Crystal Structure of the Emerging Cancer Target MTHFD2 in Complex with a Substrate-Based Inhibitor

Rapidly dividing cells depend on a high and steady supply of 10-formyltetrahydrofolate to sustain several vital anabolic reactions, for example the synthesis of purines. Targeting this pathway is one way to specifically target cancer cells, and one successful example is the antifolate drug methotrexate that has been used in cancer therapies since the 1950s (1).

In eukaryotes, the folate-dependent one-carbon metabolism is highly compartmentalized between cytoplasm and mitochondria (2, 3). These compartments are metabolically connected by the transport of the one-carbon (1C) donors serine, glycine, and formate across the mitochondrial membrane. Depending on the different redox environments in the mitochondria and cytoplasm, the metabolic flow occurs mostly in the clockwise direction in the scheme shown in Fig. 1A (3, 4). In mitochondria, the 1C unit usually derived from serine by serine hydroxymethyltransferase (SHMT) or from glycine by the glycine cleavage system, is attached to tetrahydrofolate (THF; see Fig. 1C) yielding methylene-THF (CH2-THF), which is subsequently oxidized to formate. The formate is released to the cytoplasm, where it is again attached to a THF molecule that is either used for de novo purine synthesis or reduced further and used for thymidylate or methionine synthesis (3). It has been shown that the majority of the 1C units used in the cytoplasm are derived from the mitochondria (5). The entire pathway is upregulated in cancer cells (6) as well as embryonic cells (7). It is important for maintaining the ratios of NAD⁺ to NADH and NADP⁺ to NADPH, thus affecting the redox balance of the cells and their ability to scavenge and reduce reactive oxygen species (8). One enzyme of specific interest is MTHFD2, responsible for the oxidation of methylene-THF to 10-formyl-THF in mitochondria, which is highly overexpressed in cancer cells and embryonic cells, but not in normal adult tissues. Thus, development of inhibitors targeting MTHFD2 is an attractive opportunity to specifically target cancer cells (9).

The enzyme family responsible for the conversion between methylene-THF and formate is the methylenetetrahydrofolate dehydrogenase (MTHFD) family that performs three main reactions in the 1C metabolism: the 5,10-methylene-THF (CH2-THF) dehydrogenase, 5,10-methenyl-THF (CH⁺-THF) cyclohydrolase and 10-formyl-THF (10-CHO-THF) synthetase activities (see Fig. 1B; refs. 10, 11).

In mitochondria, the 5,10-methylene-THF dehydrogenase and 5,10-methenyl-THF cyclohydrolase activities are performed by two enzymes, MTHFD2 and MTHFD2L (12). MTHFD2 was first discovered in Ehrlich ascites tumor cells (13) already in 1960 and later described as a mitochondrial NAD⁺-dependent methylene-THF dehydrogenase and cyclohydrolase expressed in embryonic and transformed cells (14–19). However, it has been demonstrated that MTHFD2 mRNA is expressed at low levels in all tissues but not confirmed to be translated (20). MTHFD2 has been shown to play an essential role in embryonic development for mammals because gene inactivation in mice resulted in embryonic lethality (21). The protein has been shown to be important for rapid growing cells, such as embryonic cells or cancer cells, mainly by supporting the high level of purine synthesis needed (21, 22). MTHFD2-null mutant fibroblasts have previously been reported to be glycine auxotrophs (23), a condition that can be rescued by expression of an NAD⁺- or NADP⁺-dependent, mitochondrially targeted methylene-THF-dehydrogenase-cyclohydrolase (24). For its dehydrogenase activity with NAD⁺ MTHFD2 has an absolute requirement for inorganic phosphate (P_i) and Mg²⁺ (14, 18, 19, 25). MTHFD2 displays activity also with NADP⁺, although lower, and in such case only requires the presence of Mg²⁺ and not P_i (25). It appears to have evolved from a tri-functional enzyme through the loss of the synthetase domain and the change of specificity from NADP⁺ to NAD⁺, P_i and Mg²⁺ (14, 26).

MTHFD2 mRNA and protein are upregulated in many cancers and their overexpression is associated with tumor cell proliferation (9). MTHFD2 depletion by RNA interference decreases cancer cell proliferation independent of tissue of origin (27) and leukemia burden in xenografts (28). Upregulation of MTHFD2 is linked to poor prognosis in breast cancer patients (27, 29) where it is associated with regulation of breast cancer cell migration and invasion (6, 30, 31). A SNP study found that MTHFD2 variants are associated with a higher risk for bladder cancer (32). MTHFD2 has been implicated in sensitivity of cancer cells toward artesunate, an anti-malarial drug (33). Recently, MTHFD2 was reported to have a nuclear localization in addition to its mitochondrial localization as well as supporting cancer cell proliferation independently of its dehydrogenase activity. Consistent with its role to support cancer cell proliferation, MTHFD2 was found to be co-expressed with proteins involved in cell-cycle progression, specifically in the S, G₂, and M phases, and often overexpressed in human tumors (34). In response to growth signals, the mTORC1 activates the ATF4 transcription factor, which stimulates expression of MTHFD2 and other enzymes of the serine synthesis and mitochondrial THF cycle, thereby increasing the production of formyl units required for de novo purine synthesis (35).

MTHFD2L, on the other hand, is described as a methylene-THF dehydrogenase and cyclohydrolase (12) that can use either NAD⁺ or NADP⁺; however, the catalytic efficiency, k_cat/K_m, is much lower than for MTHFD2 (36). MTHFD2L is highly homologous to MTHFD2 (72% identity) and is thought to be a housekeeping enzyme, because it is found in all adult tissues tested (12), as well as in embryonic cells (36).

In the mitochondria, the 10-formyl-synthetase activity is performed by MTHFD1L, which is a homolog of the cytosolic trifunctional MTHFD1 (4, 37, 38). This enzyme consists of two functional units as for MTHFD1, but several key residues for the domain responsible for dehydrogenase and cyclohydrolase activities differ, rendering it a monofunctional synthetase enzyme (37, 38). MTHFD1L is expressed in all embryonic and adult tissues examined (37, 39) and its deletion caused embryonic lethality as well as neural tube and craniofacial defects in mice (40). MTHFD1L mRNA has also been found to be upregulated in human colon adenocarcinoma (41).

In the cytosol, all three enzymatic functions are performed by MTHFD1. MTHFD1 functions as a dimer where each monomer comprises two functional units, one DC-domain containing the dehydrogenase (D) and cyclohydrolase (C) activities with a common active site (42) as demonstrated by both substrate channeling (43, 44) and X-ray crystallographic structures with NADP⁺ (10) and with folate-analogues (45). The second domain is responsible for the 10-formyl-THF-synthetase (S) activity. MTHFD1 uses NADP⁺ as cofactor for the dehydrogenase activity (10) and the rate-limiting step of the D/C activities is the cyclohydrolase (46). MTHFD1 is expressed in all adult tissues examined (47).

Computer-generated homology models of both MTHFD2 (14) and MTHFD2L (9) have been published, based on the human MTHFD1 structure and in the case of the MTHFD2 homology model, also based on the Escherichia coli (48) and Saccharomyces cerevisiae (49) homologs. So far no structure based on empirical data has been presented for any of these two proteins.

Here we identify the first MTHFD2 inhibitor LY345899 (Fig. 1D) and the target engagement of this substrate-based inhibitor, as well as present the first structure of the human mitochondrial NAD⁺-dependent methylene-THF dehydrogenase and cyclohydrolase, MTHFD2. The MTHFD2 protein was cocrystallized with the substrate-based inhibitor LY345899 and the cofactors NAD⁺ and inorganic phosphate.

For all buffers, buffer composition can be found in Supplementary Table S1.

cDNAs encoding MTHFD2 and MTHFD1 (codon optimized for E. coli expression) were purchased from Eurofins and GeneArt. Bacterial expression constructs enabling His-tag purification of human MTHFD2 lacking the N-terminal mitochondrial signal peptide (MTHFD2 AA36-350) and the dehydrogenase/cyclohydrolase (DC) domain of MTHFD1 (AA1-306) were generated.

For MTHFD2 protein expression the construct was transformed into E. coli strain BL21(DE3). The transformed cells were grown in LB-medium + 100 μg/mL ampicillin at 37°C overnight. Fresh overnight culture inoculated into LB-medium was grown at 37°C to OD₆₀₀ ∼ 1, protein expression was induced by addition of 1 mmol/L IPTG. The bacteria were harvested after 2 hours. Cells were dissolved in Lysis buffer A2. After incubation for 20 minutes at room temperature, the suspension was centrifuged and the supernatant was loaded onto Ni-Sepharose column HisTrap-HP (GE Healthcare) equilibrated with Buffer B2. The column was washed with Buffer B2 + 10 mmol/L imidazole, the bound proteins were eluted with 10–500 mmol/L imidazole gradient. The fractions containing MTHFD2 were dialyzed against Buffer C2.

His-tag was cleaved off and removed by passing the protein over a HisTrap column. MTHFD2 was dialyzed against Buffer D2 and loaded onto an anion-exchange monoQ-HP column (GE Healthcare) equilibrated with Buffer D2. The bound proteins were eluted with 20 to 800 mmol/L NaCl gradient and analyzed by SDS-PAGE. Fractions with pure MTHFD2 (Supplementary Fig. S1) were combined and protein concentration was determined by the Bradford assay.

MTHFD1 DC was expressed in E. coli strain BL21(DE3). The cells were grown in LB containing 50 μg/mL carbenicillin at 30°C until OD₆₀₀ reached 0.55. The temperature was lowered to 16°C before protein expression was induced with 1 mmol/L IPTG. The cells were harvested after overnight expression. Bacteria were resuspended in Lysis buffer A1 and homogenized using EmulsiFlex C3 (Avistin). After centrifugation, the supernatant was loaded on a 1 mL HisTrapFF column (GE Healthcare) equilibrated with Buffer B1. After washing with Buffer B1, bound proteins were eluted using a gradient to 100% Buffer C1. The imidazole was removed from the IMAC elute using a HiPrep 26/10 desalting column equilibrated and run with Buffer D1. His-tag was removed as described for MTHFD2. Finally, the nontagged MTHFD1DC pool was after desalting run on Superdex 16/60 equilibrated with Buffer E1.

LY345899 (3 mmol/L), NAD⁺ (5 mmol/L), and MgCl₂ (6 mmol/L) were added to MTHFD2 and incubated for 50 minutes. After 40 minutes, 10 mmol/L Na₂HPO₄ was added. Proteases (1:50 ratio each of trypsin, α-chymotrypsin, pepsin, papain, proteinase K, and subtilisin to MTHFD2) were added just before crystallization. For crystallization at 20°C, MTHFD2 at 5.9 mg/mL was mixed with 0.1 mol/L phosphate/citrate pH 4.1, 38% (v/v) PEG300 in 3:1 ratio. After 1 week, crystals were frozen in liquid nitrogen. Data collection were performed at beamline ID30A-3 at ESRF. Crystals diffracted to 1.9Å. Statistics can be found in Table 1. Details regarding data processing, refinement, and model building can be found in Supplementary Methods. The structure has been deposited in the protein data bank with accession code 5TC4.

To determine IC₅₀ values of LY345899, an eight-point dose–response curve with 3-fold difference in concentration between points was generated. Each assay point was run in duplicate. The starting concentration of LY345899 was 10 μmol/L for the MTHFD2 assay and 1 μmol/L for the MTHFD1 assay. Serial dilution of compound LY345899 was transferred into a Perkin Elmer 384-Proxiplate, with DMSO as negative control.

The assays were run in MTHFD2 assay buffer and MTHFD1 assay buffer, respectively. A total of 2.5 μL enzyme was preincubated with compound or DMSO for 10 minutes. The enzymatic reaction was initiated by adding 2.5 μL folitixorin. For background control, 5 μL buffer was added to the well. Final concentrations of the components in the MTHFD2 assay were 3 nmol/L MTHFD2, 5 μmol/L folitixorin and 250 μmol/L NAD⁺, and 25 nmol/L MTHFD1 DC, 30 μmol/L folitixorin, and 400 μmol/L NADP⁺ in the MTHFD1 assay. After 15-minute reaction, 5 μL NAD(P)H-Glo detection reagent (Promega) was dispensed in all wells and the plate was incubated for 60 minutes. The dose–response curve for MTHFD2 was run seven times, and twice for MTHFD1. For both assays, luminescence was measured on a Perkin–Elmer Envision reader.

Differential scanning fluorimetry (DSF) was used to detect binding of LY345899 to purified MTHFD2 protein. MTHFD2 (4 μmol/L) with LY345899 (100 μmol/L, 1% DMSO) or 1% DMSO was run with 5× SYPRO Orange dye (Ex492/Em610) in quadruplicate using DSF Buffer. The assay volumes were 20 μL in 96-well plates (HSP9655; Bio-Rad Laboratories). The temperature was increased from 25°C to 95°C with 1°C per minute. A real-time PCR (Bio-Rad Laboratories) CFX96 Optical Reaction Module, C1000T chassis, Channel 6 FRET, was used.

U2OS human osteosarcoma cells, MRC-5 human lung fibroblasts, and Hs-587T breast cancer cells were obtained from the ATCC and authenticated by the supplier using STR analysis. Cells were cultured in DMEM GlutaMAX (Life Technologies) supplemented with 10% FBS, penicillin (50 U/mL), and streptomycin (50 μg/mL) and maintained in a humidified incubator at 37°C with 5% CO₂.

U2OS cells were harvested using trypsin, upon detachment trypsination was inhibited by addition of cell media and cells spun down at 1,500 rpm, 10 minutes, 4°C. Pellet was resuspended in TBS, cells were counted and adjusted to 2 × 10⁶ cells per milliliter. Cells were lysed by a freeze–thaw cycle three times at −80°C for 3 minutes followed by 37°C for 3 minutes. Cell lysates were cleared by centrifugation at 13,000 rpm at 4°C for 20 minutes, supernatant was transferred to new tubes and 50 μmol/L LY345899 or DMSO was added and sample was incubated 30 minutes at room temperature. Cell lysates were aliquoted and heated to indicated temperatures for 3 minutes, insoluble proteins were removed by centrifugation at 13,000 rpm at 4°C for 20 minutes, supernatant transferred to new tubes and protein concentration determined using BCA assay (Pierce). Western blot analysis was performed according to standard procedures. The antibodies were mouse anti-MTHFD2 (Abcam, ab56772), mouse anti-Actin (Abcam, ab6276) followed by IRDye 800 CW donkey anti-mouse (LI-COR, 926-32212). Images were taken with Odyssey Fc imager (LI-COR; ref. 50).

U2OS cells grown to approximately 70% to 80% confluence were lysed in mammalian protein extraction lysis buffer M-PER (Thermo Scientific) supplemented with 1× complete protease inhibitor cocktail (Roche) and after centrifugation the supernatant was diluted to a final concentration of 1× TN buffer. Protein concentration of the cell lysate was determined by Bradford method. Aliquots of the cell lysate were incubated with 1% DMSO or serial dilution of compound LY345899 in 1% DMSO at room temperature for 1 hour and then with pronase (Roche) solution for 30 minutes. For the nondigested (ND) sample, TN buffer was added instead of protease. Proteins were separated using SDS-PAGE and blotted onto nitrocellulose membranes, and then probed with mouse anti-MTHFD2 antibody (Abcam, ab56772), followed by incubation with goat anti-IgG mouse antibody (Jackson ImmunoResearch, 711-035-150), and protein was visualized using SuperSignal West Femto chemiluminescence substrate (Thermo Scientific). The blot was probed with rabbit anti-GAPDH Ab (Santa Cruz Biotechnology, sc25778) followed by incubation with donkey anti-rabbit IRDye 800CW Ab (LI-COR, 926-32213) and images taken using LI-COR (51).

Structure-based sequence alignment of MTHFD1-DC (1A4I.pdb) and MTHFD2 (5TC4.pdb) was generated using PDBeFOLD (52). Graphic representation was generated using ESpript 3.0 (53).

Human MTHFD2 and the DC domain of MTHFD1 were purified from bacteria and activity was determined using the NAD(P)H-Glo assay (Promega). MTHFD2 showed clear activity and the biochemical assay gave consistent results, with Z′ = 0.8 (Supplementary Fig. S2; Supplementary Table S2). In vitro inhibition of MTHFD2 and MTHFD1 DC-domain using the substrate-based inhibitor LY345899 was shown using the same assay. IC₅₀ values of LY345899 were determined to 663 nmol/L for MTHFD2 (n = 7) and 96 nmol/L for MTHFD1 (n = 2; Fig. 2).

To further validate the interaction with MTHFD2, target engagement of LY345899 was next investigated by DSF. The temperature at which a protein unfolds is measured by an increase in the fluorescence of SYPRO Orange dye. The dye is quenched in aqueous solution but emits strong fluorescence upon binding to exposed hydrophobic parts of the protein as the protein unfolds with increasing temperature. LY345899 (100 μmol/L, 1% DMSO) was able to stabilize MTHFD2 (4 μmol/L) upon binding, increasing the protein melting point (T_m) from 44°C (1% DMSO) to 55°C (n = 4; Fig. 3A). To investigate target engagement by LY345899 in a cellular context, cellular thermal shift assay (CETSA) was performed. CETSA is based on the principle that ligand binding leads to a thermal stabilization of the target protein upon engagement, which is detected using Western blotting with antibodies specifically recognizing the protein of interest. Following treatment of U2OS cell lysates with LY345899, the MTHFD2 protein was stabilized as compared to DMSO (Fig. 3B and C). In support of this, target engagement by LY345899 in cell lysate was analyzed by drug affinity responsive target stability (DARTS). The DARTS assay is based on the principle that interaction between target protein and a small molecular ligand reduces susceptibility of the protein to protease digestion. The inhibitor LY345899 was found to protect the MTHFD2 protein from pronase digestion as compared with DMSO (Fig. 3D). However, we could not detect any target engagement by LY345899 in CETSA using intact cells, thus LY345899 does not seem to be able to enter cells under the conditions used (Supplementary Fig. S3A and S3B). Consistently, no inhibition of cell viability could be observed upon LY345899 treatment (Supplementary Fig. S3C).

To evaluate the detailed binding of LY345899 to MTHFD2, we solved the X-ray cocrystal structure of MTHFD2 with Ly345899, NAD⁺ and P_i. The structure of MTHFD2 contains one monomer of MTHFD2 in the asymmetric unit. The resolution is 1.9Å and shows clear electron density for residues Glu36-Leu332, except one break between His280 to Lys286. See Table 1 for relevant statistics of the structure.

MTHFD2 is built up of two domains connected by two long α helices (A and I in Fig. 4A) and a small helix (helix D2) ordered so that a large cleft is formed between the domains. MTHFD2 is similar to DC-domain of MTHFD1 but has several distinct features. The overall structure and secondary sequence annotation is shown in Fig. 4A. The structure-based sequence alignment to MTHFD1 DC-domain is shown in Fig. 5. 

MTHFD2, like MTHFD1, is dimeric (Supplementary Fig. S4). The monomer forms a homodimer with extensive contacts to a crystallographic symmetry related molecule (Fig. 4A), showing a buried surface area of 1609Ų.

The structure contains both the nicotine amide dinucleotide cofactor NAD⁺ as well as the inorganic phosphate needed for enzymatic activity. They are bound within the large cleft formed between the N- and C-terminal domains of MTHFD2 toward the C-terminal domain, as can be seen in Fig. 4A and B. Detailed depiction of the NAD⁺ binding site is shown in Supplementary Fig. S5A. The adenine group is bound in a crevice formed by the loops between strand g and 3₁₀-helix G and between strand f and helix F of the Rossmann domain. The classical dinucleotide binding motif GXGXXG cannot be found in the sequence of MTHFD2 but instead the residues G²⁰⁰RSKNVG making up the loop connecting strand e with helix E constitute the dinucleotide binding loop in MTHFD2. The G²⁰⁰RS²⁰²KNVG²⁰⁵ residues align structurally with the classical dinucleotide sequence GXGXXG thus making the minimal sequence of this protein GXSXXXG, as previously shown for this protein family (10).

The phosphate is bound close to the 2′-OH group of the sugar in the adenosine monophosphate moiety of NAD⁺, where the phosphate would be covalently bound in NADP⁺ as can be seen in Fig. 4B. Detailed interactions are shown in Supplementary Fig. S5B. The phosphate interacts with both monomers of the dimer. Especially the interaction to Asp216 is very short, only 2.4Å.

The inhibitor LY345899 is clearly defined by electron density (Fig. 4C) and is positioned in the large cleft between the N- and C-terminal domain with the protein bridging over the inhibitor, forming a tunnel as can be seen in Fig. 4C. The interactions with LY345899 can be seen in Supplementary Fig. S5C. The terminal glutamate moiety of LY345899 is extending through the tunnel of MTHFD2 and reaching the solvent.

The MTHFD2 protein is clearly a dimer as seen both by the extensive dimeric interface and the size exclusion chromatography described herein, which is in agreement with previous studies (Supplementary Fig. S4; ref. 18). The sulfur atom of Cys166 is within 3.4Å of its symmetry-related mate and could form a disulfide bond under oxidizing conditions.

Overlay of MTHFD2 with the structure of MTHFD1 DC-domain in complex with LY345899 published previously (45) shows considerable similarity (Fig. 5) between the structures with the same general folds being present (Fig. 6). The structure of MTHFD1 does however only have LY345899 bound in one of the monomers. In the monomer with LY345899 bound, there are only small differences in backbone between MTHFD1 and MTHFD2 with Cα-RMSD of 0.96Å, except for the insertion of a loop between helix E and strand f. For the other monomer, the changes are much larger due to the absence of bound LY345899 in MTHFD1 with an overall Cα-RMSD of 1.70Å.

The binding site for NAD⁺ (Fig. 4B) does not resemble a traditional NAD⁺ site, as many of the features described in literature are missing or more similar to a NADP⁺ site. The diphosphate binding loop sequence GXSXXXG is different from the conserved GXGXXG and has been described for the NADP⁺ binding MTHFD1 (10). The main difference from a traditional NAD⁺ site is the lack of an Asp or Glu sidechain hydrogen bonding to the ADP diol. Backbone nitrogens of Arg201 and Arg233, as well as the inorganic phosphate, instead bind the adjacent hydroxyl groups in NAD⁺ bound to MTHFD2. The presence of Arg233 in front of the A side of adenine is a usual sign of a NADP⁺ site as this residue often binds the 2′-phosphate. Interestingly, in MTHFD1 this residue is a serine, but is still binding the 2′-phosphate together with Arg173 (Arg201 in MTHFD2; ref. 44). MTHFD2 has thus adapted a NADP⁺ site to bind NAD⁺ by the use of phosphate, which binds to the ADP diol of NAD⁺. The phosphate mediates several hydrogen bonds to NAD⁺ to bind it to the protein. In the NADP⁺-dependent MTHFD1, these hydrogen bonds instead go directly to the 2′-phosphate and thus bind NADP⁺ to the site. The use of NAD⁺ as cofactor for MTHFD2 favors the production of 10-formyl-THF from methylene-THF during rapid cell proliferation, as the high ratio of NAD⁺/NADH in the mitochondria drives the reaction in this direction. MTHFD1L uses the 10-formyl-THF to produce formate, which is transported into the cytosol, where the high ratio of NADPH/NADP⁺ drives the reaction of MTHFD1 in the opposite direction forming 10-formyl-THF and methylene-THF from formate to be used for purine and thymidylate synthesis, respectively, and forcing the metabolic flow clockwise in Fig. 1A (3, 4).

The inorganic phosphate is bound by 11 hydrogen bonds (Fig. 4B and Supplementary Fig. S5B). No interaction with backbone nitrogen is observed, otherwise a common mode of phosphate binding (54). Instead all hydrogen bonds to the protein are to sidechains. The use of arginine sidechain to bind phosphates is common (54) and Arg201 has been shown to be responsible for phosphate binding because conservative mutations to lysine result in loss of all enzymatic activity (14). Asp225 from the second monomer holds Arg201 in place, suggested to be important for positioning the Arg201 sidechain correctly (14). In contrast, Arg233 can be mutated to other residues whilst retaining dehydrogenase activity. The hypothesis is that this residue helps to bind and position the phosphate but is not strictly necessary as Arg201 (14). Asp216 and His219 are in the loop between helix E and strand f and not found in MTHFD1 (Figs. 5 and 6) and binding to the phosphate from the second monomer of the dimer. Phosphate binding to histidine is common, although not as common as to arginine, binding to negatively charged residues is rare (54). When found, the bond is often short, as seen in the MTHFD2 structure. This is believed to account for the preference of dibasic phosphate present at physiological pH over tribasic phosphate or sulfate for some proteins, where it would act as a hydrogen bond acceptor (55). MTHFD2 has previously been shown to have 8% activity with sulfate instead of phosphate (18). The structure presented herein gives insights into the nature of the cofactor-binding site of MTHFD2.

In addition to phosphate, MTHFD2 requires Mg²⁺ or Mn²⁺ for activity (14, 18, 19, 25). Mg²⁺ or Mn²⁺ binding always occurs through at least one negatively charged amino acid side chain such as Asp or Glu. Mutational studies have shown that the phosphate and magnesium are likely interacting with each other (14). There are only few aspartates or glutamates in the vicinity of the inorganic phosphate, and only Asp168 (Fig. 4B) has been suggested to be conserved in related enzymes requiring Mg²⁺ (14). Furthermore, mutational studies of this position indicate that this aspartate is involved in magnesium binding (14). In the MTHFD2 structure, no magnesium is found even though 6 mmol/L is present in the crystallization conditions. We have collected anomalous data on crystals grown in the presence of Mn²⁺. The anomalous difference maps did not reveal any bound Mn²⁺. It is possible for the magnesium to bind both Asp168 and His219 as well as the inorganic phosphate (Fig. 4B). The structure presented here does not reveal why Mg²⁺ or Mn²⁺ is needed for activity, however, it does open for new approaches to investigate the reasons for the apparent metal dependence of MTHFD2.

The good resolution and well-defined electron density around LY345899 clearly shows the ligand with R configuration at the chiral carbon in the fused tricyclic system (Fig. 4C). Attempts to fit the corresponding S-diastereomer to the electron density were unsuccessful. The synthetic procedure employed to produce LY345899, starting from (S)-folic acid (Supplementary Fig. S6), in all likelihood produces a 1:1 mixture of the RS- and SS-diastereomers. Thus, MTHFD2 has singled out the RS-diastereomer as a ligand from the mixture, indicating that the SS-diastereomer probably is less active as a MTHFD2 inhibitor. Attempts to separate the diastereomers of LY345899 are currently in progress.

There are several conserved residues found in the sequence alignment of MTHFD1, MTHFD2, and MTHFD2L involved in the binding of LY345899 (Supplementary Fig. S7). An Y⁵²XXXK⁵⁶ motif has been proposed for the binding and cyclohydrolase catalytic activity of MTHFD1 (10, 45, 56), and conserved for both MTHFD2 and MTHFD2L. Tyr84 of MTHFD2 contributes with binding of the substrate analogue by pi-stacking with the p-aminobenzoate moiety. Mutations of Tyr52 in MTHFD1 still have some cyclohydrolase activity; however, mutation of Lys56 abolishes cyclohydrolase activity, showing that this position is vital in the catalytic mechanism proposed (45, 56). In the proposed catalytic mechanism for MTHFD1, Lys56 is supporting the formation of methenyl-THF by hydrogen bonding to the pteridine carbonyl of the substrate. In the cyclohydrolase reaction, Lys56 supports the attacking water by accepting and donating protons in several stages of the mechanism, thus having different roles in the two activities. Glu100 supports Lys56 by hydrogen bonding and is important for activity (56). The water between Lys88 and methenyl-THF is likely not present in the MTHFD2 structure because the carbonyl group of LY345899 displaces it. LY345899 is bound with the carbonyl directed toward Lys88, thus imitating the way a substrate methenyl-THF would be lined up for the cyclohydrolase activity. Structures of MTHFD1 with bound folate analogues, including LY345899, suggests that the pteridine moiety needs to flip 180° between the dehydrogenase and cyclohydrolase activities (45) or reorient (56). Ser49, Gln100, and Pro102 have been proposed to be important in the THF binding site of MTHFD1 and are conserved in MTHFD2 (Fig. 5 and Supplementary Fig. S7; ref. 10). The hydrophobic environment of Lys56 (Lys88 in MTHFD2) is conserved and proposed to decrease the pK_a of this residue to keep it deprotonated, which is needed for activity (56). Asp155 in MTHFD2 is contributing strongly to binding of the pteridine of LY345899 by two direct hydrogen bonds (Fig. 4C). The corresponding residue in MTHFD1, Asp125, has been shown to be critical for the binding and positioning of the substrate (56).

LY345899 has been characterized as an MTHFD1 inhibitor. To evaluate this inhibitor for MTHFD2, we synthesized LY345899 as described in Supplementary Methods. We here show that LY345899 is a potent MTHFD2 inhibitor. This is the first inhibitor to be identified for MTHFD2; LY345899 has an IC₅₀ value of 663 nmol/L. As a comparison, inhibition of the DC-domain of MTHFD1 by LY345899 was tested, showing an IC₅₀ value of 96 nmol/L, clearly showing that LY345899 has a higher affinity for MTHFD1 compared to MTHFD2. The overall binding mode of LY345899 is conserved between MTHFD1 and MTHFD2 (Fig. 6A). The reason for the difference in affinity of LY345899 to MTHFD1 and MTHFD2 likely depends on Ser83, Asn87, Phe157, Ala175, Asn204, and Leu289 that all interact with the inhibitor. These residues are not conserved between MTHFD1 and MTHFD2 (Figs. 4C and 5 and Supplementary Fig. S5C).

Using CETSA we show that LY345899 engages MTHFD2 in cell lysates (Fig. 3B and C). DSF shows that LY345899 increases the melting temperature (T_m) of MTHFD2 by 11 degrees (Fig. 3A). LY345899 is reducing the susceptibility of MTHFD2 to protease digestion (Fig. 3D). The stabilizing effect is however not seen in CETSA using intact cells, which could possibly be due to low cell permeability (57). This also explains the lack of inhibition of cell viability by LY345899 as compared to depletion of MTFHD2 by siRNA (27). LY345899 is however a good starting point for optimization toward an inhibitor with favorable ADME properties.

There is clear evidence of MTHFD2 upregulation and overexpression in cancer cells and recent data show reduced leukemia burden in xenograts upon MTHFD2 depletion by shRNA (9, 27–32, 34). The protein is expressed in embryonic cells and transformed cells independent of tissue of origin, but importantly not in adult, healthy cells (14–19). Inhibition of MTHFD2 is thus an attractive way to selectively target cancer cells to disrupt purine synthesis and 1C metabolism while sparing healthy normal cells. A limiting factor of many antimetabolite drugs is that the target enzymes are not only expressed in the cancer cells but also in healthy proliferative cells, causing adverse side-effects in patients and limiting their use. Thus, the cancer-specific expression profile of MTHFD2 holds the promise of fewer adverse side-effects in proliferative tissues compared to many current anti-metabolite drugs used in cancer therapy. Collectively, these aspects make the MTHFD2 protein highly relevant as an anticancer target.

The crystal structure of MTHFD2 and the identification of the first MTHFD2 inhibitor now provide a first starting point for developing potent and selective MTHFD2 inhibitors.

No potential conflicts of interest were disclosed.

Conception and design: R. Gustafsson, E. Homan, T. Helleday, P. Stenmark

Development of methodology: A.-S. Jemth, L. Dahllund, S. Llona-Minguez, M. Henriksson, Y. Andersson

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): R. Gustafsson, A.-S. Jemth, N.M.S. Gustafsson, O. Loseva, N. Bonagas, M. Henriksson, Y. Andersson

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): R. Gustafsson, A.-S. Jemth, N.M.S. Gustafsson, K. Färnegårdh, O. Loseva, E. Wiita, N. Bonagas, M. Henriksson, Y. Andersson, E. Homan, P. Stenmark

Writing, review, and/or revision of the manuscript: R. Gustafsson, A.-S. Jemth, N.M.S. Gustafsson, K. Färnegårdh, O. Loseva, N. Bonagas, S. Llona-Minguez, M. Henriksson, Y. Andersson, T. Helleday, P. Stenmark

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): M. Häggblad, M. Henriksson, Y. Andersson

Study supervision: T. Helleday, P. Stenmark

Other (designed, performed, and analyzed synthetic chemical experiments): K. Färnegårdh

We thank the beamline scientists at ESRF, France; Max-Lab, Sweden; BESSY, Germany; Diamond, United Kingdom; PETRA, Germany; and the Swiss Light Source, Switzerland, for their support and Biostruct-X. We thank Nina Braun for her contributions to the initial crystallizations.

This work was supported by the Swedish Research Council (T. Helleday, P. Stenmark), the Knut and Alice Wallenberg Foundation (T. Helleday, P. Stenmark), the Wenner-Gren Foundation, Åke Wiberg Foundation (P. Stenmark), the Göran Gustafsson Foundation, the Swedish Pain Relief Foundation, and the Torsten and Ragnar Söderberg Foundation (T. Helleday), the Swedish Children's Cancer Foundation (T. Helleday, N.M.S. Gustafsson), the Swedish Society for Medical Research, Karolinska Institute Foundations (N.M.S. Gustafsson), the Helleday Foundation (N. Bonagas), and the Swedish Cancer Society (T. Helleday, P. Stenmark).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. 1734 solely to indicate this fact.