Formate induces a metabolic switch in nucleotide and energy metabolism (2025)

Mathematical model

We analysed a simplified mathematical model where formate is produced from the mitochondrial one-carbon metabolism or consumed from the extracellular media (Fig. 1a, Supplementary text). The produced formate is released from cells or incorporated into adenine nucleotides. The free pools of adenine nucleotides (AMP, ADP and ATP) are established by the adenylate kinase equilibrium, ADP phosphorylation and ATPases activity.

Numerical simulations of kinetic models of glycolysis and oxidative phosphorylation indicate that the rates of ADP phosphorylation by glycolysis and oxidative phosphorylation follow effective Michaelis–Menten relationships with respect to the concentration of ADP (Fig. 1b, c). This observation is quite surprising given that both pathways have multiple reactions, with steps of ATP hydrolysis, ADP phosphorylation and oxidation/reduction reactions. These effective Michaelis–Menten models link energy and purine metabolism.

Next, we conducted numerical simulations of the full model of formate, purine and energy metabolism. The numerical simulations predict that, with increasing the rate of endogenous formate production, there is an increase in the concentration of formate and adenine nucleotides and the rates of glycolysis, oxidative phosphorylation and proliferation (Fig. 1d–j, black line). Furthermore, cells start releasing formate when the rate of serine catabolism to formate reaches the threshold rate defined by the purine synthesis rate (Fig. 1k, black line).

One-carbon units can be derived from extracellular formate and cytosolic metabolism as well. The effect of adding exogenous formate is to displace the prediction lines to the left (Fig. 1d–k, red line). To simulate the contribution of the cytosolic pathway to one-carbon unit production we added an additional flux of 10-formyl-tetrahydrofolate production, setting its rate to 50% the maximum activity of 10-formyl-tetrahydrofolate synthetase. The effect of the cytosolic pathway is similar to what observed for exogenous formate, it displaces the prediction lines to the left (Fig. 1d–k, cyan line).

In these simulations we assumed that the cell energy demand is coupled to an effective ATP consuming reaction with a Michalis-Menten dependency with the ATP levels. This assumption was motivated by the fact that some synthetases, like mammalian carbamoyl-phosphate synthetase (CAD), have a half-saturation constant on the mM ATP range²¹. An alternative hypothesis is that the proliferation rate is limited by the availability of deoxynucleotides, which in the case of adenines are derived from ADP. We have repeated the model simulations assuming that the proliferation rate follows a Michaelis–Menten relationship with the ADP levels and the outcome is the same (Fig. S1, red line). We have also repeated the simulations assuming a constant proliferation rate. In this case the transition from low to high purines is even more pronounced (Fig. S1, cyan line).

In vitro model

To validate the theoretical predictions, we selected a panel of haploid HAP1 cell lines engineered for single knockout of SHMT2 or the mitochondrial folate transporter (MFT) and double knockout of MFT and cytosolic serine hydroxymethyltransferase (SHMT1) (Fig. 2a). The intracellular formate levels are the lowest in the MFT-SHMT1 double knockout cell line, intermediate in the SHMT2 single knockout cell line and maximum in the WT parental cell line (Fig. 2b). Furthermore, supplementation of sodium formate at a concentration of 1 mM increases the intracellular formate levels of the MFT-SHMT1 and SHMT2 cell lines to values similar to what is observed in the WT cells (Fig. 2b). Based on this data, we assigned the one-carbon units availability index 0 to MFT-SHMT1, 1 to MFT and SHMT2, and 2 to MFT + 1 mM Formate, SHMT2 + 1 mM Formate and WT.

We first characterized the proteome of these cell lines using mass spectrometry (Table S1). To identify protein level changes associated with the availability of one-carbon units we calculated the slope between the protein levels and the one-carbon availability index. We observed a significant increase in the levels of proteins belonging to the minichromosome maintenance protein complex (MCM) and of ribosomal proteins (Table S2). The increase in ribosomal proteins results in a gradual increase of the total ribosomal protein mass (Fig. 2c). In contrast, we observed just a trend towards decrease in the levels of mitochondrial and vacuole proteins (Table S2). The total proteome mass associated with proteins with annotated mitochondrial localization²² is approximately constant across the different cell lines (Fig. 2d). We did not observe any significant enrichment of genes associated with metabolic pathways, except for a trend of reduced levels of glycolysis and TCA proteins (Table S3). The total proteome mass associated with proteins in the KEGG glycolysis pathway is approximately constant across the different cell lines (Fig. 2e).

Next we performed a metabolic characterization. As predicted by the model, the levels of intracellular AMP, ADP and ATP increase from the knockout cell lines to the WT cell lines and when the knockout cell lines are supplemented with formate (Fig. 2f–h). In agreement with the behaviour suggested by the computational model of glycolysis (Fig. 1b), the rate of lactate release (a surrogate of glycolysis) approximately follows a Michaelis–Menten dependency with the intracellular ADP levels (Fig. 2i). In the case of oxidative phosphorylation the data deviates from a Michaelis–Menten law suggested by the computer simulations of mitochondrial oxidative phosphorylation (Fig. 2j). Finally, as assumed in the mathematical model, the proliferation rate approximately follows a Michaelis–Menten relationship with the intracellular ATP levels (Fig. 2k).

To test the metabolic switch beyond the HAP1 background, we generated a panel of SHMT2 deficient cell lines for the colorectal cancer cell line HCT116 and breast cancer cell lines MDA-MB-231, SKB3, T47D and MDA-MB-468. The parental cell lines exhibit formate overflow and the phenotype is lost upon genetic inactivation of SHMT2 (Fig. S2a). In the HCT116 and MDA-MB-231 background the SHMT2 deficiency causes a reduction in the adenine nucleotide levels (Fig. S2b–d), in agreement with our theoretical and HAP1 genetic models. However, this is not the case in the SKB3, T47D and MDA-MB-468 cell lines. Therefore, there are additional factors that modulate the control of the adenine nucleotide pools by mitochondrial formate production.

Formate increases AICAR and reduces AMPK activity

To uncover metabolic changes not anticipated by the mathematical model, we extended the correlation analysis between the intracellular metabolites and the one-carbon availability index (Table S4). We noted a significant negative association between the one-carbon availability index and the levels of purine precursors glycinamide ribonucleotide (GAR, p = 10⁻⁶) and 5-aminoimidazole-4-carboxamide ribonucleotide (AICAR, p = 5.4 × 10⁻⁴) (Fig. 2f vs Fig. 3a, b). The elevation of AICAR in cells deficient of mitochondrial one-carbon metabolism has been observed in other cell lines^5,23. We also observe an elevation of AICAR in our panel of SHMT2 deficient cell lines (Fig. S1e). The effect being more pronounced in those cell lines where the SHMT2 deficiency is associated with a depletion of adenine nucleotides (Fig. S2b–d).

Given that both AICAR and AMP are AMPK activators²⁴, the formate-dependent decrease of AICAR while increasing AMP results in conflicting signals to AMPK. To determine which signal dominates, we quantified the level of AMPK phosphorylation using phosphoantibodies specific for pAMPK-Thr172, a canonical AMPK site that is phosphorylated under energy stress²⁴. We observed higher levels of pAMPK-Thr172/AMPK in SHMT2 deficient cells relative to WT cells (Fig. 3i, j). The noted changes in SHMT2 deficient cells are more pronounced than what was observed when treating WT cells with the AMPK activators acadesine (1 mM) or A769662 (10 μM). Finally, supplementation with 1 mM formate reduces pAMPK-Thr172/AMPK in SHMT2 deficient cells to levels between untreated SHMT2 deficient cells and WT cells, without much of an effect on WT cells.

Activated AMPK phosphorylates multiple proteins, including acetyl-CoA carboxylase (ACC) at serine 79 (Ser79)²⁴. The changes of pACC-Ser79/ACC in SHMT2 deficient cells are quite similar to those observed for pAMPK-Thr172 /AMPK (Fig. 3j, k, T2 genetic background), indicating that in SHMT2 cells the level of ACC phosphorylation is either regulated by AMPK or by an upstream kinase targeting both AMPK and ACC. In contrast, the pattern of ACC phosphorylation in WT cells is different from that of AMPK phosphorylation. A clear example is the formate-dependent reduction of ACC phosphorylation in WT cells with no significant changes in AMPK phosphorylation, suggesting an AMPK-independent mechanism (Fig. 3j, k, WT genetic background).

Taken together these data indicate that AMPK activity is repressed by formate, either produced endogenously by mitochondrial one-carbon metabolism or supplemented exogenously.

Formate increases pyrimidine nucleotides

The correlation analysis also revealed a positive association between the one-carbon availability and the levels of pyrimidine precursors dehydro-orotate and orotate and pyrimidines UMP and deoxythymidine triphosphate (dTTP) (Fig. 3e–h). Glutamine and aspartate, two precursors of pyrimidine synthesis, were rather depleted with increasing the availability of one-carbon units (Fig. 3c, d), excluding them as the cause of increased dehydro-orotate and orotate levels.

Acadesine, which is converted to intracellular AICAR after uptake and phosphorylation, was reported to increase orotate levels²⁵. To follow this lead we performed purine nucleotides supplementation experiments and quantified intracellular metabolites. We found no association between endogenous AICAR and orotate levels (Fig. 4a, b). Since the first step of pyrimidine synthesis is catalysed by the ATP-dependent activity of carbamoyl-phosphate synthetase, we hypothesized that the observed changes in orotate levels are determined by changes in ATP levels. Indeed, there is a better association between the intracellular levels of orotate and ATP, than between orotate and AICAR (Fig. 4a–c). This association is more evident in a scatter plot of orotate versus ATP levels (Fig. 4d). These data are explained by a theoretical model of orotate metabolism (Fig. 4e, Supplementary Text). When the maximum orotate production rate exceeds that of turnover the model predicts a steep increase in orotate levels as the levels of ATP approach a limiting value, as observed (Fig. 4f).

The changes in dehydro-orotate and orotate levels are recapitulated in the panel of SHMT2 deficient cell lines (Fig. S2f, g). In the HCT116, MDA-MB-231 and SKB3 backgrounds the SHMT2 deficiency causes a drop in ATP levels that is accompanied by a drop in dehydro-orotate and orotate levels. In contrast, in the MDA-MB-468 and T47D cell lines, where the SHMT2 deficiency does not decrease the ATP levels, there are no appreciable changes in the dehydro-orotate and orotate levels.

Formate supplementation

To provide additional evidence, we have titrated the amount of formate supplemented to the MFT-SHMT1 deficient cell line and quantified intracellular metabolites (Fig. 5a–j). Formate supplementation induced a dramatic difference in metabolite concentrations at a formate concentration of about 100 μM. Below this concentration the adenine nucleotide levels are low and approximately constant (Fig. 5a–c), increasing twofold or higher at formate concentrations of 500 μM or 1 mM. At low supplemented formate concentration (below 100 μM), AICAR increases with increasing the concentration of supplemented formate, then drops down to undetectable levels at the formate concentrations of 500 μM or 1 mM (Fig. 5d). Dehydro-orotate and orotate also exhibit a sharp increase above a supplemented formate concentration of 100 μM (Fig. 5e, f). Finally, there is a gradual decrease of the intracellular glucose concentration with increased concentration of supplemented formate (Fig. 5g). In contrast, the intracellular lactate levels exhibit a switch-like behaviour, with a sharp increase above a supplemented formate concentration of 100 μM (Fig. 5h). The switch-like increase in lactate levels can be explained by the association between the rate of glycolysis and the ADP levels and the switch-like increase of ADP levels induced by formate (Fig. 5b).

To search for additional metabolites that could be modulated by ATP levels we calculated the spearman correlation coefficient between ATP and intracellular metabolite levels (Table S6). As anticipated by the results described above, the top associations included a positive correlation with purines and pyrimidines and a negative association with the purine synthesis intermediate metabolites (GAR, SAICAR, AICAR). We also noted a positive correlation between ATP levels and the levels of argininosuccinate (Fig. 5i). Argininosuccinate synthetase is an ATP driven enzyme that, as carbamoyl-phosphate synthetase, exhibits cooperativity for ATP²⁶. A kinetic study of yeast argininosuccinate synthetase indicates that the enzyme kinetics is characterized by a sigmoidal dependency with respect to the concentration of ATP, with a Hill coefficient of 2. Therefore, similarly to orotate, the formate-dependent increase of argininosuccinate can be explained by the formate-dependent increase of ATP and the ATP-dependent activity of arginonosuccinate synthetase. We also noted that malate is increased following formate supplementation (Fig. 5j). These changes are consistent with the fact that fumarate is a by-product of both argininosuccinate turnover and purine synthesis and that fumarate is converted to malate by fumarate hydratase. The formate-dependent induction of argininosuccinate and malate is not recapitulated when comparing the panel of SHMT2 deficient cell lines with their parental cell lines (Fig. S2h, i).

Pharmacological inhibition

Going in the opposite direction, we tested the formate-dependent metabolic switch in the context of pharmacological inhibition using the serine hydroxymethyltransfarece inhibitor SHIN1²⁷. The data are an almost specular image of what is observed in the formate supplementation experiments (Fig. S3a–j). From this data we can conclude that inhibition of serine hydroxymethyltransfarase activity causes a systemic inhibition of cell metabolism that is mediated by the formate-dependent metabolic switch uncovered here.

In vivo validation in mouse models of cancer

To provide an in vivo validation we analysed differences between tumours and the adjacent normal tissues (Fig. 6a). As previously shown, the relative rate of serine catabolism to formate is increased in the transformed tissues relative to the adjacent normal tissues, in the APC^min/+ mouse model of colorectal adenomas and the PyMT model of breast adenocarcinoma⁷ (Fig. 6b). We have also shown that the transformed tissues have a high NAD⁺/(NAD⁺ + NADH) ratio that is similar than the adjacent normal tissue⁷ (Fig. 6c). The latter suggests that the tumour tissues have a similar redox status than the adjacent normal tissue. We have re-analysed the LC-MS data to extract the quantifications of relevant metabolites. The fraction of de novo synthesized purines, a surrogate of the purine synthesis rate, is significantly higher in the tumour tissues than in the adjacent normal tissues (Fig. 6d). The levels of ADP are increased in the transformed tissues relative to the adjacent normal tissues (Fig. 6e, trend in the APC^min/+ model and significant in the PyMT model). Furthermore, the levels of lactate are increased in the transformed tissues relative to the adjacent normal tissues (Fig. 6f, trend in the APC^min/+ and significant in PyMT models). Although these associations are not causal proof, they are consistent with our mechanistic model of increased formate production purine synthesis, ADP levels and lactate levels. In agreement with the in vitro models, there is also a significant increase in the levels of orotate in the tumour tissue relative to the adjacent normal tissue (Fig. 6g).

The AICAR levels are increased in the tumour tissue relative to the normal tissue (Fig. 6h). Based on our in vitro data this change would be the expectation if the transition happens from low to intermediate formate availability. In the genetic in vitro model, AICAR increases from the MFT-SHMT1 deficient to the SHMT2 or MFT deficient cell lines, then dropping in the WT cells (Fig. 3b). In the in vitro model of formate supplementation, AICAR increases when the MFT-SHMT1 deficient cells are supplemented with formate up to a concentration 100 μM, then dropping at 1 mM supplemented formate (Fig. 5d). We note that aspartate and glutamine, which are required as cofactors both upstream and downstream of AICAR, exhibit significantly higher levels in the tumour tissue relative to the normal tissue, while glycine is not significantly different (Fig. S4a–c). This increase in aspartate and glutamine levels in the tumour tissue may also contribute to the increased AICAR levels.

Finally, the levels of argininosuccinate and malate are significantly increased in tumour tissue of the PyMT model, but not in the APC^min/+ model (Fig. 6i, j). A similar discrepancy was observed in our in vitro models. The HAP1 cells manifest a SHMT2-dependent elevation of the intracellular argininosuccinate levels, but this is not the case for the other cell lines tested. This discrepancy is anticipated by our theoretical analysis (Supplementary Text), depending on the relative maximum activity of argininosuccinate synthesis and turnover. In turn, the lack of significant changes of malate in the APC^min/+ model could be the consequence of lack of changes in argininosuccinate, which is turned over to fumarate and subsequently to malate (Fig. 5, pathway inset).

In vivo validation in human colorectal cancer

Recently Satoh et al. have shown that colorectal cancers are characterized by a global metabolic reprogramming induced by Myc²⁸. Since MYC activates the transcription of mitochondrial one-carbon metabolism genes^3,29, we hypothesized that MYC-driven colorectal cancers should manifest the formate-dependent metabolic switch. To test this hypothesis, we first performed a gene signature analysis using the reported gene expression array data for 41 colorectal tumour samples and 39 normal colorectal samples²⁸. Using gene set enrichment analysis¹⁹ we quantified the enrichment of relevant gene signatures in the different samples (gene signature enrichment score). There are no significant differences in the enrichment scores for gene signatures of oxidative phosphorylation, HIF1α targets and glycolysis (Fig. 6k). In contrast, there is a significant increase of the mitochondrial one-carbon metabolism enrichment score signature in the tumour relative to the normal samples (Fig. 6k). The latter is also consistent with an increase of the enrichment score of a MYC targets signature in the tumour relative to the normal colorectal samples (Fig. 6k).

Next, we analysed reported metabolomic data from 275 normal and 275 tumour samples²⁸. The tumour tissues exhibit a high NAD⁺/(NAD⁺ + NADH) ratio that is not significantly different from that of the normal tissues (Fig. 6l). That together with the HIF targets signature evidence suggest that the MYC-driven colorectal tumours are of oxidative nature and that their oxidative status is not significantly different from that of normal tissues. In contrast, there is a significant increase in the levels of ADP, lactate and malate in the tumour samples relative to the normal tissues (Fig. 6m–o). Although these associations are not causal prove, they are consistent with what predicted by the formate-dependent metabolic switch.

In vivo validation following a formate bolus

To provide a direct in vivo validation of the formate-dependent metabolic switch we intraperitoneally administered a bolus of formate or vehicle to C57BL/6J mice fasted overnight (Fig. 7a). Different mice were used to collect plasma samples at 1, 2 and 4 h after the bolus injection. Plasma formate reached between 500 μM to 1 mM levels 1 h after the bolus injection, going down to μM levels 2 h after the bolus injection (Fig. 7b). At 1 h there is a significant depletion of plasma glycine (Fig. 7c) and a significant increase of plasma serine (Fig. 7d), which are consistent with the reverse activity of liver serine hydroxymethyltransferase and the administration of excess formate.

At 1 h, when the observed formate concentration was highest, there is a significant depletion of plasma glucose (Fig. 7e) and a non-significant trend towards increased plasma lactate (Fig. 7f). These changes are consistent with the formate-dependent induction of glycolysis. The lack of significant changes in plasma lactate could be due to lactate oxidation at the tissues where glucose metabolism is increased.

As observed in our in vitro models, the formate bolus induces a significant increase of plasma orotate and malate levels at the 1-h time point (Fig. 7g, h). In the case of orotate the significant increase persists 4 h after bolus injection. Since all other significant changes are absent at the 2- and 4-h time points, the simplest explanation is that the orotate turnover is slow, taking a long time to come back to control levels. In contrast, the levels of argininosuccinate does not change significantly at any time point (Fig. 7i). Finally, there are no significant changes in the levels of other amino acids implicated in purine, pyrimidine and argininosuccinate metabolism (aspartate, arginine and glutamine, Fig. 7j–l).

Therefore, a bolus of formate causes changes at the level of whole-body metabolism that are similar to what observed in our in vitro models.