A standardized quantitative analysis strategy for stable isotope probing metagenomics

Stable isotope probing (SIP) facilitates culture-independent identification of active microbial populations within complex ecosystems through isotopic enrichment of nucleic acids. Many SIP studies rely on 16S rRNA sequences to identify active taxa but connecting these sequences to specific bacterial genomes is often challenging. Here, we describe a standardized laboratory and analysis framework to quantify isotopic enrichment on a per-genome basis using shotgun metagenomics instead of 16S rRNA sequencing. To develop this framework, we explored various sample processing and analysis approaches using a designed microbiome where the identity of labeled genomes, and their level of isotopic enrichment, were experimentally controlled. With this ground truth dataset, we empirically assessed the accuracy of different analytic models for identifying active taxa, and examined how sequencing depth impacts the detection of isotopically labeled genomes. We also demonstrate that using synthetic DNA internal standards to measure absolute genome abundances in SIP density fractions improves estimates of isotopic enrichment. In addition, our study illustrates the utility of internal standards to reveal anomalies in sample handling that could negatively impact SIP metagenomic analyses if left undetected. Finally, we present SIPmg, an R package to facilitate the estimation of absolute abundances and perform statistical analyses for identifying labeled genomes within SIP metagenomic data. This experimentally validated analysis framework strengthens the foundation of DNA-SIP metagenomics as a tool for accurately measuring the in situ activity of environmental microbial populations and assessing their genomic potential. Importance Answering the question of ‘who is eating what?’ within complex microbial communities is paramount for our ability to model, predict, and modulate microbiomes for improved human and planetary health. This question is often pursued using stable isotope probing to track the incorporation of labeled compounds into cellular DNA during microbial growth. However, with traditional stable isotope methods, it is challenging to establish links between an active microorganism’s taxonomic identity and genome composition, while providing quantitative estimates of the microorganism’s isotope incorporation rate. Here, we report an experimental and analytical workflow that lays the foundation for improved detection of metabolically active microorganisms and better quantitative estimates of genome-resolved isotope incorporation, which can be used to further refine ecosystem-scale models for carbon and nutrient fluxes within microbiomes.

acids. Many SIP studies rely on 16S rRNA sequences to identify active taxa but 23 connecting these sequences to specific bacterial genomes is often challenging. Here, we 24 describe a standardized laboratory and analysis framework to quantify isotopic 25 enrichment on a per-genome basis using shotgun metagenomics instead of 16S rRNA 26 sequencing. To develop this framework, we explored various sample processing and 27 analysis approaches using a designed microbiome where the identity of labeled 28 genomes, and their level of isotopic enrichment, were experimentally controlled. With this 29 ground truth dataset, we empirically assessed the accuracy of different analytic models 30 for identifying active taxa, and examined how sequencing depth impacts the detection of 31 isotopically labeled genomes. We also demonstrate that using synthetic DNA internal 32 standards to measure absolute genome abundances in SIP density fractions improves 33 estimates of isotopic enrichment. In addition, our study illustrates the utility of internal 34 standards to reveal anomalies in sample handling that could negatively impact SIP 35 metagenomic analyses if left undetected. Finally, we present SIPmg, an R package to 36 facilitate the estimation of absolute abundances and perform statistical analyses for 37 identifying labeled genomes within SIP metagenomic data. This experimentally validated 38 analysis framework strengthens the foundation of DNA-SIP metagenomics as a tool for 39 accurately measuring the in situ activity of environmental microbial populations and 40 assessing their genomic potential. 41 42 Importance: 43 Answering the question of 'who is eating what?' within complex microbial communities is 44 paramount for our ability to model, predict, and modulate microbiomes for improved 45 human and planetary health. This question is often pursued using stable isotope probing 46 to track the incorporation of labeled compounds into cellular DNA during microbial growth. 47 However, with traditional stable isotope methods, it is challenging to establish links 48 between an active microorganism's taxonomic identity and genome composition, while 49 providing quantitative estimates of the microorganism's isotope incorporation rate. Here, 50 we report an experimental and analytical workflow that lays the foundation for improved 51 detection of metabolically active microorganisms and better quantitative estimates of 52 genome-resolved isotope incorporation, which can be used to further refine ecosystem-53 scale models for carbon and nutrient fluxes within microbiomes.

59
The explosion of environmental sequencing data in the last decade has fueled a deeper 60 understanding of the role of microbiomes in shaping human health, ecosystem function, 61 and the Earth's biogeochemical cycles (1). Further advancements in microbiome science 62 require improved experimental approaches that link genomes to their in situ activities. 63 Due to the limitations of culturing techniques, culture-independent methods that reveal in 64 situ functions and link them to taxonomic identities play a crucial role in advancing the 65 field of microbial ecology (2). Stable isotope probing (SIP) is a powerful cultivation-66 independent tool that links metabolic activity and taxonomic identity of environmental 67 microbes (3). During a DNA-SIP experiment, compounds enriched with heavy stable 68 isotopes (e.g., 13 C, 15 N, and 18 O) are added to the microbial community of interest. The 69 labeled compound is metabolized by active members of the microbial community and 70 incorporated into cellular components, including DNA, during growth (4). As a result, the 71 DNA of these active microbes becomes increasingly isotopically labeled, and, therefore, 72 'heavier' compared to the non-labeled DNA from inactive microbes (4). Isotopically-73 labeled DNA, referred to as 'labeled' from hereon, can be physically separated and 74 recovered via isopycnic centrifugation using a CsCl gradient (5). Thus, microbes 75 assimilating labeled compounds in situ can be identified through comparative sequence 76 analysis of the DNA collected at different buoyant densities (BD) along the gradient. 77 Traditional DNA-SIP studies use 16S rRNA gene sequencing to identify labeled 78 microorganisms (6, 7), and several analysis tools are available for 16S rRNA-based SIP 79 studies (8-10). In addition to identifying microbial groups as either labeled or unlabeled, 80 analysis tools such as quantitative SIP (qSIP) and delta BD (ΔBD) can also estimate the 81 extent of isotope assimilation as atom fraction excess (AFE), which is the increase in the 82 isotopic composition of DNA above background levels (11). Measurements of AFE can 83 inform in situ growth rate estimates for specific microbial populations, enabling modeling 84 of microbiome dynamics (12)(13)(14). Although 16S rRNA-based SIP analyses can 85 taxonomically classify labeled microbes, the full genomic potential of metabolically active 86 taxa are not always captured due to the difficulty in linking partial 16S rRNA gene 87 sequences to their corresponding genomes (15). Adapting SIP analysis tools for the 88 genomic level rather than the 16S rRNA gene level would enable genome-centric 89 metagenomic SIP studies and establish stronger links between genomic information and 90 in situ activity. 91 In recent years, multiple SIP studies have used metagenome sequencing in 92 addition to, or in place of, 16S rRNA gene amplicon sequencing (16)(17)(18)(19)(20)(21). We refer to this 93 general approach as "SIP metagenomics" from here on to distinguish it from traditional 94 16S rRNA-based DNA-SIP. Some recent studies have applied the qSIP approach to 95 shotgun sequencing data to estimate the isotopic enrichment of soil metagenome 96 assembled genomes (MAGs) (22)(23)(24). While these represent exciting advancements in 97 the field, SIP metagenomics faces challenges related to data analysis and interpretation. 98 For example, estimates of isotopic enrichment depend on accurate measurements of 99 absolute genome abundance, but determining genome abundance from metagenomic 100 data is difficult due to its compositional nature (25-28). In addition, outstanding questions 101 remain regarding optimal assembly strategies and the specificity and sensitivity of 102 analysis tools given varying sequencing depth and genome coverage. Empirically 103 answering these questions requires a defined experiment where the identity of labeled 104 genomes and their level of isotopic enrichment is known a priori. To date, no such 105 empirical study for validating SIP metagenomic sample processing and analysis has been 106 published. 107 Here, we explore SIP metagenomic sample processing and analysis strategies 108 using a designed microbiome where the identity of labeled genomes, and their level of 109 enrichment, were experimentally controlled. We also investigated the utility of adding 110 internal standards to monitor the quality of density gradient separations and normalize 111 genome coverage levels. With this experimental design, we were able to: a) compare 112 assembly methods for optimal genome recovery; b) determine how sequencing depth and 113 genome coverage influence the detection of labeled genomes; c) examine how different 114 approaches for measuring genome abundance impact estimates of AFE; and d) compare 115 the sensitivity and specificity of different SIP analysis tools for accurately identifying 116 labeled genomes. Based on our findings, we describe an experimentally validated 117 strategy for SIP metagenomics and provide an R package (SIPmg) that adapts SIP 118 analysis tools for shotgun metagenome sequence data, estimates absolute genome 119 abundance within each fraction using internal standards, and identifies labeled genomes. 120 121

122
To create a ground truth dataset for assessing SIP metagenomics, we generated a 123 microbial community DNA sample where the identity of labeled genomes and their level 124 of enrichment were known a priori (Fig. 1). Specifically, we combined unlabeled DNA 125 extracted from a freshwater pond with aliquots of 13 C-labeled E. coli DNA. We created 126 eight levels of E. coli labeling ranging from 0 to 36 atom% 13 C enrichment (Table S1). We 127 also added two sets of synthetic DNA oligos at two different stages of sample processing 128 to serve as internal standards (Fig. 1). The six "pre-centrifugation spike-in" standards had 129 different BDs, each reaching maximum abundance in a different and predictable region 130 of the density gradient (Table S2). Deviations from the expected distribution pattern 131 indicated possible problems, such as a disturbance of the density gradient, that might 132 compromise data quality from that sample (Fig. 2). The post-fractionation spike-ins, 133 referred to as "sequins'' hereafter (28) (Data Set S1), were added to each fraction after 134 density separation (Fig. 1)  incorporators, including qSIP (30), high-resolution SIP (HR-SIP, (8)), and moving-window 146 high-resolution SIP (MW-HR-SIP, (9)). SIPmg also implements a version of the ΔBD 147 method for estimating isotopic enrichment levels (8). To take advantage of metagenomic 148 data, and similar to Greenlon et al. (23), SIPmg updates the qSIP model to use the 149 observed GC content of assembled genomes rather than the estimated GC content used 150 in qSIP analysis of 16S rRNA data (30). Finally, to correct for multiple comparisons, i.e. 151 testing for significant isotope enrichment in multiple MAGs, SIPmg can adjust the 152 confidence intervals around bootstrapped estimates of AFE using a variation of false 153 discovery rate correction (31). With the SIPmg package, we evaluated the performance 154 of different analysis approaches using our ground truth SIP metagenomics dataset. In contrast to a typical metagenome sample, community DNA in a SIP experiment is 159 separated into multiple fractions based on BD prior to sequencing (Fig. 1). Differences in 160 GC content and levels of isotopic enrichment result in a non-random distribution of 161 microbial genomes across the density gradient and sequencing each density fraction 162 provides multiple options for assembly and binning. To determine the optimal strategy for 163 maximizing MAG recovery, we compared assembly of the intact unfractionated sample, 164 separate assemblies of each individual fraction, co-assembly of all fractions derived from 165 the same initial sample, and a massive combined assembly using MetaHipMer (32) of all 166 fractions from all samples. Each assembly was then independently binned using 167 MetaBAT2 (33). A total of 2,022 MAGs were generated across all assemblies, of which 168 248 were high-quality, 447 were medium-quality, and 1,327 were low-quality as defined 169 by the MIMAG reporting standards (34) (Data Set S2). The MetaHipMer assembly 170 produced more MAGs than any other strategy. A total of 235 MAGs were recovered from 171 the MetaHipMer assembly, of which 136 were medium-or high-quality (Fig. 3A). 172 However, estimates of average MAG completeness and contamination for each assembly 173 type were not substantially different (Fig. S1). 174 Next, we deduplicated all the medium-and high-quality MAGs recovered from all 175 assemblies to determine whether any approach generated unique MAGs that were not 176 present in other assembly types (Fig. 2B). We first grouped MAGs with average 177 nucleotide identities of ≥ 96.5 and alignment fractions of ≥ 30% into a total of 148 unique 178 clusters (35), then selected a single representative MAG for each cluster. Of these, 120 179 MAG clusters were exclusively produced by MetaHipMer. Twelve MAG clusters did not 180 include any MetaHipMer-generated MAGs, and 11 of these clusters contained at least 181 one MAG generated from the assemblies of individual fractions (Fig. 3B). Assembly of 182 the intact unfractionated mock microbiome did not produce any unique MAGs (Fig. 3B). 183 The different assembly strategies also produced MAGs with different taxonomic 184 compositions. For example, MAGs derived from the MetaHipMer assembly accounted for 185 an additional nine classes that were not present in other assemblies (e.g., Anaerolineae, 186 Andersenbacteria, Babeliae, Chlamydiia, among others) (Fig. 3C). Most MAGs that were 187 unique to the MetaHipMer co-assembly had lower coverage than MAGs recovered by 188 other assembly approaches (Fig. S2). This suggests the MetaHipMer co-assembly 189 captured more of the lower abundance MAGs in the samples than other assembly Anomalous sample detection using pre-centrifugation spike-in controls 196 As part of the quality control process, we devised an approach for detecting anomalous 197 samples whose pre-centrifugation spike-in sequences displayed aberrant distributions 198 along the BD gradient (Fig. 2C). We added six synthetic spike-ins to our samples prior to 199 ultracentrifugation, and each spike-in had a different density based either on its GC 200 content or the artificial introduction of 13 C-labeled nucleotides during oligo synthesis 201 (Table S2); therefore, each spike-in has a distinct and predictable peak in coverage along Normalizing genome coverage to quantify DNA isotope incorporation 213 Accurate abundance measurements are critical for determining levels of isotopic labeling. 214 Briefly, models such as qSIP and ΔBD estimate a taxon's AFE based on differences 215 between its weighted BD in unlabeled controls and isotope-amended treatments (8, 30) 216 (36), and weighted BD is calculated from the taxon's abundance within each density 217 fraction (see Methods equations 5 & 6). For amplicon-based qSIP studies, the relative 218 abundance of a taxon is normalized to the total number 16S rRNA gene sequences within  Table 1, Table S3). The two approaches using total DNA 248 concentrations did not produce statistically significant linear regressions (p-value > 0.05) 249 between expected and estimated AFEs (Fig. 4B, 4C, Table S3), although the sensitivity 250 for detecting labeled E. coli was the same or better than sensitivity using relative coverage 251 (Table 1). Relative coverage produced the highest specificity, although it had lower 252 sensitivity than the normalization approach using sequins ( Fig. 4D  In addition to qSIP, other analysis methods such as ΔBD (8), high-resolution SIP (HR-258 SIP, (8)), and moving-window high-resolution SIP (MW-HR-SIP, (9)) can identify labeled 259 taxa. We compared all four approaches for their ability to accurately identify isotope 260 incorporators in our defined SIP metagenomic dataset. We also compared estimates of 261 E. coli AFE predicted with the ΔBD and qSIP methods; HR-SIP and MW-HR-SIP do not 262 provide quantitative estimates of enrichment. For all methods, absolute genome 263 abundances were determined by normalization to sequins. 264 The qSIP method predicted the level of AFE for E. coli with greater accuracy than 265 the ΔBD method (Fig. 5). The qSIP approach also had higher specificity than the ΔBD 266 method, producing only 7 false positives across all conditions compared to 12 false 267 positives, respectively (Table S4). The MW-HR-SIP approach had the fewest false 268 positives, with only 4 across all conditions, while maintaining the same sensitivity as the 269 qSIP method (Table S4). The sensitivity and specificity of HR-SIP were lower than both 270 MW-HR-SIP and qSIP methods (Table S4) which an AFE could be accurately estimated, we performed qSIP and MW-HR-SIP 282 analyses after subsampling E. coli reads to 10%, 1%, 0.1%, 0.01%, and 0.001% of their 283 initial levels (Table S5). In the respective subsampled datasets, E. coli had an average 284 total coverage ranging from 0.01X to 1,400X coverage. Here, 'total coverage' refers to 285 the cumulative coverage across all density fractions of an individual sample. 286 The qSIP model consistently identified E. coli as labeled when mean total coverage 287 was ≥ 1X (Table S6). The correlation coefficient between actual and predicted AFEs was 288 0.8 within this coverage range (p-value <0.05; Fig. S6 and Table S7). However, at total 289 coverages <1X, qSIP failed to detect E. coli as labeled in several experimental conditions, 290 and the predicted AFEs were not significantly correlated to the expected AFEs (p-value 291 > 0.05) ( Fig. S6 and Table S7). The MW-HR-SIP method was also less sensitive at lower 292 coverage levels, and at 100X mean total coverage, it only detected E. coli as labeled in 3 293 out of 7 experimental conditions (Table S6). These data suggest that estimates of isotope 294 enrichment are less reliable in general when genome coverage is low. MAGs analyzed to 68. We did not test coverage limits for MW-HR-SIP because the 307 method struggled to detect E. coli as labeled when coverage dropped below 100X (Table   308 S6) and applying a threshold of 100X would have limited our analysis to only 17 genomes 309 (Table S8). These results suggest that excluding genomes with low coverage can 310 decrease false positives and increase balanced accuracy. Although the definition of "low 311 coverage" will vary based on experimental conditions and individual assessments of the 312 tradeoffs between sensitivity and specificity, these results also suggest that confidence in 313 the identification of labeled genomes should decrease along with their coverage levels. 314 We also investigated if false positives could be reduced by implementing a 315 minimum level of isotopic enrichment necessary for a genome to be considered labeled. 316 That is, rather than simply requiring genomes to be significantly greater than 0% AFE, 317 which is the default setting of the qSIP approach (30), we examined different minimum 318 AFE thresholds ranging from 2% to 12.5% (Table S9). A genome was considered to be 319 labeled if the lower bound of its AFE 95% CI was greater than the minimum AFE 320 threshold. With AFE thresholds between 2% and 6%, total false positives dropped from 321 7 to 3 across all experimental treatments, but E. coli was no longer identified as labeled 322 in one experimental condition. The balanced accuracy was also reduced from 0.925 323 without AFE thresholds to 0.856 with a 6% AFE threshold (Table S9). False positives 324 were completely eliminated with a minimum AFE threshold of 12.5%, but sensitivity was 325 so poor (0.286) that E. coli was only identified as labeled in 2 out of 7 conditions (Table   326   S9). Minimum AFE limits could not be tested with MW-HR-SIP analysis because this 327 method does not estimate levels of isotopic enrichment. Together, these results illustrate 328 a trade-off between sensitivity and specificity when increasing the minimum AFE 329 threshold above zero, and suggest that false positives can be reduced by increasing the 330 AFE threshold at the potential cost of losing sensitivity for the detection of minimally 331 labeled taxa. 332 The number and identity of false positives varied across SIP analysis methods, 333 presumably due to differences in their underlying algorithms. Therefore, we hypothesized 334 that the number of false positives might be reduced by taking the consensus of different 335 analysis methods, i.e. requiring that two separate models predict a MAG is labeled. All 336 false positive MAGs found in qSIP analysis were also false positives in ΔBD analysis, and 337 thus taking the consensus of these two methods did not produce fewer false positives 338 than qSIP alone (Table S10). In contrast, there was no overlap in the identity of false 339 positive MAGs between the qSIP and MW-HR-SIP methods, and a union of their results 340 completely eliminated false positives without producing any false negatives (Table S10). 341 However, we found it more advantageous to apply MW-HR-SIP and qSIP sequentially 342 rather than independently. MW-HR-SIP had greater specificity than qSIP, therefore it was 343 used as a first-pass filter to detect putatively labeled genomes while minimizing false Here, we found that co-assembly of all 381 density fractions generated the most medium-and high-quality MAGs, which agrees with 382 two recent SIP metagenomics studies (23, 24). However, we also found that merging 383 binning results from individual fraction assemblies and larger co-assemblies via MAG de-384 replication provided more medium-and high-quality MAGs than did co-assembly alone. 385 We posit that this approach reaps the benefits of both strategies: it provides higher read 386 recruitment for assembling rare genomes in co-assemblies and also leverages lower 387 microdiversity in individual fraction assemblies. Optimal assembly strategies may differ 388 for other environmental samples, and these strategies must be re-evaluated as 389 sequencing and assembly methods evolve, but our results suggest that SIP metagenomic 390 studies can benefit from employing multiple assembly approaches to maximize genome 391 recovery. Based on these findings, we hypothesized that adding internal standards to density 420 fractions ("sequins") could improve abundance measurements and thereby improve 421 isotope enrichment measurements. Indeed, estimates of AFE in our study were more 422 accurate using absolute abundances derived from sequin normalization compared to AFE 423 estimates using other strategies. 424 Multiple factors could explain the more accurate estimates of isotopic labeling 425 enabled by internal quantification standards. For one, sequins may have mitigated any 426 variation introduced during library creation and sequencing (28) The various SIP analysis methods examined in this study use different approaches 443 to detect labeled microorganisms, and these differences could impact the sensitivity and 444 specificity of their predictions. The accuracy of different SIP analysis methods has not 445 been assessed with metagenomic data until now, but in silico simulations of 16S rRNA-446 based SIP data revealed that MW-HR-SIP had higher balanced accuracy than the other 447 analysis methods (29). The qSIP model also generated more accurate AFE estimates 448 than the ΔBD method in those simulations. We observed similar patterns by comparing 449 analysis methods using our experimentally-designed SIP microbiome. In addition, we 450 found that the consensus of multiple approaches, i.e., MW-HR-SIP and qSIP, produced 451 higher accuracy results than any single method alone. Future SIP metagenomic studies 452 might increase confidence in identifying isotope-incorporating taxa by employing these 453 two independent strategies, although the higher confidence in true positives might come 454 at the cost of missing labeled genomes with lower coverage. Regardless of the analysis 455 tools used, analyzing more biological replicates is another simple strategy to increase 456 accuracy (41). As SIP analysis methods evolve, reassessing their performance with 457 deeper sequencing, more replicates, and an improved mock microbiome (e.g. more 458 species at different AFE levels) will provide additional insights into their accuracy and 459 limitations. 460 Altogether, we used a first-of-its-kind mock SIP metagenome to assess the 461 performance of different analysis approaches, identified a set of current best practices, 462 and established an experimentally validated workflow for SIP metagenomics. The 'wet-463 lab' aspects of the workflow include the addition of pre-centrifugation spike-ins for quality 464 control and post-fractionation sequins for genome quantitation along the BD gradient. The 465 'dry-lab' aspects entail absolute genome normalization in each density fraction, and a 466 modified qSIP model tailored to handle genome-resolved metagenomic datasets to 467 calculate AFE. We also explored strategies to more accurately identify isotope 468 incorporators, such as limiting analysis to taxa with coverage and isotope enrichment 469 levels above minimal thresholds and using the consensus of multiple SIP analysis tools 470 to detect labeling using our newly developed SIPmg package. These additional strategies 471 hold promise for improving the accuracy of SIP metagenomic results, although the 472 specifics of how and when to apply them will depend on the study design and individual 473 preferences regarding the tradeoffs between specificity and sensitivity. We believe this 474 validated analysis workflow will increase the reliability of SIP metagenomic findings, 475 enable standardization across studies, and facilitate the use of SIP data in modeling 476 microbially-mediated processes.  Table S1. These fragments were approximately 2 kbp in length with GC content of 37-63% (Table   507 S2). To change the distribution of fragments across the density gradient, some fragments 508 were artificially enriched with 13 C through PCR by adjusting the ratio of unlabeled dNTPs 509 and uniformly-labeled 13 C dNTPs (Silantes Gmhb; 120106100; >98 atom %) (Table S2).  were pooled at equal molar concentrations within the range of 400-800 bp, and the pool 557 was size selected to 400-800 bp using a Pippin Prep 1.5% agarose, dye-free, internal 558 marker gel cassette (Sage Science). For each library, 2X150 bp paired-end sequencing 559 was performed on the Illumina Novaseq platform using S4 flowcells (Table S7). Quality control of SIP data using pre-centrifugation spike-ins 596 Before performing SIP analysis, we first removed mishandled samples from our dataset.

597
For this purpose, we identified the peak of absolute concentration distributions across the 598 density gradient for each labeled pre-centrifugation spike-in. If the spike-in distribution 599 patterns did not match the expected order along the density based on the theoretical 600 estimated density of the spike-in (given its GC content and C 13 /C 12 ratio), then the sample 601 was considered potentially problematic and removed from the analysis. regression analyses, we first tested both ordinary least squares regression and robust 616 linear regression. When using ordinary least squares regression, we also tested Cook's 617 distance filtering to remove outliers at a threshold of Cook's distance < n/4 (n is the 618 number of datapoints in the regression analysis). A coefficient of variation threshold of 619 250 was employed as a quality control step in this scaling process. Due to the lower 620 number of false positives in the approach with ordinary least squares regression 621 combined with Cook's distance filtering, we continued with this approach for all analyses, 622 but also report the findings from using the robust linear regression analysis in the Table   623 S3. A detailed workflow for sequin normalization is provided in the vignette for the SIPmg 624 R package (https://github.com/ZielsLab/SIPmg).

625
In addition to sequin based normalization, we also explored genome abundance 626 estimation using: (1) unscaled coverage; (2)  Birnie and Rickwood (61). The weighted mean BDs were the same as estimated from eq. 663 5. This is a variant of ΔBD from the Pepe-Ranney and colleagues study (8) Table S1. E. coli AFE (%) in each treatment condition. 1090 Table S2. Characteristics of pre-centrifugation spike-ins. To produce distinct distribution patterns 1091 along the density gradient, some spike-ins were artificially enriched with 13 C through PCR by 1092 adjusting the ratio of unlabeled dNTPs and uniformly-labeled 13 C dNTPs. Theoretical AFE values 1093 are reported based on the ratio of labeled dNTPs, but actual AFE values were not experimentally 1094 confirmed. 1095 Table S3. Comparison of various abundance estimation strategies. All results were derived from 1096 the qSIP analysis method. Sensitivity and specificity were averaged across the seven treatment 1097 conditions. 1098 Table S4. Comparison of methods to identify isotopically labeled genomes. Evaluations were 1099 based on absolute genome abundances obtained by normalizing coverage to internal sequin 1100 standards using the sequin approach. Specificity and sensitivity were averaged across the seven 1101 treatment conditions. 1102 Table S5. Average total coverage across all fractions for E. coli in different treatment conditions 1103 after subsampling from 100% to 0.001% of the original E. coli sequence reads. 1104 Table S6. Comparison of MW-HR-SIP and qSIP methods for detecting isotopic labeling of E. coli 1105 at different levels of total genome coverage across the density gradient. 'True' indicates E. coli 1106 was correctly identified to be isotopically labeled (true positive), and 'false' indicates E. coli was 1107 incorrectly identified as unlabeled (false negative). NA corresponds to the failure of the MW-HR-1108 SIP algorithm with that dataset. 1109 Table S7. The impact of genome coverage levels on detecting isotope incorporation using the 1110 qSIP model. 1111 Table S8. Comparison of MAGs retained and the number of false positives detected using the 1112 qSIP method after applying different minimum genome coverage thresholds. MAGs were retained 1113 if their average total coverage in the unlabeled controls exceeded the coverage threshold. E. coli 1114 was the only true positive and had a coverage of 1029X, thus no false negatives were detected 1115 using the coverage thresholds below. 1116 Table S9. Comparison of specificity, sensitivity, and balanced accuracy of the qSIP method after 1117 applying minimum AFE thresholds. To be identified as isotopically labeled, the lower 95% CI 1118 interval of a genome's estimated AFE must be greater than the minimum AFE threshold. 1119 Table S10. Comparison of false positives MAGs identified by the MW-HR-SIP, qSIP, and ΔBD 1120 methods. Names of the false positive MAGs are listed in each column. 1121 Table S11. Comparison of E. coli AFE confidence intervals estimated using qSIP alone, qSIP 1122 after first applying MW-HR-SIP, and qSIP after first applying ΔBD method to identify a subset of 1123 putatively labeled MAGs. Condition B ("20pct_20ng") was removed as it E coli was never 1124 identified as an isotope incorporator in this condition. 1125 expected AFEs obtained using the ΔBD method from (a) raw coverage, (b) relative coverage, (c) 1152 multiplying relative abundance with DNA concentration following Greenlon and colleagues (23), 1153 (d) multiplying relative coverage with DNA concentration following Starr and colleagues (22), (e) 1154 Sequin approach with ordinary least squares regression without Cook's distance filtering (f) 1155