Reference data based insights expand understanding of human metabolomes

The human metabolome has remained largely unknown, with most studies annotating ∼10% of features. In nucleic acid sequencing, annotating transcripts by source has proven essential for understanding gene function. Here we generalize this concept to stool, plasma, urine and other human metabolomes, discovering that food-based annotations increase the interpreted fraction of molecular features 7-fold, providing a general framework for expanding the interpretability of human metabolomic “dark matter.”


Introduction 83
In 2016, typical MS/MS-based untargeted metabolomics studies annotated only ~2% of 84 molecules based on matches against spectral libraries, leaving the rest of the sample as 85 metabolomic "dark matter." The capture of community knowledge, accumulating public reference 86 MS/MS spectra over the past four years, has increased this baseline ~2.5-fold within the global 87 natural product social molecular networking (GNPS) infrastructure (Wang et al., 2016). This 88 growth has been even more dramatic for data from commonly-studied specimen types such as 89 human stool and plasma: 10.1 +/-4.4% of MS/MS features now match to a reference MS/MS 90 spectrum [1% FDR (Scheubert et al., 2017), n = 30, average number of unique MS/MS spectra is 91 12,889/dataset]. However, despite these advances, the vast majority of detectable spectra lack 92 any annotation. 93 This situation for MS/MS spectra is in sharp contrast to the interpretability of 94 uncharacterized portions of the human genome. For example, reference data sets for gene 95 expression, such as expressed sequence tags (an early form of RNASeq), enable the sequencing 96 of "dark matter," as opposed to monitoring the expression of a single curated gene. Such methods 97 have significantly improved interpretation by annotating genes not directly by function, but rather 98 by source (developmental stage, tissue location, organism-level, phenotype, etc.) (Bono, 2020; 99 Ono et al., 2017). Interpretation based on source has been very important for metagenomics and 100 metatranscriptomics, increasing our understanding of the structure and function of complex 101 communities by leveraging matches between genes or transcripts of known and unknown origin 102 via publicly available databases. 103 Annotation of chemicals, based on their source within publicly available complex reference 104 samples that use controlled metadata vocabularies, has not been applied to metabolomics for 105 several reasons. First, standards for annotation of molecules that are used to create spectral 106 libraries have been based on availability of individual pure, typically commercially available, 107 standards, and structural considerations such as presence of specific moieties. Many molecules 108 are observed as multiple different ion forms, such as adducts, in-source fragments, and 109 multimers. Current spectral libraries do not contain all possible ion forms of those molecules, and 110 typically only the protonated form (Schmid et al., 2020;Vinaixa et al., 2016), because reference 111 standards that run in a highly purified state that biases towards detection extraction of only specific 112 data on specific ion forms. These forms are often different from the ions associated with the same 113 molecule present in an extract from a biological matrix (e.g. proton vs sodium or even multiple 114 sodium and potassium adducts), which then cannot be matched because the relevant spectra are 115 not in the database. Second, on average, 5-10% of untargeted metabolomics data can be 116 annotated from spectral libraries: the remaining 90+% are unassignable "dark matter" in 117 metabolomics, especially when obtained from complex matrices such as human samples. Third, 118 large databases of untargeted metabolomics data with consistently annotated provenance with 119 controlled vocabularies have been neither available nor possible to effectively reuse. We recently 120 addressed this latter problem via GNPS (Wang et al., 2016), ReDU (Jarmusch et al., 2019), 121 importing data from MetaboLights into GNPS (Haug et al., 2020), with ReDU-compatible 122 metadata conversion. Finally, the availability of robust scalable analysis infrastructures and 123 algorithms, such as molecular networking, that enable the functional equivalent of reporting of 124 4/33 expressed sequence tag/RNASeq analysis, have only recently been introduced for mass 125 spectrometry (Wang et al., 2016;Watrous et al., 2012). 126 To improve interpretation of otherwise unannotated data from untargeted mass 127 spectrometry experiments, we leverage entire reference data sets with curated ontologies to 128 complement existing spectral libraries of individual molecules. Due to lack of a better term we 129 refer to this approach as interpretive metabolomics in this manuscript, and demonstrate its 130 potential by leveraging the Global FoodOmics MS/MS spectral database, which we have made 131 publicly available on MassIVE. This food reference data set will be key for enabling future insights 132 into human health given the importance of diet and the urgent need to develop additional methods 133 for empirical nutrient and diet assessments to understand acute and chronic human disease 134 (Barabási et al., 2020). We demonstrate that interpretive metabolomics can address these types 135 of knowledge gaps by showing that it not only massively expands the fraction of the data that can 136 be interpreted, but that these new insights can lead to an improved understanding of the diets 137 consumed upon co-analysis of human and food/beverage mass spectral data. 138 Results/Discussion 139 We conjectured that a major source of chemicals detected by metabolomics in human samples 140 originates from foods and beverages. We total, we report 157 metadata categories that further include a six-level food ontology, as well as 151 fermentation or organic status, land or aquatic origin, country of origin, etc. (Table S1). Foods 152 and beverages in Global FoodOmics consist of a range of items, from simple ingredients to 153 prepared meals, as well as animal feed. 154 A key benefit of interpretive metabolomics is that we consider all different ion forms 155 encountered while collecting the Global FoodOmics dataset. Within the GNPS environment, the community can also add tags to each reference 164 spectrum in the spectral library using a controlled vocabulary, including multiple per structure. An 165 InChIKey was included for 4586 of 5455 spectral matches against the reference libraries (~5% 166 5/33 annotation rate at 1% FDR), which yielded 1492 unique structures upon consideration of planar 167 structures. There were 415/1492 structures that had lifestyle tags and "food consumption" is the 168 most frequently reported with 357 entries (86%) (Figure S1a) (Bouslimani et al., 2016). Brief  169  descriptive tags provide more detail about the annotation itself, and 1131/1492 structures were  170 annotated with such tags. The most common descriptive tags were in order: "natural product" 171 (790/1131), "food" (576/1131), "human", "plant", "natural product_plant", "plant_angiospermae", 172 and "drug" (Figure S1b). Some of these associations with the category "human" may also be of 173 food origin, such as arachidonoyl carnitine, which is currently only tagged as "human," but may 174 have a variety of animal-product based food sources. Similarly, the tag "drug" includes 175 annotations such as the antimicrobial agent monensin, which is not tagged as a food molecule, 176 but is consumed with animal products from animals raised using monensin as a growth promoter. 177 Thus the Global FoodOmics reference data capture not only inherently food-derived molecules, 178 but also food-sourced exogenous compounds such as preservatives, growth enhancing 179 substances, antimicrobials, pesticides, and packaging materials. However, because the 180 annotation rate remains low, most of the data remains unused despite the informative tags. 181 In addition to annotating molecules based on matches to library spectra, spectral matches 182 to the food reference data can be obtained and visualized using MS/MS based molecular 183 networking. When applying this method to both foods and biospecimens in an experimental sleep 184 restriction and circadian misalignment study we observed connectivity of nodes within molecular 185 families representing MS/MS spectra (Figure 1a,b). Using spectral libraries the tomatidine 186 molecular family was shown to contain both annotated nodes (level 2 or 3, according to the 2007 187 metabolomics standards initiative (Sumner et al., 2007) e.g, tomatidine, solasodine and 188 sarsasapogenin (Figure 1b), as well as unannotated nodes, which are also observed with 189 molecules occurring within Nightshade (Solanaceae) samples from the Global FoodOmics data 190 set (Figure 1c). Sarsasapogenin (Figure 1c, node 1) is found in food as well as stool data while 191 the +15.996 Da, the addition of the atom "O", is only observed in stool data. However, numerous 192 other molecular families (such as Figure 1c, node 10) contain no annotation, but do have spectral 193 matches between plasma and foodsin this case features also observed in grape and fermented 194 grape samples. In other cases, a plasma metabolite is annotated and connected to unannotated 195 compounds found within the food reference samples (Figure 1c, nodes 11-14). These examples 196 highlight how molecular networking can be used to propagate potential metabolism. How potential 197 metabolism can be inferred with molecular networking is explained in (Quinn et al., 2017) and 198 (Aron et al., 2020). 199 A critical aspect of being able to leverage the food reference data, akin to expressed 200 sequence tags, is that the associated metadata can be retrieved and organized. We leverage the 201 Global FoodOmics ontology to identify different food categories in which MS/MS spectra are 202 observed. These food counts can be summarized for a dataset and then visualized as a flow chart 203 (Figure 1d). Due to the controlled research diets of the participants of the sleep and circadian 204 study in Figure 1d, we were able to report if a given food category was consumed during the 205 study. Of the 15 categories observed at level 5 of the food ontology, 8 represented direct matches, 206 3 represented fermented counterparts of consumed foods (such as yogurt and fermented grapes 207 when milk and grapes were consumed), and 4 categories were not documented to be consumed, 208 while coffee and tea were not provided to participants during this study. By and large, consistent 209 with the lack of consumption of caffeinated beverages, evidence of coffee or tea consumption 210 6/33 was only observed in two individuals. In one individual, caffeine was only detected in the first 48 211 hrs, and in the other volunteer, caffeine was observed in a single time point in the later part of the 212 study (second to last time point). Spectral matches to caffeine were not detected in any of the 213 other participants. Thus, the empirically-recovered food ontology information from metabolomics 214 data demonstrates that these matches are consistent with the food that was consumed in this 215 study. 216 217 218 Figure 1. The concept of interpretive metabolomics leveraging reference data sets. a. A schematic 219 overview of human data and reference data (e.g. data from food items) as molecular families from 220 independent data sets that are used in b. b. A schematic representation when reference data is co-221 networked with human metabolomics data. Each node represents a unique MS/MS spectrum. c.

222
Experimentally observed molecular families (sub-networks) generated from the co-analysis of stool (light 223 blue) and plasma (magenta) data from a sleep restriction and circadian misalignment study with the Global

235
To illustrate the broad utility of the Global FoodOmics reference data in enhancing the 236 information gained from untargeted metabolomics, we co-analyzed the Global FoodOmics 237 dataset with 27 human datasets (Table S2; at 1% FDR spectral matching), with the inclusion of 238 additional study specific foods (SSF) where applicable (Figure 1a). These datasets contained 239 between 5 and 2123 samples, represented multiple different biofluids and tissues, and included 240 both adult and pediatric subjects, in conditions ranging from extremely long lived, such as a 241 centenarian-enriched population in the Cilento Blue Zone in Italy, to inflammatory bowel disease, 242 the healthy young adults undergoing experimental sleep restriction and circadian misalignment 243 highlighted in Figure 1  only). Adding in additional information from molecular network connectivity, which can capture 255 metabolized versions of molecules, the fold change of interpretable data increased further to 6.8 256 +/-3.5 fold (Figure 2). The Global FoodOmics reference samples significantly increased the 257 interpretation of various human metabolome samples above the initial annotation rate by 26.8+/-258 3.3% for stool data (P = 2.8e-16, Games-Howell test), 27.5 +/-5.2% for plasma data (P = 0.0040, 259 Games-Howell test) and 41 +/-4.6% for other human data (P = 0.00020, Games-Howell test). 260 Further inclusion of connected nodes, representing potential metabolism via molecular 261 transformations, results in a total increase of 43.7 +/-3.1% (fecal; P = 6.9e-10, Games-Howell 262 test), 51.2 +/-6.9% (plasma; P = 2.8e-06, Games-Howell test), and 58.0 +/-4.2% (human other; 263 P = 1.4e-06, Games-Howell test) percent of MS/MS spectra that can now be leveraged as 264 potentially a direct empirical readout of diet. 265 For 14 of the public datasets, food samples of the region or exact dietary items frequently 266 or exclusively eaten by that particular population were also collected (study specific foods; SSF). 267 SSF and Global FoodOmics reference samples were separately (SSF; GFOP) and jointly (SSF & 268 GFOP) evaluated for changes to the interpretable fraction of MS/MS spectra (Figure 2). For 269 example, adding SSF (n=38) alone increased the percent of interpreted spectra for the 270 centenarian stool data from an initial 5.4% annotation rate against spectral libraries to 20.0% 271 interpreted (Figure 2a) and 4.9% initial to 24.4% for plasma samples (Figure 2b), and adding 272 Global FoodOmics further expanded this to 49.0% (55.0% in plasma). For the sleep restriction 273 and circadian misalignment study highlighted in Figure 1, the interpreted fraction also increased 274 from an initial 7.2% to 27.8% (n=197 food samples; 45 of which are pooled meal samples), with 275 8/33 a further increase to 46.3% when using the Global FoodOmics reference data set (7.8% to 38.9% 276 and with Global FoodOmics up to 54% for plasma). Overall, the inclusion of SSF significantly 277 contributed to the increase in dietary spectral matches in plasma (Figure 2b; P = 0.0028,  Howell test). In addition, in some cohorts the interpreted spectral rate reaches almost 80% after 279 expansion with molecular networking (Figure 2c). 280 To further demonstrate that spectral matching using reference matching reflects dietary 281 components, we performed a crossover study to test whether a mismatched SSF inventory would 282 yield similar results to the increases observed across studies with SSF (e.g. centenarian foods 283 for the sleep and circadian study cohort). Crossover revealed that the reciprocal tests 284 interpretation rates were only a few percent (5-6%) in comparison to when the correct SSF were 285 used (15-30%) (Figure 2d).

292
plasma data, and c. other human biospecimens. d. A crossover experiment between the centenarian data 293 from Italy and the sleep and circadian study from the US, for both fecal and plasma samples. Study specific 294 foods consumed by those individuals (yes) vs a different set of study specific foods (no), (Welch's t-test).

296
As the Global FoodOmics reference database expands with regionally-specific foods 297 through a continued community effort, the interpreted fraction will likely increase. For example, 298 when legume food data (15 files; SSF) similar to legumes supplemented in an infant malnutrition 299 9/33 study were included in addition to the Global FoodOmics data, the number of spectral counts for 300 legumes went from 105 to 2430 unique MS/MS spectra that matched, while other food categories 301 such as dairy and meats remained constant (level 3 food ontology; Legume supplementation, 302 urine). Regional specificity was also directly evident for plasma samples collected in Brazil for a 303 Covid-19 study, which displayed more spectral matches to a locally collected set of 60 Brazilian 304 food samples with ~35% increase than to the entire Global FoodOmics reference dataset, that is 305 dominated by US food, which only gave an ~20% increase in spectral matches (Figure 2b). Thus, 306 although there is some overlap among the data from different foods, and even overlap among 307 human-derived metabolites and the food data (e.g. many primary metabolites or those common 308 in vertebrates), a large proportion are sufficiently unique to reveal, at least in part, the dietary 309 composition in the study. To assess if interpretive metabolomics could be used to empirically establish adherence 325 to dietary recommendations using MS/MS data, we analyzed a data set from rheumatoid arthritis 326 patients (RA) asked to follow an anti-inflammatory diet (ITIS diet) (Bustamante et al., 2020). We 327 compared the per sample extracted food counts with the recommended diet alteration as well as 328 self-reported diet diary entries. The recommended diet included some foods to be avoided (such 329 as coffee, refined sugars and milk), some foods to be restricted (minimize red meat and egg 330 consumption) and some foods to be frequently consumed (such as fruits/vegetables, and plain 331 unsweetened yogurt). In total, 47 foods and beverages were observed in this project with 332 interpretive metabolomics (Figure 3a). By and large, most adhered to the recommended diet, as 333 food counts of recommended foods increased, and those of foods to avoid decreased. Although 334 there are instances when the mass spectrometry based observations did not match the 335 recommended diet regime, the self-reported dietary records matched the empirically determined 336 foods better than the recommended dietary information (Figure 3b). We further validated these 337 matches using source tracking with 16S rRNA gene amplicon data collected on ~1500 samples 338 of the Global FoodOmics foods, to predict food source contribution to the RA study stool samples. 339 We observed a highly significant correlation in the proportion change of food sources predicted in 340 the stool samples and metabolites in the plasma before and after dietary intervention (Pearson 341 r = 0.57, p-value = 0.003; Figure S2). The empirically recovered food ontology information from 342 interpretive metabolomics, in conjunction with validation with DNA sequence data, demonstrates 343 the ability to recapitulate dietary readouts from human biospecimens and assess diet adherence. 344 Interpretive metabolomics comes with several caveats to consider. We are not yet able to 345 capture a complete picture of the human diet: for example, in the RA study, the participant diet 346 diaries contained foods not yet captured in the FoodOmics database, potentially leading to an 347 underestimation of food types observed. Community expansion of the Global FoodOmics 348 database with specific foods and food ingredients will ultimately eliminate this issue. 349 Another consideration is similar to what is observed with expressed sequence 350 tags/RNASeq, where it is common to observe that there are multiple sample types, tissue 351 locations or conditions that result in misinterpretation because the same sequence occurs in 352 multiple locations. By analogy, a molecule could be produced by humans but also be part of 353 different diet sources (i.e. cholesterol produced by the human body versus consumed). However, 354 such matches still enable one to formulate a hypothesis that the observed MS/MS features from 355 the human data might originate from the reference data as source, in this case food, especially 356 when there are hundreds or thousands of signatures that point to specific foods or food groups 357 that overlap. 358 As we saw in many of the above datasets, it is not atypical to observe small numbers of 359 spectral matches to insects, rodents, fungi and worms within diet read-outs. Although data on 360 fungi, tarantula, crickets, and black ants, meant for human consumption, are included, most of 361 these samples that match human data sets are from a Global FoodOmics sampling effort at the 362 San Diego Zoo. While there is likely some overlap with molecules from these less common foods 363 to those that humans more commonly consume (e.g. certain acylcarnitines might be found in beef 364 and mice), the FDA food contamination guidelines allow for insect, fungal, worm, rodent parts and 365 fecal matter to be present in food in quantities that surprise many non-specialists (Center for Food  366 Safety and Nutrition, 2019) For example, peanut butter is allowed to have 30 or more insect 367 11/33 fragments and one rodent hair per 100 grams, and apple butter is allowed to have "5 or more 368 whole or equivalent insects (not counting mites, aphids, thrips, or scale insects) per 100 grams of 369 apple butter." As long as these dietary "additives" are added to the reference data set, they too 370 will be observed. of the samples (2399/3579)). Primarily for the initial data set these images were used as the first 584 point of reference for the collection of ancillary information about the different samples (termed 585 metadata, described in more detail below). The image archive was critical, because as the project 586 evolved and the breadth of sample types increased, new categories were added to the metadata, 587 which were then filled in weeks or even months after sample collection. 588 Samples were frozen at -80 o C within 24 h of sample collection, unless otherwise noted in 589 the metadata. Two samples were collected for each food or beverage included in the study. One 590 sample was collected as an archive and directly frozen, and a second sample was collected for 591 extraction. Food samples were collected in a tube prefilled with 1 ml 95% ethanol (Ethyl alcohol 592 (Sigma-Aldrich) and Invitrogen UltraPureTM Distilled Water), as high ethanol concentrations are 593 efficacious at preserving the sample for both DNA and metabolite analyses (Song et al., 2016). 594 Samples were collected into 2 ml round bottom microcentrifuge tubes (Qiagen) and weighed prior 595 to freezing. The pre-sample and post-sample weights as well as the weight differences were 596 recorded in the metadata. It was not possible to collect all samples at a given concentration of 597 extraction solvent (ethanol), because sampling was performed in many different environments 598 and is meant to be consistent with future crowd-based community science participation. 599 Therefore, the data can be compared qualitatively and not quantitatively, however for certain 600 subsets 50 mg of material was collected consistently. 601 Additional sets of food samples were added to the core set using the same methods as 602 outlined above when possible. Samples from Venezuela were collected whole in absolute ethanol 603 >=99.8% (Sigma Aldrich) and the extract was processed directly. 604 The experimental protocol for the sleep restriction and circadian misalignment study has 605 been described previously ( Community-based science collection 620 The first sample collected was a carrot from a home garden. The participant was interested in 621 how the soil conditions from prior tenants would impact the chemistry of the carrot, since the 622 gardening practices of the prior tenant were unknown (organic or not, pesticide usage, etc.). In 623 addition, home grown foods often taste different than store bought, likely reflected in the food 624 metabolome. 625 During the course of sampling, samples were received from over 50 different individuals 626 in California as well as from different states as well as countries (such as Venezuela, Italy and 627 most recently Brazil). Contributions from individuals ranged from produce from home gardens, 628 home fermented products (yogurt, kombucha, sauerkraut), meat and dairy from private farms, to 629 items individuals had purchased that were of interest to them. 630 We were also directly invited to sample at local stores and organizations, including 631 Venissimo cheese, Good Neighbor Gardens, and the San Diego Zoo and San Diego Zoo Safari 632 Park, as well as local supermarkets such as Sprouts Farmers Market, Whole Foods Market, and 633 Ralphs. We were invited by San Diego Fermenter's Club founder Austin Durant to the San Diego 634 Fermenter's Club meeting and sampled from multiple vendors at both the Oregon Fermentation 635 Festival in 2017 as well as the San Diego Fermentation Festival in 2018. We also received citrus 636 samples from a farm at the US-Mexico border, with visibly dark skin due to air pollution, a 637 particular concern of the farmer. Other sampling occurred in conjunction with study design, as 638 was the case for the Rheumatoid arthritis cohort and the Covid-19 study. In total we engaged with 639 a broad range of individuals, organizations, businesses and scientists, to generate this dataset of 640 3579 samples (for future use this is already expanded beyond this number due to the collection 641 of sets of SSF). A predominance of foods included in this initial dataset were sampled and/or 642 purchased in California, leaving room for much further expansion and the inclusion of a crowd-643 sourced community science initiative to expand the array of samples. 644 The sample set contains a broad set of simple foods including fruits, vegetables, grains, 645 as well as raw meat and fish, which build the foundation of many food products. In addition, we 646 have 1133 fermented samples. This subcategorization of foods is made possible by the metadata 647 collected on these samples, described in the Metadata Curation section. we generated a specific complex food ontology based on the known presence of common 704 categories (corn, dairy*, egg*, fruit, fungi, fish*, shellfish*, meat, peanut*, seaweed, soy*, tree nut*, 705 vegetable/herb, wheat* (*designates known food allergen)). These categories reflect the main 706 food groups and some of the most common allergens (US FDA Food Allergen Labeling And 707 Consumer Protection Act of 2004) (Sicherer and Sampson, 2006), items which are of interest 708 when correlating food metabolome data with other datasets, such as human fecal material (where 709 the foods eaten are known or unknown). 710

Fermented foods 711
Preservation and processing method are included in the metadata. However, due to the potential 712 importance of fermentation in the alteration of the food metabolome, and the potential health 713 benefits that have been ascribed to fermented foods, several categories were included to highlight 714 this feature: fermented or not, whether it contains live active cultures, whether it contains 715 chocolate (which then was cross checked with the fermented category, as chocolate is a 716 fermented food). The list of fermented foods crosses many of our sample types as it includes 717 fermented dairy (yogurt, cheese), fermented meat/fish (salami, fish sauce), fermented vegetables 718 (kimchi, sauerkraut), fermented fruit (chocolate, coffee), and fermented grains/legumes (bread, 719 tempeh). 720

Food specific categories 721
Certain individual food categories also necessitated creation of specific categorization.

Sample Preparation 737
A sterile stainless steel bead was added to each sample collected in ethanol and the samples 738 were thawed on ice for 30 min. Samples were homogenized at 25-30 Hz for 5 min using a tissue 739 homogenizer (QIAGEN TissueLyzer II, Hilden, Germany). Samples were swabbed with sterile 740 dual tip swabs (BD swubes) and frozen immediately at -80°C until DNA extraction.  The Deblur 150-bp observation table  768 consisting of 1511 food samples was used as the set of source environments and the Rheumatoid 769 Arthritis (RA) data set consisting of 49 fecal samples was used as the sink. All source and sink 770 samples were rarefied to 2000 sequences per sample before source-tracking and doubleton 771 ASVs were removed. Leave-one-out cross-validation was used to predict the source samples with 772 heterogeneity from all other food categories. After source sample filtering a total of 346 samples 773 representing a total of 25 broad food categories were retained. Food microbial source 774 contributions were then predicted for RA samples and the difference in food contribution before 775 and after diet intervention was calculated and compared by diet recommendations. Germany) and transferred to a 96-well shallow well plate, and diluted either 5x or 10x to avoid 785 saturating the MS detector. 786 Liquid Chromatography -Mass Spectrometry 787 Food extracts were analyzed using a UltiMate 3000 UHPLC system (Thermo Scientific, Waltham, 788 Ma) equipped with a reverse phase C18 column, prepended with a guard cartridge (Kinetex, 100 789 x 2.1 mm, 1.7 μm particles size, 100 Å pore size; Phenomenex, Torrance, CA, USA), at a column 790 compartment temperature of 40°C. Samples were chromatographically separated with a constant 791 flow rate of 0.5 ml / min using the following gradient:  consensus spectra that contained less than 2 spectra were discarded. A network was then 825 created where edges where filtered to have a cosine score above 0.65 (slight variation per study 826 based on FDR calculation) and more than 5 matched peaks. Further edges between two nodes 827 were kept in the network if and only if each of the nodes appeared in each other's respective top 828 10 most similar nodes. The spectra in the network were then searched against the GNPS spectral 829 libraries. The library spectra were filtered in the same manner as the input data. All matches kept 830 between network spectra and library spectra were required to have the same cosine score and 831 minimum matched peaks as for library search. Version release 18 was used to process all studies 832 with the exception of the Covid-19 dataset, which was processed with identical methods and 833 version 23. 834 Molecular networks were visualized in the GNPS browser as well as with the freely 835 available program Cytoscape (version 3.5.1) (Shannon et al., 2003). 836 Interpreted spectral rate calculation 837 The levels of interpretation are delineated as follows: A spectral match between an MS/MS 838 spectrum from human or food data with a library spectrum constitutes a molecular ID and 839 determines the initial percent of interpreted spectra, which is also equivalent to the annotation 840 rate of the dataset. A spectral match between MS/MS spectra in human and reference samples 841 (by performing molecular networking of the datasets together and identifying nodes with overlap 842 between the two groups) indicates a potential source. Matches between human and food data 843 therefore implicate food as the potential source of the molecule. Food reference data are referred 844 to in two main categories: the Global FoodOmics dataset (GFOP; broad range of foods and 845 23/33 beverages) and study specific food (SSF; foods and/or beverages known to be consumed by 846 some participants). The last level of interpretation is based on connectivity within a molecular 847 family, which allows us to infer structural relatedness or possible metabolism of food derived 848 compounds. 849 Food reference data and human data were organized into separate groups in the 850 molecular networking analysis. The annotation and interpreted spectral rates were calculated 851 using R (3.6.3) and the tidyr and dplyr packages. We first calculated percent annotation rate, or 852 molecular ID, for all studies (stool, plasma, etc.) (i.e. # of stool nodes with a molecular ID / total # 853 of stool nodes). Spectral matches between food reference data and human MS data (overlap 854 between the two groups) provides the next level of information, referred to as the interpreted 855 spectral rate (i.e. # of nodes found in food and stool data / total # of stool nodes), indicating a 856 potential food source. 857 For molecules without annotations to reference libraries, we wanted to measure the 858 potential to explain their presence using molecular networking. By removing single loops in each 859 dataset and comparing metabolites that shared a component index with an annotated compound, 860 we were able to identify molecules that belong to the same molecular family to infer their potential 861 classification, and calculate the interpreted spectral rate by dividing unannotated molecules that 862 network with annotated ones by total metabolites within each sample type. Overlap between 863 sample types was again assessed to understand contributions due to co-networking of molecules 864 across sample types, increasing our ability to explain unannotated molecules found in our 865 datasets. Visualizations were generated using graphics and beeswarm packages, and significant 866 differences were calculated using Welch's t-tests (stats::t.test), Welch's F-test 867 (onewaytests::welch.test), and Games-Howell (rstatix::games_howell_test) for multiple 868 comparisons, as appropriate, with multiple comparisons correction using Tukey's method. All data 869 are expressed as the mean ± standard error and considered significant if P < 0.05 unless 870 otherwise stated. 871 For example, for GNPS molecular networking analyses test datasets were consistently 872 placed in group 1 (G1) (and G2 for paired datasets, such as stool and plasma) and Global 873 FoodOmics data were placed in group 4 (G4). SSF were consistently placed in G3 when used. 874 The common nodes between G1 and G4 represent the overlap and potential enhancement of 875 information, directly from the reference dataset. The improvement is thus measured by the 876 difference in the overlap of G1 and G4 divided by the total nodes in G1 versus the # of annotations 877 in G1 divided by the total nodes in G1. The "propagation" refers to the counting of nodes within 878 connected components in molecular families which capture three types of additional information: 879 1) unannotated compounds found only in G1 that network with an annotated compound found in 880 G4 (could be an annotated molecule observed only in G4 or in G4 and G1), 2) unannotated 881 compounds found only in G1, but in the same molecular family with an unannotated food 882 compound (G4), or 3) unannotated compounds found only in G1, but in the same molecular family 883 with an annotated food compound (G4 in writing, based on the printed text of the Free and Informed Consent Form, which contained the 938 general proposal of the study, the procedures for obtaining the samples, the risks and benefits. 939 In addition, they were assured about confidentiality of their name, personal data and the possibility 940 of giving up their participation at any time. Following the signature, patients received a copy of 941 the informed consent form. The following were included: 1) Patients diagnosed with Covid-19 in 942 moderate, severe or critical forms and in need of hospital treatment; 2) Over 18 years old; 3) At 943 least 50 kg of body weight; 4) Admission electrocardiogram without changes in rhythm and with 944 QT interval <450 ms; 5) normal serum levels of Ca 2+ and K + ; 6) If a woman, between 18 and 50 945 years old, negative β-HCG test on admission. Patients were excluded who: 1) have the mild forms 946 of SARS-CoV-2; 2) pregnant; 3) unable to understand the information contained in the Free and 947 Informed Consent Form (ICF). 948 Sample preparation: For the Covid-19 plasma samples, aliquots of 20 μL were transferred 949 to eppendorf tubes and 120 μL of cold extracting solution, MeOH: MeCN (1: 1, v/v) was added. 950 After orbital shaking for 1 min (Gehaka AV-2 Shaker, São Paulo, Brazil), the samples were left at 951 -20 o C for 30 minutes and then centrifuged for 10 min at 20000 × g at 4 o C (Centrifuge Boeco  952 Germany M-240R, Germany). An aliquot of the organic phase (120 μL) was transferred to another 953 eppendorf tube and evaporated to dryness in a rotary vacuum concentrator for 60 min, at 30 o C 954 (Analitica, Christ RVC2-18, São Paulo). The residues were resuspended in 80 μL of H2O and 955 centrifuged (10 min, 5000 ×g, 4 o C), an aliquot of 5 μL was injected. 956 Mass spectrometry data collection plasma sample extracts were chromatographically 957 separated with anHPLC (Shimadzu, Tokyo, Japan), coupled with a micrOTOF-Q II mass 958 spectrometer ( Figure S1. GNPS tag and GNPS Lifestyle Tag distribution for the Global FoodOmics reference 994 data set (GNPS task ID: f1a1f3a61aca416a9b3687d72488da7f). Annotated MS/MS spectra were 995 assigned planar InChIKeys, and at least one tag. Spectra can be assigned multiple tags, 996 indicating multiple potential sources. 1131 total unique planar InChIKeys with at least one GNPS 997 tag. a. Lifestyle tags and b. GNPS tags. 998 999 1000 1001 1002 1003 Figure S2. Linear regression scatter plot of difference in food contributions for metabolite spectral 1005 match (x-axis) and microbes by source tracking prediction (y-axis) before vs. after diet intervention 1006 compared by diet recommendation of avoid (orange) or include (blue). Correlation evaluated by 1007 Pearson correlation coefficient. 1008