeled with heavy isotopes of 64 carbon, nitrogen, oxygen and hydrogen [2,7,10,11] . When spatial organization of samples is 65 important, approaches are available to image labeling outcomes [12,13] . When it is unknown in 66 advance which species or pathway might be involved in a target process, labeling can be 67 combined with untargeted metagenomics and metaproteomics analyses. 68 Recently, we developed an algorithm (Calis-p 1.0) to estimate natural isotope abundances 69 (stable isotope fingerprints, SIF) of carbon isotopes of individual species within complex 70 microbial communities using metaproteomics [14] . In nature, 13 C and 12 C occur side by side at a 71 ratio of approximately 0.011 13 C/ 12 C. For microbial biomass, very subtle changes to this ratio, as 72 little as 0.0001, already provide information about carbon assimilation pathways and carbon 73 sources used. Our algorithm, which modeled mass spectra of individual peptides using Fast 74 Fourier Transformations (FFTs), was able to detect these subtle changes. In the present paper we 75 further develop this extremely sensitive approach to also work for stable isotope probing (SIP)  76 experiments. This enables us to detect and quantify the assimilation of heavy isotopes by 77 individual species in complex microbial communities using metaproteomics (Protein-SIP). 78 Protein-SIP differs from other metabolic labeling approaches in that the heavy isotopes 79 from the substrate are incorporated into protein through de novo synthesis of amino acids from 80 the substrates via biosynthetic pathways, rather than directly in the form of labeled amino acids. 81 Such labeled amino acids are used, for example, in the "Stable Isotope Labeling by Amino Acids 82 in Cell Culture" (SILAC) approach [15] . The "random" incorporation of label into various amino 83 acids and ultimately into peptides makes data analysis much more complicated in Protein-SIP, at 84 least compared to the predictable exact mass shifts resulting from direct assimilation of labeled 85 amino acids in SILAC. 86 Protein-SIP approaches have been successfully developed before, but these approaches 87 have their challenges (for an overview see introduction of [10] ). Metaproteomics relies on high-88 resolution mass spectrometry to detect, identify and quantify peptides, which are then used for 89 protein identification and quantification [16] . Using the same mass spectra already used for peptide 90 identification to also quantify abundances of heavy isotopes in these peptides appears a 91 straightforward add-on, as these spectra resolve the peptide isotopes and provide their intensities. 92 However, unknown amounts of heavy isotopes shift peptide mass peaks by unknown numbers of 93 mass-units, which makes the identification of peptides based on masses computationally 94 challenging. The existing Sipros algorithm solved this problem with brute force by coupling the 95 detection of labeled peptides with the initial peptide identification. Sipros predicts the most 96 abundant peptide masses and isotopic distributions of b and y ions in an isotope atom% range of 97 0 -100% in 1% increments [17] . This approach makes Protein-SIP experiments computationally so 98 expensive that dedicated smaller protein sequence databases have to be constructed for 99 determination of stable isotope content of peptides [18] and even then the approach still requires a 100 supercomputer to work. For example, one study using the Sipros approach had to invest around 101 500,000 CPU hours for a study with less than 10 labeled samples [10] . The MetaProSIP [19] and 102 SIPPER [20] algorithms overcame the problem by using spectra of unlabeled peptides as a starting 103 point for computations. In case of MetaProSIP these unlabeled peptides can be derived from the 104 SIP experiment itself if a portion of the original unlabeled proteins is still present, or, 105 alternatively, from a control sample that was incubated without label. MetaProSIP then detects 106 the labeled peptides corresponding to the unlabeled peptides and computes the relative isotope 107 abundance and labeling ratio based on the comparison of the labeled and unlabeled form of 108 peptides [19] . Because MetaProSIP requires a labeled peptide's spectrum to be shifted away from 109 the mono-isotopic mass, it has been speculated that it requires relatively heavy labeling (e.g. 110 >20.24 atom% for 13 C and >73.1 atom% for 15 N [21] ). In case of SIPPER the isotopic patterns for 111 unlabeled peptides are generated in silico and subtracted from the experimental isotope patterns 112 of peptides. Remaining peak intensities after subtraction are used for estimating isotope content. 113 SIPPER is designed to detect small changes in isotopic profiles of complex mixtures after short 114 exposure to 13C label, with proposed scoring schemes to reduce the rate of false discoveries. 115 While the identification challenges can be solved by clever algorithms, underneath these 116 challenges hides a more fundamental problem. Figure 1 shows the expected mass spectra of three 117 E. coli peptides after 1/8 generation of labeling with 13 C-glucose. The figure illustrates the 118 problem with these data: Assimilation of heavy isotopes into peptides leads to broadening of 119 spectra. Thus, a peptide's matter gets divided over ever more peaks, reducing sensitivity. Also, 120 because many peptides get injected into the mass spectrometer simultaneously, especially for 121 complex samples such as a microbial community, the probability of the peptide's spectrum 122 overlapping with another spectrum increases as it broadens, reducing data quality. Heavy 123 peptides are especially sensitive to these issues. Counter-intuitively, for Protein-SIP, sensitivity 124 is highest when using small amounts of label. 125 126 Figure 1: Modeled spectra of three E. coli peptides after ⅛ generations of growth on 1% (left) and 10% 127 (right) 13 C1-6 glucose (13C/12C 0.02 and 0.11 respectively). Assimilation of 13C into peptides leads to a 128 shift of matter away from from the monoisotopic mass (shown as *). The resulting peak intensity changes 129 are shown in red -for peaks with decreased intensity -, and blue -for peaks with increased intensity after

134
Our previous algorithm was developed to estimate slight differences in isotopic content 135 based on peptide mass spectra, to determine natural carbon isotope abundances [14] . It made use of 136 the fact that in nature, heavy isotopes are distributed randomly, yielding spectra that are perfect 137 Poisson distributions. This enabled us to reduce the noisiness of the data by identification and 138 rejection of imperfect spectra. Spectra in Protein-SIP experiments do not have such conveniently 139 predictable properties. With labeled samples, the shape of spectra cannot be predicted using FFT, 140 because these spectra become mixtures of spectra associated with labeled and unlabeled 141 peptides. Both the proportion of heavy isotopes in the labeled peptides and the extent of labeling 142 -the relative abundances of labeled versus unlabeled populations of peptides -are unknown in 143 advance. For analysis of these data we therefore developed rigorous noise filtering and estimated 144 isotopic content based on neutron abundance, requiring no assumptions about a spectrum's 145 shape. 146  We present new algorithms and software for sensitive and quantitative estimation of  147  isotopic content of individual species in stable isotope probing experiments with complex  148  microbial communities. The new algorithms have been integrated into the Calis-p software  149  together with the SIF algorithms, and the software was completely re-written to enable Protein-150 SIP (new version is Calis-p 2.1). The software decouples peptide identification from label 151 detection and is thus compatible with most standard peptide identification pipelines. 152 Computation of label content is very fast, a high-end desktop computer only needs one minute 153 for processing ~1 Gb of data, corresponding to ~10,000 MS1 spectra or ~40 min of Orbitrap 154 runtime. Using pure cultures of bacteria and mock communities, we show that Protein-SIP with 155 Calis-p yields best results when substrates are partially labeled. For example, for carbon the 156 fraction of heavy atoms should make up <10% of the total. For abundant organisms, 157 assimilation of label (such as 13 C) into protein can be quantified within minutes after adding the 158 label, within 1/16 of a generation. Even for rare organisms making up ~1% of a community, a 159 single generation of labeling is sufficient for robust detection of label assimilation. We believe 160 these advances will be helpful to microbiome researchers and microbial ecologists seeking to 161 assign functions and activities to taxa, to track metabolic pathways and for resolving trophic 162 relationships among species. 163

164
Previously we presented algorithms and software for estimating natural isotope fingerprints from 165 peptide mass spectra [14] . Our previous algorithm made use of the stochastic distribution of 166 isotopes in nature and mass spectra that can be modeled by Fast Fourier Transformations. 167 Quality control is intrinsic to that approach, as poor quality spectra cannot be modeled with FFT 168 and can be rejected. Examples of low quality spectra are spectra that overlap with other spectra 169 or low intensity spectra that are affected by noise. Feeding microbes labeled substrates for 170 Protein-SIP experiments leads to peptide mass spectra with irregular shapes that cannot be 171 modelled with FFT, as explained in the introduction. Isotopic composition of such spectra can 172 still be inferred, by adding up the mass intensities of all peaks in the spectrum according to 173 Equation 1 in the Methods (implemented as "neutron abundance" model in Calis-p). 174 Unfortunately, that approach does not enable rejection of low quality spectra. Therefore, we 175 implemented a simple noise filter based on unsupervised Markov clustering of all spectra 176 associated with a single peptide (see Materials and Methods for details). The assumption 177 underlying this approach is that most spectra are relatively unaffected by noise and will form the 178 largest cluster. Spectra outside the largest cluster should be rejected for being of lower quality. 179 The performance of this filter was benchmarked using previous natural-isotope abundance data 180 of pure cultures and mock communities of microbes (Suppl. Results & Discussion, Fig. S1, 181 Tables S1 & S2). The FFT estimates of 13 C/ 12 C ratios for filtered spectra was as good or better 182 than reported previously without filtering [14] . Even better, after filtering the estimates of 13 C/ 12 C 183 ratios according to Equation 1 (see Materials and Methods) were now almost as good as for FFT. 184 The average difference between the actual and estimated median δ 13  between peptides identified and extra computation time needed. This strategy adds six custom 210 "post-translational" modifications to the protein identification search. These modifications 211 generated more PSMs by enabling addition of one to three neutron masses at the N-terminus of a 212 peptide and four to six neutron masses to its C-terminus. We observed a strong difference in how 213 label amount impacted the number of PSMs between B. subtilis and E. coli. For B. subtilis a 214 small amount of added label strongly increased the number of PSMs, which then sharply 215 dropped at 1% label. We currently have no good explanation for this phenomenon.

235
After assigning peptides to spectra with the improved peptide identification strategy, low 236 quality spectra were rejected using the filter described above. For the remaining spectra, the 237 number of neutron masses added as custom modifications during the identification step already 238 provided a qualitative, or at best semi-quantitative, measure for label incorporation (Fig. 3, 239 Suppl. Table S3). However, inference of the 13 C/ 12 C ratios by Equation 1 was much more 240 precise, even for minimally (0.01%) labeled cells providing a limit of detection <0.01% label in 241 most cases (see supplementary text). Precision and especially label recovery were both higher 242 when using glucose labeled at only a single position rather than with fully labeled glucose. For 243 the latter, the recovery was only 75-79%, meaning that the 13 C/ 12 C ratio was 21-25% lower than 244 expected. Potentially, this was caused by broadening of spectra with fully labeled glucose ( Fig.  245 1). As explained in the introduction, broader spectra reduce sensitivity. Interestingly, the breadth 246 of spectra could be used to infer to what degree 13 C carbon was assimilated in clumps of multiple 247 atoms (pie charts in Fig. 3). This approach, which only works when all atoms in a substrate are 248 labeled and when cells are labeled to saturation, could be used to infer the number of carbon 249 atoms in substrates that a given species is assimilating. In other words, Protein-SIP can provide 250 hints on whether a species is autotrophic or heterotrophic. We carried our similar labeling to 251 saturation experiments with 15N (E. coli labeled to saturation with 2.5% 15N ammonium) and 252 obtained similar results (data not shown, but available via PRIDE see data accessibility 253 statement). 254 257 fully labeled ( 13 C1-6) glucose. The 13 C/ 12 C ratio in the substrate was varied. Note that unlabeled glucose 258 (0% added 13 C glucose) has a natural 13 C content of around 1.1%. Each orange circle is the median 259 13 C/ 12 C ratio of all peptides measured in one replicate incubation (on average 2758 peptides per 260 replicate). Determined 13C/12C ratios increased linearly with substate 13 C/ 12 C ratios (R 2 >0.999). Almost 261 100% of the substrate 13 C was recovered in protein for 13 C2 glucose labeled cells. Recovery was lower for 262 13 C1-6 glucose. The proportion of neutron masses detected via the improved peptide identification strategy 263 using N-and C-terminal modifications (yellow circles) increased with substrate 13 C/ 12 C ratios, but at low 264 linearity and sensitivity. The number of Calis-p filtered Peptide Spectrum Matches (PSM) decreased for 265 13 C/ 12 C ratios above 2.5% (insets) as expected based on Figures 2 & S2. Assimilation of carbon into 266 amino acids in clumps of multiple 13 C atoms was detectable in peptide spectra of cultures fed with 13 C1-6 267 glucose as shown in pie charts for experiments fed with 13 C/ 12 C 1% above natural background. The 268 detailed data for this figure can be found in Suppl. Table S3.

269
The data of Figure 3 are not yet a meaningful representation of what an actual Protein-270 SIP experiment would look like. In practice, we would always avoid labeling a microbial 271 community to saturation, because all community members would end up being labeled equally, 272 providing no new information on elemental fluxes and substrate uptake in the community. To 273 mimic an actual Protein-SIP experiment, we mixed labeled and unlabeled cells of E. coli at 274 different ratios, leading to compound spectra as shown in Figure 1. 275 The results indicated that estimation of 13 C/ 12 C ratios with this type of compound spectra was 276 more challenging (Fig. S3, Suppl. Table S4). Also, the difference between single labeled and 277 fully labeled glucose was more pronounced, with the former yielding much better sensitivity and 278 label recovery than the latter. For example, at 1% label recovery was 92% for single labeled 279 glucose, while it was 80% for glucose with 6 labeled carbons.We also compared the performance 280 of two center statistics for 13 C/ 12 C ratios, the intensity-weighted mean and the median. The 281 intensity-weighted mean displayed higher sensitivity and precision than the median in these 282 experiments (for contrasting results for community samples see below). However, both with 1% 283 and 10% single labeled glucose, even the median 13 C/ 12 C ratios accurately quantified label 284 assimilation within 1/16 of a generation (simulated by mixing of labeled and unlabeled cells), 285 corresponding to as little as 1-2 min of growth for E. coli. 286 Next, we investigated whether our approach was capable of detecting label assimilation 287 in the context of a microbial community. For this, we used a previously described mock 288 community, comprising >30 microbes, including gr+ and gr-bacteria, an archaeum, a eukaryote 289 (algae) and several phages [22] . This community also included E. coli K12, at ~6% abundance. 290 Here, we mixed cells of E. coli labeled with 1%, 5% and 10% 13 C1-6-glucose into the unlabeled 291 mock community at a ratio corresponding to one generation of growth for E. coli. Quantification 292 of 13 C assimilation was straightforward and linear (R 2 0.99, Fig. 4A, Suppl. Table S5). This was 293 perhaps not surprising because a relatively large amount of label was used and the relative 294 abundance of E. coli in the mock community was high, i.e. ~12% after addition of the labeled 295 cells. 296   relative abundance of these unlabeled organisms was between 0.1% and 7% . The determined 314 13 C/ 12 C ratios for the >20 other members of the mock community are reported in supplementary 315 table S5. We found that the choice of center statistic used has a major impact on the false 316 positive detection of label incorporation. When using the median, the overall (i.e. all unlabeled 317 species in all replicates) False Positive Rate (FPR) of label detection for populations with nine or 318 more peptides (after filtering) was 3.4% and for populations with eight or fewer peptides it was 319 45%. In contrast, when using the weighted mean the FPR was 51% for populations with nine or 320 more peptides and 50% for populations with eight or less peptides. In our dataset, the nine 321 peptide threshold corresponded to ~1% relative abundance of strains/species within the mock 322 community. We investigated the massive differences in FPRs between the two center statistics 323 by manually checking spectra causing false positives and found that low-intensity peptide 324 spectra associated with less abundant populations were often affected by the overlap with 325 broadened spectra of a labeled, more-abundant population. Therefore, we concluded that for 326 label detection in microbial communities the median should be used (see detailed discussion in 327 supplementary methods). Figure 4 shows examples of false-positive inferences for 328 Pseudomonas pseudoalkaligenes. 329 To investigate whether label assimilation can be correctly inferred for less abundant 330 populations, we downsampled (bootstrapped, up to ten times) the set of >6,000 peptides 331 collected for E. coli, using the peptides of each other organism as templates. In the resulting 332 datasets, each E. coli peptide was matched to a peptide of the other organism with a similar 333 intensity. Based on inferences for these bootstrapped datasets shown in Figure 4, label 334 assimilation could be robustly estimated, at least for populations associated with nine or more 335 peptides, corresponding to ~1% abundance. This number of peptides is much smaller than the 336 ~30 peptides needed for estimation of natural carbon isotope content in a species using Protein-337 SIF (Suppl. Results & Discussion, Fig. S1). 338 Next, we analyzed how well we could detect incorporation of label into individual 339 proteins based on how many peptides passed the Calis-p quality filters for a protein. For this we 340 analyzed the Calis-p reported 13 C/ 12 C ratios for proteins from the mock communities with 5% 341 labeled E. coli spiked-in and without spiked in E. coli. 13 C/ 12 C ratios in E. coli proteins from the 342 5% spike-in samples were on average much higher than the ratios for proteins from the unlabeled 343 mock communities and the unlabeled mock community members in the 5% spike-in samples 344 (Fig. 5a). Even for proteins for which only 1 peptide passed the Calis-p quality filters, this 345 pattern was observed. This indicated that label incorporation into individual proteins can be 346 detected with as few as 1 peptide. For some proteins from unlabeled organisms, for which 3 or 347 less peptides passed the Calis-p quality filter, 13 C/ 12 C ratios that were above the expected value 348 of 0.011 (Fig. 5a)

392
To compare the performance of Calis-p with existing Protein-SIP approaches four of the above described 393 datasets with labeled E. coli spiked into a mock community were processed by expert operators of the 394 SIPPER and MetaProSIP workflows. We would like to highlight here that these two approaches were 395 developed for higher label amounts (MetaProSIP) and low level labeling after shot label exposure 396 (SIPPER) and they may well outperform Calis-p under specific conditions. However, we focused our 397 comparison on the low label amounts for which the ultra-sensitive Protein-SIP within Calis-p was 398 developed and a comparison at high label amounts (>10% label) was outside the scope of this study.

399
The three approaches differed strongly in the number of peptides for which label content was 400 quantified (Fig. 6b). Part of this was due to the fact that for each of the three approaches different peptide

411
The three approaches also differed in the number of false positive detection of above natural 412 abundance 13C in peptides (Fig. S4, Tables S7 to S10) and species (Figs. 6a & S4). Calis-p 413 overestimated 13C content only in a small number of peptides and in none of the species (i.e. median 414 13C content was at or below expected value). MetaProSIP also did not overestimate label content in any 415 of the species, instead it had the tendency to underestimate isotope content of unlabeled species.

416
SIPPER had a high rate of false positive detections, while for the 1% labeled sample all 13C content 417 estimates (including the estimate for the labeled E. coli) were strongly underestimated. 13C content of 418 unlabeled species for which enough peptides were quantified (9 peptides) were strongly overestimated.

419
This false positive label detection is likely due to the underlying principle of SIPPER, which is to identify 420 labeled peptides, while not trying to classify unlabeled peptides. This means that truly unlabeled peptides 421 are not available for calculation of label content of species and a small number of false positive label 422 detection in peptides can lead to miss estimation of label content in species and proteins.

423
Finally, the three approaches also differed in the accuracy of 13C content estimates for the labeled E.

429
In summary, while MetaProSIP outperformed Calis-p in terms of quantity of peptide isotope quantification,

430
Calis-p was more sensitive and accurate. The sensitivity of MetaProSIP was, however, much better

431
(detection down to at least 5% 13C) than recently suggested by Starke [21] . Both MetaProSIP and Calis-p 432 showed high specificity; the lower specificity shown by SIPPER was likely caused by SIPPER's primary   Figure S4 and tables 451 S7 to S10.

452
Case Study: Differential heavy water incorporation reveals activity changes 453 for intestinal microbiota species in response to dietary changes 454 To demonstrate the power of the Calis-p approach and to test our approach for additional 455 elements we analyzed data from a complex microbial community grown with two types of heavy 456 water [11] .  S5) and we were able to quantify heavy water incorporation (>9 peptides passing Calis-p 466 filters for species) for 21 to 30 species per sample (Mean = 27, Median = 28). We found that 467 overall incorporation of 18 O was much higher than incorporation of 2 H (Fig. S4). The low 468 measured incorporation of 2 H can potentially be attributed to variation in retention times of 469 isotopically different forms of a deuterated peptide in reversed-phase chromatography [24,25] . Such 470 retention time variation can generate distinct isotope patterns for the same peptide at different 471 retention times, which would lead to failure to cluster by the Markov clustering in Calis-p during 472 spectrum filtering. Since the amount of 2H used in this experiment was relatively high we do 473 expect relevant retention time shifts of deuterated peptides. Additional factors that might explain 474 low measured incorporation of 2 H are the known strong fractionation of hydrogen isotopes in 475 organisms [26] , the fact that many hydrogen atoms on peptides can freely exchange with water 476 leading to loss of label during sample preparation [27] , the dilution of 2H in stable C-H bonds in de 477 novo synthesized amino acids by hydrogens derived from organic growth substrates [23,28] , and the 478 known toxicity of deuterium to many organisms slowing down their growth rates and thus 479 reducing the rate of incorporation, which however usually occurs at higher concentrations 480 (>50%) of deuterium than used in this experiment [23,29]  water [27] . For hydrogen to be in positions with low exchangeability, amino acid de novo synthesis 487 is required, because the necessary carbon-hydrogen bonds are only generated then [27] . 488 On the whole community level, label incorporation was significantly higher in 489 communities grown with high protein as compared to high fiber (Fig. S5). On the level of single 490 species we observed that responses to a change in "diet" were species-specific with some species 491 such as Akkermansia muciniphila, Bacteroides ovatus, and Clostridium bolteae incorporating 492 significantly more label under high protein conditions, while other species, such as Alistipes 493 onderdonkii, Clostridium lavalense, and Flavonifractor plautii, showed no change or non-494 significant trends towards higher incorporation under high fiber conditions (Fig. 7) protein, is the limiting nutrient for the intestinal microbiota [31,32] . This indicates that mixed 501 community bioreactors can be a useful analog to the intestinal tract for studying specific 502 ecological factors (such as nitrogen limitation) driving community function. Surprisingly, 503 although typically described as fiber degrading specialists [33][34][35] , we saw a significant increase of 504 activity in several Bacteroides species in the high-protein medium relative to the high-fiber 505 medium. This shows that it is critical to assess nitrogen/protein supply when analyzing fiber 506 dependent growth of intestinal microbes. Furthermore, it suggests that nitrogen/protein supply is 507 critical to consider when developing fiber based prebiotics to manipulate intestinal microbiota 508 species [36,37] , which to our knowledge has not been considered so far. In summary, our results 509 show that the use of heavy water for Protein-SIP allowed us to detect changes in the activity of 510 microbiota members in response to changes in complex substrates. 511 input to Calis-p, (5) isotope pattern extraction and computation of isotope content in Calis-p, and 525 (6) analysis and interpretation of data provided by Calis-p (Fig. 8). The provision of isotopically 526 labeled substrates in experiments can take many forms, such as addition of substrate to 527 incubations of enrichment cultures/bioreactors [11] , addition to animal feed [38,39] , CO2 in plant 528 incubation chambers [40] or as 15 N in plant fertilizer, and in situ incubations [10] . For the Protein-529 SIP approach presented here, substrate should be supplied with 1-10% of the total substrate 530 containing the heavy isotope (label). Please note that this range refers to 13 C, for other elements, 531 such as N, which make up a smaller portion of atoms in a peptide, a higher amount of label can 532 be used, as the associated peptide mass shifts are smaller. If the substrate is a small molecule 533 (e.g. glucose), but contains multiple atoms per molecule of the element to be labeled, ideally 534 only one of the atoms is labeled (or a small portion of atoms if it is a very large molecule) to 535 avoid isotope "clumping", as this can lead to a reduction in sensitivity (Fig. 3). Calis-p can, 536 however, handle "clumped" data if needed. Similarly, if a complex substrate is used (e.g. 537 complete plant leaves) ideally the complex substrate should only be partially labeled (e.g. by 538 growing plants in an atmosphere with 10% of the CO2 being labeled) rather than using fully 539 labeled substrate. 540 Other considerations for the labeling experiments include the number of replicates that 541 are required, which depends on the biological question of the experiment, if a time course or a 542 single time point will be sampled, and if a control with unlabeled substrate will be carried out, 543 which is not needed for Calis-p, but can be helpful in data interpretation. Generally, we 544 recommend to carry out a feasibility study, if at all possible, to determine the correct amount of 545 label that works for the study system and time points that need to be sampled. Measurement of 546 bulk label incorporation using an isotope ratio mass spectrometer can be useful in determining if 547 an experiment worked prior to starting sample preparation for Protein-SIP. 548 The produced samples should be processed with a standard metaproteomic sample 549 preparation method tuned to the particular sample type. In contrast to the protein-SIF method [14] , 550 which requires calibration for a small isotope offset caused by the instrument, no calibration 551 reference material needs to be prepared for Protein-SIP. The produced peptide mixtures need to 552 be analyzed by 1D or 2D liquid chromatography (LC) and tandem mass spectrometry (MS/MS) 553 using a high-resolution Orbitrap mass spectrometer with standard metaproteomic LC-MS/MS 554 approaches (see Methods and e.g. [41] ). One important consideration for the data acquisition in 555 the mass spectrometer is the choice of resolution particularly for experiments involving 15 N 556 labeling (see Suppl. Results and Discussion). 557 The steps for data preparation for Calis-p and the computational steps implemented in 558 Calis-p are described in detail in the Methods and on the Calis-p software repository website 559 (https://sourceforge.net/projects/calis-p/). 560

568
The developed Protein-SIP approach provides a means to detect and quantify the incorporation 569 of stable isotopes from labeled substrates into many individual species in microbial communities 570 in one LC-MS/MS measurement and with minimal computational cost. Our approach has many 571 advantages over other SIP approaches and previously developed Protein-SIP approaches. First, 572 the approach allows for high throughput, as compared to most other stable isotope probing 573 methods, such as DNA/RNA-SIP and nanoSIMS because as little as 2 hours of LC-MS/MS time 574 will allow to quantify label incorporation for a good number of the more abundant species in a 575 sample. For example, in bioreactors with 63 species we were consistently able to obtain 576 sufficient measurement depth to quantify heavy water incorporation in >=20 species (Fig. 6). In 577 contrast, nanoSIMSonly allows for measurement of isotope incorporation into a limited number 578 of individual cells of very few species (2-3) in this time frame as species assignment of cells 579 depends on species specific probes, and DNA-SIP is limited by the number of samples that fit 580 into the ultracentrifuge rotor (usually six) needed for fine scale separation of heavy and light 581 DNA and the cost associated with sequencing of a great number of individual density gradient 582 fractions per sample. Second, our approach is a departure from previously developed Protein-SIP 583 approaches in that it is highly sensitive and affords a large dynamic range of three orders of 584 magnitude detecting label incorporation in the range of 0.01-10% of added label, while previous 585 Protein-SIP detection principles require much higher label amounts to enable detection and 586 usually offer only a dynamic range of one order of magnitude [21] . Similar is true for DNA/RNA-587 SIP based approaches, which require at least 20% label for detection [8] . The high sensitivity, 588 accuracy and large dynamic range of our approach brings numerous advantages, including 589 significant cost reduction due to lower use of often very expensive isotopically labeled 590 substrates, the ability to work with much shorter labeling times, and simultaneous detection of 591 label incorporation in slow and fast growing microorganisms. A massive cost reduction 592 compared to other SIP approaches is possible, because most existing (Protein-)SIP approaches 593 use upward of 20% (most often 100%) labeled substrate, while when using the Calis-p approach 594 experiments with 1-10% label can be done cutting the experimental isotope use by 50-99%. 595 Using shorter incubation times is possible because incorporation of labels into proteins does not 596 require for replication to occur, which is the case for DNA-SIP. It is important to note here that 597 the 13 C-label content of the substrate needs to be kept at 10% or below for our approach to work 598 (higher percentages can be used for other elements see Box). A short labeling pulse with a 599 substrate with higher label percentage would generate a heavy peptide population that would be 600 completely mass shifted away from the unlabeled peptide population and thus become 601 undetectable by Calis-p. Such strongly mass shifted peptide populations would be detectable 602 with the MetaProSIP [19] and SIPPER [20] softwares. Third, we developed our approach to work 603 with stable isotopes of all elements present in proteins, which allows tracking of assimilation of a 604 large diversity of simple and complex substrates, as well as general activity markers such as 18 O 605 water. Based on our 2 H and 18 O case study results, we would recommend to use 18 O water as the 606 activity marker if compatible with the experimental design, as the current Calis-p version showed 607 much higher sensitivity with the 18O data. More testing and optimization of Calis-p will be 608 needed in the future for deuterated water using data to be generated with lower 2H labeling 609 amounts. Fourth, Protein-SIP does not require isotope based separations of biological material 610 such as the density gradient centrifugation used for DNA/RNA-SIP. That approach requires large 611 amounts of material and sequencing of multiple fractions per sample. For this reason, Protein-612 SIP can be done with very small amounts of sample with an ideal starting amount of 1 mg or 613 more of wet weight cell mass [14] . However, we have achieved good isotope estimates with as 614 little as 50 µg using Calis-p for stable isotope fingerprinting [42] . 615 Currently, Protein-SIP only allows for labeling with one isotope per sample as changes in 616 peptide isotope patterns cannot be attributed to specific elements. However, in the future it might 617 be possible to develop Protein-SIP approaches that allow for parallel measurement of 15 N and 618 13 C incorporation in a single sample, because added neutron masses for 15 N and 13 C are 619 sufficiently different from each other -due to differences in nuclear binding energy-to allow for 620 their separation in ultra-high resolution mass spectrometers (Suppl. Results and Discussion). The 621 current limitation for generating ultra-high resolution data suitable for separating peptide carbon 622 and nitrogen isotopes is that higher resolution comes at slower mass spectrometric acquisition 623 time. Thus, there is a tradeoff between ultra-high resolution data acquisition and obtaining a 624 large number of MS 2 spectra for peptide identification. Instruments with faster acquisition times 625 and potentially alternative data acquisition modes such as data-independent acquisition (DIA) 626 metaproteomics could make dual-label Protein-SIP feasible in the next few years. 627

628
Generation of labeled pure culture samples 629 The following steps were followed for single-carbon labeled and six-carbon labeled 13  of overnight culture were spun down at 18,000 g for five minutes, the supernatant was discarded 648 and pellets were washed twice with PBS to remove unlabeled glucose. Pellets were resuspended 649 in 1 ml PBS. Labeling: Ten milliliters of liquid media without glucose were aliquoted into a 650 total of 24 serum bottles per strain (triplicate bottles for each of the eight 12 C/ 13 C glucose mixes). 651 200 µl of the 12 C/ 13 C mixes and 10 µl of overnight culture were added into the serum bottles. 652 The bottles were then crimped, the headspace was flushed three times with CO2-free air and 653 cultures were incubated overnight at 37˚C while shaking at 100 RPM. Sample processing: 654 Serum bottles were depressurized by inserting a sterilized needle into the septum to release air. 655 Ten milliliters of culture from each bottle were spun down at 18,000 g for five minutes. The 656 supernatant was discarded and the pellet resuspended in 2 ml of PBS to make two 1 ml aliquots. 657 50 µl of 1%, 5% and 10%-labeled glucose grown cells were used for cell counts using a 658 Neubauer counting chamber. Cells were pelleted at 10,000 g for five minutes, the supernatant 659 was discarded and pellets were flash-frozen in liquid nitrogen before being transferred to -80°C. 660

661
The generation of the mock community (UNEVEN type) is described in Kleiner et al. (2017) [22] . 662 We mixed E. coli cells grown in 1, 5 and 10% 13 C6-labeled glucose containing media into three 663 replicate samples of this mock community. We mixed the labeled E. coli cells in Once input files and optional parameters are provided Calis-p extracts isotope patterns for all 717 identified peptides using a procedure optimized for Protein-SIP. The isotope patterns are 718 extensively filtered for quality and high quality patterns are used for calculation of peptide 719 isotope content using three different models. The "default" model developed for Protein-SIF, the 720 "neutron abundance" model, which usually works best for Protein-SIP, and the "clumpy" model 721 (see Methods). Calis-p automatically provides output files for all three models for taxa, proteins 722 and peptides in a tabulated format that can subsequently be used in statistical and other data 723 analysis softwares such as R. 724 SIP computation algorithms and computational improvements to increase 725 speed and accuracy of isotopic pattern extraction 726 As a starting point for estimation of stable isotope composition of isotopically labeled samples, 727 we augmented the Calis-p software previously developed for estimation of 13 C at natural 728 abundance [14] . For estimating natural 13 C abundance the software uses a model that assumes 729 random distribution of 13 C atoms in peptides, leading to peptide spectra with predictable isotope 730 patterns. These isotope patterns are modelled in Calis-p with Fast Fourier Transformations 731 (FFT). With labeled samples, the shape of spectra cannot be predicted using FFT, because these 732 spectra become mixtures of spectra associated with labeled and unlabeled peptides. Both the 733 proportion of heavy isotopes in the labeled peptides and the extent of labeling -the relative 734 abundances of labeled versus unlabeled populations of peptides -are unknown in advance. . 735 Therefore, we used the following more general equation to infer the number of neutrons from 736 peptide isotope patterns to implement a "neutron abundance" model: 737 With, on the left, considering an isotope pattern of size n peaks, p is the peak number, and I is the 739 intensity of peak p. On the right, for each isotope, e is its element [C,H,O,N,S], n the number of 740 additional neutrons, φ its abundance (fraction), and a the number of atoms of the element in the 741 peptide associated with the spectrum.
The second SIP computation algorithm was implemented in Calis-p as the "clumpy label" model. 752 When labeling with substrates that contain multiple isotopically labeled atoms, for example fully 753 labeled 13 C1-6 glucose, this can lead to assimilation of clumps of labeled atoms into a single 754 amino acid. For example, fully labeled glucose will be converted to fully labeled pyruvate, 755 which, in turn, will be converted to fully labeled alanine, which will be incorporated into protein.

756
This leads to peptide spectra that display higher-than-expected intensity at a higher isotopic peak 757 numbers. To estimate the "clumpiness" of heavy isotopes in peptides, we developed the 758 following procedure (detailed explanation in Figure S6): First, only the monoisotopic peak (A = 759 +0) and A+1 peaks of the spectrum are used to estimate the fraction assimilated in clumps of 760 one heavy atom (e.g. 13 C). Next, the experimental intensity of the A+2 peak is compared to its 761 expected intensity assuming all label was assimilated in clumps of one heavy atom. Any 762 additional intensity of the A+2 peak is assigned to assimilation of clumps of two heavy atoms. 763 This way, all peaks up to A+6 are inspected. The algorithm assumes the peptides are completely 764 labeled, i.e. labeled to saturation. Usually, stable isotope probing experiments do not proceed that 765 long, but doing so would enable determination of the number of labeled atoms in the substrate 766 assimilated by each species via this procedure. 767 In typical proteomics data, tens to hundreds of MS1 spectra are collected for each 768 detected peptide, at different elution times and mass over charge ratios. MS1 spectra can be 769 crowded, especially for samples from more complex microbial communities. Unfortunately, 770 overlap between isotopic patterns associated with different peptides can lead to overestimation of 771 labeling. We have added new filtering routines, which remove such compromised isotopic 772 patterns in two steps. First, any isotopic patterns with uneven spacing between peaks (which 773 could indicate overlap with another spectrum) are discarded. On average, the peaks that form an 774 isotopic pattern associated with a given peptide are separated by 1.002 Da, divided by the charge 775 z of the peptide. If a pattern's median peak spacing was <1.000/z or >1.004/z, or if the average 776 sum of squares of the difference between the actual spacings and the median spacing was > 777 1×10-5, the entire isotopic pattern was discarded. 778 Next, remaining spectra are filtered out by unsupervised Markov clustering of all 779 remaining spectra associated with a peptide [47] . The premise of this filtering approach is that 780 clean spectra will be similar to each other, while spectra affected by noise are likely to be more 781 different from each other. After filtering, all remaining spectra are truncated to the most common 782 number of peaks, and spectra with fewer peaks are discarded. Spectra are then normalized to a 783 total intensity of 1, and an average (weighed by total spectral intensity) normalized spectrum was 784 calculated for each peptide. The averages are weighed by intensity because high intensity spectra 785 are more accurate and less noisy. 786 The normalized spectrum of each peptide is used to estimate the peptide's isotopic 787 composition using the original "Fast Fourier Transformations" based model (also called 788 "default"), as well as the new "neutron abundance" (Equation 1) and "clumpy label" models.. 789 For each species and protein in the sample, two center statistics are calculated based on all 790 peptides associated with a species or protein: the median and the intensity-weighted average. The 791 supplementary methods provide a detailed discussion of which center statistic to use when. 792 Other improvements of the Calis-p software 803 In addition to expanded functionality with regard to filtering of peptides and labeling, the 804 software was also improved in many other ways: It now computes isotopic content of peptides 805 with post-translational modifications and peptides containing sulfur peptides. It finds many more 806 MS1 spectra for each peptide by searching for spectra at additional mass to charge ratios. Next to 807 tab-delimited text PSM files exported from Proteome Discoverer, it now also parses open source 808 mzidentml XML files (http://www.psidev.info/mzidentml). Finally, code efficiency 809 improvements and implementation of multi-threading led to much faster computation, requiring 810 less than one minute to process all spectra recorded during a 2 h run on a QExactive Plus 811

Generating an additional label incorporation measure and Increasing
Orbitrap mass spectrometer, using 10 threads. Source code and more details about algorithms 812 and procedures can be found at http://sourceforge.net/projects/calis-p/. 813

814
To benchmark and compare Calis-p against existing Protein-SIP tools under optimal 815 operation conditions, we invited developers/operators of Sipros [17] , MetaProSIP [19] and SIPPER 816 [20] to participate in a tool comparison. Drs. Sachsenberg (MetaProSIP) and Tolić (SIPPER) 817 joined our effort. We were unfortunately not able to find an operator for Sipros and were also 818 unable to get the tool to work on our own. For the comparison we used four raw files from the 819 mock community spike-in experiments including unlabeled, 1% label, 5%, label, and 10% label 820 in the spiked in E. coli and the corresponding protein sequence database. For processing with 821 Calis-p we used the optimized settings from this study. 822 The SIPPER tool was developed at the Pacific Northwest National Laboratory and is 823 available as an open source software written in C#. SIPPER requires an unlabeled incubation MS 824 dataset as a reference for extracting target peptide IDs to which all stable isotope incubated 825 datasets are compared. To generate peptide-spectrum matches to be used for isotope estimates 826 the MSGF+ search algorithm [48] was used to search the unlabeled sample against the protein 827 sequence database. The precursor mass tolerance was set to 20 ppm and oxidation of M and N-828 terminal acetylation were included as dynamic modifications and carbamidomethylation of C as 829 static modification -, The search provided 112,580 identified target sequences (including 830 contaminant IDs). The parameters for the SIPPER isotope calculation run included summing 7 831 precursor spectra around each target scan number, a 10 ppm tolerance for mass accuracy, 10% 832 tolerance for normalized elution time, and filter confidence ID criteria outlined in the 833 manuscript. 834 The MetaProSIP tool is integrated into the OpenMS open source software GUI [49] . An 835 OpenMS workflow including a database search with Comet followed by the MetaProSIP tool 836 was built according to the recommended parameters from the original publication with the minor 837 modification that we activated the MetaProSIP setting to subtract the mono isotopic peak value. 838 A reference (unlabeled) incubation was not needed to perform isotope calculations, because we 839 expected part of the E. coli peptide population in the sample to be unlabeled due to the presence 840 of unlabeled E. coli in the mock community thus all four data files were run through the 841 workflow. The database search parameters were 10 ppm mass tolerance for precursor ions, up to 842 two missed cleavages, carbamidomethylation of C as static modification, and oxidation of M as 843 dynamic modification. Results were reported as relative isotope abundances (RIA) per ID. The 844 MetaProSIP algorithm can split the RIA distributions of peptides with higher abundances into 845 multiple isotope abundance clusters, leading to multiple reported RIAs per peptide. This 846 happened on average for ~9% of peptides in our datasets (0.3% in unlabeled, 6% in 1% labeled, 847 14% in 5% labeled, and 13% in 10% labeled). If the algorithm provided multiple RIAs for a 848 peptide, we only used the highest RIA value in all downstream calculations. 849 To make the output data from both SIPPER and MetaProSIP comparable to the Calis-p 850 output we used the tables from both approaches that report 13C content per PSM or peptide. 851 SIPPER 13C content was reported in terms of percent carbon and percent peptides labeled, 852 MetaProSIP 13C content was reported as relative isotope abundance, and Calis-p 13 C/ 12 C ratios 853 were converted to 13C atom percent. We filtered the data to only include distinct protein unique 854 peptides then calculated the 13C content for each species by taking the median. For visual 855 representations (Fig. S4), we additionally filtered the data for which 13C values were provided 856 for at least 9 peptides per species (equivalent to filter used for Calis-p). 857