Charting unknown metabolic reactions by mass spectrometry-resolved stable-isotope tracing metabolomics

Abstract

Metabolic reactions play important roles in organisms such as providing energy, transmitting signals, and synthesizing biomacromolecules. Charting unknown metabolic reactions in cells is hindered by limited technologies, restricting the holistic understanding of cellular metabolism. Using mass spectrometry-resolved stable-isotope tracing metabolomics, we develop an isotopologue similarity networking strategy, namely IsoNet, to effectively deduce previously unknown metabolic reactions. The strategy uncovers ~300 previously unknown metabolic reactions in living cells and mice. Specifically, we elaborately chart the metabolic reaction network related to glutathione, unveiling three previously unreported reactions nestled within glutathione metabolism. Among these, a transsulfuration reaction, synthesizing γ-glutamyl-seryl-glycine directly from glutathione, underscores the role of glutathione as a sulfur donor. Functional metabolomics studies systematically characterize biochemical effects of previously unknown reactions in glutathione metabolism, showcasing their diverse functions in regulating cellular metabolism. Overall, these newly uncovered metabolic reactions fill gaps in the metabolic network maps, facilitating exploration of uncharted territories in cellular biochemistry.

Introduction

Metabolic reactions in which metabolites are biosynthesized or degraded, lay the foundation of cellular biochemistry by fulfilling energy to cellular functions, metabolizing nutrients to building blocks for macromolecule biosynthesis, and producing messengers for cellular signaling^1,2,3. In living organisms, metabolites are connected to each other through one or multiple metabolic reactions and form the metabolic reaction network that affords the biochemical architecture of cellular metabolism⁴. Dedicated databases such as Kyoto Encyclopedia of Genes and Genomes (KEGG)⁵ and REACTOME⁶ have collected known metabolic reactions of various organisms, providing valuable resources to understand cellular metabolism. In our previous work, we demonstrated that metabolic reaction network is essential for both known and unknown metabolite discovery^7,8. However, more than 30% of gene sequences are functionally unannotated with a large portion of them being considered to encode metabolic enzymes^9,10. Thus, the metabolic network interwoven with different metabolic reactions is largely incomplete, which restricts our knowledge on the scope of cellular biochemistry.

To this end, both computational and experimental approaches have been developed to expedite the discovery of previously uncharacterized metabolic reactions^11,12. Computational prediction plays a crucial role in prediction of hypothetical metabolic reactions by leveraging existing knowledge of metabolic networks, enzyme functions, and chemical transformations, which includes the homology-based enzyme function annotation¹³, constraint-based genome-scale metabolic modeling¹⁴, and machine learning-based prediction^15,16. While its strength in a high-coverage of predicted reactions, only a very small portion of them can be experimentally validated in organisms of interest. In addition, computational prediction usually confines to known metabolites, which restricts the scope of previously unknown reactions being predicted. As a comparison, in vitro enzyme activity-based metabolomics profiling represents one of the most common experimental technologies for discovering previously unknown metabolic reactions^12,17. For example, Sévin et al. implemented this technology to characterize 1275 uncharacterized proteins in Escherichia coli, and discovered 241 enzymes that catalyze metabolic reactions. Among them, 29 metabolic reactions with 12 enzymes were experimentally validated¹⁷. However, a major drawback of this approach is the low recovery rate with less than 20% of enzymes being successfully characterized, due to the prerequisites of purifying proteins, concentrations, and availabilities of metabolites¹⁷. In addition, non-enzymatic reactions which are integral parts of the metabolic networks with essential roles in cellular metabolism, cannot be captured by the enzyme activity-based metabolomics approaches.

Alternatively, stable-isotope tracing metabolomics involves introducing isotopically labeled substrates (e.g., ¹³C-glucose) into a living organism^18,19,20. This allows for the tracking of the incorporation of stable-isotopes (e.g., ¹³C) into metabolites produced by cellular metabolism²¹. Although in piecemeal ways, this technology has increasingly become the method of choice to verify the newly discovered metabolites and metabolic reactions based on preconception of metabolic routes^22,23. For example, Fox et al. recently used stable-isotope tracing metabolomics with ¹³C₅-valine confirming multiple previously unknown conjugations between 3-hydroxypropionate and amino acids, and discovered a shunt-within-a-shunt metabolic pathway for propionate degradation in Caenorhabditis elegans²⁴. In practice, stable-isotope tracing metabolomics generally requires a prior knowledge of possible metabolite targets, leading to the ab initio elucidation of previously unknown metabolic reactions a great challenge^21,25. Specifically, the implementation of stable-isotope tracing metabolomics with a data-driven and hypothesis-free strategy for discovering unknown metabolites and associated metabolic reactions in one study has not been realized by far. Under stable-isotope tracing experiments, metabolic reactions that take place in a living organism result in specific isotope labeling patterns in metabolite products, termed as isotopologue patterns^20,21,25,26. The substrate and product metabolites within one metabolic reaction usually share similar isotopologue patterns (Fig. 1a), which leaves ample potential for previously unknown reaction discovery. Thus, we reason that previously unknown metabolic reactions can be deduced by comparing the isotopologue pattern similarity between labeled substrate and product metabolites.

Fig. 1: Reaction-paired metabolites tend to share similar isotopologue patterns in stable-isotope tracing metabolomics.

a Schematic illustration of generation of metabolic reaction-paired metabolites and calculation of their isotopologue pattern similarity. b Numbers of labeled metabolites and reaction pairs found in 293T cells labeled with [U-¹³C]-glutamine or [U-¹³C]-glucose for 17 h. c Isotopologue pattern similarity scores (S_ISO) of reaction-paired metabolites (asparagine/aspartate and aspartate/orotate) in 293T cells labeled with [U-¹³C]-glutamine (n = 6 biological replicates per group). Values represent means ± SEM. d Isotopologue pattern similarity scores (S_ISO) of reaction-paired metabolites (glutamate/α-ketoglutarate and α-ketoglutarate/cis-aconitate) in 293T cells labeled with [U-¹³C]-glucose (n = 6 biological replicates per group). Values represent means ± SEM. e Percentages of high isotopologue similarity scores of two metabolites (S_ISO ≥ 0.7) in reaction pairs (RPs) and non-reaction pairs (non-RPs), respectively, in 293T cells labeled with [U-¹³C]-glutamine or [U-¹³C]-glucose. The metabolite pairs with more than 7 steps of reactions in metabolic reaction network were defined as non-reaction pairs. Source data are provided as Source Data files for Fig. 1b–e.

In this work, we develop an approach, namely isotopologue similarity networking (IsoNet), for charting unknown metabolic reactions through mass spectrometry-resolved stable-isotope tracing metabolomics. Using IsoNet, we construct the isotopologue similarity networks for labeled metabolites in live cells and mice, and discover hundreds of putative unknown metabolic reactions. Specifically, we complete the metabolic reaction network associated with glutathione metabolism, which includes 10 metabolic reactions. Most importantly, we uncover a previously uncharacterized transsulfuration reaction by which γ-Glu-Ser-Gly was synthesized directly via transsulfuration of glutathione. These findings of unknown reactions fill the previously unexplored metabolic network, complementing our understanding of the metabolic maps in cellular biochemistry.

Results

Reaction-paired metabolites tend to share similar isotopologue patterns

The basic principle of our method is that the reaction-paired metabolites have higher tendency to be similar in isotopologue patterns in the mass spectrometry-resolved stable-isotope tracing metabolomics. To demonstrate this principle, we first analyzed 293T cell samples labeled with isotopic tracers such as [U-¹³C]-glutamine and [U-¹³C]-glucose using a high-resolution mass spectrometer. For ¹³C-labeled metabolites, we retrieved their metabolic reaction relationship from metabolic reaction network curated using the KEGG database, which results in reaction pairs (RPs) between labeled metabolites (Fig. 1a). Then, we extracted corresponding isotopologues for individual metabolites and characterized the isotopologue pattern similarity for the reaction-paired metabolites (Fig. 1a). To calculate isotopologue similarity scores (S_ISO), we developed a scoring algorithm applicable for three types of metabolic scenarios which are based on the numbers of carbon and labeled isotopologues between metabolite pairs (Supplementary Fig. 1a–d; see “Calculation of isotopologue pattern similarity” section in “Methods”). As a result, we identified 174 ¹³C-labeled metabolites within 199 reaction pairs in 293T cells labeled with [U-¹³C]-glutamine (Fig. 1b and Supplementary Data 1). Likewise, when [U-¹³C]-glucose was used for labeling, 242 ¹³C-labeled metabolites and 445 reaction pairs were characterized in 293T cells (Fig. 1b and Supplementary Data 1). Further examination showed that reaction-paired metabolites, such as asparagine and aspartate, aspartate and orotate, displayed similar isotopologue patterns with S_ISO being calculated as 0.83 and 0.87, respectively (Fig. 1c). This observation was also seen in cells labeled with [U-¹³C]-glucose. For examples, reaction-paired metabolites such as glutamate and α-ketoglutarate, α-ketoglutarate and cis-aconitate in TCA cycle, had isotopologue similarity scores of 0.81 and 0.96, respectively (Fig. 1d).

Next, we extended the isotopologue similarity calculation to all ¹³C-labeled metabolites in RPs and non-RPs (see “Isotopologue similarity between reaction-paired and non-reaction paired metabolites” section in “Methods”). In cells labeled with [U-¹³C]-glutamine, more than 60.7% of labeled metabolites in reaction pairs have an isotopologue similarity score larger than 0.7, while only 18.6% of labeled metabolites in non-reaction pairs met this criterion (Fig. 1e). The results in the [U-¹³C]-glucose labeling experiment showed the similar conclusion wherein ¹³C-labeled metabolites in RPs (45.7%) had higher chances of similar isotopologue patterns than those in non-RPs (14.6%). Closer examination of the relationship between isotopologue similarity and metabolic reaction steps revealed that the isotopologue similarity score decreased as steps of reactions increased, either in [U-¹³C]-glutamine or [U-¹³C]-glucose labeling experiments (Supplementary Fig. 1e). Overall, these data demonstrated that the reaction-paired metabolites in principle tend to share similar isotopologue patterns in stable-isotope tracing metabolomics.

Discovery of previously unknown metabolic reactions through isotopologue similarity networking

We further developed an isotopologue similarity networking approach, namely IsoNet, by combining mass spectrometry-resolved stable-isotope tracing metabolomics and isotopologue similarity networking between labeled metabolites (Fig. 2a). We demonstrated that IsoNet enables to discover previously unknown metabolic reactions in living organisms. The detailed description of the IsoNet algorithm is provided in “The workflow of IsoNet” section of “Methods”. In brief, IsoNet includes four major steps: (1) annotation of labeled metabolites; (2) isotopologue similarity networking; (3) unknown reaction discovery; and (4) elucidation of unknown reactions and structures (Supplementary Fig. 2a–d). To demonstrate the IsoNet approach, we analyzed 293T cells labeled with a mixture of ¹³C-tracers including [U-¹³C]-glutamine, [U-¹³C]-glucose, and [U-¹³C]-acetate. Then, we constructed an isotopologue similarity network containing 841 nodes and 1544 edges (Fig. 2b). Specifically, one node represents one ¹³C-labeled metabolite, while an edge represents that two labeled metabolites shared similar isotopologue patterns and MS/MS spectra (i.e., structural similarity). Among all labeled metabolites, 109 of them were known metabolites, while the rest 732 ones were unknowns (Fig. 2c and Supplementary Data 2).

Fig. 2: Discovery of previously unknown metabolic reactions through isotopologue similarity networking.

a Workflow for the isotopologue similarity networking approach. Red dot represents known metabolites and blue dot represents unknown metabolites. b The isotopologue similarity network in 293T cells labeled with mixed tracers ([U-¹³C]-glutamine, [U-¹³C]-glucose, and [U-¹³C]-acetate) for 17 h. Nodes represent labeled known (red) and unknown (black) metabolites; edges represent that two metabolites shared similar isotopologue patterns and MS/MS spectra. c Numbers of known and unknown labeled metabolites in the network. d Numbers of known and unknown reactions in the network. The reactions between known and unknown metabolites were defined as unknown I. The reactions between unknown metabolites were defined as unknown II. Red dot represents known metabolites and blue dot represents unknown metabolites. e Retrieval of 86 known reactions from KEGG: the recall rate (left) and reaction step distribution (right). f Statistics of reaction classes for unknown reactions. g Top 6 ranked reaction types with indicated atom exchanges in unknown reactions. Source data are provided as Source Data files for Fig. 2c–g.

To elucidate metabolic reactions revealed by the isotopologue similarity networking, we curated a metabolic reaction network, a delta mass library, and a reference reaction library from the KEGG database (Supplementary Fig. 2e and Supplementary Datas 3 and 4). First, 86 reactions were identified as known reactions which were consisted of two known metabolites (Fig. 2d and Supplementary Data 2). To validate these known reactions, we retrieved their metabolic routes from KEGG metabolic reaction network. The results showed that 91% of known reactions (n = 78) were successfully recalled (Fig. 2e). In addition, we revealed that the majority of these reactions were found to take place within 1–3 steps of metabolic reactions (Fig. 2e). For example, an isotopologue similarity subnetwork, which consisted of 15 nodes and 25 edges, uncovered known metabolic reactions among uridine diphosphate (UDP)-glucuronate, UDP-glucose/UDP-galactose, uridine triphosphate (UTP), and cytidine triphosphate (CTP) (Supplementary Fig. 3). These results demonstrated that the isotopologue similarity networking strategy enables to identify metabolic reaction relationships with a high accuracy in living cells.

Next, we matched the experimental mass differences (Δm/z) between paired metabolites in the isotopologue similarity network (n = 1544) with the delta mass library. Each delta mass in the library is associated with atom differences and reaction classes (Supplementary Fig. 2e and Supplementary Data 3). Thus, we successfully annotated a total of 322 unknown metabolic reactions with information on atom differences and reaction classes (Fig. 2d and Supplementary Data 2). They were classified as two types: unknown I and unknown II. Unknown reaction type I is of unreported metabolic relationship between known metabolites and unknown metabolites, while unknown reaction type II is that between unknown metabolites. As a result, 99 and 223 reactions were characterized as unknown reaction type I and unknown reaction type II, respectively (Fig. 2d). To elucidate these unknown reactions, we first examined their reaction classes obtained from the delta mass library. Reaction classes including redox reaction, transfer reaction, hydrolysis, lytic reaction, isomerization, ligation reaction, and others, were identified both in unknown type I and type II reactions (Fig. 2f). In particular, the class of redox reaction was the most common one, with 42 and 73 redox reactions being identified in unknown reactions type I and type II, respectively (Fig. 2f). Then, the atom difference was used to search the reference reaction library (Supplementary Data 4) to retrieve the relevant reference reactions. Consistently, closer examination of atom differences between the reaction-paired metabolites revealed that the most common unknown reaction involved the exchange of an oxygen atom (e.g., hydroxylation; Fig. 2g). Other reactions, such as exchanges of atoms of [H + N] (e.g., amination), [H + N − O] (e.g., hydrolysis of amide), [2H + C] (e.g., methylation), [2H + O] (e.g., hydration) and [2H] (e.g., reduction), were also among the top-ranked unknown reactions (Fig. 2g).

Collectively, these results proved that the isotopologue similarity networking approach enables to discover and elucidate unknown metabolic reactions from stable-isotope tracing metabolomics data, which provides a great potential to characterize previously unknown metabolic routes of endogenous metabolites. To further demonstrate the applicability of the methods in other biological systems, the results of IsoNet approach on global stable-isotope tracing metabolomics datasets of immortalized bone marrow-derived macrophage (iBMDM) cells and mouse liver were also provided. Their isotopologue similarity networks and discovered unknown reactions were also demonstrated using IsoNet (Supplementary Fig. 4 and Supplementary Datas 5 and 6).

Glutathione-associated subnetwork reveals previously unknown metabolic reactions

Glutathione (GSH) metabolism is instrumental for maintenance of cellular redox. Aberrant glutathione metabolism leads to many diseases such as cancer, Alzheimer’s disease, and diabetes²⁷. IsoNet analysis of 293T cells revealed a glutathione-associated isotopologue similarity subnetwork in the stable-isotope tracing metabolomics data collected in positive ionization mode (Fig. 3a). In this subnetwork, two nodes are identified as known metabolites (glutathione and S-lactoylglutathione), while the rest three nodes are unknown metabolites which were elucidated through our IsoNet approach. In the glutathione-associated subnetwork, we observed a key unknown reaction in which the thiol group of glutathione was converted into a hydroxyl group of M292T447 (Fig. 3b). To elucidate this unknown reaction, we first examined the metabolite pair of glutathione and M292T447 that they had a high isotopologue similarity score (S_ISO) of 0.94 and a mass difference (Δm/z) of 15.9771 Da (Fig. 3c). Matching the mass difference to the delta mass library suggests that the atom difference between two metabolites is [S − O]. Given the atom difference, the related reference transsulfuration reaction (R03923) was retrieved from the reference reaction library, and used to elucidate the unknown transformation between glutathione and M292T447. In the reference reaction, L-cysteine is converted into L-serine, wherein a thiol group being converted into a hydroxyl group (Fig. 3c). Such a transsulfuration reaction also results in an atom difference of [S − O] and a delta mass of 15.9772 Da. Inspired by this reference reaction, we therefore speculated a similar transsulfuration reaction that taken place between glutathione and the unknown metabolite M292T447 (Fig. 3c).

Fig. 3: Glutathione-associated isotopologue similarity subnetwork reveals previously unknown metabolic reactions.

a The glutathione-associated isotopologue similarity subnetwork in positive ionization mode. The grey shad highlights the edge between glutathione and M292T447. b The unknown metabolic reaction between glutathione and M292T447. Orange points represent the ¹³C-labeled atoms. c Elucidation of the metabolic reaction between glutathione and M292T447 (γ-Glu-Ser-Gly). The red shads highlight the variable substructures. d Interpretation of MS/MS spectra between glutathione and γ-Glu-Ser-Gly. The red shads highlight the variable substructures. e Validation of γ-Glu-Ser-Gly using chemical standard (Std): retention time match (top) and MS2 spectral match (bottom). f The relative levels of γ-Glu-Ser-Gly in various mouse tissues (n = 3–4) quantified by LC − MS. g Isotopologue pattern similarity between glutathione and γ-Glu-Ser-Gly in mouse liver tissue. Mice were administered with [U-¹³C]-glucose for 24 h (n = 4 biological replicates per group). h Intracellular γ-Glu-Ser-Gly levels in 293T and iBMDM cells after the treatment of buthionine sulfoximine (BSO; 100 μM) for 12 h or 10 h, respectively (n = 6 biological replicates for 293T cells per group, p = 0.0070 for glutathione, p < 0.0001 for γ-Glu-Ser-Gly; n = 5 biological replicates for iBMDM cells per group p = 0.0093 for glutathione, p < 0.0001 for γ-Glu-Ser-Gly). i Quantitative real-time PCR analyses of mRNA levels of GCS gene expression in control group and the siRNA-transfected 293T cells (n = 3 biological replicates per group; p = 0.023, ctrl vs siRNA1; p = 0.025, ctrl vs siRNA2; p = 0.031, ctrl vs siRNA3). j, k Intracellular levels of glutathione (p = 0.000064, ctrl vs siRNA1; p = 0.0019, ctrl vs siRNA2; p = 0.0015, ctrl vs siRNA3) ( j) and γ-Glu-Ser-Gly (p = 0.021, ctrl vs siRNA1; p = 0.011, ctrl vs siRNA2; p = 0.015, ctrl vs siRNA3) (k) in control group and siRNA-transfected 293T cells (n = 3 biological replicates per group). P-values were determined by a two-tailed Student’s t-test. *, p-value < 0.05; **, p-value < 0.01; ***, p-value < 0.001; ****, p-value < 0.0001. Values represent means ± SD. Source data are provided as Source Data files for Fig. 3c and f–k.

To validate the reaction product, we interpreted MS/MS spectra of glutathione and M292T447, and jointly inferred the structure as γ-Glu-Ser-Gly (Fig. 3d). Specifically, red, green, and blue fragments shifted the similar delta mass as precursors (Δm/z = 15.9771 Da), which was used to determine the variable substructure highlighted as the red shads. Further, we synthesized the chemical standard of γ-Glu-Ser-Gly, and matched retention times (RT) and MS/MS spectra between M292T447 and the synthesized chemical standard to confirm its identity (Fig. 3e). In addition to 293T cells, we also measured the levels of γ-Glu-Ser-Gly across a variety of mouse tissues (brown adipose tissue, brain, heart, intestine, liver, lung, muscle, spleen, and white adipose tissue), with the spleen being the most abundant organ (Fig. 3f). Subsequently, we conducted an in vivo stable-isotope tracing experiment in mice using [U-¹³C]-glucose and collected liver tissue to examine the incorporation of ¹³C-tracer into the glutathione-associated reaction network. As a result, γ-Glu-Ser-Gly in mouse liver was found as being isotopically labeled and shared a high isotopologue similarity with glutathione, which proved that γ-Glu-Ser-Gly could be endogenously synthesized and had a close metabolic relationship with glutathione (Fig. 3g). To verify that the newly found metabolic reaction links to glutathione metabolism, we next treated cells with buthionine sulfoximine (BSO), which is an inhibitor of gamma-glutamylcysteine synthetase (GCS) in glutathione metabolism pathway²⁸. As expected, BSO treatment effectively decreased levels of glutathione and γ-Glu-Ser-Gly in both of 293T and iBMDM cells (Fig. 3h). To confirm that γ-Glu-Ser-Gly is synthesized directly from glutathione, we further generated GCS knockdown 293T cells deficient in glutathione biosynthesis using RNA interference (Fig. 3i). As results showed in Fig. 3j, k, GCS knockdown dramatically decreased cellular levels of not only glutathione but also γ-Glu-Ser-Gly, demonstrating the direct link between γ-Glu-Ser-Gly biosynthesis and endogenous glutathione level. Collectively, these results demonstrated a tangible transsulfuration reaction of glutathione that generated γ-Glu-Ser-Gly in the glutathione metabolism network.

Validation of a previously uncharacterized transsulfuration reaction with glutathione

As in the reference transsulfuration reaction (R03923), cysteine is converted to serine by transferring the sulfur to tRNAs and producing thiolated tRNAs, which play essential functions in regulating cellular translational capacity and metabolic homeostasis²⁹. In the previously unknown transsulfuration reaction discovered above, we found that glutathione is converted to γ-Glu-Ser-Gly, and also acts as a “sulfur-donating” metabolite (Fig. 4a). The newly identified metabolite γ-Glu-Ser-Gly belongs to the family of γ-glutamyl tripeptide. However, the biosynthesis of γ-glutamyl tripeptide is commonly reported through the ligation of glutamate with various amino acids via gamma-glutamylcysteine synthetase (GCS) and glutathione synthetase (GSS)²⁸ (Fig. 4b). To demonstrate the finding of a previously unknown transsulfuration reaction distinct from the known ligation reaction, we first examined the isotopologue patterns of key metabolite substrates such as glutamate, glycine, serine, and cysteine in 293T cells labeled by a mixture of isotopic tracers ([U-¹³C]-glutamine, [U-¹³C]-glucose, and [U-¹³C]-acetate) (Fig. 4c). Joint probability can calculate the likelihood of different events in one time when random variables are given (Supplementary Fig. 5a). With inputs of experimental isotopologue patterns for metabolite substrates, we then calculated the predicted isotopologue patterns of γ-Glu-Ser-Gly in the scenarios of two different reaction pathways using the joint probability (Fig. 4d, e). If a ligation reaction is carried out, γ-Glu-Ser is first synthesized from glutamate and serine followed by glycine ligation to produce γ-Glu-Ser-Gly mainly with isotopologues of M + 5, M + 6, M + 7, M + 8, and M + 10 (Fig. 4e), which significantly differs from its experimental isotopologue pattern. The similarity score between experimental isotopologue pattern and predicted isotopologue pattern for γ-Glu-Ser-Gly is only 0.52 (Fig. 4d). As a comparison, in the scenario of transsulfuration reaction, glutathione (M + 5 and M + 7) is first biosynthesized via the ligation of glutamate with cysteine and glycine. Then, transsulfuration of glutathione occurs which results in the metabolite product γ-Glu-Ser-Gly mainly with isotopologues of M + 5, and M + 7 (Fig. 4e). Next, we calculated the similarities between experimental isotopologue patterns and predicted isotopologue patterns for glutathione and γ-Glu-Ser-Gly in the previously unknown pathway. The generated S_ISO values were calculated as 0.67 and 0.89 for glutathione and γ-Glu-Ser-Gly, respectively (Fig. 4e). These results demonstrated that the transsulfuration reaction from glutathione is used for γ-Glu-Ser-Gly biosynthesis instead of the conventional ligation reaction.

Fig. 4: Validation of a previously uncharacterized transsulfuration reaction with glutathione.

a Biosynthesis of γ-Glu-Ser-Gly via the transsulfuration reaction from glutathione. b Biosynthesis of γ-Glu-Ser-Gly via the ligation reaction. c The isotopologue patterns of glutamate, glycine, serine, and cysteine in 293T cells labeled with mixed tracers of [U-¹³C]-glutamine, [U-¹³C]-glucose, and [U-¹³C]-acetate for 17 h. d The predicted isotopologue pattern of γ-Glu-Ser-Gly by joint probability based on ligation reactions in (b). e The predicted isotopologue patterns of glutathione (left) and γ-Glu-Ser-Gly (right) by joint probability based on transsulfuration reaction in (a). f, g The experimental isotopologue patterns of glutathione (red) and γ-Glu-Ser-Gly (red) and predicted γ-Glu-Ser-Gly (blue and green) in 293T cells labeled with [U-¹³C]-serine (f) or [U-¹³C]-cysteine (g) for 6 h. Values represent means ± SEM (n = 6 biological replicates per group). Source data are provided as Source Data files for Fig. 4c–g.

Further, we carried out distinct stable-isotope tracing experiments in 293T cells to validate the newly found glutathione transsulfuration route using [U-¹³C]-serine, [U-¹³C]-cysteine, and [U-¹³C]-glycine as tracers individually. Specifically, in [U-¹³C]-serine labeling experiment, M + 3 labeled serine was quickly converted into M + 2 labeled glycine (Supplementary Fig. 5b), which was further used to synthesize glutathione. Therefore, the newly synthesized glutathione had the expected M + 2 isotopologue from labeled glycine (Fig. 4f), which is consistent with the isotopologue pattern of glutathione labeled with [U-¹³C]-glycine (Supplementary Fig. 5d, e). Importantly, γ-Glu-Ser-Gly was measured only in M + 2 isotopologue form, while M + 3 or M + 5 labeled isotopologues were absent (Fig. 4f), suggesting that the newly labeled γ-Glu-Ser-Gly was directly derived from glutathione through transsulfuration but not from ligation with M + 3 labeled serine. Also, [U-¹³C]-cysteine labeling demonstrated the incorporation of labeled cysteine into both glutathione and γ-Glu-Ser-Gly in M + 3 forms (Fig. 4g and Supplementary Fig. 5c). Thus, cysteine but not serine was involved into γ-Glu-Ser-Gly biosynthesis. Additionally, comparative analyses demonstrated a high consistency between experimental isotopologue pattern and predicted isotopologue pattern via the transsulfuration reaction for γ-Glu-Ser-Gly. However, experimental isotopologue pattern of γ-Glu-Ser-Gly was remarkedly different from the predicted isotopologue pattern via the ligation reaction (Fig. 4f, g). Finally, we demonstrated that the previously unknown transsulfuration reaction from glutathione for γ-Glu-Ser-Gly synthesis was also present in iBMDM cells through [U-¹³C]-cysteine labeling experiment (Supplementary Figs. 5f and 5g). Additionally, γ-Glutamyl transferase (GGT), another enzyme that is involved in the biosynthesis of γ-glutamyl tripeptide, catalyzes the transfer of γ-glutamyl group of glutathione to amino acids or peptides. The likelihood of this γ-glutamyl transfer reaction was excluded by [U-¹³C]-serine labeling experiment (Supplementary Fig. 5h). Collectively, above results disclosed a previously uncharacterized transsulfuration reaction from glutathione which revised the reported biosynthetic routes of γ-glutamyl tripeptide.

IsoNet fulfils the glutathione metabolic reaction network

Next, we combined the isotopologue subnetworks of both positive and negative ionization modes to chart an integrated reaction network for glutathione metabolism (Fig. 5a). In total, 11 nodes and 12 edges were included in the network with three known metabolites (glutathione, glutathione disulfide, and S-lactoylglutathione). We therefore attempted to deduce the rest nodes and edges using IsoNet. In addition to γ-Glu-Ser-Gly, seven nodes were further identified as S-acetylglutathione (Supplementary Fig. 6), glutathione sulfinamide (Supplementary Fig. 7), deaminated glutathione (Supplementary Fig. 8), γ-Glu-3-sulfamoyl-Ala-Gly (Supplementary Fig. 9), glutathione sulfinic acid (Supplementary Fig. 10), glutathione sulfonic acid (Supplementary Fig. 11), and S-(2-succinyl)glutathione (Supplementary Fig. 12) by the comparison of isotopologue similarity, relevance to reference reactions, and interpretation of MS/MS spectra. These metabolite identities were also validated with synthesized chemical standards or products from in vitro reactions (Supplementary Figs. 6–12). Most of these metabolites were not included in neither KEGG nor HMDB databases (Supplementary Table 1). In the network, a total of 10 reactions were identified, which includes distinct reaction classes such as redox reaction, transfer reaction, hydrolysis reaction, and ligation reaction. Among them, two reactions were recorded in the KEGG database, five reactions were previously reported, and three previously unknown reactions were firstly reported and characterized in this study (Supplementary Table 2). Three previously unknown reactions included the transsulfuration reaction to synthesize γ-Glu-Ser-Gly, the acetylation reaction to synthesize S-acetylglutathione, and the oxidation reaction to synthesize γ-Glu-3-sulfamoyl-Ala-Gly. Most importantly, γ-Glu-3-sulfamoyl-Ala-Gly is a previously unreported metabolite that not included in PubChem, KEGG or HMDB databases (Supplementary Table 1). Finally, we also confirmed their metabolism linked to glutathione metabolism pathway using in vivo stable-isotope tracing metabolomics in live mice and BSO inhibition in different cell lines (Supplementary Figs. 6–12).

Fig. 5: IsoNet fulfils the glutathione metabolic reaction network.

a An integrated glutathione metabolic network combined with datasets of positive and negative ionization modes. b The glutathione metabolic reaction network including 10 metabolic reactions. c PCA analysis of individual metabolite treatments. Metabolite intensities were normalized to their respective control group. d Significantly enriched pathways of individual metabolite treatment experiments. Pathway enrichment p-values were calculated using the hypergeometric test (unadjusted). The circle size represents the number of metabolite hits in the pathway. The color represents raw p-value of enrichment. Source data are provided as Source Data files for Fig. 5c, d.

Further, we verified the metabolic reactions using stable-isotope tracing technology and in vitro chemical reactions (Supplementary Fig. 13). For example, S-acetylglutathione was found to be synthesized through an acetyl transfer reaction between glutathione and acetyl coenzyme A (acetyl-CoA) in live cells (Supplementary Fig. 13a, b). Labeling cells with [U-¹³C]-pyruvate and [U-¹³C]-acetate, which serve as the precursors of the acetyl group in acetyl-CoA, produced S-acetylglutathione with isotopologue of M + 2 (Supplementary Fig. 13a). Also, we replicated the reaction by directly mixing glutathione and acetyl-CoA without enzymes, and observed a time-dependent production of S-acetylglutathione (Supplementary Fig. 13b). Glutathione sulfinamide has been reported as a specific biomarker for the exposure to HNO^30,31. Indeed, we observed a significantly increased level of glutathione sulfinamide in the redox reaction system involving glutathione and Angeli’s salt (HNO source) (Supplementary Fig. 13c). Additionally, we found that γ-Glu-3-sulfamoyl-Ala-Gly was synthesized as a product in this reaction system (Supplementary Fig. 13c). This result suggests that glutathione sulfinamide could be metabolized into γ-Glu-3-sulfamoyl-Ala-Gly via an oxidative reaction. Also, in the presence of reactive oxygen species (ROS), glutathione undergoes oxidation to produce glutathione sulfinic acid and glutathione sulfonic acid³². The two oxidative reactions were confirmed with glutathione sulfinic acid and glutathione sulfonic acid being detected in the reaction system of glutathione and hydrogen peroxide (Supplementary Fig. 13d). The ligation reaction between glutathione and fumarate producing S-(2-succinyl)glutathione was also validated as a non-enzymatic reaction³³ (Supplementary Fig. 13e). S-(2-succinyl)glutathione was reported act as an alternative substrate to glutathione reductase to decrease NADPH levels and boost mitochondrial ROS and HIF-1 activation³³. The discovery of this reaction enhances our understanding of how fumarate functions as a proto-oncometabolite. In summary, these results demonstrated that the isotopologue similarity networking approach can resolve both previously characterized and uncharacterized metabolic reactions associated with glutathione metabolism, charting a complete glutathione-centered metabolic reaction network (Fig. 5b).

To explore the biochemical effects of these reactions, we treated 293T cells with S-acetylglutathione, S-lactoylglutathione, glutathione sulfinic acid, γ-Glu-Ser-Gly, γ-Glu-3-sulfamoyl-Ala-Gly, and glutathione sulfonic acid, respectively (Fig. 5c). Comprehensive untargeted metabolomic analyses using LC–MS were performed on individual treatments and controls (Supplementary Data 7). Metabolite treatment increased the endogenous levels of individual treated metabolites in each treatment group compared to the control group (Supplementary Fig. 14a). Next, principle component analysis (PCA) was conducted using 345 metabolites detected in all groups (Fig. 5c and Supplementary Data 7). The results showed that the metabolic changes were similar between treatments with S-acetylglutathione and S-lactoylglutathione. Similar changes were also observed between treatments with γ-Glu-3-sulfamoyl-Ala-Gly and glutathione sulfonic acid. However, the changes induced by γ-Glu-Ser-Gly and glutathione sulfinic acid were exceptionally distinct among all of them. Further pathway enrichment analysis also demonstrated disparate impacts on cellular metabolism elicited by different glutathione metabolites (Fig. 5d and Supplementary Table 3). For examples, S-acetylglutathione and S-lactoylglutathione affected sphingolipid metabolism and unsaturated fatty acid synthesis, leading to a decrease of sphingosine, phytosphingosine, arachidonic acid, eicosapentaenoic acid, and eicosatrienoic acid (Fig. 5d and Supplementary Fig. 14b). γ-Glu-Ser-Gly treatment significantly decreased levels of nucleotides and nucleotide derivatives (Supplementary Fig. 14c). Glutathione sulfinic acid and γ-Glu-3-sulfamoyl-Ala-Gly primarily influenced amino acid metabolism (Supplementary Fig. 14d, e). Specifically, glutathione sulfonic acid treatment impacted not only amino acid metabolism but also glyoxylate and dicarboxylate metabolism (Supplementary Fig. 14f). In conclusion, these previously uncharacterized metabolites and reactions nested in the glutathione metabolic reaction network serve as important metabolic niches modulating cellular biochemistry.

Itaconate-associated isotopologue similarity subnetwork disclose previously unknown metabolites

Itaconate is a crucial anti-inflammatory metabolite, which modulates immune responses upon infections and the pathogenesis of inflammatory diseases^34,35. Next, we constructed an isotopologue similarity network in the lipopolysaccharide (LPS)-treated iBMDM cells, uncovering previously unknown reactions and metabolites related to inflammation (Supplementary Data 8). In the itaconate-associated isotopologue similarity subnetwork, we identified two previously unknown metabolites and associated reactions (Supplementary Fig. 15a, b). The metabolites produced by these reactions were identified as itaconate-cysteine (Supplementary Fig. 15c–g) and sulfoitaconate (Supplementary Fig. 15h–l) by isotopologue similarity comparison, relevance to reference reactions, interpretation of MS/MS spectrum, and validation using chemical standards. Neither itaconate-cysteine nor sulfoitaconate is recoded in the KEGG or HMDB databases. Moreover, both itaconate-cysteine and sulfoitaconate were found to increase dramatically by LPS stimulation in iBMDM cells, which were even more profound than the elevation of itaconate (Supplementary Fig. 15f and k), suggesting their relevance with inflammation. In conclusion, IsoNet facilitates the discovery of previously unknown metabolites linked to inflammation, offering valuable potentials for understanding anti-inflammatory processes.

Discussion

Cellular metabolism is a complex network intertwined with a myriad of metabolic reactions which support various cellular processes such as cell proliferation, signaling transduction, and growth. However, complementing the map of metabolic reaction network remains non-trival due to the intrinsic complexity of cellular metabolism and the lack of powerful technologies. Combining global stable-isotope tracing metabolomics and isotopologue similarity networking, the IsoNet technology developed in this study allowed to uncover hundreds of putative unknown metabolic reactions in cellular metabolism. The findings of unknown reactions fill the previously unexplored networks and complement our understanding of the metabolic maps in cellular metabolism. This technology is particularly useful towards reaction discovery in the case of unknown reaction type I. In this scenario, IsoNet establishes the connection between a known metabolite and an unknown metabolite. Given the experimental mass differences, IsoNet deduces the atom difference thus a substructure of the unknown metabolite, and then a potential reaction taken place by matching against the reference reaction library. Indeed, we performed manual inspection to deduce metabolic reactions between the labeled metabolite pairs. With the advancement of artificial intelligence models in metabolomics, it is plausible that in the near future, we may be able to achieve automated deduction of unknown metabolic reactions with the reaction-paired metabolites obtained from of IsoNet.

IsoNet calculates isotopologue pattern similarity based on biochemical principles which directly aligns isotopologue pattern segments (defined as “motifs”). The motif-based alignment offers better interpretability compared with point-to-point alignment approaches such as dynamic programing, since the isotopologue pattern represents a specific substructure of metabolites from metabolic reactions. However, in principle, isotopologue pattern similarity calculation using IsoNet is highly similar to conventional dynamic programming. In addition, IsoNet tends to discover large and composite unknown metabolites. This is because that large, composite metabolites with multiple substructures tend to display more diverse isotopologue patterns in the stable-isotope labeling. Consequently, the motif-based isotopologue pattern similarity networking is more likely to connect these metabolites. Additionally, the diverse isotopologue patterns facilitate the discovery and annotation of previously unknown metabolites and related biochemical reactions. The discovery of glutathione-associated previously unknown metabolites and reactions serves as a prominent example. For small metabolites, such as glycolytic intermediates and amino acids, IsoNet can also connect them through isotopologue pattern similarity networking (see examples in Fig. 1c, d). However, due to the limited structural information that can be obtained from the isotopologue patterns or MS/MS spectra of small metabolites, the annotation of unknown metabolites within these subnetworks remains a challenge. We anticipate that the annotation functions within IsoNet can be further improved to continue filling gaps in the metabolic network. The coverage of metabolites and reactions in the isotopologue similarity network is influenced by the MS ionization mode. To address this, we therefore acquired data using both positive and negative ionization modes to improve the coverage of detected metabolites. Then, we used IsoNet and constructed an integrated isotopologue similarity network which combines data from individual ionization modes. To further expand coverage of metabolites in the isotopologue similarity network, methods such as chemical derivatization and multidimensional chromatography separation can also be employed. In our study, our original focus was primarily on metabolomics. We believe that with suitable experimental conditions designed for lipidomic analysis (e.g., optimized measurement settings and labeling protocols), IsoNet can also be applicable for the discovery of unknown lipid species. Furthermore, the reference reaction library within IsoNet was primarily designed for metabolites and should be further developed to account for lipid metabolism and reactions.

A key finding using IsoNet strategy is the discovery of a previously uncharacterized transsulfuration reaction by which γ-Glu-Ser-Gly is synthesized directly via glutathione transsulfuration. Cellular transsulfuration play fundamental roles in the maintenance of redox homeostasis and sulfur balance^36,37. Abnormalities in transsulfuration reactions contribute to a broad spectrum of diseases, such as neurodegenerative diseases, autism, and vascular dysfunctions^38,39. The sulfur-containing metabolite cysteine has been canonically regarded as the central metabolic hub for transsulfuration reactions in cells. Our findings demonstrated that glutathione acts as a “sulfur-donating” metabolite for a previously unknown transsulfuration reaction with the metabolite γ-Glu-Ser-Gly being biosynthesized. Inspired by the reaction products of cysteine transsulfuration, we reasoned that the sulfur of glutathione is probably transferred to tRNAs which produces thiolated tRNAs. In human, there is a set of tRNA thiolation enzymes such as mitochondrial tRNA-specific 2-thiouridylase 1 (MTU1) and cytoplasmic tRNA 2-thiolation protein 1/2 (CTU1/2), catalyzing the sulfur modification on tRNAs^40,41,42. More importantly, the human tRNA-modification genes remain largely unexplored. Approximately 23% of the modification genes in human remain unknown⁴². Given that glutathione is a critical cellular sulfur donor, it is highly plausible that these unidentified genes and enzymes may also be involved in tRNA thiolation and the transsulfuration reaction of glutathione. Thiolated tRNAs are indispensable modules for cell growth through regulating cellular translational capacity²⁹. Importantly, metabolic homeostasis is tightly controlled by thiolated tRNAs whose levels reciprocally regulate nucleotide synthesis, amino acid metabolism and carbohydrate metabolism^29,43. Indeed, we observed consistent metabolic changes in these pathways, in particular the decreased levels of nucleotides in cells with γ-Glu-Ser-Gly treatment. Therefore, we surmise that the previously unknown transsulfuration reaction with glutathione functions is potentially in synergy to determine thiolated tRNAs levels, thus modulating cellular metabolism. Undeniably, more mechanistic studies are warranted to understand how the previously unknown glutathione transsulfuration reaction modulates tRNA thiolation and balances translational capacity and cellular redox homeostasis.

IsoNet prioritizes reaction discovery based on stable-isotope tracing metabolomics, but rather activity-based enzyme screening systems. One limitation of our method is that IsoNet can not characterize specific metabolic enzymes responsible for the previously unknown reactions discovered in this study. This could be addressed by integrating IsoNet analysis with functional metabolomics approach developed by Sévin et al. who incubated purified protein and protein-overexpressing cell lysate in metabolite cocktails. As such, metabolic enzymes associated with reactions in specific models may be targeted to complete the metabolic reaction network. Nevertheless, these findings of previously unknown reactions though IsoNet fill the previously unexplored metabolic network, and complement our understanding of the metabolic maps in cellular metabolism.

Methods

Ethical statement

The animal experiments were compliant with the ethical guidelines of the Institutional Animal Care and Use Committees of Interdisciplinary Research Center on Biology and Chemistry, Shanghai Institute of Organic Chemistry, Chinese Academy of Sciences (approval research project number: ECSIOC_2023-23).

Chemicals and standards

LC–MS grade water (H₂O) was purchased from Honeywell (Muskegon, MI, USA). LC–MS grade acetonitrile (ACN) was purchased from Merck (Darmstadt, Germany). Ammonium hydroxide (NH₄OH) and ammonium acetate (NH₄OAc) were purchased from Sigma (St. Louis, MO, USA). Stable-isotope tracers including [U-¹³C]-glucose, [U-¹³C]-glutamine, [U-¹³C]-acetate, [U-¹³C]-serine, [U-¹³C]-cysteine, [U-¹³C]-glycine, and [U-¹³C]-pyruvate were purchased from Cambridge Isotope Laboratories (MA, USA). DL-buthionine-sulfoximine (BSO) and N-ethylmaleimide (NEM) were purchased from Sigma-Aldrich (St. Louis, MO, USA). Catalase (C100456) and Crotalusadamanteus L-amino acid oxidase (A128538), Angeli’s salt (A332354), and S-lactoylglutathione (L121369) were purchased from Aladdin Bio-Chem Technology Co. LTD (Shanghai, China).

Stable-isotope tracing in cells and sample preparation

Cells (293T, iBMDM cell lines) were plated in 6-cm dishes and cultured in Dulbecco Modified Eagle’s Medium (DMEM) containing 10% dialyzed fetal bovine serum (dFBS) and 1% penicillin/streptomycin (PS). When cells were grown to 80% confluence, the culture medium was changed to a fresh medium solution with the following stable-isotope tracers for indicated times. For labeling experiments with a mixture of tracers, the fresh medium contained 25 mM [U-¹³C]-glucose, 4 mM [U-¹³C]-glutamine, 5 mM [U-¹³C]-acetate in glucose-free and glutamine-free DMEM. For labeling experiments with [U-¹³C]-glucose, the fresh medium contained 25 mM [U-¹³C]-glucose in glucose-free DMEM. For labeling experiments with [U-¹³C]-glutamine, the fresh medium contained 4 mM [U-¹³C]-glutamine in glutamine-free DMEM. For labeling experiments with [U-¹³C]-serine, the fresh medium contained 0.6 mM [U-¹³C]-serine in DMEM. For labeling experiments with [U-¹³C]-glycine, the fresh medium contained 0.6 mM [U-¹³C]-glycine in DMEM. For labeling experiments with [U-¹³C]-cysteine, the fresh medium contained 0.4 mM [U-¹³C]-cysteine in DMEM. For labeling experiments with [U-¹³C]-pyruvate, the fresh medium contained 4 mM [U-¹³C]-pyruvate in glucose-free DMEM. For labeling experiments with [U-¹³C]-acetate, the fresh medium contained 5 mM [U-¹³C]-acetate in DMEM.

Fast extraction of metabolites in cells was performed as follows. In brief, the culture medium was quickly removed, and cells were washed with PBS twice. Cell dishes were placed on dry ice and the precooled metabolite extraction solution (MeOH:ACN:H₂O = 2/2/1, v/v/v, 800 μL) was added to dishes to quench metabolism. The dishes were then incubated at −80 °C for 40 min. The cell contents were scraped and transferred to a 1.5-mL Eppendorf tube. Another 400 μL extraction solution was added to wash dish and transferred to the same EP tube. The samples were vortexed for 1 min and centrifuged for 10 min at 16,200 × g and 4 °C to precipitate insoluble materials. For the measurement of cysteine levels if needed, 50 μL supernatant was taken out and 50 μL extraction solvent (MeOH: H₂O, 4:1, v/v, containing 25 mM NEM and 10 mM ammonium formate, pH 7.0) was added followed by incubation on ice for 30 min^28,44. The NEM-derivatized metabolite extracts were then analyzed by LC-MS. The rest supernatant was taken to a new 1.5-mL Eppendorf tube and evaporated to dryness at 4 °C using a vacuum concentrator. The dried extracts were kept in −80 °C. Before LC–MS analysis, the dried extracts were reconstituted in 100 μL of ACN:H₂O (1:1, v/v), sonicated for 10 min, and centrifuged for 15 min at 16,200 × g and 4 °C to remove insoluble debris. The supernatant was then transferred to HPLC vials for LC–MS analysis.

Stable-isotope tracing in mice and sample preparation

12-week-old male mice (C57BL/6J; n = 3/4) were group-housed in a barrier facility at room temperature of 22 °C with 50% humidity and 12 h light/12 h dark cycles. Mice first received a liquid diet with free access to drinking water for one week. The liquid diet was composed of glucose (30 g), soy protein (10 g), coconut milk (34.3 mL) from Nature’s Charm, and water per 100 mL volume. Then, mice were fasted overnight prior to stable-isotope tracing. Tracing experiments were conducted by switching to a liquid diet with replacement of unlabeled glucose to [U-¹³C]-glucose at 9 am. Mice were sacrificed at 9 am on the second day after 24 h dietary [U-¹³C]-glucose tracing. Brown adipose tissue (BAT), brain (cortex), heart, intestine, kidney, liver, lung, muscle, spleen and white adipose tissue (WAT) were dissected and quickly frozen in liquid nitrogen immediately and stored at −80 °C until metabolite extraction. The animal experiments were compliant with the ethical guidelines of the Institutional Animal Care and Use Committees of Interdisciplinary Research Center on Biology and Chemistry, Shanghai Institute of Organic Chemistry, Chinese Academy of Sciences (approval research project number: ECSIOC_2023-23). Sex was controlled by exclusively using male mice, thereby avoiding metabolic variability introduced by periodic hormonal changes in females.

The mouse tissues were transferred into homogenizer tubes and homogenized with H₂O at a ratio of 200 μL H₂O per 20 mg tissue and ceramic beads using a homogenizer (JXFSTPRP-CL, Shanghai Jingxin Experimental Technology) at the low-temperature condition. 200 μL homogenized solution was taken out and 800 μL extraction solution (ACN: MeOH = 1:1, v/v) was added for metabolite extraction. The mixture solution was vortexed for 30 s, and sonicated for 10 min at 4 °C water bath. After incubation for 1 h at −20 °C, the sample was centrifuged for 15 min at 16,200 × g and 4 °C. The supernatant was taken to a new 1.5-mL EP tube and evaporated to dryness at 4 °C in a vacuum concentrator. The following extraction processes were the same as those for cell samples.

Metabolite treatment in cells

293T cells were plated in 6-cm dishes and cultured in Dulbecco Modified Eagle’s Medium (DMEM) containing 10% fetal bovine serum (FBS) and 1% penicillin/streptomycin (PS). When cells were grown to about 80% confluence, the culture medium was changed to a fresh medium solution with the following metabolites for 12 h. For treatment with γ-Glu-Ser-Gly, the fresh medium contained 0.15 mM γ-Glu-Ser-Gly in DMEM. For treatment with S-acetylglutathione, the fresh medium contained 0.5 mM S-acetylglutathione in DMEM. For treatment with S-lactoylglutathione, the fresh medium contained 0.5 mM S-lactoylglutathione in DMEM. For treatment with γ-Glu-3-sulfamoyl-Ala-Gly, the fresh medium contained 0.5 mM γ-Glu-3-sulfamoyl-Ala-Gly in DMEM. For treatment with glutathione sulfinic acid, the fresh medium contained 1 mM glutathione sulfinic acid in DMEM. For treatment with glutathione sulfonic acid, the fresh medium contained 2 mM glutathione sulfonic acid in DMEM. The fast extraction of metabolites was used the same method as that of stable-isotope tracing experiments in cells.

RNA interference, transfection, and quantitative real-time PCR analysis

siRNA duplexes were obtained from Genomeditech (Shanghai, China). The siRNA sequences are designed to target GCLC (glutamate-cysteine ligase catalytic subunit), the catalytic subunit of GCS. The detailed sequences were 5′-GAAGGAGGCUACUUCUAUAtt-3′ (siRNA1), 5′-GGAUCAUAUUUACAUGGAUtt-3′ (siRNA2), and 5′-GAGCCAUUGAAGAACAAUAtt-3′ (siRNA3). The siRNA (5′-UUCUCCGAACGUGUCACGdTdT) was used as negative control. Approximately 5 × 10⁵ cells were plated the day before transfection in 6-well plates and collected 2 mL of fresh growth medium prior to transfection. siRNA was transfected using Lipofectamine 2000^TM Transfection Reagent (Thermo Fisher Scientific) according to the manufacturer’s protocol. The final concentration of siRNAs was 100 nM. Cells were harvested 48 h later and processed further for real-time PCR and LC-MS analyses.

The 293T cells were directly lysed and used for RNA extraction. Total RNA was isolated using RNA Easy Fast Tissue/Cell Kit (TIANGEN, DP451) according to the manufacturer’s instruction. The concentration and purity of the extracted RNA was determined using NanoDrop (ThermoFisher Scientific). First-strand cDNA was synthesized from 1 μg of RNA using Reverse Transcriptase M-MLV (RNase H-) (Takara, Japan). The cDNAs were analyzed by quantitative real-time PCR with the following primers: GCS, 5′-GCTGTTGCAGGAAGGCATTG-3′ and 5′-AGTTTGGAGGAGGGGGCTTA-3′; actin, 5′- CTTCGCGGGCGACGAT-3′ and 5′- CCACATAGGAATCCTTCTGACC -3′. Real-time PCR analysis was performed using the QuantStudio 6 Flex real-time PCR system (ThermoFisher Scientific) with 2× SYBR Green qPCR Master Mix (SelleckChem). The 2^−ΔΔCT method was used to calculate the gene expression data, with the actin gene serving as the reference control.

LC−MS-based metabolomics

The metabolomics analysis protocol followed our previous publication with minor modification^7,8. Metabolomics data of biological samples were acquired using a Vanquish UHPLC coupled to an Orbitrap Exploris 480 (ThermoFisher Scientific, United States) using Xcalibur (version 4.4.16.14, Thermo Fisher Scientific, USA). A Waters ACQUITY UPLC BEH amide column (particle size, 1.7 μm; 100 mm (length) × 2.1 mm (i.d.)) and Phenomenex Kinetex C18 column (particle size, 2.6 μm; 100 mm (length) × 2.1 mm (i.d.)) were used for LC separation and the column temperature was kept at 25 °C. The injection volume was 2 μL. For the amide column, mobile phase A was water with 25 mM ammonium hydroxide (NH₄OH) and 25 mM ammonium acetate (NH₄OAc), and B was ACN for both positive (ESI+) and negative (ESI−) ionization modes. The flow rate was 0.5 mL/min and the gradient was set as follows: 0–0.5 min, 95% B; 0.5–7 min, 95% B to 65% B; 7–8 min, 65% B to 40% B; 8–9 min, 40% B; 9–9.1 min, 40% B to 95% B; 9.1–12 min, 95% B. For the C18 column, mobile phase A was 0.01% acetic acid in 100% water, and B was acetonitrile/isopropanol (1/1; v/v) for both positive (ESI+) and negative (ESI−) ionization modes. The flow rate was 0.3 mL/min and the gradient was set as follows: 0.0–1.0 min, 1% B; 1.0–8.0 min, 1% B to 99% B; 8.0–9.0 min, 99% B; 9.0–9.1 min, 99% B to 1% B; 9.1–12 min, 1% B.

ESI source parameters of the Orbitrap Exploris 480 were set as follows: spray voltage, 3000 V or −3000 V, positive (ESI+) and negative (ESI−) ionization modes, respectively; vaporizer temperature, 400 °C; sheath gas, 50 arb; aux gas, 15 arb; sweep gas, 2 arb; ion transfer tube temperature, 350 °C. LC–MS data acquisition was operated in full scan with polarity switching mode for all samples. Additional ddMS2 scans were performed on QC samples to acquire MS/MS spectra. The full scan was set as: orbitrap resolution, 60,000; AGC target, 1e6; maximum injection time, 100 ms; scan range, 70–1200 Da. The ddMS2 scan was set as: orbitrap resolution, 30,000; AGC target, 1e5; maximum injection time, 60 ms; scan range, 50–1200 Da; top N setting, 6; isolation width, 1.0. The collision energy was set as SNCE 20-30-40%. Dynamic exclusion duration was set as 4 s and isotope exclusion was on.

Data processing of metabolite-treatment untargeted metabolomics

The metabolite annotation steps followed previous publication²¹. Briefly, the raw data (.raw) was converted to.mzXML (for full scan mode) and.mgf (for ddMS2 mode) format using ProteoWizard (version 3.0.20360). Then the mzXML data files were grouped for peak detection and alignment using R package “xcms” (version 3.12.0). the generated MS1 peak table and MS2 files were uploaded to MetDNA2⁸ (version 1.4.4; [http://metdna.zhulab.cn/]) for metabolite annotation.

In each metabolite-treatment group, Unpaired two-tailed Student’s t-test in R was performed to compare the treatment group with the control group (Supplementary Data 7). Metabolites with a p-value < 0.05 were selected as significantly changed metabolites. The KEGG ID of those significantly changed metabolites was inputted for pathway analysis. The pathway analysis was performed by MetaboAnalyst 6.0 [https://www.metaboanalyst.ca/]. Enrichment method was set as “Hypergeometric Test”. Topology analysis was set as “Relative-betweeness Centrality”. The pathway was considered significant at a threshold of p-value < 0.05.

Calculation of isotopologue pattern similarity

We calculated isotopologue pattern similarity scores of metabolite pairs stratified by the numbers of carbons and labeled isotopologues (Supplementary Fig. 1a). In the type I scenario, metabolite pairs had the same number of carbons in their structures and an isotopologue pattern similarity score was calculated by normalized Manhattan distance Eq. (1) (Supplementary Fig. 1b).

$${S}_{{{\rm{ISO}}}}=1/\left(1+{\sum }_{i=0}^{n}\left|{I}_{a,i}-{I}_{b,i}\right|\right)$$

(1)

Where ${I}_{a,i}$ is the labeled fraction of isotopologue ${M}_{i}$ from metabolite/motif A, ${I}_{b,i}$ is the labeled fraction of isotopologue ${M}_{i}$ from metabolite/motif B, ${{\rm{n}}}$ is the minimum carbon number of the metabolite A or B.

In the type II scenario, metabolite pairs had different carbon numbers and only one isotopologue was labeled (Supplementary Fig. 1c). For example, the only isotopologue of M + 3 was observed in both metabolites A and B. Then, a motif was generated in which the carbon number of motifs was the minimum number of that in metabolite pairs. Finally, an isotopologue pattern similarity score S_ISO from the motif match was calculated by Eq. (1).

In the type III scenario, metabolite pairs had different carbon numbers and more than one isotopologue were labeled (Supplementary Fig. 1d). First, the metabolite with less carbon number (e.g., metabolite A) was used as a reference to generate motifs. Motifs used for match follows three criteria: (1) exclusion of M0; (2) carbon numbers of motifs should be larger or equal to half of that for the metabolite with more carbon number (e.g., metabolite B); and (3) the maximum generated motif number is 999. Then, the generated motifs (motifs in A) and the metabolite with more carbon number (B) formed motif-metabolite pairs. Each motif-metabolite pair was matched sequentially starting from M1, with a step length of 1. Isotopologue pattern similarity scores (S_ISO) from the motif-metabolite match were then calculated by Eq. (1). Finally, the maximum S_ISO and the optimal motif were exported.

The workflow of IsoNet

The IsoNet workflow includes four steps: (1) annotation of labeled metabolites; (2) isotopologue similarity networking; (3) unknown reaction discovery; and (4) elucidation of unknown reactions and structures (Supplementary Fig. 2a–d).

Annotation of labeled metabolites

Raw mass spectrometry data files (.raw) acquired from the orbitrap mass spectrometer were first converted to.mzXML (for full scan mode) and.mgf (for ddMS2 mode) format using ProteoWizard (version 3.0.20360). The R package “xcms” (version 3.12.0) was used for peak detection and alignment of mzXML data files of unlabeled samples. Key parameters were set as follows: method, “centWave”; ppm, 10; mzwid, 0.006; snthresh, 6; peakwidth, c(5, 30); minfrac, 0.5. The intermediate xcmsSet object from ‘xcms’ after peak detection was exported as a xcmsSet file (.Rda). The generated MS1 peak table and MS2 files were then uploaded to MetDNA2 (version 1.4.1; [http://metdna.zhulab.cn/]) for metabolite annotation^7,8. Key parameters in MetDNA2 were set as follows: instrument, ThermoExploris; column, “HILIC” or “RP” according to LC separation; ce, SNCE20_30_40%; method_lc, “Amide12min” or “RP12min” according to LC separation; mz_tol, 15 ppm. Metabolite annotations with metabolomics standards initiative (MSI) level 1 and level 2 were initially assigned as known metabolites. For remaining unknown features, peak annotation from CAMERA⁴⁵ and the peak correlation network in MetDNA2 were first used to remove redundancies including isotopes, adducts, in-source fragmentation (ISF), and neural losses. Then, formulas of the remaining features were predicated by Genform⁴⁶ (download on June 28th, 2020; [https://sourceforge.net/projects/genform/]). Key parameters of Genform were set as: adduct forms [M + H]⁺ in positive and [M-H]^- in negative modes; tolerance of m/z error, 5 ppm; elements included, C, H, N, O, P, S. The resultant top five formula candidates were kept. Finally, a metabolite annotation table (.csv file) including both known and unknown metabolites was generated as a target list for subsequent isotopologue extraction. MetDNA also outputted a MS/MS spectral data file (.msp).

The above metabolite annotation table, a previously generated xcmsSet file, unlabeled data (.mzXML), and labeled data (.mzXML) were subjected to our recently published R package “MetTracer” for global extraction of labeled metabolites²¹. The parameters for MetTracer were set as follows: rt.extend, 15; value, “maxo”; equipment, “Orbitrap”; ppm, 10; res.define, 200; resolution, 60000; d.extract, “labelled”; correct.iso, “TRUE”; adj.contaminate, “TRUE”. In the labeled samples, if the labeled fraction of one isotopologue (except M0) in one metabolite is larger than 0.02 in >50% of samples, the metabolite was considered to isotopically labeled. For data processing with manual check, the m/z and RT information of each isotopologue, and the raw data files (.raw) were imported into Skyline (v21.1.0.278). Integration range for each isotopologue was manually adjusted to ensure accurate quantification. Key parameters in Skyline were set as: ion match tolerance, 0.01 m/z; precursor mass analyzer, orbitrap; resolving power, 60000 at 200 m/z. Then, the quantification result was corrected using a R package “Accucor” (v0.3.0)⁴⁷ for natural isotope correction. Key parameters in Accucor were set as: resolution, 60000; resolution_defined_at, 200. Finally, labeled metabolites with normalized intensities of isotopologues were obtained as a csv file.

Isotopologue similarity networking

To perform isotopologue pattern similarity networking, we developed an R package “IsoNet”. The input data for IsoNet analysis includes a metabolite annotation table (.csv), a quantification table of labeled metabolites(.csv), and a MS/MS spectral data file (.msp). First, IsoNet calculates the isotopologue pattern similarity scores between two labeled metabolites by the method as described above. The MS/MS spectral similarity scores were calculated via a modified cosine score function reported in GNPS⁴⁸. The tolerance of m/z error was 15 ppm and the least matched fragment was 1. Key parameters for IsoNet include: mid_cutoff, 0.7; mid_fc, 20; mid_isoDegree, 0.1; mid_min_motifLen, 0.5; ms2_score_cutoff, 0.5; mid_max_motif, 1000; mass_diff_freq_cut_off, 4; ignore_max, TRUE. Finally, the isotopologue similarity network was generated in which one node represents one isotopically labeled metabolite and an edge represents the two metabolites had an isotopologue pattern similarity score >0.7 and a similar MS/MS spectral similarity score >0.5. The isotopologue pattern network was visualized in Cytoscape (v3.9.1) and the subnetworks containing lipids were removed.

Unknown reaction discovery

We categorized metabolic reactions in the isotopologue pattern network into three types, including known reactions, unknown reaction type I, and unknown reaction type II. Among them, known reactions are of characterized metabolic relationship between two known metabolites. For known reactions, their reaction steps were calculated based on the knowledge-based metabolic reaction network curated from the KEGG database reported in our previous work of MetDNA⁸. This reaction network comprises 6397 metabolites and 8129 reactions (Supplementary Fig. 2e). The reaction step represents the minimum number of metabolic reactions occurred between two metabolites in the metabolic reaction network. Calculation of reaction distance between two metabolites was realized by using a distance function in R package “igraph” (v1.2.9), which determines the number of connections between any two nodes within the network. By inputting KEGG IDs, the number of reaction steps between two metabolites can be derived from the MRN network. To annotate unknown reactions, we curated a delta mass library (Supplementary Fig. 2 and Supplementary Data 3). We matched the experimental mass differences (Δm/z) of metabolite pairs in the isotopologue pattern network with a list of delta masses that are associated with atom differences and reaction classes in the delta mass library. The tolerance of delta mass match was set as 2 mDa. Annotated information for each unknown reaction included the atom difference and the reaction class. These annotated unknown reactions were classified into type I and type II. Unknown reaction type I is of unreported metabolic relationship between one known metabolite and one unknown metabolite, while unknown reaction type II is that between two unknown metabolites.

Elucidation of unknown reactions and structures

We used the known metabolic reactions in the reference reaction library as references to deduce the structures of unknowns in the unknown metabolic reactions. For each unknown reaction, the annotated atom difference was first searched against the reference reaction library (Supplementary Data 4). Then, potential chemical transformations were summarized including reaction groups and reaction positions. With this reaction information, for unknown reaction type I, since one metabolite has known structure, we enabled to elucidate the structure of the other unknown metabolite in this reaction. For unknown reaction type II, once one of the metabolites is deduced in other reactions, the other unknow metabolite of this unknown reaction type II could be potentially deduced in the same way. For verification, we also compared the MS2 spectra between the reaction-paired metabolites. Shared fragments with same m/z values in two MS2 spectra often indicate their shared substructures. In addition, we assessed fragments and neutral loss (NL) with annotated delta masses between the two MS2 spectra, which often indicate variable substructures. We deduced these fragments into substructures by the fragmentation rules summarized from known metabolites using CFM-ID [https://cfmid.wishartlab.com/], mzCloud [https://www.mzcloud.org/] and SIRIUS (v5.5.5). Finally, we enabled to confirm the putative structures of the unknowns in the unknown reactions.

Curation of a delta mass library and a reference reaction library

We employed the metabolic reaction network (MRN) from the KEGG database for curation of a delta mass library and a reference reaction library. The MRN was obtained from our previous publication of MetDNA⁸. For each reaction pair in MRN, we retrieved their associated metabolic reactions in the KEGG database to form a reference reaction library. Given that one reaction pair could be matched with multiple reactions, for example, the reaction pair of NAD⁺/NADH, we performed redundancy removal based on the reaction classes. As a result, one metabolic reaction was kept for one metabolite pair in each reaction class. A total of 8704 reactions were obtained after the processing. The detailed information on each metabolic reaction was summarized in Supplementary Data 4. Then, seven reaction classes were classified, including redox reaction, transfer reaction, hydrolysis, lytic reaction, isomerization, ligation reaction, and others (Supplementary Fig. 2f). Then we calculated the mass difference and the corresponding atom difference for each reaction pair of reference reaction. As a result, 1123 unique delta masses with information on atom differences were obtained. For each delta mass, we defined its reaction class with the most records. The detailed information on the delta mass library was provided in Supplementary Data 3.

Isotopologue similarity between reaction-paired and non-reaction paired metabolites

For each pair of two isotopically labeled known metabolites in 293T cells labeled with [U-¹³C]-glucose or [U-¹³C]-glutamine, we retrieved their metabolic reaction relationship from the metabolic reaction network. Metabolites that were bridged by one or two reaction steps were defined as reaction pairs (RPs). Metabolites that were connected by more than seven reaction steps were defined as non-reaction pairs (non-RPs). Next, isotopologue similarity scores were calculated for the RPs and non-RPs using the method described above.

Prediction of isotopologue patterns

Joint probability can calculate the likelihood of different events in one time when random variables are given. With inputs of experimental isotopologue patterns for metabolite substrates, we used the joint probability theory to predict isotopologue patterns for the specific metabolite of interest. First, we calculated all possible isotopologue combinations of different isotopically labeled metabolite substrates as a data frame by the “expand.grid” function from R package “base” (v4.0.3). In the data frame, each row was a specific combination of isotopologues from labeled metabolites. Then, the probability of individual combination was calculated by multiplying the labeled fractions of isotopologues. The number of labeled carbons was calculated as the sum of carbons from different isotopologues in the combination. Next, the probabilities of rows with the same number of labeled carbons were summed up as the final predicted probability of one labeled isotopologue in the metabolite. Finally, the labeling pattern for the metabolite of interest was predicted.

Structural validation of metabolites with chemical standards

Previously unknown metabolites discovered in this study were validated using chemical standards. The structures of these chemical standards were confirmed with nuclear magnetic resonance (NMR) spectroscopy (Supplementary Figs. 16–22). The γ-Glu-Ser-Gly was synthesized by MuJin Biotech Inc., Shanghai, China. The chemical structure was confirmed by nuclear magnetic resonance spectroscopy (Supplementary Fig. 16). ¹H NMR (400 MHz, D₂O, ppm): δ = 4.38 (t, J = 5.2 Hz, 1H), 3.94–3.84 (m, 3H), 3.76 (d, J = 5.2 Hz, 2H), 2.48 (td, J = 7.4, 2.2 Hz, 2H), 2.16–2.06 (m, 2H); ¹³C NMR (101 MHz, D₂O, ppm): δ = 174.6, 173.1, 172.3, 172.3, 61.0, 55.5, 52.7, 41.1, 30.9, 25.5. Glutathione sulfonic acid was purchased from Bidepharm, Shanghai, China (product no. BD132132). The chemical structure was confirmed by nuclear magnetic resonance spectroscopy (Supplementary Fig. 17). ¹H NMR (400 MHz, CD₃OD, ppm): δ = 4.86–4.84 (m, 1H), 4.04 (t, J = 6.2 Hz, 1H), 3.92 (s, 2H), 3.40–3.32 (m, 1H), 3.21–3.10 (m, 1H), 2.56 (t, J = 6.4 Hz, 2H), 2.39–2.27 (m, 1H), 2.18–2.05 (m, 1H); ¹³C NMR (101 MHz, CD₃OD, ppm): δ = 173.0, 171.6, 171.3, 170.5, 52.1, 51.0, 50.8, 40.6, 31.1, 25.6. Glutathione sulfinic acid was purchased from Comparison, Beijing, China (product no. GGT160-06). The chemical structure was confirmed by nuclear magnetic resonance spectroscopy (Supplementary Fig. 18). ¹H NMR (400 MHz, D₂O, ppm): δ = 4.62–4.53 (m, 1H), 3.74 (s, 2H), 3.70 (t, J = 6.2 Hz, 1H), 2.80–2.69 (m, 1H), 2.59–2.50 (m, 1H), 2.48–2.39 (m, 2H), 2.14–2.03 (m, 2H); ¹³C NMR (101 MHz, D₂O, ppm): δ = 175.5, 174.6, 173.9, 172.5, 61.4, 54.1, 49.9, 42.9, 31.3, 26.0. S-Acetylglutathione was purchased from Bidepharm, Shanghai, China (product no. BD150717). The chemical structure was confirmed by nuclear magnetic resonance spectroscopy (Supplementary Fig. 19). ¹H NMR (400 MHz, D₂O, ppm): δ = 4.62 (dd, J = 7.7, 5.1 Hz, 1H), 3.94 (s, 2H), 3.80 (t, J = 6.3 Hz, 1H), 3.46–3.38 (m, 1H), 3.22–3.12 (m, 1H), 2.51–2.41 (m, 2H), 2.36 (s, 3H), 2.19–2.08 (m, 2H); ¹³C NMR (101 MHz, D₂O, ppm): δ = 199.9, 174.7, 173.5, 173.4, 172.0, 53.7, 52.8, 41.4, 31.1, 29.9, 29.8, 25.9. The γ-Glu-3-sulfamoyl-Ala-Gly was synthesized by MuJin Biotech Inc, Shanghai, China. The chemical structure was confirmed by nuclear magnetic resonance spectroscopy (Supplementary Fig. 20). ¹H NMR (400 MHz, CD₃OD, ppm): δ = 5.04–4.95 (m, 1H), 4.10–4.00 (m, 1H), 3.93 (s, 2H), 3.75–3.62 (m, 1H), 3.51–3.37 (m, 1H), 2.63–2.50 (m, 2H), 2.34–2.22 (m, 1H), 2.21–2.09 (m, 1H); ¹³C NMR (101 MHz, CD₃OD, ppm): δ = 173.1, 171.5, 170.3, 170.3, 54.7, 52.1, 49.6, 40.5, 31.0, 25.5. S-(2-succinyl)glutathione was synthesized by MuJin Biotech Inc., Shanghai, China. It’s a stereoisomeric mixture. The chemical structure was confirmed by nuclear magnetic resonance spectroscopy (Supplementary Fig. 21). ¹H NMR (400 MHz, D₂O, ppm): δ = 4.67–4.57 (m, 1H), 3.98 (s, 2H), 3.90 (t, J = 6.4 Hz, 1H), 3.81–3.72 (m, 1H), 3.27–3.14 (m, 1H), 3.10–2.91 (m, 2H), 2.85–2.74 (m, 1H), 2.60–2.45 (m, 2H), 2.25–2.11 (m, 2H); ¹³C NMR (101 MHz, D₂O, ppm): δ = 175.7, 175.6, 174.7, 174.7, 174.6, 174.5, 173.0, 172.7, 172.4, 172.3, 53.1, 52.6, 42.6, 41.8, 41.2, 36.3, 36.0, 32.6, 32.2, 31.0, 25.7, 25.7. Sulfoitaconate was synthesized by MuJin Biotech Inc., Shanghai, China. The chemical structure was confirmed by nuclear magnetic resonance spectroscopy (Supplementary Fig. 22). ¹H NMR (400 MHz, D₂O, ppm): δ = 3.28–3.20 (m, 1H), 3.19–3.11 (m, 1H), 3.10–3.03 (m, 1H), 2.77 (d, J = 6.7 Hz, 2H); ¹³C NMR (101 MHz, D₂O, ppm): δ = 176.84, 175.50, 51.29, 37.96, 34.78.

The enzymatic synthesis and validation of deaminated glutathione

For deaminated glutathione, the reference sample was prepared as previous publication⁴⁹. Briefly, a 150 mM solution of GSH, adjusted to pH 7, was incubated overnight with 4 units/mL catalase and 2 mg/mL of Crotalus adamanteus L-amino acid oxidase at 37 °C. 10 μL mixture was then deproteinized by adding 40 μL extraction solution (ACN: MeOH = 1:1, v/v). The mixture was incubated at −20 °C for at least 40 min. After incubation, the mixture was vortexed for 1 min, centrifuged for 15 min at 16,200 × g and 4 °C to precipitate insoluble materials. The supernatant was analyzed by LC-MS. To further confirm the presence of deaminated glutathione, 10 μL solution from the injection vial was mixed with 40 μL of derivatization reagent (MeOH: H₂O, 4:1, v/v, containing 25 mM NEM and 10 mM ammonium formate, pH 7.0) followed by incubation on ice for 30 min and LC-MS analysis. Subsequently, the retention time and MS/MS spectra were compared with cell samples that underwent the same derivatization.

The validation of glutathione-related previously unknown reactions in vitro

For S-acetylglutathione, glutathione (1 mM) was reacted with Acetyl-CoA (1 mM) in a buffer solution containing 10 mM Tris and 1 mM MgCl₂. At 20, 100, and 480 min, 10 μL reaction solution was taken, respectively, and mixed with 40 μL methanol. Then, the solution was analyzed by LC-MS. For glutathione sulfinamide and γ-Glu-3-sulfamoyl-Ala-Gly, glutathione was reacted with Angeli’s salt³⁰. Specifically, 20 μL aliquots of a 20 mM Angeli’s salt stock solution were added to glutathione solution (10 mL; 50 μM) every 10 min for five times in total. During reaction, the solution was incubated at 37 °C. Then, the generated solution was analyzed by LC-MS. The product in the mixture was also used as in-house synthesized standard for structural validation of glutathione sulfinamide. For the S-(2-succinyl)glutathione, glutathione (10 mM) was reacted with fumarate (5 mM) for about 3 h under pH 7 and 37 °C. Subsequently, the mixture was analyzed by LC-MS. For glutathione sulfinic acid and glutathione sulfonic acid, glutathione (10 μM) was reacted with H₂O₂ (9 M) for about 30 min and analyzed by LC-MS. For the control groups, the substrates such as Acetyl-CoA, Angeli’s salt, fumarate, and H₂O₂ were replaced with water, and the same procedures were followed.

LPS treatment, stable-isotope tracing experiment of iBMDM cells and data processing

The cell culture conditions were consistent with methods described above (see “Stable-isotope tracing in cells and sample preparation” section of “Methods”). When cells were grown to 80% confluence, the culture medium was changed as follows: for the unlabeled group, the fresh medium contained 100 ng/ml LPS in DMEM; for the labeled group, the fresh medium contained 100 ng/ml LPS, 25 mM [U-¹³C]-glucose, 4 mM [U-¹³C]-glutamine, 5 mM [U-¹³C]-acetate in glucose-free and glutamine-free DMEM. Both the LPS treatment and tracing time was 10 h. The sample preparation and data preprocessing were performed using the methods described above. Key parameters for IsoNet: ms2_score_cutoff, 0.4; mass_diff_freq_cut_off, 1.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.