Hydropersulfides inhibit lipid peroxidation and ferroptosis by scavenging radicals

Ferroptosis is a type of cell death caused by radical-driven lipid peroxidation, leading to membrane damage and rupture. Here we show that enzymatically produced sulfane sulfur (S0) species, specifically hydropersulfides, scavenge endogenously generated free radicals and, thereby, suppress lipid peroxidation and ferroptosis. By providing sulfur for S0 biosynthesis, cysteine can support ferroptosis resistance independently of the canonical GPX4 pathway. Our results further suggest that hydropersulfides terminate radical chain reactions through the formation and self-recombination of perthiyl radicals. The autocatalytic regeneration of hydropersulfides may explain why low micromolar concentrations of persulfides suffice to produce potent cytoprotective effects on a background of millimolar concentrations of glutathione. We propose that increased S0 biosynthesis is an adaptive cellular response to radical-driven lipid peroxidation, potentially representing a primordial radical protection system.

In this study, we explored the possible connection among S 0 biology, LPO and ferroptosis. The biology of S 0 species (that is, hydropersulfides, hydropolysulfides and polysulfides) is a relatively new field of study. Nevertheless, it is now well-established that all cells can produce persulfides and polysulfides endogenously, through several enzymatic pathways and mechanisms, Cys being the ultimate source of sulfur 14 . Previous studies have indicated that S 0 species can have potent anti-oxidative and cytoprotective properties 15,16 . These studies have mostly focused on the reactions between S 0 species and two-electron oxidants. However, a puzzling aspect of S 0 biology is the low intracellular concentration of persulfides and polysulfides (low micromolar range) 15 relative to glutathione (millimolar range), which makes it difficult to rationalize an efficient scavenging role for S 0 species. Furthermore, the influence of biological S 0 species on one-electron oxidants (radicals) has so far received little attention. Although it has been shown in earlier studies that polysulfides can prevent lipid oxidation in vitro 17 , the prevailing view has been that persulfides and polysulfides react with product peroxides, thus preventing the initiation of new chain reactions. However, work in the 1990s and during the last few years has shown that hydropersulfides are excellent hydrogen atom transfer agents, directly engaging in one-electron reactions [18][19][20] , although the exact mechanisms and their relevance to biological systems remain to be clarified.
Here we show that biologically relevant S 0 species are potent radical scavengers and chain terminators, in vitro and inside living cells, and are important for the protection of membranes against LPO and, thus, are also relevant to ferroptosis sensitivity. Specifically, we report the following. (1) Pro-ferroptotic conditions trigger an increase in intracellular S 0 levels, likely constituting an adaptive cellular response. (2) Increased cellular uptake of Cys can S 0 -generating/degrading enzymes modulate LPO and ferroptosis. Next, we asked if ferroptosis sensitivity could be modulated by enzymes involved in the generation and degradation of S 0 species. To this end, we genetically manipulated the expression levels of key enzymes in HeLa cells. To assess the effect of these manipulations, we monitored the loss of membrane integrity under conditions of ferroptosis induction. As expected, the loss of membrane integrity in wild-type HeLa cells upon treatment with RSL3 or CHP was mostly prevented by liproxstatin-1 (Extended Data Fig. 2a). We first modulated levels of the H 2 S-generating enzyme cystathionine γ-lyase (CSE, also known as CTH). As expected, the depletion of CSE (Extended Data Fig. 2b) lowered endogenous H 2 S levels (Extended Data Fig. 2c), and the overexpression of CSE (Extended Data Fig. 2d) increased endogenous H 2 S levels (Extended Data Fig. 2e).
Having observed the influence of S 0 -metabolizing enzymes on ferroptotic cell death, we next asked about their influence on LPO.
To this end, we induced ferroptosis with RSL3 and measured LPO using BODIPY-C11 and coumarin hydrazide (CHH). Depletion of CSE elevated RSL3-induced LPO ( Fig. 2d and Extended Data Fig.  4a), whereas depletion of ETHE1 lowered it ( Fig. 2e and Extended Data Fig. 4b). Taken together, these results suggest that endogenously produced persulfides protect against LPO and consequently prevent ferroptosis.
Exogenously supplied S 0 eliminates intracellular radicals. We then asked if exogenously supplied membrane-permeable hydropersulfide donors could also protect cells against LPO. Indeed, addition of either diallyl tetrasulfide (DATS) or cysteine trisulfide (CSSSC) to the cell culture medium suppressed RSL3-induced LPO (Extended Data Fig. 5a). Having established that S 0 contributes to preventing LPO and ferroptosis, we next set out to understand the underlying molecular mechanism. Because LPO is driven by radical chain reactions, we considered the possibility that S 0 species act by quenching endogenous radicals. To investigate this possibility, we developed a chemogenetic system that allowed us to generate endogenous radicals in a controlled manner. To this end, we stably expressed the engineered heme peroxidase APEX2 (ref. 23 ) in either the cytosol or mitochondria of HeLa cells. APEX2 uses H 2 O 2 to generate phenoxyl radicals from phenolic substrates. Co-treatment of APEX2-expressing cells with H 2 O 2 and the phenolic peroxidase substrate HPI (1-(p-hydroxyphenyl) imidazole) enhanced LPO in a synergistic manner ( Fig. 3a and Extended Data Fig. 5b). To directly measure the formation of radicals in this system, we incubated cells with the membrane-permeable spin probe DEPMPO and subjected them to electron spin resonance (ESR) spectroscopy. We expected the resulting phenoxyl radicals to be largely reduced by GSH, in turn forming glutathionyl radicals (GS•), which should be trapped as characteristic DEPMPO adducts (Fig. 3b). The observed ESR signal was strictly dependent on the combined presence of APEX2, H 2 O 2 and HPI, confirming the specificity of the system and the absence of background activities (Fig. 3c). The shape of the ESR spectrum indicated formation of mainly GS• by cytosolic or mitochondrial APEX2 (Extended Data Fig. 5c). Pre-treatment of cells with membrane-permeable inorganic or organic polysulfides, confirmed to generate endogenous GSSH (Extended Data Fig. 5d), suppressed the ESR signal in a concentration-dependent manner (Fig. 3d). To further consolidate these findings, we aimed to follow radical formation and elimination in real time. To this end, we used luminol instead of the spin trap (Fig. 3e). Radicals generated by APEX2 are expected to oxidize luminol to luminol radicals, which further react with molecular oxygen to emit blue light (Extended Data Fig. 5e). As expected, the addition of H 2 O 2 to cells pre-incubated with the phenolic substrate and luminol induced transient luminescence (Fig. 3f) in a strictly APEX2-dependent manner (Extended Data Fig. 5f) and proportional to the amount of H 2 O 2 added (Extended Data Fig. 5g).
We then found that pre-treatment of cells with a membrane-permeable persulfide donor (Na 2 S 2 ) suppressed H 2 O 2 -induced luminescence in a concentration-dependent and time-dependent manner ( Fig. 3g and Extended Data Fig. 5h). Similar effects were recapitulated with the organic persulfide donor CSSSC (Fig. 3h). In sum, we found that exogenously supplied hydropersulfide donors lower endogenous radical levels.  Endogenously produced S 0 eliminates intracellular radicals. Next, we asked if endogenously produced persulfides are able to eliminate intracellular radicals. To this end, we manipulated the enzymes involved in S 0 metabolism (Fig. 4a) in APEX2-expressing cells. The exogenous addition of substrates for CSE (Cys), MPST (3MP) or SQR (H 2 S) lowered endogenous radical load, in line with the notion that enzymatically generated S 0 species contribute to radical scavenging ( Fig. 4b and Extended Data Fig. 6a). Indeed, overexpression of CSE decreased radical load (Fig. 4c), whereas its depletion led to an increase (Fig. 4d). Conversely, overexpression of ETHE1 increased radical load (Fig. 4e), whereas its depletion resulted in a decrease (Fig. 4f). Depletion of MPST also increased radical load, albeit to a lesser extent than depletion of CSE (Extended Data Fig. 6b). Similar results were obtained when APEX2 was expressed in the mitochondrial matrix instead of the cytosol (Extended Data Fig. 6c).
Hydropersulfides are potent radical scavengers. To obtain further mechanistic insight, we studied the GSSH-generating GSSSG/GSH system and its interaction with radicals in vitro. We used recom-  GSH. Adding the DEPMPO spin trap, the ESR spectrum confirmed formation of the glutathione adduct (Fig. 5a, 24 (Extended Data Fig. 7a). As expected, resorufin formation was inhibited under hypoxic conditions or in the presence of the spin trap (Extended Data Fig. 7b). The addition of increasing micromolar amounts of GSSSG inhibited resorufin formation for increasing periods of time (Fig. 5b, left panel). Based on resorufin absorbance and the observed inhibition time, the amount of scavenged radicals was calculated, indicating that low amounts of GSSSG scavenge >10 times the amount of radicals (Fig. 5b, right panel). Addition of the inorganic persulfide donor Na 2 S 4 instead of GSSSG achieved similar results (Extended Data Fig. 7c    that H 2 S (as potentially formed by GSH-mediated GSSH reduction) is unlikely to be involved in radical scavenging: Na 2 S inhibited resorufin formation only weakly (Extended Data Fig. 7d), and this inhibition is likely due to polysulfide contaminations typically accompanying commercial Na 2 S preparations 25 . Additional experiments confirmed that persulfides and polysulfides do not block APEX2 activity (Extended Data Fig. 8a) and that neither GSH nor GSSH can reduce compound I of APEX2 (Extended Data Fig.  8b). Similarly to the resazurin/resorufin system, GSSSG accelerated GSH-dependent 1e − reduction of cytochrome c (cyt c) at sub-stoichiometric concentrations (Fig. 5c), whereas H 2 S did not facilitate any cyt c reduction (Extended Data Fig. 8c). Likewise, addition of sub-stoichiometric amounts of CSSSC accelerated the reduction of TEMPOL radicals by GSH (Fig. 5d), without involvement of superoxide (Extended Data Fig. 8d).
GSSH catalyzes GSH-dependent radical reduction. Having observed that radicals are efficiently scavenged in the presence of sub-stoichiometric amounts of persulfides, we considered the possibility that persulfides couple radical reduction to glutathione oxidation in a self-regenerating manner. We, therefore, considered an autocatalytic cycle in which GSSH reduces radicals to form perthiyl radicals (GSS•), which recombine to form glutathione tetrasulfide (GSSSSG), which is then reduced by GSH to regenerate GSSH (Fig. 6a). Using cyclic voltammetry, we confirmed that GSSH has a much lower reduction potential than GSH (Fig. 6b). In line with this result, quantum mechanical (QM) calculations showed the unpaired electron of GSS• to be de-localized between the two sulfur atoms and confirmed the higher stability of GSS• relative to GS• (Extended Data Fig. 9a). The nitric oxide (NO) donor GSNO (facilitating S-nitrosylation of GSSH) completely abolished GSSSG-dependent cyt c reduction (Fig. 6c), without inhibiting cyt c (Extended Data Fig. 9b). We furthermore confirmed by mass spectrometry (MS) that the reaction of GSSH with either ferric cyt c or ABTS radicals indeed generates GSSSSG ( Fig. 6d and Extended Data Fig. 9c), implying the recombination of GSS• radicals. When we used GS 34 SSG to generate singly labeled GS 34   ). b, Influence of na 2 S (5 µM, substrate for SQR), 3MP (100 µM, substrate for MPST) or Cys (1 mM, substrate for CSE) on the radical load of aPEX2-expressing HeLa cells, added 5 minutes before triggering radical generation with luminol, substrate and H 2 O 2 (50 µM). n = 3. P = 0.0011, 0.0011 and 0.0136. c-f, Influence of CSE overexpression (c), CSE depletion (d), ETHE1 overexpression (e) and ETHE1 depletion (f) on the luminescence profiles recorded from aPEX2-expressing HeLa cells (left panels). Cells were incubated with luminol and substrate (250 µM each), and radical generation was triggered with H 2 O 2 (50 µM). normalized aUC (right panels). EV, empty vector. n = 4. P = 0.0056 (c). n = 3. P = 0.0009 (d). n = 4. P = 0.0001 (e). n = 4. P = 0.0016 (f). Data are presented as mean values. Error bars represent s.d. * P ≤ 0.05; ** P ≤ 0.01; *** P ≤ 0.001; and **** P ≤ 0.0001 based on a two-tailed unpaired t-test. a.u., arbitrary units; OE, overexpressing. radical recombination ( Fig. 6e and Supplementary Fig. 2). As predicted by the model, the catalytic cycle can also be initiated by providing GSSSSG instead of GSSSG ( Fig. 6f and Extended Data Fig. 9d). Furthermore, the model predicts competition between radical-dependent and GSH-dependent GSSH reduction. Indeed, we found that H 2 S formation is suppressed until cyt c is completely reduced (Fig. 6g and Extended Data Fig. 9e). Finally, QM calculations support the notion that the proposed persulfide cycle is thermodynamically favorable (Extended Data Fig. 9f). In conclusion, several independent lines of evidence support the concept of an autocatalytic persulfide cycle driven by radicals.

Discussion
In this study, we explored connections among Cys availability, hydropersulfide levels, free radical levels, LPO and ferroptosis sensitivity. Our starting point was the observation that cells primed to undergo ferroptosis increase their endogenous S 0 levels. This finding prompted us to consider the possibility that S 0 species (typically persulfides and polysulfides) somehow act against ferroptosis and that their upregulation is an adaptive response to pro-ferroptotic conditions. This hypothesis seemed reasonable, as S 0 species have previously been observed to be upregulated under oxidative stress conditions 26 and to exert cytoprotective effects 16,27 .
Cys is known to be the ultimate source of S 0 in persulfides and polysulfides 14,21 . Hence, we reasoned that Cys uptake may counteract ferroptosis not only through the well-established Cys-GSH-GPX4 axis but also by enabling biosynthesis of S 0 species. Indeed, we found that increased intracellular Cys availability can suppress lipid oxidation (and, hence, ferroptosis) independently of the canonical GPX4 system.
If persulfides naturally play a role in limiting LPO, the expression levels of enzymes involved in persulfide generation and turnover can be expected to influence lipid oxidation levels and ferroptosis sensitivity. To this end, we mainly focused on the manipulation of two enzymes. One enzyme, CSE, supports the biosynthesis of persulfides, primarily by generating H 2 S from Cys, but potentially also by generating cysteine persulfide from cystine 14 . The other enzyme, ETHE1, degrades persulfides (GSSH 14 . We found the two enzymes to have opposing effects on membrane integrity (Fig. 2c) and LPO (Fig. 2d,e). The pro-ferroptotic effect of CSE depletion aligns with a recent study reporting that depletion of cystathionine β-synthase (CBS), another enzyme capable of supporting persulfide biosynthesis, promotes ferroptosis in breast cancer cells. CBS depletion was observed to diminish persulfide levels, but not Cys or GSH levels, thus implicating its non-canonical H 2 S/S 0 -generating function in ferroptosis protection 28 . A potential caveat in interpreting CSE or CBS effects is that SQR-dependent oxidation of H 2 S (as generated by CSE or CBS) not only yields GSSH but also feeds electrons into the mitochondrial CoQ 10 pool. However, recent work on the CoQ oxidoreductase FSP1 has provided some clear indications that it is the extra-mitochondrial CoQ 10 pool at the plasma membrane that is relevant for ferroptosis protection 7,8 . Assessing the relative significance of individual S 0 -generating enzymes (including CSE, CBS, SQR, MPST and CARS) in ferroptosis, depending on cell type and metabolic context, will require further analysis.
A central finding of our study is that exogenously supplied or endogenously produced persulfides facilitate the rapid removal of endogenously produced free radicals. We used the engineered heme peroxidase APEX2 as a chemogenetic tool to generate free radicals and drive lipid oxidation. Spin trapping in living cells revealed that mostly glutathionyl radicals, previously shown to facilitate LPO 29 , were generated in this setting. Using luminescence-based real-time monitoring of radical levels, we observed that low concentrations of exogenously supplied membrane-permeable persulfide donors (CSSSC, Na 2 S 2 ) accelerate endogenous radical removal (Fig. 3). Manipulating CSE and ETHE1 levels in both directions, we found a consistent pattern in that the two enzymes mirrored each other: silencing CSE yielded the same effect (slower radical removal) as overexpressing ETHE1, and overexpressing CSE yielded the same effect (faster radical removal) as silencing ETHE1 (Fig. 4). In addition, we depleted another persulfide-generating enzyme, MPST, and also observed slower radical removal.
These observations raised the question of how hydropersulfides actually eliminate radicals. To approach this question, we studied persulfide-radical interactions in vitro. To our surprise, we observed non-stoichiometric effects-that is, more radicals were eliminated than persulfides were added to the system (Fig. 5). This led us to consider a three-step autocatalytic cycle in which persulfides are regenerated (Fig. 6a). We further realized that a catalytic role for persulfides in 1-electron reactions had already been proposed more than 50 years ago. In 1971, it was reported that, in the presence of excess GSH, one molecule of CSSSC led to the reduction of at least 25 molecules of cytochrome c, implicating hydropersulfides as catalysts 30 .
The first step of the proposed cycle is the reduction of radicals by hydropersulfides. Our in vitro experiments strongly suggest that hydropersulfides are highly efficient in directly reducing radicals. This conclusion is in line with and supported by recent findings. Hydropersulfides were found to be excellent H-atom donors to radicals, including alkyl, alkoxyl, peroxyl and thiyl radicals 31 . Their inherent reactivity to chain-carrying peroxyl radicals is four orders of magnitude greater than for thiols and essentially the same as that of α-tocopherol. However, hydropersulfides are superior to α-tocopherol owing to their low H-bond acidity, a unique attribute of hydropersulfides, which makes their reactivity largely insensitive to changes in the surrounding medium 20 . It may be argued that radical scavenging by hydropersulfides is nevertheless limited by their inherent instability, as they are subject to thiol-mediated reduction in the presence of millimolar amounts of GSH (GSSH + GSH → GSSG + H 2 S). However, hydropersulfides are continuously re-synthesized by enzymatic activity and available at micromolar steady-state concentrations 15 . When radicals appear, their reaction with persulfides should outcompete GSH-dependent persulfide reduction, which is a comparably slow process. Indeed, we observed in our in vitro experiments that the decay of persulfides to H 2 S was strongly suppressed in the presence of radicals. Hence, in the presence of radicals, other persulfide-degrading processes likely become kinetically irrelevant. The direct reduction of radicals by hydropersulfides leads to the formation of perthiyl radicals. These have unique properties. They are extraordinarily stable in terms of reactivity toward other molecules. They have a very low tendency to pass on oxidizing equivalents by extracting H-atoms from other molecules 31 . Nevertheless, perthiyl radicals have a fleeting (almost diffusion-limited) existence, as they have a strong tendency to recombine with each other. In other words, perthiyl radicals persist until they encounter another perthiyl radical with which they recombine to form tetrasulfides 18 . Perthiyl radicals are usually not observed to react with other radicals or with O 2 (refs. 32,33 ), probably because these reactions are outcompeted by rapid self-recombination. The second step in the cycle is, thus, the recombination of perthiyl radicals at a rate close to the diffusion limit 18 . Perthiyl radicals seem to be unique among all other biological radicals by almost exclusively engaging in self-recombination. Hence, the formation and recombination of perthiyl radicals efficiently terminates radical chains 18 . In contrast, the thiyl radical propagates radical chain reactions 34,35 , and even the long-lived ɑ-tocopherol radical does so 36 . The failure to directly detect GSS• by us and others 31,32 is in line with their stability toward other molecules (including spin traps) on the one hand and their rapid disappearance (through self-recombination) on the other. Accordingly, our evidence for their formation is indirect. However, we detected the product of recombination, GSSSSG, and confirmed that it was formed in a radical-dependent manner. Furthermore, isotopic labeling of GSSH led to the formation of doubly labeled GSSSSG, in support of perthiyl recombination (Fig. 6e).
The third step in the cycle is the reduction of GSSSSG to regenerate GSSH. In the presence of millimolar amounts of GSH, GSSSSG is subject to reduction. GSSSSG may, in principle, regenerate two GSSH molecules (net reaction: GSSSSG + 2 GSH → GSSG + 2 GSSH). It is conceivable that GSSSSG reduction can be catalyzed by enzymes. Glutathione reductase (GR) is known to accept GSSSG as a substrate 37 and may also act on GSSSSG. Additionally, GSSSSG may also act as a radical scavenger by undergoing replacement reactions with radicals (GSSSSG + R• → GSSR + GSS•) 38,39 , thus regenerating perthiyl radicals, which, again, recombine to yield GSSSSG. Irrespective of the detailed steps, the recycling of S 0 in a persulfide cycle may explain why low micromolar amounts of GSSH can be efficient radical scavengers in the presence of a large excess of GSH (Fig. 6h).
Another key question is how small molecule persulfides, such as GSSH, which are mostly hydrophilic, can intercept radical chain reactions taking place within the hydrophobic environment of a lipid bilayer. We can see several possibilities, which are not mutually exclusive. First, water-soluble hydropersulfides may reduce initiator radicals formed within the aqueous phase-for example, thiyl radicals (GS•), which are known to initiate LPO 29,40 . Intracellular APEX2 mostly generates GS• (Extended Data Fig. 3c) and, at the same time, drives lipid oxidation (Fig. 3a). In the same system, we observe the elimination of GS• in the presence of S 0 species (Fig. 3d,g). These findings are compatible with the idea that persulfides intercept initiator radicals before they can start new chain reactions. Second, water-soluble hydropersulfides may reduce α-tocopherol and/or ubiquinone (CoQ 10 ) radicals, which are accessible for interactions on the membrane surface. The coupling of lipophilic and hydrophilic radical scavengers is a well-known principle in redox biology. Tocopherols reduce lipid radicals within the membrane, and the resulting tocopherol radicals are then directly re-reduced by ascorbate 41 . Another example is lipid radical scavenging by CoQ 10 , which can also couple to water-soluble reductants. Indeed, the oxidoreductase FSP1 protects against LPO and ferroptosis by transferring electrons from NAD(P)H to oxidized CoQ 10 (refs. 7,8 ). Thus, it seems plausible that hydropersulfides such as GSSH have access and the ability to reduce tocopherol and/or ubiquinone radicals. It has been reported that GSH can couple to the tocopherol system by directly reducing the tocopherol radical 41 . Given the 100-mV lower redox potential of GSSH relative to GSH (Fig. 6b), GSSH would be a much better reductant for tocopherol radicals. If GSH can interact with tocopherol on the membrane surface, the same should be true for GSSH. Third, water-soluble hydropersulfides may directly reduce chain-propagating lipid radicals (LOO•/LO•) if the oxidized fatty acid chain swings out to the membrane surface, due to its increased polarity, as suggested previously 42 . Fourth, there is also the possibility that hydrophilic persulfides/polysulfides give rise to lipophilic persulfide species. However, it is currently unknown if lipophilic persulfides are produced inside mammalian cells.
Taken together, our results support the notion that hydropersulfides provide protection against free radicals and associated chain reactions, as they react rapidly and in a potentially self-regenerating manner. Notably, the hydropersulfide/perthiyl system appears to be distinguished by an exceptional combination of chemical properties. By avidly reducing radicals, hydropersulfides essentially act as a sink for unpaired electrons. The resulting perthiyl radicals are resonance stabilized and are, therefore, unlikely to pass on the unpaired electron status to other molecules. Instead, they have a strong tendency to pair their unpaired electrons among each other (that is, by dimerization), thus eliminating unpaired electrons in the most direct way. Given its simplicity, the hydropersulfide/ perthiyl system may represent an evolutionary ancient cellular radical removal system.

online content
Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/ s41589-022-01145-w.
Plasmids. The open reading frame (ORF) encoding APEX2 was obtained from Addgene (79057). Gateway full ORF clones encoding CSE and ETHE1 were obtained from the DKFZ Genomics & Proteomics Core Facility (Supplementary Table 1). All expression plasmids were constructed using the Gibson Assembly Cloning Kit (New England Biolabs). Primers for Gibson Assembly were designed using the NEBuilder Assembly tool. Expression plasmids used in this study were: pLPCX_APEX2-GFP, pLPCX_Su9-APEX2, pcDNA3.1_CSE, pcDNA3.1_ETHE1 and pTRC-APEX2.
Overexpression of enzymes. In total, 250,000 cells per well were seeded into a six-well plate 1 day before transfection. For each transfection, 5 µl of Lipofectamine 2000 (11668019, Invitrogen) was added to 100 µl of OptiMEM (Gibco), vortexed briefly and incubated for 5 minutes at room temperature. In parallel, plasmid DNA (2.5 µg for CSE and 1 µg for ETHE1) was added to 100 µl of OptiMEM. Empty vectors were used as negative controls. The two solutions were combined and vortexed for 20 seconds. The mixture was incubated at room temperature for 25 minutes and then added to the cells. After 6 hours, the medium was replaced with fresh medium.

Generation of HeLa cells stably expressing APEX2.
For generation of stable cell lines, the pLPCX retroviral expression vector encoding APEX2-GFP or mito-APEX2 (Su9-APEX2) was transfected into the packaging cell line Phoenix-AMPHO. After 24 hours, viral supernatant was collected, filtered through a 0.45-μm cellulose acetate filter and used to infect freshly thawed HeLa cells. Transduced cells were selected with puromycin, expanded and frozen for later use. Cells expressing APEX2-GFP were enriched by fluorescence-activated cell sorting (FACS), and cells expressing mito-APEX2 were enriched by puromycin selection.
Immunoblot analysis of protein expression. Cells were lysed with 0.1% Triton X-100 in TBS (10 mM Tris-HCl and 150 mM NaCl, pH 7.4) in the presence of protease inhibitors (complete, 04693132001, Roche). The lysate protein content was estimated using the BCA assay (23225, Thermo Fisher Scientific).
Lysate samples were dissolved in SDS-PAGE sample buffer containing 10 mM DTT. Samples containing 20 µg of protein were run on an SDS-PAGE gel and transferred to polyvinyl difluoride (PVDF) membranes (Immobilon-F, Millipore) using a transfer tank (TE22, Hoefer). Membranes were probed with appropriate antibodies, HRP-conjugated secondary antibodies and chemiluminescent substrate (SuperSignal West Femto, Thermo Fisher Scientific).
PrestoBlue cell viability assay. In total, 1,000 cells per well were seeded into black 96-well plates with a transparent bottom (655090, Greiner). After 96 hours, 10 µl of the PrestoBlue reagent (A13261, Invitrogen) was added to each well, and plates were incubated at 37 °C for 20 minutes. Fluorescence (ex 545 nm/em 600 nm) was recorded with a microplate reader (CLARIOstar, BMG).
Real-time monitoring of intracellular radical load. In total, 25,000 APEX2-expressing HeLa cells, suspended in FluoroBrite (Gibco, A1896701) with 2% FCS, were seeded per well into opaque white 96-well plates (Thermo Fisher Scientific, 136101). Stock solutions were prepared by dissolving luminol (2.5 mM) in 50 mM borate buffer (H 3 BO 3 , pH 9) and the peroxidase substrate HPI (2.5 mM) in PBS with 30% DMSO. After 18 hours, luminol and HPI were added to the wells, both to 250 µM final concentration. Luminescence was recorded with a microplate reader (PHERAstar, BMG). Twenty minutes after the start of the measurement, 50 µM of H 2 O 2 was added. After the measurement, cells were lysed with 0.1% Triton X-100 and stained with SYBR Green (S9430, Sigma-Aldrich) in the well. SYBR Green fluorescence (ex 480 nm/em 520 nm) was recorded with a microplate reader (PHERAstar, BMG). Luminescence was normalized to SYBR fluorescence. For quantitative comparisons of intracellular radical load, the area under the curve (AUC) was calculated with GraphPad Prism 8.
Expression and purification of recombinant APEX2. APEX2 was expressed from a plasmid (72558 pTRC-APEX2, Addgene) in Escherichia coli BL21(DE3) (EC0114, Thermo Fisher Scientific), as described previously 23 . In brief, 500 ml of Luria broth (LB) with 10 μg ml −1 of ampicillin was inoculated with a single colony. The culture was grown at 37 °C to an OD 600 of 1. Then, protein expression was induced with IPTG (420 μM). 5-aminolevulinic acid hydrochloride (1 mM) was added to promote heme biosynthesis. The culture was continued overnight at room temperature and then centrifuged for 10 minutes at 4,000 relative centrifugal force (RCF) and 4 °C. The dried bacterial pellet was solubilized in 20 ml of B-PER (78243, Thermo Fisher Scientific), supplemented with protease inhibitors (1 µg ml −1 of leupeptin, 1 µg ml −1 of AEBSF HCl and 1 µl of benzonase) and 5 mM imidazole and then transferred to a 50-ml centrifuge tube. The lysate was carefully mixed for 10 minutes at 4 °C and then centrifuged at 16,000 RCF for 30 minutes at 4 °C. The supernatant was collected. Ni-NTA agarose beads (30210, Qiagen) were washed three times with NH 4 HCO 3 buffer (50 mM NH 4 HCO 3 , pH 7.4). The bacterial lysis supernatant was added to 1 ml of Ni-NTA agarose beads and incubated for 1 hour at 4 °C under gentle rotation. The Ni-NTA bead suspension was transferred to 5-ml disposable chromatography columns (29922, Thermo Fisher Scientific). Beads were washed three times with 5 ml of wash buffer (50 mM NaH 2 PO 4 , 300 mM NaCl, 20 mM imidazole, 1 µg ml −1 of leupeptin, 1 µg ml −1 of AEBSF HCl and 1 µl benzonase, pH 8.0). The protein was eluted in three steps, first with 5 ml of elution buffer containing 50 mM imidazole (50 mM NaH 2 PO 4 , 300 mM NaCl and 50 mM imidazole, pH 8.0) and then with 5 ml of elution buffer containing 100 mM imidazole and finally with 5 ml of elution buffer containing 500 mM of imidazole. Eluates containing pure protein (as determined by SDS-PAGE) were combined. The combined protein sample was placed in a 10-kDa cutoff dialysis cassette (66382, Thermo Fisher Scientific) pre-equilibrated with PBS buffer for 5 minutes. The cassette was then incubated in 2 L of PBS buffer under constant stirring. Dialysis was performed overnight at 4 °C. Finally, 1-ml aliquots of dialyzed protein were stored at −80 °C. The final protein concentration was estimated using the BCA assay.
Electron spin resonance spectroscopy. For the analysis of in vitro samples, reactants, including the spin trap, were mixed in PBS containing diethylenetriaminepentaacetic acid (DTPA, 50 µM) to a final volume of 50 µl, using low protein binding tubes (90410, Thermo Fisher Scientific). For the analysis of intact cells, 5 × 10 6 cells were resuspended in PBS containing 50 µM DTPA, using low protein binding tubes (90410, Thermo Fisher Scientific). The spin trap and other reagents (peroxidase substrate, S 0 species and H 2 O 2 ) were added as described in the figure legends to a final volume of 50 µl. Samples were transferred to capillaries (ring caps, NOXygen), sealed with Critoseal (Noxygen) and immediately inserted into the cavity (ST9010) of the electron paramagnetic resonance (EPR) spectrometer (ESP300e, Bruker). Spectra were recorded and pre-processed with proprietary software (Lila-X and Medeia, provided by Gerhard Bracic). Spectrometer settings: microwave attenuation 10 dB (=20 mW microwave power); modulation amplitude 0.1 mT; receiver gain 60 dB; modulation frequency 100 kHz; conversion time 40 ms; time constant 20 ms; center field 339.5 mT; sweep width 14.0 mT. The microwave frequency was measured with an HP 5010 frequency counter. When necessary, spectra were accumulated to improve the signal-to-noise ratio. Control experiments were run with different combinations of reactants to identify unwanted side reactions. Spectra were simulated with the EasySpin package for MATLAB 47 .
Cyclic voltammetry. Ferrocene (0.5 mM) was mixed with GSH and GSSSG in N 2 -saturated Britton-Robinson buffer (0.04 M boric acid, 0.04 M phosphoric acid and 0.04 M acetic acid) at pH 9. Measurements were conducted with an Autolab PGSTAT12 potentiostat-galvanostat electrochemical system (Eco Chemie) using a conventional three-electrode setup. An Ag/AgCl (3 M KCl) electrode was used as a reference electrode, and a platinum wire served as the counter electrode. The working electrode was a glassy carbon electrode (Metrohm, 6.1204.300) with a diameter of 3 mm. It was cleaned before each experiment by polishing it with aluminum oxide powder (grain size 0.3 µm) (Alfa Aesar, 14558) for 30 seconds and then rinsing it with ethanol and deionized water. The potential step was 1 mV, and the scan rate was 10 mV s −1 or 20 mV s −1 . Measurements were carried out in de-aerated (N 2 -saturated) solutions at room temperature under quiescent conditions. Data were analyzed with General Purpose Electrochemical System (GPES) version 4.9 (Eco Chemie) software.
Amplex Red autoxidation assay. Amplex Red autoxidation (resorufin formation) was measured as described previously 24 . Reactants (as indicated in the figure legends) were mixed in PBS. Resorufin formation was monitored in 96-well plates by measuring absorbance at 562 nm (FLUOstar, BMG) or fluorescence (ex 571 nm/ em 585 nm) (CLARIOstar, BMG) at room temperature.
Cyt c reduction assay. Ferric cyt c (24 µM) was mixed with GSH (1 mM) and variable amounts of polysulfides (as indicated in the figure legend) in PBS. Absorbance at 550 nm was recorded with a microplate reader (FLUOstar, BMG) at room temperature.
TEMPOL reduction assay. TEMPOL (18 mM) was mixed with GSH (20 mM) and variable amounts of CSSSC (as indicated in the figure legend) in PBS containing DTPA (50 µM). Absorbance at 430 nm was recorded with a microplate reader (FLUOstar, BMG) at room temperature.
Analysis of GSSSSG generation. Ferric cyt c (400 µM) or ABTS (400 µM) was mixed with GSSSG or GS 34 SSG (400 µM), NADPH (400 µM) and GR (1 U ml −1 ) in a total volume of 100 µl in PBS. After 5 minutes, the reaction was stopped by the addition of 100 µl of MBB in MeOH (10 mM). The samples were incubated for 5 minutes in the dark, and then 200 µl of CHCl 3 was added to remove GR by precipitation. After centrifugation at 300 RCF, the upper phase was removed for further analysis. The sample (10 µl) was injected into an Agilent 1260 Infinity LC system attached to an Agilent 6120 Single Quadrupole MS with ESI source and evaporative light scattering detector (ELSD). Separation was performed on a Kinetex 2.6 μm C18 100 Å LC column (50 × 2.1 mm) at 40 °C using a flow rate of 0.6 ml min −1 . Solvent ' A' was 0.01% HCOOH in water; solvent 'B' was 0.01% HCOOH in MeCN. The method was: 100% A for 2 minutes, then from 100% to 10% A in 10 minutes and then 1% A for another 10-12 minutes. Data were processed with MestReNova (14.2.1) software.
Electrode measurement of H 2 S. Reactions were performed in 1 ml of PBS containing 1 mM of GSH, 50 µM of ferric cyt c and 10 µM of GSSSG. H 2 S release was recorded with a hydrogen sulfide electrode (TBR 4100, WPI) by submerging the sensor tip into the well of a 12-well plate. The plate was agitated on a rotary shaker (Titramax 101) with 150 r.p.m. to prevent the buildup of a diffusion layer around the electrode. Data were analyzed with LabScribe software. Synthesis of CSSSC. CSSSC was synthesized as described previously 48 . In brief, 1 g of cystine (CSSC) was dissolved in 1 M H 2 SO 4 and reacted with 2 equivalents of peracetic acid on ice for 2 hours. The product cystine-S-oxide was precipitated by neutralizing the solution with pyridine and thoroughly washed with EtOH and THF. Cystine-S-oxide was redissolved in 1 M H 2 SO 4 and reacted with 2 equivalents of NaSH for 2 hours at room temperature. The final product CSSSC was precipitated by neutralizing the solution with pyridine, thoroughly washed with EtOH and THF and then lyophilized, with a yield of approximately 25%. High-performance liquid chromatography-mass spectroscopy (HPLC-MS) analysis indicated >98% purity (Supplementary Note 1). QM calculations. Geometries were optimized, and Gibbs free energies were calculated with Gaussian 09 (ref. 49 ), using 6-31 + G(d) as the basis set. The Polarizable Continuum Model with water was used as the solvent model. For the rest, error bars represent SD, ** P ≤ 0.01 based on a two-tailed unpaired t-test.