Main

The major source of folate one-carbon (1C) units in mammalian cells is the amino acid serine1,2,3. Transfer of the 1C unit of serine to tetrahydrofolate (THF) can occur in either the cytosol or the mitochondrion, via the enzyme SHMT1 or SHMT2, respectively10 (Fig. 1a). Evidence from stable isotope tracing indicates that cancer cells predominantly use SHMT2 to catabolize serine, exporting the resulting 1C units to the cytosol to support nucleotide synthesis2,9. The extent to which 1C unit production via SHMT2 is also important to support mitochondrial health has yet to be determined.

Figure 1: Mitochondrial respiratory chain function is dependent on SHMT2 catalytic activity.
figure 1

a, 1C pathway and known mitochondrial products. b, Lactate secretion of HCT116-knockout cell lines (n = 6). ΔSHMT2-A and ΔSHMT2-B denote two separate SHMT2-knockout lines. WT, wild type. c, Oxygen consumption rate measured by Seahorse XF analyser (n = 3). FCCP denotes a mitochondrial uncoupling agent. Oligom., oligomycin; Rot./antim., rotenone/antimycin. d, Immunoblot for mitochondrial respiratory complex I and II (CI and CII) proteins (NDUFS4 and SDHA, respectively), 1C enzymes, and a marker of mitochondrial mass (VDAC1). e, Basal respiration (n = 3) upon re-expression of wild-type or catalytically deficient mutant forms of SHMT2 in HEK293T knockout cell lines. Data are mean ± s.e.m. n indicates the number of biological replicates, which for the Seahorse experiments refers to independent plates on separate days. *P < 0.01, two-tailed Student’s t-test (see Supplementary Table 7 for exact P values). Cat. inact., catalytically inactive SHMT2; PLP bind., PLP binding-deficient SHMT2.

PowerPoint slide

Source data

Full size image

When characterizing a set of human HCT116 colon cancer CRISPR-deletion cell lines that lack folate 1C enzymes, we discovered that the loss of SHMT2 induces a change in media colour indicative of enhanced extracellular acidification (Extended Data Fig. 1a). Quantitative analysis of the media confirmed increased glucose uptake and lactate secretion. This effect was specific to SHMT2: loss of other core 1C enzymes, including SHMT1 and mitochondrial enzymes such as MTHFD2 and MTHFD1L, did not induce glycolysis (Fig. 1b, Extended Data Fig. 1b).

A common cause of increased glycolytic flux is respiratory deficiency11. Loss of SHMT2 reduced both basal respiration and maximal respiratory capacity and decreased the NAD+/NADH ratio in several HCT116 (Fig. 1c, Extended Data Fig. 1c) and HEK293T SHMT2-knockout clones. Knockout cell lines lacking other core folate enzymes did not show impaired respiration (Extended Data Fig. 1d, e). Consistent with respiratory chain deficiency, the loss of SHMT2 decreased glucose flux into the tricarboxylic acid (TCA) cycle intermediate citrate, with an increased fraction of citrate instead being produced by reductive carboxylation12. As reported recently in other models of mitochondrial damage13,14, the pool size of TCA cycle metabolites and associated amino acids was also decreased (Extended Data Fig. 1f, g). To identify the cause of respiratory deficiency, we examined the abundances of several mitochondrial proteins, and found decreased abundances of complex I, IV and V subunits with retained levels of complexes II and III and markers of mitochondrial mass (Fig. 1d and Extended Data Fig. 1h, i). Thus, SHMT2 is required to maintain the levels of several mitochondrial respiratory chain proteins.

Given that the loss of SHMT2, but not of the immediate downstream enzymes of mitochondrial 1C metabolism, caused impaired oxidative phosphorylation, we wondered whether the phenotype reflected a requirement for the catalytic activity of SHMT2, or alternatively a non-catalytic role of SHMT2, perhaps related to its reported interaction with the mitochondrial nucleoid15. Accordingly, in the SHMT2-knockout background, we stably re-expressed catalytic inactive SHMT2 (p.Glu98Leu/p.Tyr106Phe), pyridoxal 5′-phosphate (PLP)-binding mutant SHMT2 (p.Lys280Gln) or wild-type SHMT2 protein (Extended Data Figs 2, 3a). Re-expression of wild-type protein, but not the catalytically inactive mutants, rescued the oxidative phosphorylation deficiency (Fig. 1e and Extended Data Fig. 3b) and normalized glycolytic flux (Extended Data Fig. 3c). In addition, SHMT2 re-expression rescued the growth defect of SHMT2-knockout cells9 (Extended Data Fig. 3d) and normalized 1C metabolism (Extended Data Fig. 3e, f). Thus, mitochondrial SHMT catalytic activity is crucial to sustain oxidative phosphorylation.

Two compartment-specific uses of mitochondrial folate 1C units have been reported: the local biosynthesis of deoxythymidine triphosphate (dTTP)16,17 and of N-formylmethionine (f-Met)18,19 (Fig. 2a). The production of dTTP requires 5,10-methylene-THF (methylene-THF), whereas f-Met requires 10-formyl-THF (formyl-THF). SHMT2 is upstream of both compounds. By contrast, MTHFD2 sits between methylene-THF and formyl-THF. The lack of an oxidative phosphorylation phenotype with MTHFD2 knockout led us to hypothesize that methylene-THF is the required 1C species. Consistent with this, SHMT2-knockout cell lines showed unchanged N-terminal f-Met levels of the mitochondrially translated COX1 peptide (encoded by the MT-CO1 gene)19 (Extended Data Fig. 4a). To confirm that methylene-THF is the required species, we generated SHMT2/MTHFD2 double-deletion cells and supplemented them with methylglycine (sarcosine), which can produce mitochondrial methylene-THF via sarcosine dehydrogenase. Although sarcosine is not a preferred 1C source and its feeding was insufficient to restore oxidative phosphorylation in SHMT2 single gene knockout cell lines, it fully restored oxidative phosphorylation in the SHMT2/MTHFD2 double-knockout cells, in which drainage of methylene-THF to formyl-THF is blocked (Fig. 2b and Extended Data Fig. 4b). Thus, mitochondrial methylene-THF is required to maintain respiratory capacity.

Figure 2: SHMT2-knockout-induced respiratory chain deficiency is caused by mitochondrial methylene-THF depletion but is unrelated to dTTP synthesis.
figure 2

a, Sarcosine serves as an SHMT2-independent source of mitochondrial methylene-THF. b, NAD+/NADH ratio (n = 6) and NDUFS4 (complex I) protein expression upon sarcosine supplementation (1 mM) in SHMT2 single-knockout (ΔSHMT2) and SHMT2/MTHFD2 double-knockout (ΔSHMT2/ΔMTHFD2) cell lines compared to wild-type cells. c, d, Functional readouts for mitochondrial dTTP status based on mitochondrial DNA (mtDNA) levels (n = 3) determined by quantitative PCR (qPCR; c) and gene expression determined by RNA-seq (d) in SHMT2-knockout and wild-type HEK293T cells. RPKM, reads per kilobase per million mapped reads. In d, each data point represents the mean gene expression of two biological replicates of two independent knockout clones (n = 4) and two wild-type replicates (n = 2). Genes linked to OXPHOS function37 are highlighted in red (nuclear-encoded) or blue (mitochondrial-encoded). Significantly differentially expressed genes are shown in Supplementary Table 2. Data are mean ± s.e.m. n indicates the number of independent biological replicates. *P < 0.01, two-tailed Student’s t-test (see Supplementary Table 7 for exact P values).

PowerPoint slide

Full size image

Depletion of mitochondrial dTTP and/or accumulation of uridine nucleotides have been shown to induce mitochondrial respiratory chain deficiency by promoting mitochondrial DNA damage17,20. However, SHMT2-knockout cells showed no evidence of altered mitochondrial DNA copy number (Fig. 2c and Extended Data Fig. 4c), deletions (Extended Data Fig. 4d), or mutations (Extended Data Fig. 4g, h and Supplementary Table 1). Moreover, whole-cell RNA sequencing revealed normal transcript levels for both nuclear and mitochondrial-encoded respiratory chain protein subunits (Fig. 2d, Extended Data Fig. 4e, f). Thus, the dependence of oxidative phosphorylation on SHMT2 reflects a requirement for mitochondrial methylene-THF for a purpose other than supplying local dTTP to maintain mitochondrial DNA.

Whereas the vast majority of the approximately 1,100 mitochondrial proteins are imported from the cytosol, 13 essential respiratory chain subunits are locally transcribed and translated21. These include components of complexes I, III, IV and V, but not complex II. On the basis of the normal mitochondrial transcript abundances and complex II protein levels, we hypothesized that mitochondrial methylene-THF is required for local translation. Indeed, [35S]methionine incorporation assays showed decreased synthesis of certain complex I and IV subunits (Extended Data Fig. 5a). To probe mitochondrial translation further, we developed a protocol for mitochondrial ribosome profiling based on digesting unprotected mRNA with micrococcal nuclease, enriching the 55S mitochondrial ribosome, and sequencing the protected footprints (Fig. 3a and Extended Data Fig. 5b). This approach achieved more than 90% average mitochondrial transcript sequence coverage, with an average depth of at least 80 reads per codon (Supplementary Table 3 and Extended Data Fig. 5c). In SHMT2-knockout cells, the distribution of ribosome-protected footprints showed pronounced stalling at defined codon positions (Fig. 3b and Extended Data Fig. 6a). This resulted in relatively fewer actively translating ribosomes (that is, bound and not stalled) for certain subunits of respiratory chain complexes I, IV and V (Extended Data Fig. 6b). Consistent with the ribosome profiling data, enzymatic assays revealed decreased activity of complexes I, IV and V (Extended Data Fig. 6c).

Figure 3: Mitochondrial ribosome profiling reveals that SHMT2-knockout cells are deficient in translating specific guanosine-ending codons.
figure 3

a, Workflow of mitochondrial ribosome profiling. Translation was halted using chloramphenicol and immersion into liquid nitrogen, cells were lysed and RNA was digested using micrococcal nuclease (MNase). After sucrose-gradient enrichment for mitochondrial ribosomes (shaded in red), protected fragments were sequenced. b, Mean cumulative ribosome density along selected mitochondrial transcripts. Additional transcripts are given in Extended Data Fig. 6a. RPF, ribosome-protected fragment; RPM, reads per million. c, Mean codon-specific mitochondrial ribosome occupancy in HCT116 cells (∆SHMT2/wild type). Red data points correspond to codons that are decoded by tRNAs carrying the 5-taurinomethyluridine (τm5U) modification. Red labels correspond to the subset of codons that end in guanosine and thus require wobble-base pairing. Methionine codons are highlighted in blue and show no increased codon occupancy. The insert shows mean normalized ribosome density relative to UUG and AAG codon position. aa, amino acids. Data in b and c represent two technical replicates of two independent samples.

PowerPoint slide

Source data

Full size image

We next aimed to determine the cause of ribosomal stalling. The aminoacyl-tRNA acceptor-site (A-site) coordinates of stalled ribosomes revealed notable ribosome accumulation in SHMT2-knockout cells at particular lysine and leucine codons: LysAAG and LeuUUG (Fig. 3c). This did not seem to reflect a shortage of these amino acids or tRNAs in mitochondria, as much less stalling was observed at the corresponding LysAAA and LeuUUA codons. Instead, it appeared to relate to difficulty in reading the 3′ codon guanosine of certain codons: those in which the 3′ position identity (purine versus pyrimidine) determines the encoded amino acid (‘split codon boxes’; Extended Data Fig. 6d). Increased codon occupancy was observed also for TrpUGG, GluGAG and GlnCAG, but not the corresponding codons with 3′ adenosine. The sole exception was MetAUG/AUA, in which no stalling was observed. Decoding of A/G-ending codons in split codon boxes is facilitated by methyl-derivative base modifications of the tRNA anticodon 5′ nucleotide, allowing non-Watson–Crick base-paring with the codon 3′ base22 (Fig. 4a). The mitochondrial tRNAMet anticodon has a 5′ cytidine that is formylated with the 1C unit derived from S-adenosyl-methionine (SAM)23. By contrast, the mitochondrial tRNAs for Lys, Leu1 (one of two mitochondrial leucine tRNAs), Trp, Glu and Gln have uridine at the 5′ anticodon position24. Mammalian cytosolic tRNAs with uridine at the anticodon 5′ position are modified to produce 5-methoxycarbonylmethyluridine (mcm5U), with the methoxy carbon derived from SAM25,26. Corresponding mitochondrial tRNAs are 5-taurinomethyl modified at the 5′ anticodon uridine (τm5U at position 34 of the tRNA)27,28,29. As SHMT2/MTHFD2 double-knockout cell lines supplemented with sarcosine showed rescue of codon-specific stalling on mitochondrial ribosome profiling (Extended Data Fig. 6e), we hypothesized that the τm5U modification is dependent on mitochondrial methylene-THF and that the observed oxidative phosphorylation defect is a consequence of impaired translation due to defective tRNA modification.

Figure 4: MTO1/GTPBP3-dependent tRNA methylation requires mitochondrial methylene-THF.
figure 4

a, Interaction of tRNA position 34 anticodon loop modified base with mRNA codon 3 position A/G, forming a non-Watson–Crick base pair. b, Total ion chromatogram of 5-taurinomethyluridine monophosphate (p-τm5U) (m/z = 460.043) from digested mitochondrial tRNAs. 5-formylcytidine monophosphate was not altered (Extended Data Fig. 7a). c, Mean codon-specific mitochondrial ribosome occupancy for MTO1-knockout (ΔMTO1) HCT116 cell lines and primary patient-derived fibroblasts carrying MTO1 mutations or the MT-TL1 m.3243A>G MELAS variant (n = 2). Corresponding immunoblots are shown below. Individual patient data are in Extended Data Fig. 9a. d, Basal respiration rates measured using the Seahorse XF analyser. Data were collected after growth in the absence (−) of folate for 5 passages or in the presence of the indicated methotrexate (MTX) concentration for 96 h (n = 3, except HCT116 WT n = 4 and MTX 50 nM n = 6). HT, 100 μM hypoxanthine and 16 μM thymidine. Data are mean ± s.e.m. n indicates the number of biological replicates. *P < 0.01, two-tailed Student’s t-test (see Supplementary Table 7 for exact P values).

PowerPoint slide

Source data

Full size image

To test this hypothesis, we established a liquid chromatography–mass spectrometry (LC–MS) method for detection of modified mitochondrial tRNA bases. In SHMT2-knockout cell lines, levels of formylcytidine were unchanged (Extended Data Fig. 7a). By contrast, τm5U and its 2-thio derivative (in which the uracil 2-position oxygen is replaced with sulfur) were depleted to undetectable levels (Fig. 4b and Extended Data Fig. 7b). The depletion was not due to a lack of taurine, the cellular levels of which tended to be higher (Extended Data Fig. 7c). The loss of τm5U and its 2-thio derivative was reversed upon re-expression of wild-type SHMT2 and also upon sarcosine supplementation of SHMT2/MTHFD2 double-knockout cells (Extended Data Fig. 7d, e). Thus, SHMT2 or an alternative source of mitochondrial methylene-THF is required to produce τm5U.

To verify that the methyl group of τm5U is coming from methylene-THF made by SHMT2, we conducted stable isotope tracing with 13C-labelled serine (the SHMT2 substrate) versus [13C]methionine (which feeds into SAM). As expected, methionine labelled formylcytidine. By contrast, serine labelled τm5U, but methionine did not (Extended Data Fig. 7f). Thus, SHMT2 is required for oxidative phosphorylation because it produces mitochondrial methylene-THF for tRNA taurinomethylation. To our knowledge, this is the first direct evidence of folate-dependent macromolecule modification in mammalian cells.

The enzyme complex that catalyses the τm5U base modification in mitochondria comprises two proteins, MTO1 and GTPBP3, both of which are encoded by genes that are mutated in human mitochondrial diseases30,31. Orthologues of the enzyme complex, forming the prokaryotic tRNA-modifying MnmE/GidA complex, have been shown to use THF-bound 1C units32. We therefore explored whether the mitochondrial translation defect upon SHMT2 loss matches human diseases proposed to affect taurinomethylation of mitochondrial tRNAs27,28,30,31. Indeed, MTO1 knockout in HCT116 cells caused loss of τm5U (Extended Data Fig. 7e). Mitochondrial ribosome profiling of these engineered MTO1-deletion cells showed an increase in codon occupancy at the same codons as with SHMT2 knockout (Fig. 4c). In both cases, the strongest stalling occurred at LysAAG and LeuUUG, potentially because these codons have, in addition to the wobble base pairing, only A–U base pairings versus the stronger C–G base pairings in GluGAG, GlnCUG and TrpUGG codons33. Mitochondrial ribosome profiling of primary fibroblasts of two patients carrying different MTO1 missense mutations revealed increased codon occupancy selectively at LysAAG, but not LeuUUG. Thus, these hypomorphic mutations seem to cause a selective defect in taurinomethylation of the Lys tRNA. Of all adenosine-ending codons, LysAAA showed the greatest increase in codon occupancy across all cell lines, being most pronounced in the MTO1 patients (Fig. 4c).

Notably, in both SHMT2 and MTO1 deletion cells, stalling did not affect all AAG and UUG codons uniformly, but occurred most strongly at the same specific gene locations (Extended Data Fig. 7g). Mapping of AAG and UUG codons at stalling sites relative to mRNA secondary structure34 did not reveal any clear pattern (Extended Data Fig. 8a, b and Supplementary Tables 4, 5). Mapping onto the structure of the protein being synthesized (Extended Data Fig. 8c), however, showed a trend towards stalling at transitions between transmembrane helices and non-membrane domains (Supplementary Table 5). Thus, stalling due to the defective codon–anticodon interaction might be exacerbated by particular protein sequence features.

Defined mutations in the Lys and Leu1 mitochondrial tRNAs (which decode the most strongly affected LysAAG and LeuUUG codons) result in a tRNA-specific τm5U modification defect, causing the human mitochondrial disorders MERRF and MELAS27,28. Mitochondrial ribosome profiling of fibroblasts from two patients with MELAS due to the m.3243A>G mutation in the MT-TL1 gene revealed, as expected35, increased occupancy at LeuUUG but not LysAAG or LeuUUA (Fig. 4d). For unknown reasons, increased ribosome occupancy for either SerAGU or ThrACG was also observed in individual patients (Extended Data Fig. 9a). The extent of stalling and complex I depletion was less than with nuclear SHMT2 or MTO1 mutations, presumably owing to the heteroplasmic nature of the mitochondrial tRNA mutation (Fig. 4c and Extended Data Fig. 9b). Collectively these observations highlight a common biochemical mechanism that links mitochondrial folate metabolism with a considerable fraction of inborn errors of mitochondrial metabolism27,28,30,31.

On the basis of these observations, we wondered whether folate deficiency, which in pregnancy causes neural tube defects1,2,3, could also result in mitochondrial impairment. Consistent with such a possibility, growth of HCT116 wild-type cells in folate-deficient media resulted in decreased levels of complex I enzymes and a reduced basal respiration rate (Fig. 4d and Extended Data Fig. 10a), without affecting mitochondrial DNA content (Extended Data Fig. 10b). In individuals with adequate nutrition, functional folate deficiency can be caused by antifolate therapy for autoimmunity or cancer. Low doses of the antifolate methotrexate, resulting in up to 100 nM circulating drug concentrations, are commonly used to treat autoimmune diseases such as rheumatoid arthritis4,36. In cell culture, nanomolar concentrations of methotrexate resulted in decreased levels of complex I enzymes and oxygen consumption (Fig. 4d and Extended Data Fig. 10c). Thus, the essential role for folate metabolism in mitochondrial translation may contribute to the clinical manifestations of folate deficiency and the clinical efficacy of antifolate therapies.

Methods

Cell lines and growth conditions

HCT116 (CCL-247) and HEK293T/17 (CRL-11268) were purchased from ATCC. Generation of a subset of clonal CRISPR–Cas9 knockout cell lines and detailed characterization has been reported previously9. Additional clonal knockout cell lines (Supplementary Table 6a) were established following the protocol published previously38. In brief, exon-targeting guide RNAs (Supplementary Table 6b) were designed against genes of interest and cloned into an expression vector containing the double nicking Cas9 variant (Addgene). Cells were transiently transfected using Lipofectamine 2000 (Life Technologies) (HEK293T) or Fugene HD (Promega) (HCT-116) and selected for 48 h with 2 μg ml−1 puromycin. Single clones were isolated using serial dilution into 96-well plates.

Stable SHMT2 re-expression was achieved by transfecting HEK293T knockout cell lines with NM_005412.5 cDNA (GE Healthcare) cloned into pCMV-Tag8 vector (Agilent) and selection for three weeks with 200 μg ml−1 hygromycine B (Sigma-Aldrich). Catalytic inactive (p.Glu98Leu/p.Tyr106Phe)39,40 and PLP-binding deficient (p.Lys280Gln)41 mutants were obtained following the QuickChange II protocol (Agilent). Knockout and re-expression cell lines were functionally verified by immunoblotting followed by targeted genomic sequencing (Supplementary Fig. 9) and, in the case of SHMT2 cell lines, also by tracing of [2,3,3-2H]serine labelling into dTTP. MTO1 and MT-TL1 patient fibroblasts and controls were provided by the Department of Paediatrics, Salzburger Landeskliniken and Paracelsus Medical University, Salzburg. Studies with primary human cell lines were approved by the local ethics-committee and informed consent was obtained from all subjects. The genotypes of MTO1-deficient patients (GenBank NM_012123.3) were as follows: patient-A c.[1261-5T>G];[1430G>A], (p.[?];[Arg477His]); patient-B c.[1222T>A];[1222T>A], (p.[Ile408Phe]; [Ile408Phe]). Patient-B MTO1 has been reported before42. Both MELAS patients carried the common MT-TL1 m.3243A>G mutation with the heteroplasmy rate reported in this study (Extended Data Fig. 9b). All cell lines tested negative for mycoplasma and were cultured in DMEM without sodium pyruvate (Sigma-Aldrich) supplemented with 10% dialysed fetal bovine serum (dFBS, GE Healthcare) in a 5% CO2 incubator at 37 °C. No antibiotics were used.

Glucose uptake and lactate secretion

Cells were seeded in 6-well plates 24 h before the start of an experiment. After reaching 50% confluency, plates were washed with PBS (GE Healthcare) and 3 ml of fresh media was added. Glucose uptake and lactate secretion were determined using an YSI 2900D Biochemistry Analyzer (Xylem Analytics) and normalized to cell growth as determined by microlitre packed cell volume (PCV; as measured using packed cell volume microfuge tubes from TPP).

Proliferation

Proliferation assays were conducted in 96-well plates and relative cell number was measured using resazurin sodium salt. Approximately 5,000 cells were plated in each well with 150 μl DMEM supplemented with 10% dFBS. Cell growth at each day was read as fluorescence intensity using a Synergy HT plate reader (BioTek Instruments).

Oxygen consumption

Oxygen consumption rates were measured on a XF24 extracellular flux analyser (Agilent) following the manufacturer’s instructions. In brief, XF24 cell culture microplates were coated with fibronectin (Sigma-Aldrich) and cells were seeded at 5 × 104 (HEK293T) and 7 × 104 (HCT116) cells per well. After reaching 70–90% confluency, cells were equilibrated for 1 h in XF assay medium supplemented with 10 mM glucose, 1 mM sodium pyruvate, and 2 mM glutamine in a non-CO2 incubator. Oxygen consumption rates were monitored at baseline and throughout sequential injections of oligomycin (1 μM), carbonyl cyanide-4-(trifluoromethoxy)phenylhydrazone (1 μM) and rotenone/antimycin A (0.5 μM each). Data for each well were normalized to cell number as determined by CyQUANT Cell Proliferation Assay Kit (Invitrogen) and to the on-plate wild-type control. For absolute oxygen consumption rate values, the protein concentration was determined by BCA protein assay (Thermo Fisher Scientific).

Immunoblotting

Cells were cultured to sub-confluency in 6 cm plates. After removal of media, cells were rinsed with 4 °C PBS and lysed in radio-immunoprecipitation assay (RIPA) buffer with phosphatase and protease inhibitors (Roche). Lysates were cleared by 10 min centrifugation at 16,000 g and quantified using a BCA assay (Pierce). Samples were resolved by SDS–PAGE on precast gels (Bio-Rad) and transferred to a nitrocellulose membrane using the Trans-Blot Turbo system (Bio-Rad). After overnight incubation with primary antibodies, bands were visualized with horseradish peroxidase-conjugated secondary antibodies (Cell Signaling Technologies). ChemiDoc XRS+ system was used for image acquisition. The following antibodies were used according to their manufacturer’s directions: anti-SHMT1 (12612), anti-SHMT2 (12762), anti-MRPL11 (2066), anti-S6RP (2217) and anti-b-actin (5125) were from Cell Signaling Technologies; anti-MTHFD2 (ab151447), anti-NDUFS4 (ab139178), anti-SDHA (ab14715) and anti-VDAC (ab14734) were from Abcam; anti-MTHFD1L (HPA029041) was from Sigma-Aldrich.

Analysis of mitochondrial specific translation

35S-labelling of mitochondrial proteins was performed following the method described previously43. In brief, cells were grown on 6 cm plates for 48 h to sub-confluency. Then media was changed to DMEM with 10% dFBS without methionine (MP Biomedicals). After a 30 min incubation, cytosolic translation was inhibited by emetine hydrochloride (0.1 mg ml−1, Sigma-Aldrich) and labelling was conducted for 1 h after the addition of 500 μCi [35S]methionine (EasyTag L[35S]-Methionine, Perkin Elmer). Protein lysates (30 μg) were then separated on a 15% polyacrylamide gel (8.3 × 7.3 cm) and dried using a 443 Slab Dryer (BioRad). The dried gel was exposed to a storage phosphor screen (GE Healthcare) and imaged on a Typhoon FLA 9500 (GE Healthcare). Equal sample loading was confirmed by Coomassie brilliant blue staining (BioRad).

Metabolite concentrations and labelling patterns

Cells were grown in 6-cm dishes for at least 48 h and collected at 75% confluency. Media was replaced every 24 h and additionally 6 h before collection. Metabolism was quenched and metabolites were extracted by aspirating media and immediately adding 1 ml of 80:20 methanol:water at −80 °C. Plates were kept on ice, scraped and non-soluble debris was pelleted at 18,000 g for 10 min. Samples were directly analysed by hydrophilic interaction chromatography coupled with negative-mode electrospray-ionization high resolution mass spectrometry on a quadrupole-orbitrap scanning from m/z 73 to 1,000 at 1 Hz and 140,000 resolution (Q Exactive Plus, Thermo-Fisher). Liquid chromatography separation was achieved on a XBridge BEH Amide column (2.1 mm × 150 mm, 2.5 μm particle size, 130 Å pore size; Waters) using a gradient of solvent A (20 mM ammonium acetate + 20 mM ammonium hydroxide in 95:5 water:acetonitrile, pH 9.45) and solvent B (acetonitrile). Flow rate was 150 μl min−1. The gradient was: 0 min, 85% B; 2 min, 85% B; 3 min, 60% B; 9 min, 60% B; 9.5 min, 35% B; 12 min, 35% B; 12.5 min, 0% B; 18 min, 0% B; 18.5 min, 85% B; 23 min, 85% B. Data were processed and analysed using MAVEN software44. All isotope tracer experiments were conducted at isotopic steady state: [2,3,3-2H]serine was traced into dTTP for a minimum of 6 h and [3-13C]serine and [U-13C]methionine were traced into tRNA for a minimum of 4 days. Isotopic tracers were purchased from Cambridge Isotope Laboratories. Isotopically labelled media was prepared from scratch and supplemented with 10% dFBS.

RNA sequencing

RNA was isolated from cell lines using RNeasy Plus kit (Qiagen) according to the manufacturers’ recommendation. After the depletion of ribosomal RNA, libraries were prepared according to the TruSeq Stranded Total RNA protocol (Illumina) and sequencing was performed on a HiSeq 2500 (Illumina). Analysis was performed using the Galaxy system45 and the R software package46. Adaptor sequences were trimmed using Cutadapt (Galaxy version 1.6)47 and the trimmed reads were then mapped with TopHat (Galaxy Version 0.9)48 to the GRCh38 reference using ENSEMBL version 80 genes as known splice junctions. The read counts per gene were determined using htseq-count (Galaxy Version 0.6.1galaxy1)49 in ‘union’ mode. Differential expression analysis was performed in R using DESeq2 1.12.3 package50. SHMT2-knockout gene expression (log2 RPKM) was graphed relative to the wild-type and the re-expressed cell lines.

Mitochondrial DNA content and integrity

To analyse mitochondrial DNA content, total DNA was extracted from 7 × 106 cells using Gentra Puregene Cell Kit (Qiagen) after freezing the cell pellet at −80 °C for 1 h and overnight digestion with Proteinase K (Roche Diagnostics). qPCR (ViiA 7, Applied Biosystems) was performed using primers targeting the mitochondrial ND2 locus (also known as MT-ND2 pseudogene 1) (forward: 5′-TGTTGGTTATACCCTTCCCGTACTA-3′; reverse: 5′-CCTGCAAAGATGGTAGAGTAGATGA-3′) and a nuclear ALU repeat sequence (forward: 5′-CTTGCAGTGAGCCGAGATT-3′; reverse: 5′-GAGACGGAGTCTCGCTCTGTC-3′) as published earlier51. The relative mitochondrial DNA content was determined using the ∆∆Ct method. Each independent sample given in the figures represents the mean of 6 technical replicates.

The mitochondrial genome was screened for deletions by long range PCR with two primer pairs spanning the whole coding region (forward_1: CCAACCAAACCCCAAAGAC, reverse_1: TACTGCGACATAGGGTGCTC; forward_2: CACCAGCCTAACCAGATTTCA reverse_2: tggtacccaaatctgcttcc) and products run on a 1% agarose gel52. DNA from a mitochondrial DNA deletion patient was used as positive control.

For mitochondrial genomic sequence analysis, mtDNA was enriched using the multiple displacement amplification strategy (REPLI-g, Qiagen) and sequencing was performed on a HiSeq 2500 after library preparation following the Nextera library prep kit protocol (Illumina). Reads were mapped to GRCh38 using Bowtie2 (Galaxy Version 0.6)53 with default settings. Coverage plots were generated using DeepTools bamCoverage (Galaxy Version 2.3.6.0)54. The data were normalized to 1× coverage using an effective genome size of 16,569. Freebayes (Galaxy Version 0.4.1)55 with frequency-based pooled settings was used to generate the variant data. The figures were generated in R using Gviz 1.18.056.

The MT-TL1 mutation load was determined using primers specifically spanning the m.3243 position for targeted enrichment (forward: 5′-AATGATACGGCGACCACCGAGATCTACACNNNNNNGCCTTCCCCCGTAAATGATA-3′, reverse: 5′-CAAGCAGAAGACGGCATACGAGATCGTCAGCGAAGGGTTGTAGT-3′) followed by sequencing on a MiSeq nano flow cell using a custom sequencing primer (sequence: TATTATACCCACACCCACCCAAGAACAGGGTTTGTTAAG). Alignment to GRCh38 was performed using Bowtie2 (Galaxy Version 0.6)53 at default settings and position-specific mutation load was derived from the Integrative Genomics Viewer.

Mitochondrial ribosome profiling

Development of our ribosome profiling method was based on concepts reported previously57,58,59. For mitochondrial ribosome profiling, cell lines were grown on 15-cm plates to 70–85% confluency. Sarcosine rescue of ribosome stalling in the SHMT2/MTHFD2 double-knockout background was assessed after growth in the presence of 1 mM sarcosine for 5 days. After removal of media, plates were rapidly rinsed with ice-cold PBS containing chloramphenicol (100 μg ml−1) (Sigma-Aldrich) and cycloheximide (100 μg ml−1) (Sigma-Aldrich) followed by immediate immersion into liquid nitrogen. Plates were then transferred to wet ice and 1 ml of 1.5× lysis buffer was added and the lysate was collected using a cell scraper. Lysis buffer contained the following: 1.5% Triton X-100 (Sigma-Aldrich), 0.15% NP40 (Sigma-Aldrich), 1× complete phosphatase and protease inhibitors (Roche), and 30 U ml−1 DNase 1 (Roche) in buffer base (20 mM Tris–HCl pH 7.8 (Ambion), 100 mM KCl (Ambion), 10 mM MgCl2 (Ambion), 100 μg ml−1 chloramphenicol, 100 μg ml−1 cycloheximide). 1.6–1.8 ml were recovered per plate and homogenized by passing three times through a 32G needle at 4 °C. Non-soluble debris was pelleted at 5,000 g for 10 min and 1,520 μl supernatant was used for digestion with 7,500 U ml−1 micrococcal nuclease (Roche) after adding 40 μl SUPERaseIN (Ambion) and 5 mM CaCl2 (Ambion). Digestion was stopped after 1 h gentle shaking at 25 °C using a final concentration of 6 mM EGTA.

Buffer base was used to make 5%–45% sucrose gradients (Gradient Master, Biocomp). After cooling to 4 °C, samples were separated in an ultracentrifuge using the SW-41Ti rotor at 210,000g for 2.5 h. Live UV absorption at 254 nm was used to track the mitochondrial 55S monosome enriched fractions (Extended Data Fig. 5b). The 55S fractions were pooled and mixed with 57 μl 20% SDS per millilitre sample before performing acid phenol chloroform RNA extraction. RNA was precipitated using 300 mM sodium acetate pH 5.5 and equal volume isopropanol and run on a 15% TBE-urea gel (Invitrogen) at 210 V for 1 h for size selection. Gels were stained with Sybr Gold (Invitrogen) and RNA fragments corresponding to mitochondrial ribosome footprints (approximately 28–40 nucleotides) were cut and recovered from the gel using the crush and soak method. After sodium acetate/isopropanol precipitation, library preparation was conducted following the TruSeq Ribo Profile (Illumina) protocol.

Sequencing of ribosome protected footprints (RPFs) was performed on a HiSeq 2500 in rapid mode followed by adaptor trimming using Cutadapt (Galaxy Version 1.6)47. Reads were mapped to the human genome reference GRCh38 using BWA (Galaxy Version 0.9)48 with those mapping to the mitochondrial protein-coding genes included in the subsequent analysis. The Plastid package60 and customized Phyton and R46 scripts were used for analysing mitochondrial ribosome profiling data. Alignment was performed from 3′ of reads which was reported to yield superior results after digestion with Micrococcal nuclease61,62. Each read, corresponding to a mitochondrial ribosome protected fragment (mtRPF), was assigned to a nucleotide position representing the respective ribosomal A-site as determined by metagene analysis60. mtRPF counts were then normalized to reads per million (RPM) mapped reads within each sample and single nucleotide positions were grouped by codon index. This transformation allows for relative quantification of bound ribosomes for each nucleotide triplet along a transcript. Stalling plots were created by plotting the mean cumulative mtRPF count along each mitochondrial open reading frame.

Codons were defined as stalling sites when the normalized counts mapped to the specific codon (mtRPF codon/mtRPF gene median) exceeded 2 s.d. from all codons in the genome. The relative abundance of actively translating ribosomes (that is, not stalled) was calculated by subtracting mtRPF counts in stalled regions from the total sum of ribosome footprints for each gene as ΣmtRPFactive = ΣmtRPFtotal − ΣmtRPFstalled. Then the gene specific ratio was plotted as ΣmtRPFactive,ΔSHMT2/ΣmtRPFactive,WT. Stalling sites specific to the SHMT2-knockout condition were identified using the ratio of occupancy at each codon position relative to wild type. Specifically, codons were defined as SHMT2-specific stalling sites when the normalized counts in the mutant relative to wild type (ΣmtRPFcodonΔSHMT2/ΣmtRPF codonWT) exceeded 2 s.d. (or, as indicated, 3 s.d.) from this ratio as determined for all codons in the genome, and the site also met the general stalling site criterion.

To determine the relative abundance of mitochondrial ribosomes bound to each nucleotide triplet, codon-specific occupancy ratios were calculated. For each codon (codoni = 1–64), the gene-specific ratio between experimentally measured ribosome density and expected density (which is proportional to codon frequency) was calculated. Codon occupancy (COi = 1–64) for each codon is the mean of the ratios from all 13 genes. The relative codon occupancy (COi = 1–64,ΔSHMT2)/(COi = 1–64,WT) was plotted with error bars representing s.d. across replicates after error propagation. To investigate the ribosome distribution relative to the major stalled codons (AAG and UUG), ribosome densities flanking the codons of interest within 25 amino acids were selected. Each selected fragment was first normalized to its total count so every codon of interest from the genome is weighted equally. The mean value from each position was plotted.

Investigating protein secondary structure effects on stalling at AAG and UUG

On average stalling was most pronounced at AAG and UUG codons, but not all codons of the same sequence were equally affected. We therefore investigated whether the positioning relative to protein secondary structures (transmembrane helices) influences the extent of stalling. As micrococcal nuclease treatment induces imprecision in A-site mapping owing to sequence-biased digestion63,64, individual positions identified as SHMT2-specific stalling sites were first grouped to the adjacent codons decoded by 5-taurinomethyluridine-modified tRNAs. Then amino acid residues corresponding to the AAG and UUG codons were mapped to Bos taurus crystal structures of mitochondrial respiratory chain complex proteins using iCn3D65. Structure data was retrieved using the following Protein Data Bank (PDB) accession codes: ATP6: 5ARA_W; MT-CO2: 2Y69_B; MT-CYB: 1QCR_C; and MT-ND6: 5LDW_J (underscored suffixes denote the respective subunits of the supercomplexes). In addition, a hidden Markov model based algorithm for transmembrane helices (TMHMM 2.066) was used to predict α-helical transmembrane domains in the Homo sapiens sequences. This method assigns each codon a probability for transmembrane helix localization which was then used for genome wide assessment of AAG and UUG localization relative to transmembrane helices. AAG and UUG codons were defined to be at a transition between a transmembrane helix and a non-membrane region if, within the five flanking codons, probabilities >0.5 and <0.5 for being in a transmembrane helix are found. AAG and UUG codons were defined as stalling sites based on the 3 s.d. cut-off as per Extended Data Fig. 8b. In total, 4 out of 5 stalling AAG and UUG codons, and 7 out of 23 non-stalling AAG and UUG codons were at a membrane transition (P = 0.04 by chi-square test).

Evaluation of mRNA secondary structure effects on ribosome stalling

To study a potential effect of mRNA secondary structure67 (that is, base pairing) on ribosome stalling, we used the previously published dimethyl sulfate sequencing datasets on human K562 cell lines34 to identify structured regions in mitochondrial transcripts. Following the methods described in the manuscript for nuclear transcripts, identification of sites with secondary structure was performed on mitochondrial transcript data34. In brief, FASTQ files (accession numbers GSM1297495 and GSM1297493) were retrieved from sequence read archive and mapped to GRCh38 with BWA (Galaxy Version 0.9)48. Reads were assigned to the nucleotide at the 5′ end with no offset using Plastid60. R values (cut-off 0.75) and Gini differences (cut-off 0.1) between the in vivo and denatured dataset were calculated for the complete mitochondrial transcriptome for a window size of 50 adenosine/cytosine nucleotides and a step size of 10. This provided a list of structured mitochondrial transcript regions, with most mRNA regions unstructured. The list of structured regions was compared to the SHMT2-specific stalling sites for potential co-localization. No stalling site mapped to a structured region.

Mitochondrial enzyme activities

Activities of individual OXPHOS complexes I–IV, ATP synthase and citrate synthase (which is nuclear encoded and was used as a marker of mitochondrial mass) were spectrophotometrically measured as previously described (Uvicon 922, Kontron)68,69,70. Measurements were performed with 2 μl of mitochondria isolated by differential centrifugation71 (except for complex I, in which 10 μl was used).

Citrate synthase (EC 2.3.3.1) activity was determined following extinction dynamics at 412 nm, indicating the cleavage of Elman’s reagent (0.2 mM) after addition of oxaloacetate (0.5 mM) to the buffered reaction solution containing acetyl-CoA (0.15 mM). Rotenone-sensitive complex I (NADH:decylubiquinone oxidoreductase, EC 1.6.5.3) activity was measured by adding NADH (0.2 mM) and monitoring at 340 nm for the reduction of decyl-ubiquinone (50 μM). Complex II (succinate:ubiquinone-oxidoreductase, EC 1.3.5.1) was measured at 600 nm by monitoring the reduction of 2,6-dichlorophenol-indophenol (80 μM) after addition of succinate (10 mM). The reaction mixture to determine complex III activity (coenzyme Q:cytochrome c–oxidoreductase, EC 1.10.2.2) contained cytochrome c (100 μM) and decyl-ubiquinol (200 μM) and was measured at 550 nm. After inhibition by addition of antimycin A (1 μM), the insensitive activity was subtracted to calculate specific complex III activity. The enzyme activity of complex IV (ferrocytochrome c:oxygen oxidoreductase, EC 1.9.3.1) was read as the oxidation rate of reduced cytochrome C (60 μM) at 550 nm. Complex V (F1F0 ATP synthase, EC 3.6.3.14) was indirectly measured as oligomycin-sensitive ATPase activity in a reaction mixture containing 0.5 mM ATP. Formed ADP was coupled to a pyruvate kinase reaction, using phosphoenolpyruvate (2 mM) to generate ATP and pyruvate. The latter is then used by lactate dehydrogenase in the oxidation of NADH (0.2 mM) that served as readout (340 nm). Reagents were obtained from Sigma-Aldrich.

Mitochondrial tRNA modifications

HCT116 or HEK293T cell lines and sub-clones were grown to 70–85% confluency and collected for mitochondrial extraction (1 × 108–2 × 108)71. Mitochondrial tRNAs were extracted using the MirVana miRNA Isolation kit (Ambion) for isolation of small RNAs followed by 10% TBE-urea gel purification and extraction as described above (tRNA fraction 65–85 base pairs). Quantitative analysis of 5-taurinomethyluridine monophosphate (p-τm5U), 5-taurinomethy-2-thiouridine monophosphate (p-τm5s2U), 2-thiouridine monophosphate (p-s2U) and 5-formylcytidine monophosphate (p-f5C) was performed using high-resolution mass spectrometry following an adapted previously published protocol25,72. In brief, 100 ng tRNA were digested at 37 °C for 2 h by nuclease P1 (2 U) in 30 μl of 100 mM ammonium acetate and quenched by adding 60 μl 50/50 methanol/acetonitrile followed by centrifugation at 10,000 g for 10 min. 5 μl of sample was injected for LC–MS analysis. Nucleoside monophosphates were analysed on a quadrupole-orbitrap mass spectrometer (Q Exactive plus, Thermo Fisher Scientific) operating in negative-ion mode coupled to hydrophilic interaction chromatography via electrospray ionization and used to scan in SIM mode from m/z 459 to 482 (τm5U and τm5s2U) or m/z 338 to 355 (s2U and f5C) at 1 Hz and 140,000 resolution. Consistent loading was ensured by measuring f5C levels (which are not altered by SHMT2 knockout) in the same sample as the τm5U and τm5s2U. The mass spectrometry standard for 5-taurinomethyluridine was synthesized as described previously73 and was used to confirm peak identity in digested tRNA samples after an additional treatment with 1 U of alkaline phosphatase (Roche) in 100 mM ammonium carbonate at 37 °C for 2 h72.

Folate depletion and methotrexate treatment

Cells were grown for five passages in folic acid-deficient DMEM with 10% dFBS (US Bio), either with or without hypoxanthine and thymidine supplementation (Gibco; final concentrations of 100 μM sodium hypoxanthine and 16 μM thymidine). To evaluate the effect of targeting 1C-metabolism on mitochondrial function, methotrexate (Sigma-Aldrich) was used at 25, 50 and 75 nM concentration. Chloramphenicol (100 μg ml−1 = 310 μM) and ethidium bromide (100 ng ml−1 = 250 nM) served as positive controls for the inhibition of mitochondrial translation and mtDNA depletion respectively. At each time point, protein and DNA samples were collected for immunoblot and mtDNA content analysis. Basal respiration was assessed using the Seahorse XF analyser with measurements conducted as described above.

N-terminal protein formylation

N-formylmethionine modification on COX1 was assayed by mass spectrometry following the protocol described previously19. In brief, mitochondria were isolated from cells by differential centrifugation71 and complex IV was immunoprecipitated (ab109801, Abcam) before separation on a 4–20% polyacrylamide gel. A band running at the same molecular mass as a band reactive with an anti-COX1 antibody (ab14705, Abcam) was excised and lysed using 1.5 μg LysC (Wako) as described74. Samples were dried in a speedvac and re-suspended with 15 μl of 0.1% formic acid pH 3. Per run, 5 μl was injected using an Easy-nLC 1000 UPLC system. Samples were loaded directly onto a 45 cm × 75 μm nano-capillary column packed with 1.9 μm C18-AQ (Dr. Maisch) mated to a metal emitter (Thermo Scientific) in-line with a Thermo Orbitrap Elite or Thermo Orbitrap Lumos. The mass spectrometer was operated in data-dependent mode with the 120,000 resolution MS1 scan (400–1,800 m/z) in the Orbitrap followed by up to 20 MS/MS scans in the ion trap. Raw files were searched using MSAmanada75, Sequest HT76 and Byonic77 algorithms and validated using the Percolator algorithm78 within the Proteome Discoverer 2.1 suite (Thermo Scientific). 10 p.p.m. MS1 and 0.6 Da MS2 mass tolerances were specified. Carbamidomethylation of cysteine was used as fixed modification and oxidation of methionine as dynamic modification. In addition, acetylation, formylation and loss of methionine were specified as potential modifications at the N terminus of proteins. The resulting msf file was used to construct a spectral library (percolator peptide q > 0.95) and extract MS1 ion chromatographs in Skyline79,80. The fraction of modified N-terminal peptides of COX1 was calculated as the area of the formylated peptide divided by the sum of the areas of all the n-terminal peptides in that sample.

Statistics and reproducibility

Significance was determined by two-tailed Student’s t-test comparing the indicated condition to the corresponding wild type or control. Asterisks denote P < 0.01. Exact P values for individual comparisons are given in Supplementary Table 7. Small filled circles are individual data points. All results have been independently replicated at least twice. This includes figure panels where representative data are shown. For quantitative measurements, n is provided in the figure legends. For western blots and DNA gels, we show representative data of multiple independent replicates: Fig. 1c (n = 2), Fig. 2b (n = 3), Fig. 4c (n = 3); Extended Data Fig. 1a (n = 2), Extended Data Fig. 1h (n = 2), Extended Data Fig. 1i (n = 2), Extended Data Fig. 2b (n = 2), Extended Data Fig. 2c (n = 2), Extended Data Fig. 3a (n = 2), Extended Data Fig. 4d (n = 2), Extended Data Fig. 5a (n = 2), Extended Data Fig. 5b (n = 3), Extended Data Fig. 7f (n = 2), Extended Data Fig. 10a (n = 3) and Extended Data Fig. 10c (n = 2). No statistical methods were used to predetermine sample size. The experiments were not randomized, and investigators were not blinded to allocation during experiments and outcome assessment.

Code availability

All code used to generate the data in this manuscript is publicly available from GitHub (https://github.com/R-J-Morscher).

Data availability

Raw sequencing data are available from Sequence Read Archive (SRA) under BioProject accession number PRJNA419990. Source data for Figs 1c, 3c, 4c, Extended Data Figs 3d, 6e are provided with the paper. Uncropped versions of blots are provided in Supplementary Figs 1–8; Sanger sequencing traces are provided in Supplementary Fig. 9. Other data that support the findings of this study are available from the corresponding author upon request.