Cancer metabolism in a snapshot: MRS(I)

The contribution of MRS(I) to the in vivo evaluation of cancer‐metabolism‐derived metrics, mostly since 2016, is reviewed here. Increased carbon consumption by tumour cells, which are highly glycolytic, is now being sampled by 13C magnetic resonance spectroscopic imaging (MRSI) following the injection of hyperpolarized [1‐13C] pyruvate (Pyr). Hot‐spots of, mostly, increased lactate dehydrogenase activity or flow between Pyr and lactate (Lac) have been seen with cancer progression in prostate (preclinical and in humans), brain and pancreas (both preclinical) tumours. Therapy response is usually signalled by decreased Lac/Pyr 13C‐labelled ratio with respect to untreated or non‐responding tumour. For therapeutic agents inducing tumour hypoxia, the 13C‐labelled Lac/bicarbonate ratio may be a better metric than the Lac/Pyr ratio. 31P MRSI may sample intracellular pH changes from brain tumours (acidification upon antiangiogenic treatment, basification at fast proliferation and relapse). The steady state tumour metabolome pattern is still in use for cancer evaluation. Metrics used for this range from quantification of single oncometabolites (such as 2‐hydroxyglutarate in mutant IDH1 glial brain tumours) to selected metabolite ratios (such as total choline to N‐acetylaspartate (plain ratio or CNI index)) or the whole 1H MRSI(I) pattern through pattern recognition analysis. These approaches have been applied to address different questions such as tumour subtype definition, following/predicting the response to therapy or defining better resection or radiosurgery limits.


| CANCER PRODUCES A SERIES OF ANATOMICAL AND METABOLIC CHANGES: EMERGING HALLMARKS
Cancer development and progression depends on the metabolic reprogramming caused by oncogenic mutations. 4 Cancer cells are known to be highly proliferative, and the metabolism adaptation of cancer cells has the ultimate goals of survival and sustaining proliferation. This includes the ability to gather nutrients from a sometimes limited environment for viability maintenance and new biomass building. Pavlova and Thompson 5 have organized known cancer-associated metabolic changes into six hallmarks: (a) deregulated uptake of glucose and amino acids, (b) use of opportunistic modes of nutrient acquisition, (c) use of glycolysis/TCA (tricarboxylic acid) cycle intermediates for biosynthesis and NADPH (nicotinamide adenine dinucleotide 2′-phosphate, reduced form) production, (d) increased demand for nitrogen, (e) alterations in metabolite-driven gene regulation and (f) metabolic interactions with the microenvironment. The specific hallmarks exhibited by an individual tumour may contribute to better classification and help in treatment monitoring due to the relationship between malignant transformation, response to treatment and changes observable in tissue metabolism. Needless to say, the neoplastic process is also accompanied by changes at the level of microenvironment, cellular microstructure and peripheral vascular networks. The ability to sample all or most of these transformations is of crucial importance in a cancer patient's management, not only for early diagnosis but also for the non-invasive evaluation of the tumour response to a given therapy.
It is worth stressing that many therapies can target cancer metabolism, and the changes triggered in endogenous metabolites should be detectable before changes in tumour volume are observed. Last but not least, it is important to bear in mind that MR techniques are essentially sampling a volume of tissue that is not entirely composed by tumour cells. [6][7][8] Indeed in brain tumours such as glioblastoma (GB), the percentage of tumourassociated macrophages can be as high as 30% of the abnormal mass, 9 constituting a cell population with its own metabolic features and differentially affected by therapy.

| ADVANTAGES OF SAMPLING THE METABOLOME BY MRI/MRS(I)
Magnetic resonance is a non-invasive, non-ionizing, non-destructive technique, offering a high level of anatomic detail (MRI) and information about metabolomics/chemical environment (MRS(I)), being in general the method of choice for in vivo tumour diagnosis and follow-up. As an advantage over methods involving ionizing radiation, repeated MR measurements can be made in longitudinal experiments without cumulative harm to subjects. However, in vivo MR studies are only sensible if in vitro evaluations are not feasible or relevant, because the resolution and sensitivity level obtained are not comparable. In fact, most of the time ex vivo studies (i.e. high resolution magic angle spinning (HR-MAS) of biopsies) are performed to confirm or explain changes seen in in vivo studies. 3,[10][11][12][13][14][15][16][17][18][19] However, interest in MRS(I) is increasing for clinical applications and research due to its ability to provide information about tissue spatial heterogeneity through collection of spectroscopic data and spatial distribution of metabolites, as well as their changes. In this respect, MRS(I) protocols can provide relevant information on surrogate biomarkers of disease diagnosis and prognosis, helping surgery and radiation planning, as well as predicting survival and treatment response.
Next, we will consider some specific aspects in which cancer metabolism studied by MR is contributing to increase our understanding of its progression and helping to control and eventually cure it.

| Carbon consumption
Cancers are highly glycolytic; they can be highly glutaminolytic too, and this usually requires high lactate dehydrogenase (LDH) content and activity. [20][21][22] Accordingly, one would expect a tumour to demonstrate increased LDH-catalysed substrate flow through the enzyme, from pyruvate (Pyr) to lactate (Lac), with concomitant NAD + (nicotinamide adenine dinucleotide, oxidized form) regeneration, when compared with normal tissue or when progressing to higher grade. This metric is a usually accepted surrogate biomarker of increased LDH activity and has been described in Reference 23 in a preclinical model of genetically engineered mice (GEMs) that spontaneously develop pancreatic cancer, undergoing several intermediary stages (Figure 1). Hyperpolarized 13 C MRS(I) has been also used to monitor the increase in the Lac/Pyr ratio with progression to tumour, 24 which can be transformed onto magnetic resonance spectroscopic imaging (MRSI)-derived images, for example in pancreatic cancer (Supplementary Figure 2).
Conceptually similar observations have been described for the TRAMP preclinical model of prostate cancer 25 with significantly increased Pyr to Lac flux (a parameter known as k PL ) in high-grade compared with low-grade tumours monitored by hyperpolarized 13 C MRSI after the injection of hyperpolarized [1-13 C]Pyr. This correlated with a more than twofold proliferation increase (percentage of Ki67 + cells) and close to twofold lactate dehydrogenase isoform A (LDHA) expression increase in high-grade versus low-grade tumours, ex vivo.
In a first-in-human study performed with 31 prostate cancer patients receiving an escalation dose design with injected hyperpolarized [1-13 C] Pyr, 26 a high Lac/Pyr ratio identified cancer in one of the bilaterally hyperpolarized 13 C MRSI-detected regions that were inconspicuous by classical MR approaches (T 2 -weighted (T2W) or diffusion-weighted imaging) or 1 H-based MRSI (Supplementary Figure 3).
Accordingly, a tumour responding to a particular therapy should show decreased LDH activity when injecting hyperpolarized Pyr. This effect has been shown 27 in a case report of a patient with prostate-specific antigen of 25.2 ng/mL and Gleason 4 + 5 prostate adenocarcinoma on biopsy, who showed response to androgen deprivation therapy after 6 weeks of treatment, when only modest changes of abnormal prostate volume tissue and apparent diffusion coefficient (ADC) could be seen ( Figure 2).
These results are in agreement with previous findings on preclinical GB models, 28 for which the Lac/Pyr ratio after injection of hyperpolarized Pyr was seen to significantly decrease after combined dual PI3K/mTOR (mammalian target of rapamycin) inhibitor and temozolomide (TMZ) treatment, and before changes in tumour size or total choline (tCho) in the 1 H MRS pattern could be detected. Decreased flow through the glycolytic pathway due to treatment with TMZ in a preclinical GB model had also been detected, 29 but in this instance the reduced Lac/Pyr ratio after injection of hyperpolarized Pyr was caused not by LDHA expression changes but by downregulation of the M2 isoform of the pyruvate kinase enzyme (PKM2). Furthermore, mTOR inhibition alone in a preclinical xenograft model of sarcoma also resulted in a reduced Lac/Pyr ratio after injection of hyperpolarized Pyr, 24 h after the mTOR inhibitor administration (rapamycin) and prior to any anatomic detectable changes. 30 Still, there seem to be exceptions to high Lac/Pyr ratio in hyperpolarized 13 35,36 when a variety of preclinical tumours (glioma, mammary carcinoma, ovarian, lung carcinoma) respond to therapy. In References 33, 35 and 36 , the therapeutic agent used displayed anti-angiogenic or vascular disrupting activity, while in Reference 34 the drug used was an activator of PKM2. The rationale to explain this initially conflicting evidence is essentially put forward in References 33 and 36 : anti-angiogenic agents will produce hypoxic regions in the tumour, causing increased glycolytic flow to maintain required ATP (adenosine triphosphate) production by the tumour cell, overriding the decreased demand of carbon precursors for cell duplication by the growth-arrested tumour cells. Thus, in the case of therapeutic drugs working through this anti-angiogenic pathway, Lac/Pyr or k PL would not be an adequate biomarker of therapeutic success and other strategies may be required to asses it. In this respect, both Park et al and Iversen et al 35,36 point to a different ratio, hyperpolarized Lac/bicarbonate, as a better sensor of response to antiangiogenic agents (Supplementary Figure 4).
Take home messages. The Lac/Pyr ratio reflects the glycolytic state of tumours and could have potential for assessing tumour progression and also response to therapy, correlating with proliferation. The ratio should not be used during therapy with antiangiogenic agents as they produce hypoxic regions within tumours, and in this case the Lac/bicarbonate ratio would be more suitable.

| pH changes
The fact that actively proliferating tumours display a reversed pH gradient, i.e. acidic extracellular pH (pH e ) and slightly basic intracellular pH (pH i ), has been known for some time. 37 This is caused by active glycolysis (producing lactic acid) and oxidative metabolism (producing carbonic acid) in well perfused tumours on one hand, and active extrusion of protons and carbon dioxide/bicarbonate away from tumour cells, 38 on the other. Only recently has 31 P MRSI of GB patients been able to provide non-invasive images of this slightly basic pHi in proliferating GB and its changes upon transient response to therapy or in relapse. Wenger et al 39 used 31 P MRSI at 3 T to prove that good quality spectra can be obtained with this technique (Supplementary Figure 5) and pH i maps produced ( Figure 3). In these MRSI grids, it can be shown that untreated GB has a more basic pH i than contralateral brain (mean 7.110 for tumour and mean 7.017 for contralateral brain), but also that upon transiently successful antiangiogenic therapy (bevacizumab) pH i decreases (about 7.090 at best response time), and at relapse it slightly increases again. Moreover, the region where a new tumour mass eventually appeared on relapse had mean pH i = 7.065 before therapy and before MRI signs of tumour FIGURE 1 Hyperpolarized 13 C MRSI spectra acquired at different stages of progression in GEMs prone to develop pancreatic cancer. The predominant lesion histology pattern is shown on the left. The 13 C hyperpolarized spectra on the right were selected from an MRSI grid following the injection of hyperpolarized [1-13 C]Pyr. Both lac/Pyr and lac/alanine ratios increase with progression. PDA, pancreatic ductal adenocarcinoma; mPanin, murine pancreatic intraepithelial neoplasia. Modified from reference 23 (https://creativecommons.org/licenses/by/4.0/) presence, already significantly higher than contralateral (7.017); then pH i decreased, pH i = 7.044, at the time of best main mass response, and pH i increased, pH i = 7.085, at MRI detectable progression time. The higher pH i (7.065 versus 7.017) in the region where tumour relapsed later on could suggest that tumour cells were already proliferating there, but without major MRI detectable changes. In this respect pH i could predict the site of future progression in recurrent GB.

FIGURE 3
Brain and tumour pH i maps were calculated using the inorganic phosphate signal chemical shift of multivoxel 31 P MRS data and were colour coded (bottom row). The area of interest for voxel selection defined on standard MRI sequences is demonstrated with arrows (parieto-occipital selection, red, marks subsequent tumour position; frontal selection, green, marks control). A small number of spectra outside the ROI, close to the skull/brain interface, suffered from artefacts and were removed. Taken from reference 39

FIGURE 2
Representative axial T2W anatomic image, corresponding water ADC image and T2W image with an overlaid Pyr-to-lac metabolic flux (k PL ) image and corresponding hyperpolarized 13 C MRSI spectral array (SA) for a 52-year-old prostate cancer patient with extensive highgrade prostate cancer before therapy A, and 6 weeks after initiation of androgen ablation and chemotherapy B, before treatment, the region of prostate cancer can be clearly seen (red arrows) as a reduction in signal on the T2W and ADC images, and increased hyperpolarized lac and associated k PL flux on hyperpolarized 13 C MRSI. After initiation of androgen deprivation therapy there was a significant reduction in hyperpolarized lac and k PL fell to normal levels, with only a modest treatment effect on prostate volume and ADC. Taken from reference 27 13 C MRSI using hyperpolarized substances (bicarbonate) has also been used for pH imaging 40 although this approach does not seem to be in use yet in the field of cancer global MR. Concomitantly, other hyperpolarized probes, such as 13 C-labelled zymonic acid, have been shown to be able to produce fast and robust pH e imaging in preclinical cancer models in vivo. 41,42 Take home messages. Local pH changes, essentially measured with 31 P spectroscopy, can reflect cellular environment and could be useful for predicting the site of future progression in recurrent GB. In vivo measurement of pH is also possible by 13

| Oncometabolites
2-hydroxyglutarate (2HG) is produced due to a neomorphic mutation, 43,44 either in the moonlighting* IDH1 gene (NADP + (nicotinamide adenine dinucleotide 2′-phosphate, oxidized form) dependent, cytoplasmic) or in IDH2 (NADP + dependent, mitochondrial). 45 2HG is defined as an oncometabolite 44 : in IDH1 mutants, 2HG has been reported to be mostly derived from glutamine (80%), 13 and to trigger metabolic reprogramming. 46,47 2HG inhibits histone demethylation 48 through inhibition of α-KG-dependent dioxygenases, promotes nuclear dedifferentiation into a stem-cell-like state 49 and downregulates choline kinase and ethanolamine kinase, 50,51 among other effects. 2HG has been the subject of many studies 52 by the MRS community, since IDH1 or IDH2 mutant gliomas accumulate 2HG in the millimolar range of concentration. In 2012, several publications showed that the metabolite is detectable by MRS in vivo with 3 T clinical scanners, 53 and optimally when acquisition parameters are adjusted 54 spectroscopic imaging can be performed 55,56 to detect 2HG hot-spots and to monitor disease progression. 57 Most spectral analyses are still focused on quantification of the 2HG resonances, in particular those aimed to check whether the patient is a mutant for IDH or not in a non-invasive way 58,59 or to find an optimal combination of acquisition and quantification, 60 and only recently are metabolomics approaches starting to be used. In Reference 61, a cohort of 47 patients was analysed with a conventional PRESS T E 35 ms SV sequence at 3 T followed by LCModel quantification and metabolomics analysis, finding that the combination of 2HG and Glx (glutamate and glutamine) values could provide better discrimination between wild-type (wt) and IDH1 mutant patients than either of them alone. In Reference 62 a small cohort of 14 glioma patients was studied by a SEMI-LASER acquisition at 7 T coupled to a non-targeted metabolomics approach, again with the goal to distinguish wt from mutant patients. Despite the pervading observations that the IDH mutation not only causes an increase in 2HG but also a change in several other metabolites, either at the initial diagnosis 63,64 (Figure 4) or during assessment of response to treatment ( Figure 5), the change in the metabolomic pattern seems not to be regarded with much interest in the literature. This seems surprising because 2HG is a non-invasive diagnosis and prognosis biomarker for IDH mutation, 66 and it is druggable. 58,65 Perhaps small cohort sizes collected by each individual study and the difficulty of integrating these data with other studies using different acquisition and processing parameters may provide an explanation for this low uptake by clinical centres. In this respect, it has been shown in ex vivo 31  heterogeneity and low sample size for proper stratification into grades, and to poor resolution and instrumental problems, to explain these discrepancies. In this respect, the sensitivity of 2HG detection has been shown to be dependent on voxel size (good for tumours >8 mL and bad for tumours <3.4 mL) 58 on the one hand, and on the other phosphocholine (PCho) and phosphoethanolamine have been shown to be downregulated in a 2HG-dependent way. 50 Mutations in the gene for succinate dehydrogenase (SDH; another moonlighting protein) cause accumulation of succinate due to impairment of its oxidation to fumarate. Succinate is nowadays regarded as an oncometabolite: its accumulation inhibits prolyl-hydroxylases in the cytosol, 69 stabilizing and activating hypoxia-inducible factor (HIF)-1 alpha, and induces inflammation through this HIF-1 alpha signalling, among other roles in the cancer immune-cycle cross-talk 70 ( Figure 6). Succinate can be detected with MRS, and recently the SUCCES (Succinate Estimation by FIGURE 5 A, treatment-induced changes in mutant IDH patients. Boxplot of the fractional changes for the mean values averaged over the tumour ROI. The box size represents the first and third quartiles, and the horizontal red line indicates the median. Statistically significant changes are indicated by asterisks (P < 0.05). B, regional metabolic changes induced by treatment in a responding patient with the IDH mutation (patient 3, Pt 3). Adapted from reference 65 (https://creativecommons.org/licenses/by/4.0/) FIGURE 6 Succinate homeostasis is controlled by multiple factors in the cancerimmunity cycle cross-talk. Succinate elevations through upstream SDH mutations in cancer cells can promote downstream HIF-1α stabilization, which induces the key proinflammatory cytokine, IL-1β. IL-1β can polarize macrophages to the M1 sub-type, which gives them the capacity to damage tumour cells. HIF-1α driven by elevated succinate can promote glycolysis in cancer cells. Also, glutamine contributes to succinate levels through the GABA (gamma aminobutyric acid) shunt pathway, although glutamine has opposite roles in cancer cells and macrophages. Taken from reference 70 Spectroscopy) sequence (SV, PRESS, 144 ms, 3 T) has been applied to identify carriers (about 40%) of the SDHx gene mutation among patients affected by the rare tumours pheochromocytoma or paraganglioma. Interestingly, high choline (Cho) was also reported to be associated with the succinate peak detection in patients with such a mutation (Figure 7). 71 Although not yet at clinical level, mutations in fumarate hydratase are also a hallmark of some renal tumours, causing accumulation of fumarate. In addition to its role as an oncometabolite, it has also been shown that fumarate is an epigenetic modifier able to trigger the epithelial-to-mesenchymal transition. 72 It had previously been shown 73 that hyperpolarized fumarate can be used to monitor treatment response in murine EL-4 lymphomas through the assessment of malate production. Recently, co-polarized Pyr and fumarate have been used to monitor Lac export and necrosis in tumours. 74 Fumarate has also been found to be one of the discriminant metabolites when using an untargeted metabolomics approach to assess tamoxifen treatment resistance in breast cancer cell lines. 75 Take-home message. Not only oncometabolite accumulation but also the reprogrammed metabolome can be measured with MRS(I), extending the traditional single-metabolite biomarker concept.

| Cho metabolism and proliferation
Abnormal Cho metabolism has been identified in different cancer types. Choline-containing compounds are involved in the biosynthesis and degradation of phospholipids from cell membranes. Glunde et al proposed possible causes for this abnormal metabolism, 76 including changes in kinases (choline kinase-α, ethanolamine kinase-α), phospholipases (phosphatidylcholine-specific phospholipases C and D) and glycerophosphocholine (GPC) phosphodiesterases, as well as in Cho transporters. As a result of these metabolic changes, we usually expect to detect high content of PCho and tCho in actively growing tumours. In some cancer types, a low GPC content can also be observed, although it is also known that the actual tumour microenvironment can influence Cho metabolism. 77 FIGURE 7 SUCCES in a patient with an SDHB gene mutation. A, 1 H-MRS spectra of the right cervical paraganglioma (PGL) of patient 1. A succinate peak was detected in the cervical PGL with two different averages (1024 and 512) and two different VOI sizes (4.5 and 1.5 cm 3 ). B, applying the PRESS sequence to the abdominal tumour mass of the same patient permits detection of a succinate peak. C, in the healthy liver, the absence of a succinate peak demonstrates the specificity of the method. D, patient with a pheochromocytoma (PCC) without the SDHx mutation. Adapted from reference 71 A high tCho is considered the most useful MRS(I) parameter indicating high cellular proliferation. This tCho peak consists of water-soluble choline-containing compounds, such as PCho, GPC and free Cho. In addition to this, phosphatidylcholine in cellular membranes may contribute to the tCho peak in HR-MAS ex vivo spectra at physiological recording temperature 78,79 or have minor contributions to in vivo spectra in the case of small intracellular freely tumbling vesicles. 80 The most relevant applications in which the tCho signal can be of help in diagnosis, prognosis or therapy response are in prostate, breast and brain cancer, and some of these applications are highlighted in the following subsections.

| Prostate tumours
MRS allows the analysis of levels of choline-containing compounds, but also polyamines (PAs: spermine-Spm-and spermidine), total creatine (tCr) and citrate (Cit). 81 Cit is secreted in the luminal space by normal prostate epithelial cells and accumulates at high concentrations. Levels of Cit are reduced in prostate cancer and thus may act as an in vivo biomarker to discriminate cancer from normal uninvolved prostate tissue. Regarding PAs, about 50-90% in the prostate is Spm, and they are also secreted in the luminal space by ductal cells, as is Cit. Some or all of these have been jointly used for prognosis evaluation, showing good correlation with Gleason score. 82,83 Moreover, good correspondence between MRSI features and clinical outcome has been suggested, 84,85 also supported by ex vivo HRMAS studies. 18,86 Due to the multifocal nature of prostate cancer and its intragland heterogeneity ( Supplementary Figures 3 and 7), most recent cohort studies to detect, grade or even assess response to therapy use MRSI rather than SV MRS. 11,81,82,84,[87][88][89] The information contained in the spectral pattern may vary quantitative and qualitatively according to the optimized pulse sequence and echo time used for acquisition ( Figure 8).
Still, there is general consensus on the fact that metrics based on the changes in the MR-visible content of certain metabolites allow discrimination between normal tissue and prostate cancer. Thus, Cit and PAs (e.g. Spm) are high in normal tissue and choline-containing compounds (tCho) are high in cancer, while tCr does not change much. Then, different content-like ratios have been used in the past to visualize the proposed tissue type being sampled by MRSI in each volume element (Cho/Cit; [Cho + Cr (creatine)]/Cit; Cho/(Cit + PA)), although perhaps the most robust ratio, due to sometimes overlapping resonances of Cho, PA and Cr, may be (Cho + PA + Cr)/Cit. 82 These content/area/peak height ratios have been transformed using diverse thresholds in predicted tissue abnormality maps or hot-spots for prostate cancer detection (Figure 9).
With respect to the clinical interest of such metrics, a retrospective study by Zakian et al 84 on 262 patients at their institution concluded that prostate MRSI could help to predict outcome in radical prostatectomy patients. Additionally, Hansen et al 11 found that the Cit/Cre ratio was significantly different between erythroblast transformation-specific transcriptional factor-related gene (ERG)-high and ERG-low prostate cancer patients, which could bear in vivo biomarker interest for patient progression risk stratification. Moreover, a systematic literature review of the contribution of MR-based methodologies to the non-invasive detection of prostate cancer concluded that MRSI provides added value on sensitivity and/or specificity values when combined with classical T2W MRI explorations. 87 Still, the complexity of acquiring, post-processing and interpreting MRSI maps of patients is causing slow uptake of the use of such maps by clinical MR-based multi-parametric prostate cancer detection systems (see Reference 91 for further comments on this). Using higher recording fields (7 T) 89 or facilitating post-processing and display of multi-parametric datasets 92 may provide a handle to improve this uptake, competing for this with the hyperpolarized 13 C MRSI approach using 13 C-labelled Pyr discussed above.

| Breast tumours
Breast tumours present with high incidence (about 15% of all new cancer cases, according to NIH data), and early diagnosis and proper therapy response follow-up can improve patient outcome and prognosis. Breast cancer diagnosis is mostly performed using either mammography, ultrasound or MRI. However, other approaches can also provide insight into cellular processes, and their role in breast imaging is being currently evaluated. Regarding breast MRS, the spectrum of healthy breast is predominantly formed by signals from lipid protons with little contribution from water. In addition, resonances that correspond to choline-containing compounds (GPC, PCho and Cho, forming a joint peak of tCho at 3.2 ppm) are also seen. 93 The Cho signal is found to be elevated in malignant breast tumours, probably due to increased tumour proliferation, and can help in distinguishing malignant from benign lesions. 94 A meta-analysis performed by Baltzer and Dietzel 95 found high specificity and lower, variable sensitivity (88% and 73% respectively) of breast MRS for distinguishing lesions from uninvolved tissue. See Figure 10 for examples of breast MRSI and MRS applications.
Breast spectroscopic information can also be helpful for tumour grading, therapy monitoring and predicting response to treatment. Cho levels have been related to histologic grade and biological aggressiveness. 94,96 In treated breast cancer tumours, an early decrease of Cho levels is consistent with tumour response and is even more sensitive than anatomic or other functional criteria. 94 A recent review on in vivo and HRMAS of breast tissue metabolism by MRS provides further information on the subject. 16

| Brain tumours
A metabolite-ratio-derived strategy for gaining information from brain tumours by MRS(I) has been pioneered and used for some time already by Sarah Nelson's group at UCSF. These researchers used images of metabolic 'abnormality' at and around MRI-detected abnormal brain masses by calculating the tCho-to-N-acetylaspartate (NAA) ratio index, normalized to the same ratio in the normal contralateral brain of the same patient. This is known as CNI, where CNI values above a certain threshold would indicate presence of high-grade glioma (e.g. a z-score above 2; see References 97 and 98 for details of CNI and z-score). Other authors have used similar CNI-like indexes to simply mean direct tCho-to-NAA area ratios ( Figure 11). 97,98 These 'metabolic abnormality regions' have been shown to extend beyond structural MRI-defined abnormality boundaries. [99][100][101] This is helping to better define resection margins, 98 radiosurgery limits 102 and malignant transformation in recurrent glioma, 14 as well as to associate CNI metrics to survival after various therapeutic protocols, 103,104 and to identify tumour regions at risk for progression. 105 Take-home message. The Cho peak is useful to detect proliferative alterations in various tumour tissues (prostate, breast, brain), and has potential for decision making (response assessment and/or treatment planning). Representative spectra from normal appearing tissue acquired with PRESS A, and GOIA-sLaser E, and tumour tissue acquired with PRESS B, and GOIA-sLaser F, illustrate the differences in metabolite signal intensities between healthy and tumour tissue. The PRESS data were measured at T R /T E = 1020/145 ms and the GOIA-sLASER data was measured at T R /T E = 1020/88 ms. indicated are the main prostate metabolites: Cit, Cho, Spm and Cr. Healthy spectra A and E have the same scaling, and tumour spectra B and F have the same scaling, showing the increase in SNR of the GOIA-sLASER sequence compared with PRESS. Note that, because of their different distances from the endorectal coil used, scalings between tumour and healthy voxel were different. Taken from reference 81 (https://creativecommons.org/licenses/by-nc-nd/4.0/) FIGURE 10 A, MRSI grid for a breast evaluation. B, overlay of MRSI tCho map superimposed on an MRI image of a breast with a large tumour. Adapted and modified from reference 93. Voxel 1 inside the tumour presented a high Cho level, while voxel 2 (outside the tumour) had no Cho signal detectable. C, D, breast MRS from a 67-year-old woman with a suspicious right breast lesion (solid arrow in C) that shows high vascularization in maximum relative enhancement parametric map and maximum intensity projection images derived from DCE-MR imaging (C) and a marked Cho peak in SV MRS (D), both findings consistent with malignant origin. Notice the positioning of the VOI in C (middle and right images). Papillary breast carcinoma was confirmed on core biopsy. Taken from reference 94

FIGURE 9
A, T2W image of the prostate of a 57-year-old man with biopsy-proven prostate cancer in the left peripheral zone. The MRSI grid is plotted over the prostate, as are the outer volume saturation slabs. B, an example of a tumour probability map based on MRSI data that could be presented to a clinician. In this case, the probability map is based on the Cho + Cr to Cit ratio (colour coded bar on right). Taken from reference 90

| MRS CANCER PATTERN
The observation that tumours have a distinctive spectral pattern can nowadays be considered as 'old': as early as 1992 William Negendank reviewed the literature on 31 P, 1 H, 13 C, 23 Na and 19 F MRS in vivo in human cancers, 106 and among other conclusions stated the need for improvement of the diagnostic specificity and statistical analysis of multiple features, as well as trials in large populations in well-defined clinical settings.
As can be noticed from previous sections, even today, pattern recognition is not much in use for the diagnostic or prognostic evaluation of cancer, with most studies focusing on single metabolites (Cho, 2HG) or particular metabolite ratios ((Cho + PA + Cr)/Cit, 107 Glx/NAA, 101,108 ) or indices (CNI), the paradigm being 2HG, a single-metabolite MRS-visible biomarker that allows clinicians to perform patient stratification and follow-up of response to therapy in carriers of the IDH mutation. However, as elaborated in Section 4.3, even in that situation other MRS-visible changes have been observed, which taken individually are non-significant. Perhaps their joint use in pattern recognition strategies could improve discrimination of the condition of interest. Furthermore, it remains to be assessed whether acquiring larger, multicentre patient cohorts, using compatible acquisition and processing protocols, could help achieve higher sensitivity and specificity (in terms of diagnostic tests) values for detection of the IDH mutation despite the low sensitivity of the MRS technique (about 1mM for 2HG and voxel sizes below 8 mL not optimal).
Recent pattern recognition studies have been dealing with two aspects: first, diagnosis, performed mainly with SV, and second, delimitation of the abnormal area, performed with MRSI. One of the main reasons for this thematic restriction is that SV MRS can address a clinically relevant question, namely the added value of MRS to MRI for brain tumour type and grade assessment. [109][110][111][112][113] Whatever the aspect addressed, all recent

FIGURE 11
Case of a GB patient investigated at 3 T for maximizing resection of the metabolically abnormal region. The process of converting and merging 3D 1 H-MRS images (T E = 135 ms) with structural images is summarized. A, 3D 1 H-MRS revealed different metabolic margins generated for CNIs (defined as total tCho/NAA area ratios as in reference 97) of 0.5, 1.0 and 1.5, which were transferred into iPlan 3.0 cranial. The pseudocolour maps were then generated. B, postoperative MR image. C, functional MRI and diffusion tensor imaging located the hand motor areas and pyramidal tracts (PTs). D, E, the contour line for a CNI of 1.0 was adjacent to the PTs. red, 1.5; yellow, 1.0; blue, 0.5. To achieve a maximal safe resection, the extent of resection was carried as close as possible to the metabolic margin generated for a CNI of 1.0. F, sagittal view revealing the relationship of the PTs and metabolic margins. A safe margin of 0.8 mm was set to avoid further neurological deficits, even though this margin was near the PTs. a subtotal resection was achieved in this patient. Taken from reference 98 studies use a 'targeted metabolomics' approach, i.e. quantification of a number of metabolites, which is set a priori, except that already mentioned by Emir and colleagues, 62 where they use an untargeted approach (although in the restricted specific spectral region [1.6, 3.1] ppm).
With respect to diagnosis, the need for multicentre studies seems well established since the INTERPRET experience on preoperative brain tumour typing/grading at 1.5 T. 110,114 The fact that the curated INTERPRET dataset has been made accessible to research groups allows them to develop classifiers comparing single-centre cohorts with that dataset, for example to differentiate abscesses from tumours at 3 T. 115 Diagnosis of paediatric brain tumours has also been the subject of two recent works at 1.5 T on a large cohort of 90 patients 109 and at 3 T in a multicentre study on 41 patients, 113 to distinguish among the three main paediatric brain tumours (medulloblastoma versus pilocytic astrocytoma versus ependymoma). Despite this, single-centre studies are still reported. 108 Recent studies are using pattern recognition of MRS as one of the modalities in the multiparametric assessment of tumour regions, again using a targeted metabolomics approach where pre-quantified metabolites or ratios are the input features for the classifiers. Two quantified metabolite ratios ((NAA + NAAG)/(Cr + PCho), (GPC + PCh)/(Cr + PCho)) were used in Reference 116, as well as other image features (K trans , K ep etc), to train a classifier to distinguish between enhancing/non enhancing tumour/non-tumour. In Reference 117 the problem addressed was the distinction between brain tumour and necrotic areas, and in Reference 118 the focus was to segment the prostate into tumour and normal tissue, providing the user with a 'nosological image'. Transformation of the metabolomic pattern information into nosological images of tissue/tumour type (Supplementary Figure 8) is an approach that was proposed 18 years ago by the group in Grenoble. 119 A limitation of this approach will always be the validation of these metabolic segmentations in humans, in particular for the brain. In the case of prostate cancer, the availability of some prostatectomy specimens is an advantage and can partially help to assess the validity of the models, although a frequent problem is that the matched specimen is not available in all cases, for obvious reasons. 118,120 In the brain the situation is worst, as the ground truth used for validating the metabolic segmentation must be the MRI set of images, pre-segmented by a radiologist, 117,121,122 or by software. 116 This means that normally the metabolic information is 'trained' to fit the anatomy, despite the fact that, in gliomas at least, it has been shown that the metabolic abnormality extends past the T 2 abnormal region by a couple of centimetres. 123 In this sense, the use of a semi-supervised approach (telling the system which class each spectrum belongs to, but not forcing it to classify it as one or other class) has provided robust results for prediction of response to therapy in preclinical models, 124 as will be discussed in subsequent sections. During the period reviewed, methodological developments have been a primary focus: multiparametric fusion, 92,121 sometimes scaling MRSI to MRI resolution by linear interpolation, 116 or studying the efficacy of different implementations of non-supervised pattern recognition techniques. 117,118,121,122,124,125 The problem of imbalanced classes has also been tackled with the use of AdaBoost. 109 Supervised pattern recognition techniques such as support vector machines, 109,115,116 linear discriminant analysis 62,126 or several different algorithms, 109 and classifier fusion, 127 are also used in the latest literature, mainly for diagnostic questions. These methodologies are, generally too sophisticated to be put in practice by purely clinical groups, and when methodologies are tested the norm is availability of very small datasets, although there are exceptions such as Reference 113 or 109 . On the other hand, groups from clinical centres tend to use simpler approaches on larger series of patients from one clinical setting. In Reference 126, Crain et al measured five peak heights with ImageJ directly over the image provided by the scanner software, to apply a linear discriminant classifier on SPSS, testing different ratio combinations, to differentiate between post-radiation effects and recurrence in gliomas, however with low success.
In Reference 128, principal component analysis, a feature extraction method, is used to successfully visually distinguish one genetically different type of medulloblastoma of the four already accepted existing subtypes. 129 When the multiparametric approach is tackled, methods such as logistic regression find a place, for example to predict whether the patient has prostate cancer or not, 130 then restricting the MRS to the use of a ratio value (Cit/(Cho + Cr)) as one of the parameters of the predictive equation.
Take-home message. The potential of the whole spectral pattern is still underused, in particular in untargeted metabolomics approaches.

| Tumours are heterogeneous tissues
MRS detects signal proportionally to the concentration of nuclei of the investigated type/molecule present (weighted by relaxation time and multiplicity effects on possible phase modulation of the signal). Simplifying a lot, we could state that MRS(I) signal intensity detected is directly proportional to the amount of the substance investigated contained in the sampled volume. In solid tumours in vivo, cancer cells constitute only a fraction of the tissue volume sampled by MR. Accordingly, between 30 and 50% of the tumour tissue (in tumours such as gliomas) can be noncancer cells 6 (mostly endothelial cells, fibroblasts and infiltrating immune cells 131 ). Furthermore, extracellular volume and, broadly speaking, 'interstitium' 132 may also account for a relevant large amount of tumour tissue volume (e.g., more than 40% of tumour tissue in preclinical GL261 GB transiently responding to the TMZ therapy has been reported to be 'acellular' 133 ). In addition to this, dynamic changes in the 13 C-labelled metabolite pattern of tumours with time (see Section 4.1 on carbon consumption by tumours) or in a steady state metabolomic pattern (see later in this section) may originate or have relevant contributions from non-cancerous cells. In this respect Guglielmetti et al 134 have described an increased Lac/Pyr ratio after injection of hyperpolarized Pyr in a mouse model of multiple sclerosis with respect to control brain. The relevant point here is that Guglielmetti et al demonstrated that this increase in the Lac/Pyr ratio was due to the infiltration of the investigated brain region by the M1 subtype of proinflammatory macrophages, which are highly glycolytic cells. 135 In this instance, those macrophages were shown to upregulate PDK1 (pyruvate dehydrogenase kinase 1), resulting in inhibited PDH and reduced entrance of Pyr into the TCA cycle, possibly facilitating Pyr flow through LDH, which did not seem to be upregulated in their investigated tissue. Infiltration of tumours by innate and adaptive immune system cells is a well-known fact. 9,[136][137][138] Accordingly, the possible contribution of the metabolomic pattern of immune system cells and their changes with time/therapy to the MRS(I)-based investigation of tumours may need to be considered (see Section 4.4).
On top of the contribution of non-tumour cells to the MR-detected pattern, one additional factor to take into account is the widely recognized clonal heterogeneity of tumours, [139][140][141][142] for example, in brain 143 (Supplementary Figure 9), prostate 141 (Supplementary Figure 7) and breast 144 tumours. This will increase the variability of their metabolic pattern if large regions of interest (ROIs) are evaluated, as in SV MRS, while more relevant information with respect to this heterogeneity may be potentially sampled by multivolume MRSI approaches.

| Tumour spectral pattern is also heterogeneous
Radiological and clinical guidelines are the current standards to evaluate therapy response in brain tumours, particularly through acquiring MRI and dynamic contrast-enhanced MRI data and the application of the RANO criteria. This uses 2D single slice data. However, MRSI might be a useful supplementary tool for this purpose allowing to monitor and follow-up brain tumour molecular properties and heterogeneity 14,103,104,107,133,145,146 for the whole tissue (3D). Most of the time, authors focus their analysis on a few metabolites or metabolite ratios (the CNI in brain tumours, the (Cho + PA + Cr)/Cit ratio in prostate cancers), neglecting the rest of the spectral pattern, as has been discussed in Section 4.2. This has been evolving, and recently classic 'single ratio' groups have been considering additional metabolite-derived metrics: edited Lac, Cho/NAA and lipid/NAA indexes. 103 According to the molecular/cellular heterogeneity described for tumours in the previous section, the spectral pattern can also be heterogeneous through the tumour volume, reflecting differences in molecular signatures. Recent studies in human GB tend to focus on metabolite ratios in the peritumoral area. 101 This is indeed true not only for a snapshot of a tumour at a given time point, but also informative in longitudinal studies, for example, to follow treatment response. This has been extensively studied in a preclinical model of GB, using pattern recognition approaches with histopathological validation. 124,133,145 The spectroscopic differences between treated and untreated GL261 GB tumours were used to train a mathematical classifier that was further tested in new animals (treated and untreated), and results were validated against histopathology. Transient response producing significant survival improvement induced by TMZ was found to correlate with decreased proliferation measured by Ki67 immunostaining in histopathological slides. Prototypical spectra (source) signals were used with a semi-supervised methodology within the training set, which represented the three tissue types studied: normal brain parenchyma, GL261 tumours treated with TMZ mostly showing response or stable disease, and untreated (control) GL261 tumours (i.e. actively proliferating). The changes observed in the spectral pattern of treated tumours were mainly related to mobile lipid (ML; 0.9 and 1.3 ppm) and polyunsaturated fatty acid in ML (2.8 ppm) resonances. Still, other contributions (e.g. Lac (1.3 and 4.1 ppm), glutamine, glutamate, alanine (3.8 ppm) and myo-inositol/glycine (3.5 ppm)) could be helping discrimination between tumour sources, although it is worth noting that the isolated analysis of chosen metabolites or their ratios did not prove useful for discriminating response, the analysis of the whole spectral pattern being the most informative and robust method of discrimination. 124 The correlations between the sources and the MR spectrum of each voxel in the MRSI grids were used for calculating nosologic images. Supplementary Figure 10 illustrates the three groups of nosologic patterns to be recognized after training by the source methodology in the training set for the region recognized as tumour: tumours with a clear control proliferating pattern, tumours clearly responding to treatment and treated or control tumours with a partial or heterogeneous pattern of response. After the histopathological validation, the methodology was also applied to mice studied in longitudinal studies until endpoint, showing a good correlation between the nosological imaging and tumour behaviour, even predicting the outcome before evident changes were seen in their volumes. 124 Furthermore, a 3D-like multislice MRSI acquisition was acquired in a new cohort of animals 133 in order to assess heterogeneity along the whole tumour volume instead of a single 2D slice, and the same classifier was applied. The presence of a responding pattern was confirmed to correlate with histopathological features such as decrease of the Ki67 + determined proliferation rate, while lymphocyte-like cells were also observed in the responding tumours. This led Arias-Ramos et al 133 to propose a relationship between the oscillating pattern of response percentage measured in nosological images seen in GL261 GB tumours treated with TMZ and longitudinally followed until endpoint ( Figure 12) and the 'immune cycle' (Supplementary Figure 11). Immune cycle-associated changes being detected by MRSI has been given additional support by other work on the GL261 GB tumour model under other therapeutic strategies (cyclophosphamide, metronomic therapy administration, 147,148 and also by the duration of the immune cycle in mouse brain described in Reference 149 -6dwhich is in agreement with the oscillation period recorded for the response pattern by Arias-Ramos et al. 133 Take-home message. Tumours are intrinsically heterogeneous, their response is heterogeneous too, and this heterogeneity can be sampled by whole spectral pattern analysis of MRS(I). There is a non-negligible contribution from non-tumoral cells such as tumourassociated macrophages.

| What are the limitations of MRS(I)?
Some of the main disadvantages or limitations of MRS (mainly MRSI) are longer acquisition times needed for data collection, compared with classic structural MRI, lack of adequate signal-to-noise ratio (SNR), sometimes improper water/lipid signal suppression and limitations in spatial coverage.
Moreover, it is worth noting that an MRS(I) exploration is equivalent to a snapshot of a given system, and most of the in vivo processes, mainly the metabolic ones, are dynamic. This means that some processes can be either missed or improperly sampled, because in long explorations the final output will be an average of the processes taking place during the overall acquisition time.

| The range of metabolites that can be studied in tumours is limited
MRS (in either its single or multi-voxel modality, respectively SV-MRS or MRSI) has the ability to detect changes in the relative content of metabolites, even allowing the calculation of absolute concentrations. Still, only small and/or mobile molecules, in the millimolar concentration range, are directly observed in an in vivo MR spectrum. This is extremely low as compared with tissue water (about 55 M). Therefore, proper water suppression (and also lipid suppression in the case of tumour surrounded by adipose tissue, e.g. subcutaneous fat, scalp lipids) is required for confident metabolite observation and quantitation. The biological significance of the major compounds found in tissues, for example in brain spectra and tumours, is reviewed in References 10,17 and. [150][151][152] Other minority substances are hardly detected due to either extremely small or overlapping signals, although spectral editing may allow their observation and quantification. 153

| The resolution achieved is usually lower than MRI resolution
As the concentrations of the MRS-detected metabolites are much lower than that of water, larger voxel sizes are needed to detect enough signal for MRS(I) with respect to MRI. This prevents MRSI from providing comparable resolution to MRI (although further post-processing, i.e. interpolation, can partially address this limitation). The lower spatial resolution of MRSI compared with anatomical imaging makes consideration of the contribution of the spatial point spread function (signal bleeding into the surrounding voxels, producing a blurring effect 154 ), proportionally more relevant than for MRI.

FIGURE 12
Graphical representation of the tumour volume evolution and the percentage of responding 'green' pixels obtained after source analysis of MRSI data acquired in the multislice set for case C974 (intermediate response case). The green shaded columns indicate TMZ administration periods. For chosen time points, the evolution of the nosological images obtained with the semi-supervised source extraction system is shown in four rows of colour coded grids, superimposed on the T2W MRI for each slice. Green brackets indicate TMZ administration periods. A cyclical pattern of response was observed, marked by the red arrows in the graphical representation. Taken from reference 133 (https:// creativecommons.org/licenses/by/4.0/)

| There is important sensitivity to magnetic field variation
Another important point is the MRSI sensitivity to magnetic field inhomogeneity across the volume of interest (VOI), which may require avoiding anatomical regions prone to present strong magnetic field inhomogeneity such as tissue-bone or tissue-air interfaces in order to avoid line broadening and spectral artifacts. This is one of the reasons why MRSI has been most successful in brain tumour acquisitions: more favourable field homogeneity conditions and sensitive signal detection due to the use of surface coil arrays close to the volume being sampled. Accordingly, MRSI has been extensively applied for diagnosis and treatment follow-up of brain tumours. Regarding prostate tumours, MRSI acquisition is usually performed with an endorectal coil for signal reception, to achieve good SNRs. However, it also implies longer time needed for preparation, as well as patient discomfort. Tayari et al 91 describe MRSI acquisition approaches without an endorectal coil obtaining good quality MRSI data, which would probably contribute to increasing its incorporation in clinical pipelines for prostate cancer diagnosis and follow-up in the future. Finally, regarding the breast, the relevant magnetic field inhomogeneities observed, as well as the small water peak in adipose tissue, make shimming quite challenging in these anatomical locations.

| Conclusions
Regarding cancer metabolomics analysis, MRSI should be the standard choice for in vivo (pre) clinical MRSI due to its ability to provide spatially resolved information, and consequently details of relevance with respect to tumour tissue heterogeneity, which are not provided by SV MRS.
However, this is restricted to applications in which limited volume coverage and spatial resolution with respect to MRI/MRS are acceptable, and also where the achievable homogeneity of the magnetic field is not an issue. SV MRS will remain an important tool for evaluating metabolite content in regions where magnetic field inhomogeneity degrades spectral quality in MRSI or when single biomarker detection may be of interest (2HG detection in LGG brain tumours). Although most clinical studies focus into 1 H-MRS(I), there are different studies using 13 C and 31 P, which could be also of help either in grading or therapy response assessment by monitoring carbon consumption and local pH changes. The study of single metabolite biomarkers (e.g. Cho, 2HG) has also shown its potential in the diagnostic and prognostic assessment of tumours. However, the study of single metabolites or metabolite ratios offers a restricted picture of the metabolic reprogramming that occurs in tumours, which is being acknowledged in recent literature and could be better tackled by a pattern recognition approach. Recent pattern recognition studies are primarily focused on methodological developments rather than multicentre validation of preclinical or clinical study findings based on small one-centre cohorts. Additional steps toward simplification and standardization of processing files from different vendors would probably be needed in this sense, as well as familiarization of clinicians and researchers with pattern recognition tools. Novel dynamic metabolic approaches such as using hyperpolarized substrates will probably have considerable impact on future applications of MRSI and technology development, provided that large animal/patient cohorts are explored at more than one site with comparable data acquisition protocols.