Raman micro-spectroscopy reveals the spatial distribution of fumarate in cells and tissues

Aberrantly accumulated metabolites such as fumarate elicit intra– and inter-cellular pro-oncogenic cascades, yet current methods to measure them require sample perturbation or disruption and lack spatio-temporal resolution, limiting our ability to fully characterize their function and distribution in cells and within a tissue. Raman spectroscopy (RS) is a powerful bio-analytical tool that directly characterizes the chemical composition of a sample based solely on the optical fingerprint of vibrational modes. Here, we show for the first time that RS can directly detect fumarate in living cells in vivo and animal tissues ex vivo. Using the observed linear relationship between Raman scattered intensity and fumarate concentration, we demonstrate that RS can distinguish between Fumarate hydratase (Fh1)-deficient and Fh1-proficient cells based on their fumarate concentration. Moreover, RS reveals the spatial compartmentalization of fumarate within cellular organelles: consistent with disruptive methods, in Fh1-deficient cells we observe the highest fumarate concentration (37 ± 19 mM) in the mitochondria, where the TCA cycle operates, followed by the cytoplasm (24 ± 13 mM) and then the nucleus (9 ± 6 mM). Finally, we apply RS to tissues from an inducible mouse model of FH loss in the kidney, demonstrating that RS can accurately classify FH status in these tissues. These results suggest that RS could be adopted as a valuable tool for small molecule metabolic imaging, enabling in situ dynamic evaluation of fumarate compartmentalization.


Introduction
Reprogramming of cellular metabolism is a key hallmark of cancer. 1 Perhaps the most prominent example of this is the existence of oncometabolites, aberrantly accumulated metabolites with pro-oncogenic capabilities. 2 Mutations in the gene encoding fumarate hydratase (FH) lead to profound cellular metabolic alterations and fumarate accumulation, which predispose to the rare hereditary leiomyomatosis and renal cell cancer (HLRCC) syndrome. 3 FH loss or transcriptional downregulation is also well established in a range of other cancers, 4-7 implying a key role for fumarate in tumorigenesis.
The accumulation of fumarate leads to diverse biological consequences. For example, accumulated fumarate can react with cysteine residues of proteins, leading to the post translational modification known as succination. Protein succination is a robust marker of FH loss in HLRCC tumours and it affects hundreds of proteins with important pathophysiological consequences, which are compartment specific. 8 In the mitochondria, fumarate-induced succination of the family of Fe-S cluster biogenesis proteins Iscu, Bola, and NFU1 decreases their availability for respiratory chain complexes, causing mitochondrial dysfunction. 9 In addition, mitochondrial fumarate can promote the vesicular release of mtDNA, causing the activation of innate immunity. 10 In the cytosol, fumarate-induced succination inactivates Keap1, leading to the subsequent stabilization of the antioxidant master gene NRF2, 11 or it can form succinic glutathione, leading to oxidative stress and senescence. 12 Fumarate can also inhibit α-ketoglutarate (αKG)-dependent dioxygenases such as prolyl hydroxylases, leading to the aberrant stabilization of HIF (Hypoxia-Inducible Factor) under normal oxygen levels. Nuclear accumulation of fumarate has a profound impact on cellular function, through inhibition of histone and DNA demethylases, leading to substantial epigenetic changes. 13 Fumarate is also implicated in genome stability and DNA repair; the ability of FH-deficient cells to respond to DNA damage is compromised, making them more sensitive to DNA damage. 13 Interestingly, the function of FH in DNA repair can be replaced by high concentrations of fumarate, partially compensating for the loss of FH. 13 Taken together, these lines of evidence indicate that fumarate exerts multiple biological functions in different cellular compartments, eliciting distinct cellular responses. To precisely understand how FH-deficient cells cope with the excess accumulated fumarate and unveil how diverse signals elicited by fumarate are coordinated, it is crucial to quantify fumarate concentrations dynamically and visualize its intracellular spatial distribution. Current methods to measure the intracellular concentration of fumarate rely on liquid chromatography -mass spectrometry (LC-MS), reporting average cellular concentrations of ~9 mM in FH-4 deficient cells. 14 Yet, LC-MS metabolite extraction protocols disrupt the intracellular architectures of a cell and require lengthy subcellular fractionations to interrogate concentrations in different compartments of the cell, hence are not well-suited to dynamic studies in living cells or tissues. Furthermore, spatial information within these compartments is lost. Small molecules such as fumarate are not amenable to staining for confocal fluorescence microscopy, since fluorescent labels are typically larger than the molecule under study and thus interfere with metabolism. 15 Recently, in vivo metabolic magnetic resonance imaging [16][17][18] has been used to track the production of [2,3-2 H2]-or [1,[4][5][6][7][8][9][10][11][12][13] C] malate in tumours after injection of labelled fumarate, however, this approach cannot probe the endogenous fumarate pool. 19 Knowledge of the spatiotemporal dynamics of fumarate accumulation in, and trafficking between, cellular compartments thus remains elusive, making it difficult to understand how the many signals elicited by fumarate are coordinated.
Raman Spectroscopy (RS) is a label-free technique that distinguishes chemical compounds by the optical signature of their vibrational modes and has been widely applied in living cells. [20][21][22][23] In particular, Raman spectra from living cells are sensitive to changing concentrations of lipids, 24-27 proteins, 28 carbohydrates and nucleic acids, 20,23 enabling identification and spatial localization of the main cellular compartments. 21,28 Time-lapse studies can be applied to discriminate the emergence of oxidative stress 29 , the internalization and spatial distribution of therapeutic compounds, 30,31 as well as the progress of the apoptotic cascade. 32 Nonetheless, while surface-enhanced Raman methods have shown promise in measuring metabolites in biofluids 33,34 , spontaneous RS of small molecules in intact cells has thus far remained beyond the limit of detection. 35 Here, we demonstrate that RS can directly detect and map the oncometabolite fumarate in living cells. We first examine the fumarate Raman spectra in detail to understand the chemical origin of the observed vibrational modes. We then use Fh1-deficient andproficient cell lines and mouse models to demonstrate the limits of detection and evaluate the performance of RS as a metabolic imaging tool. We show that RS can discriminate Fh1deficient and -proficient cells and tissues, while revealing the spatial distribution of fumarate concentrations according to cellular compartment. Our findings open an exciting new avenue to metabolic imaging of small molecules, enabling the study of in situ dynamics in living cells and tissues.

DFT calculations
Density functional theory (DFT) calculations were performed on major TCA cycle metabolites (including different hydrolysis states of fumarate, succinate, malate, α-ketoglutarate and oxaloacetate, all in vacuum) using the hybrid B3LYP exchange-correlation energy functional. [36][37][38][39] Starting structures were generated using Chem3D 16.0 software (CambridgeSoft). Gas-phase geometry optimisations were performed using the Def2GFP basis set, except for disodium fumarate where the Popletype 6-311G* basis set was used.
Raman frequency calculations were then performed on the optimised structures using the Gaussian '09 ab initio program package. 40

Preparation of chemical compounds for RS
To measure metabolites in solution, powders were dissolved in deionized water (18.2 mΩ cm) to a concentration of 200 mM to remain below saturation. Concentration dilution series of fumarate and succinate were prepared as follows. First, powders were dissolved to a 1 M 6 stock solution (total volume: 100 mL), after which dilutions in steps of 100 mM were made by combining stock solution with deionized water to a total volume of 10 mL in polypropylene centrifuge tubes (Fisherbrand). A total volume of 100 mL was prepared of the 100 mM solution, which was similarly used as a stock solution to prepare lower concentrations in steps of 10, 5 or 2 mM.

Monolayer Cell Culture and Cell Preparation for RS
We used Fh1-proficient and -deficient cell lines, both of which are adherent epithelial murine kidney cells originally derived as described in [ 41 ]. Briefly, in this method, immortalized Fh1- For mitochondrial labelling, compound 1 described in ref. [ 42 ], referred to hereafter as mitokyne, was prepared and provided by Prof. Duncan Graham at the University of Strathclyde. The compound was dissolved in DMSO to a concentration of 1 mM. 10 µL of mitokyne solution was added to a well containing a quartz cover slip with cells of interest and 2 mL medium, followed by incubation for 30 mins. The cover slip was rinsed twice with imaging medium before placing in the cell chamber.

Animal handling and tissue harvesting
All animal experiments were conducted in accordance with the United Kingdom Animals For tamoxifen preparation and treatment, 5 mL ethanol were added slowly to 1g tamoxifen (Sigma-Aldrich) in a 50 mL Falcon tube. The tamoxifen/ethanol mix was sonicated at 40% amplitude in 20 s pulses until the Tamoxifen was completely dissolved. 50 mL of corn oil (Sigma-Aldrich) pre-heated at 50 °C were added immediately to the tamoxifen/ethanol mix to obtain a 20 mg/mL Tamoxifen stock solution. The tube was then vortexed and incubated for up to 12 h in an orbital shaker at 50 °C to ensure adequate solubilisation. Aliquots of 5 mL were then stored indefinitely at -20 °C. Prior to injection, the mixture was allowed to warm to room temperature. Age-matched mice (between 10 and 12 weeks old) were used in all experiments. Each animal received three doses of 2 mg tamoxifen each via intraperitoneal injection. Tamoxifen was administered every other day in order to allow animals to recover between doses. At day 10 post-induction, animals were killed by neck dislocation and the kidneys were speedily collected and snap frozen in liquid nitrogen. Tissues were frozen-8 sectioned onto Thermo Scientific SuperFrost Plus slides, with a thickness of 6 µm and a minimum separation of 300 µm between sections.

Confocal Raman micro-spectroscopy
RS was performed on an inverted confocal Raman microscope (Alpha 300M+, WITec GmbH, Germany) equipped with a 532 nm laser (WITec GmbH, Germany) and a 785 nm single mode diode laser (XTRA II; Toptica Photonics Inc., USA). The output signal was coupled into a 300 mm triple grating imaging spectrometer (Acton SpectraPro SP-2300; Princeton Instruments Inc., USA) with a 600 l/mm grating (500 nm blaze) and a 1200 l/mm grating (750 nm blaze), coupled into a 100 μm single-mode fibre. Spectra were recorded on a thermoelectrically cooled CCD camera (DU401A-BV; Andor, Ireland) using Control FOUR of the WITec Suite software (WITec GmbH, Germany).
The spectral resolution of the Raman spectra was 3.61 cm -1 (532 nm laser, 600 l/mm grating) and 1.65 cm -1 (785 nm laser, 1200 l/mm grating). A 60x water immersion objective with cover glass correction collar (Nikon Plan Apo IR 60x, NA 1.27) was placed against the cover slip for imaging. For this objective, a 532 nm laser has a theoretical lateral resolution of 214 nm and spot size of 511 nm. The spatial resolution was determined by scanning over the edge of a silicon wafer. A silicon wafer can be split cleanly along a crystal plane, to create an edge response. Differentiating the edge response gave the line spread function (LSF), and the spatial resolution of the microscope was taken to be the full width at half maximum (FWHM) of the LSF. A spatial resolution (0.35 ± 0.04)•10 3 nm at 532 nm incidence and (1.5 ± 0.2)•10 3 nm at 785 nm incidence ( Figure S1) were confirmed. The wavenumber scale was calibrated in the fingerprint region using known peak positions of biphenyl-4-thiol (powder, 97%, CAS 19813-90-2, Aldrich). Before measuring each sample, a silicon wafer was measured to ensure reproducible intensities, ensuring (8.0 ± 0.3) x 10 2 counts mW -1 s -1 for the primary band.
To measure dilution series of fumarate, 1 mL solution was placed onto a quartz cover slip in a cell chamber. The objective was focused into the solution just above the quartz cover slip.
For spectra of dry metabolites and 200 mM solutions, a few mg of powder or 1 mL of solution was placed in a well of a µ-Slide 8 Well plate (Ibidi, with #1.5 polymer cover slip). Spectra were acquired by focusing beyond the cover slip of the 8-well plate onto the powder or into the solution, moving the objective until the polymer cover slip's fingerprint spectrum had disappeared.
To enable live cell imaging, the microscope was equipped with a custom chamber to control the temperature (37 °C) and CO2 level (5%) (Digital Pixel). Cells were seeded onto unmodified fused quartz cover slips, which only display intrinsic Raman signals only below 500 cm -1 , outside of the wavenumber range of interest for live cells. 45 Additionally, these cover slips are thin enough (0.17 mm, #1.5) to use with a short working distance water immersion objective, in contrast to widely used calcium fluoride windows. For tissue imaging, sections were imaged through a quartz cover slip placed directly on top of the defrosted tissue.
Data acquisition for spontaneous RS requires a careful trade-off between resolution, step size, integration time, and total imaging time. To determine the optimal conditions for mapping, we compared 532 nm and 785 nm laser wavelengths available on our instrument. Similar laser wavelengths, laser power and integration times have been used successfully in literature to map cells. 30,46,47 Raman cross sections scale to the fourth power with laser frequency. 32 Integration times of ~30 s were required to achieve sufficient signal-to-noise ratio for peak fitting at 785 nm excitation at 90 mW power (on sample, as determined with an optical power meter, Thorlabs PM100D), which led to a scan time of ~120 min to cover the area of a single cell. At 532 nm excitation at 26 mW power, an integration of 0.3 s was found to provide comparable signal-to-noise ratio, enabling single cell coverage in ~10 min with a spatial step size of 0.5 μm, albeit with a slightly poorer spectral resolution compared to 785 nm due to the grating characteristics. The shorter imaging duration is highly advantageous, being less susceptible to cellular motion and enabling a higher throughput of data acquisition.
Nonetheless, illumination at 532 nm has a higher potential to cause long-term damage to the cells with extended illumination at ≥5mW. 48,49 In our study, cell morphology was unchanged during <15 min scanning time and no change in chemical composition was observed. To confirm the expected damage threshold in the cell line under study, we illuminated the same spot on one cell repeatedly at 0.5 s acquisition time with 0.5 s pause, observing spectral changes only after 200s, suggesting that our experimental parameters are below the threshold at which biological changes would be expected ( Figure S2). For PLS-DA, data consisted of line scans taken through a major axis of the cell in 20 steps at 5s acquisition time (532 nm incidence, 26 mW power) or 10 steps at 30s acquisition time (785 nm incidence, 120mW power), to maximise signal-to-noise ratio while still representing the spectral contributions from the major cellular compartments.

Data processing and availability
Data processing was performed in Igor Pro (version 8.0.4.2). Cosmic ray removal was applied to all spectra before analysis. To analyse fumarate concentration dilution series, a flat baseline was subtracted (the average number of counts between 1324 -1379 cm -1 , the wavelength range preceding the fumarate band of interest) and the area under the curve (between 1399 -1419 cm -1 ) was integrated, leading to a calibration curve as a function of fumarate concentration. For baseline subtractions in cellular spectra, we used an iterative algorithm fitting a polynomial ( Figure S3). To visualize differences between the fingerprint areas of different cell lines, map-scans of each cell were averaged and then baseline subtracted with a tenth order polynomial (10 iterations). To extract fumarate concentrations from individual cellular spectra, we applied a 4 th order polynomial baseline subtraction (7 iterations with a 'tolerance' of 15 counts, which was taken as twice the experimental noise floor, Figure S4), and integrated the area under the baseline-corrected curve. No Savitzky-Golay filter 50 was applied before processing, since such smoothing was found to affect extracted concentrations depending on chosen smoothing window. Alternatively, a Gaussian peak shape was fitted to the spectrum without baseline subtraction, and the fitted peak height was used as a measure for fumarate concentration.

for WT cells and for KO cl.1 cells, 6 for KO cl.19 cells) was such that 4-5 classes
remained for the cellular spectra corresponding to major organelles.

Theoretical and experimental Raman spectra of fumarate indicate a suitable limit-of-detection for application in cell studies.
Density functional theory (DFT) simulations were performed to estimate the anticipated Raman cross sections of fumarate and comparable TCA cycle metabolites (succinate, malate, α-KG and oxaloacetate) in vacuum ( Figure S5). While DFT simulations cannot provide a quantitative prediction of the spectral peak positions, due to the various ways in which a molecule can ionize, hydrogen-bond with water, or complex with available ions, they do give a general indication of relative cross sections and aid assignment of peaks to specific vibrational modes. All metabolites displayed prominent peaks in the 400-1800 cm -1 (cellular "fingerprint" region) and 2800-3200 cm -1 ("long wavenumber" region), covering carbon-carbon bond vibrations and C-H bond vibrations respectively. Encouragingly, fumarate displayed up to 5-fold higher peak intensities in the fingerprint region, suggesting a lower limit-of-detection (LOD) compared to other TCA intermediates. DFT calculations examining a range of fumarate ions further show a general trend towards peaks at lower wavenumbers and with lower intensities during the transition from a non-hydrolysed state via monovalent ions to divalent ions of fumarate ( Figure S6).
Next, we sought to compare the simulations to experimental Raman spectra acquired from powders and aqueous solutions. Comparing the sodium salts of primary TCA cycle intermediates ( Figure 1A,B), we see the highest Raman peak intensities for fumarate, indicative of a larger Raman cross section and in line with DFT simulations.
The fumarate powder spectrum is dominated by four peaks in the low wavenumber range at 913±2 cm -1 , 1296±2 cm -1 , 1431±2 cm -1 and 1657±2 cm -1 . Spectral peak positions and intensities are expected to change in solution, as ionization (depending on the cellular pH) and hydrogen bonding affect vibrational modes. We prepared aqueous solutions of metabolites using their disodium salts, which have a higher solubility than their acid counterparts. In these aqueous solutions, for many of the TCA cycle intermediates, the powder Raman peaks are absent or not distinguishable above the background ( Figure 1C). Again, fumarate stands out, displaying > 2-fold higher peak intensities in the fingerprint wavenumber range compared to metabolites. In solution, fumarate displays three dominant peaks at 1277±2 cm -1 , 1401±2 cm -1 and 1652±2 cm -1 ( Figure 1D), indeed shifted to lower wavenumbers compared to the powder spectra, as predicted by DFT. The 1277 cm -1 band is a C-H deformation (δ) mode. The 1401 cm -1 peak arises from a fully symmetric CO2stretch (ν) vibration, and the 1652 cm -1 band 13 corresponds to the C=C stretch/CO2symmetric bending mode with a small shoulder peak belonging to C=C stretch/CO2asymmetric bending mode (see Supporting Videos SV1-4, c.f. refs. 53,54 ). Finally, we examined a concentration dilution series to establish the sensitivity and limit-ofdetection (LOD) for fumarate on our microscope. By integrating the area under curve for 14 fumarate bands, their Raman intensity (in cts cm -1 s -1 ) can be extracted as a function of fumarate concentration, serving as a calibration curve ( Figure 1E,F). A sensitivity of 19.2±0.1 cts cm -1 s -1 mM -1 was recorded for the 1401 cm -1 peak for the current experimental setup at 26 mW incidence power. The LOD is defined here as the fumarate concentration above which the concentration series can no longer be fitted with a horizontal line; at a dwell time of 0.3 s, this occurs at 8 mM ( Figure 1F). Since signal-to-noise improves with longer dwell times, the LOD can be brought down to 4 mM for 5 s dwell time ( Figure S7)

Fumarate can be resolved in Raman spectra of Fh1-KO cells.
To examine the potential of RS to resolve fumarate accumulation, we used two previously characterised 2,41,55,56 mouse FH-deficient cell lines (Fh1 -/-Cl1 and Fh1 -/-Cl19 ) and their isogenic control (Fh1 fl/fl ), which will be referred to as 'knock-out' (Fh1-KO) and 'wild-type' (Fh1-WT), respectively. Initially, we performed area scans of 20 individual cells per condition and obtained the average spectra across these scans (Figure 2A). Raman spectra display four regimes: the 'fingerprint' region rich in Raman peaks of DNA, proteins and lipids (500-1800 cm -1 ); a Raman silent region (1800-2750 cm -1 ); the 'long-wavenumber' region of protein and lipids peaks (2800-3000 cm -1 ); and two broad peaks caused by the symmetric and asymmetric stretch vibrations of water (>3000 cm -1 ). We focus here on the fingerprint region, which is of greatest interest for evaluation of fumarate concentrations based on our DFT and solution spectra results.
The fingerprint region is dominated by the CH2 deformation mode 32 associated with lipids, as well as by the protein Amide I and III bands and DNA bases (see annotations on Figure   2A). 57,48,58,32,59 Significant contributions also arise from phenylalanine and cytochrome C. 30 While the 1652 cm -1 band of fumarate overlaps with the Amide I band and the bending mode of water, the 1401 cm -1 mode is expediently located in the valley adjacent to the CH2 deformation mode, apparently isolating it from other prominent peaks in the cell. Examining the 1401 cm -1 mode, being the preferred band for fumarate detection, the average spectra of both Fh1-KO clones indeed displayed a higher Raman intensity than Fh1-WT cells (Figure 2A inset) and the difference spectra confirm a significant contribution at 1401 cm -1 ( Figure S8A).

15
The higher fumarate peak intensity persists when corrected for the generally higher Raman intensities of KO cells (Figure 2A) after normalizing by the area under the curve in the fingerprint region ( Figure S8B). Using a PLS-DA model to classify the cells, 1280 cm -1 , 1401 cm -1 and 1652 cm -1 fumarate peaks feature prominently in the variable importance projection scores ( Figure S9) demonstrating that accumulated fumarate is the main discriminator between the cell lines, with the model showing over 97% sensitivity and specificity for Fh1-WT and more than 93% sensitivity and specificity for the Fh1-KO clones (Supplementary Table 1).
These findings were also confirmed (using longer scan times) at 785 nm ( Figure S10, S11), where the 1401 cm -1 peak also becomes more pronounced due to the higher spectral resolution.
To further test whether the observed peak at 1401 cm -1 stems from fumarate, Fh1-WT and Fh1-KO cells were cultivated in medium with isotopically labelled glutamine-13 C5, which generates fully labelled 13 C4-fumarate. 41 DFT calculations show that the 1401 cm -1 peak would be expected to shift to 1373 cm -1 for fumarate-13 C4 ( Figure S12). The Fh1-KO cells cultivated with L-glutamine-13 C5 lacked the 1401 cm -1 band, with a minor increase in the 1373 cm -1 peak observed ( Figure 2B). The relatively modest increase at 1373 cm -1 may be due to the formation of a range of isotopically labelled fumarate species, each with their own characteristic vibrational modes. Taken together, these findings demonstrate the accumulation of fumarate upon Fh1 loss is resolvable using RS.

Fumarate concentrations can be spatially resolved and vary significantly between cellular compartments.
Comparing cell spectra (Figure 2A) with fumarate calibration curves ( Figure 1E,F) enables the extraction of cellular fumarate concentrations ( Figure 3A). In WT cells, the bulk of fumarate appears to lie below the LOD, hence cannot be accurately determined. Average fumarate concentrations in Fh1-KO cells derived from these histograms are 10 ± 2 mM (mean ± standard error among cells for Fh1-KO cl.1) and 20 ± 4 mM (for Fh1-KO cl. 19), which are in the range previously reported. Maximum apparent concentrations reach up to 60 mM in Fh1-KO cl.19.
K-means clustering was applied to all area scans per cell line to partition the spectra by cellular compartment. The k-means cluster maps and their associated loadings (here called set K1) are mostly determined by the long-wavenumber bands, which are most intense ( Figure 3B,C).
The cluster maps visualize the spatial organisation of the cell into nuclear, cytoplasmic, mitochondrial, and membrane regions, as confirmed by performing a peak assignment.
Primarily, the k-means loadings of the mitochondria are characterised by prominent peaks of cytochrome C ( Figure S13), which is indeed located in the mitochondrial membrane 30,31,60,61 .
For WT cells, a faint band at 1401 cm -1 is present throughout and a second faint band at 1388 cm -1 becomes more pronounced from nucleus to mitochondria. These faint bands can be attributed to the oxidized and reduced states of cytochrome C, as previously observed in mitochondria via SERS 61 (see also the shoulder peak in Figure S13). Thus, the 1401 cm -1 band of fumarate is not completely isolated from underlying cellular bands as anticipated from the average cell spectra. A further potential confounder could be the complexes formed by excess fumarate with cysteine and the cysteine-containing protein glutathione, which could alter the fumarate vibrational modes. S-(2-succinyl)glutathione was observed to display a band at 1413 cm -1 of negligible intensity compared to fumarate itself even for a >100 mM solution ( Figure S14). Therefore, we deem it unlikely this fumarate adduct is contributing to signals in the cell, but it is possible that such complex formation could reduce observed fumarate concentrations measured by the 1401 cm -1 band. Nonetheless, the fumarate band at 1401 cm -1 is intense and becomes increasingly visible in the loadings from nucleus to mitochondria for Fh1-KO cells, giving a first insight into its distribution. Having confirmed the k-means cluster assignments using a targeted mitochondrial Raman probe (see Supplementary Note 1 and Figure S15), we then mapped the spatial fumarate concentrations in the cell. In a qualitative comparison of the fumarate locations ( Figure 3D-E, left) and k-means maps for KO cells ( Figure 3D-E, right), fumarate is located preferentially outside the nucleus. By comparing k-means cluster maps and fumarate concentration maps, the average fumarate concentration in each identified organelle was calculated (Table 1).
Although average fumarate concentrations differ slightly per cell, they tend to increase from relatively lower levels in the nucleus and cell membrane to the cytoplasm and still further in the mitochondria (Figure 3F-H). Similar concentrations were found when extracted by fitting a Gaussian peak shape to the 1401 cm -1 band, which reduces the influence of parameter choices as it does not require baseline fitting, but is more susceptible to noise (Figure S16, Supplementary   (Table 1).

RS reveals fumarate in excised tissues from an inducible mouse model of FH-loss in the kidney.
To assess whether fumarate can also be detected in tissues, we examined tissues from an inducible transgenic mouse model of Fh1-loss 10

Discussion
We show here the potential of RS to detect fumarate in situ, identifying the Raman fingerprint of the molecule and showing a linear concentration dependence down to an LOD of ~8 mM. We found that the distinctive Raman peak arising from the symmetric CO2stretch vibration (1401 cm -1 ) could be distinguished in live murine kidney Fh1-knock out cells but was not apparent in wild-type cells. Moreover, the peak height varied by cellular compartment, Additionally, while the 1401 cm -1 peak can be distinguished, it sits on the shoulder of an adjacent peak at 1445 cm -1 , associated with CH2 mode in DNA, making quantification challenging. We also identified potential confounders at that peak location, with weak contributions possible from cytochrome C and S-(2-succinyl)glutathione, which likely underpin at least some of our measured signal from Fh1-WT cells. Stable isotope probing techniques may add value in this context, which we already examined by the application of L-glutamine-could be to use deuterium labelling, since deuterium peaks typically appear in the silent region of the cellular Raman spectrum, thus improving signal-to-noise ratio.
Finally, fumarate accumulation induces a host of cellular responses that could also be detected by RS, for example, oxidative stress, 29,70-72 which we did not address in this study.
To tease apart these different contributions to the Raman spectra, machine learning methods that have already shown promise in a range of RS applications could be applied. 73 For example, non-negative matrix factorization and principal component analysis have assisted in analyzing radiation response biomarkers and hypoxia indicators in cells and tissues, 74 while a convolutional neural network has been applied for metabolomics on liver carcinoma tissue. 75 Adopting machine learning may also aid in detection of other oncometabolites with less expediently located vibrational modes, such as succinate.
Overall, this study introduces RS as a tool to enable in situ mapping of fumarate, opening new avenues for the study of fumarate accumulation and spatial compartmentalization in cells and within tissues in pathophysiological processes.

Data availability
All code and raw data associated with this manuscript will be made available at