Soft tissue sarcomas (STSs) are rare, highly malignant mesenchymal tumours that account for roughly 1% of all adult cancers and about 15% of paediatric solid tumours. In Europe, the incidence stands at 4.7 cases per 100,000 per year; in China it is reported at 2.9 per 100,000. Despite this low overall frequency, STS carries a 5-year survival rate of only 50-70%, and a substantial fraction of patients still die from distant metastases, most commonly to the lungs. The extremities account for approximately 60% of primary tumour sites, with the trunk (25%) and retroperitoneum (15%) following.
Histological and molecular complexity: STS encompasses over 70 histologically distinct subtypes, each with unique morphological features and biological behaviours. At the genomic level, STSs exhibit substantial complexity: chromosomal rearrangements, copy number alterations, and somatic mutations in oncogenes and tumour suppressor genes. Specific fusion oncogenes are hallmarks of particular subtypes, including EWSR1-FLI1 in Ewing sarcoma and SS18-SSX in synovial sarcoma. Key signalling pathways that are dysregulated across subtypes include PI3K/AKT/mTOR, MAPK, and Wnt/beta-catenin, as well as receptor tyrosine kinases such as PDGFR, VEGFR, and c-Met.
Limitations of tissue biopsy: Surgery remains the primary curative modality, while chemotherapy efficacy remains uncertain outside the metastatic setting. Standard tissue biopsy captures only a single spatial and temporal snapshot of tumour biology, making tumour heterogeneity a leading cause of therapeutic failure. Liquid biopsy offers a complementary approach, sampling tumour-derived or tumour-associated components circulating in blood, including circulating tumour cells (CTCs), cell-free DNA, cancer-associated proteins, and dysregulated metabolites, without the discomfort or access challenges of repeat tissue sampling.
The metabolomics opportunity: Metabolomics, particularly in untargeted form, captures a broad spectrum of low-molecular-weight compounds that reflect real-time metabolic state across all active biochemical pathways. Because cancer cells fundamentally rewire metabolism to sustain rapid proliferation, metabolite profiling of serum can reveal pathway-level alterations that tissue biopsy cannot dynamically monitor. This study applies untargeted NMR-based serum metabolomics to a cohort of 75 STS patients to discover circulating biomarkers relevant to diagnosis, prognosis, and potential therapeutic targeting.
The study enrolled 75 STS patients from the Oncology and Reconstructive Orthopedics SOD at AOU Careggi in Italy, all diagnosed by an expert pathologist after providing written informed consent. Both adult and paediatric patients were eligible. The cohort consisted of 56% male and 44% female subjects, with a mean age of 64 years (range 22-92). The distribution of histotypes mirrored the real-world STS landscape: pleomorphic sarcoma (29%), liposarcoma (25%), myxofibrosarcoma (16%), leiomyosarcoma (9%), synovial sarcoma (7%), rhabdomyosarcoma (6%), dermatofibrosarcoma protuberans (4%), and fibrosarcoma (4%). The primary tumour site was the thigh in 65% of cases. Tumour grade was high in 72%, intermediate in 8%, and low in 20% of patients. An age- and sex-matched healthy control group of 85 subjects was included for comparison.
Sample preparation and NMR acquisition: Serum was prepared from peripheral blood, centrifuged, and stored at -80 degrees Celsius until analysis. For NMR measurement, 300 microliters of serum were combined with 260 microliters of TRIS-d11 buffer (150 mM) and 40 microliters of calcium formate (45 mM) as an internal quantification standard. All spectra were acquired on a Bruker Avance III 600 spectrometer operating at 14.1 Tesla. The PROJECT pulse sequence was employed with an echo time of 0.3 milliseconds and 128 loops to achieve an optimised T2 filter, which suppresses signals from large macromolecules and retains only small-molecule metabolite signals. Each spectrum was acquired with 32 scans over a 6 kHz spectral width, digitised over 32k data points, and zero-filled to 256k points. Solvent suppression was achieved via presaturation during the repetition delay of 4 seconds.
Data processing and metabolite identification: Raw NMR spectra were processed using Chenomx 10 software, which uses a library-based deconvolution approach to identify and quantify individual metabolites. Formate served as the internal concentration standard. The Chenomx library enabled identification and quantification of 63 metabolites spanning amino acids, amines, short-chain fatty acids, ketones, alcohols, nucleobases, nucleosides, and carbohydrates. Four metabolites were excluded prior to analysis: mannitol (a drug constituent), ethanol and isopropanol (equipment pollutants), and urea (unreliably quantified by NMR at the concentrations present in serum).
Statistical and machine learning approach: Multivariate analysis began with principal component analysis (PCA) and partial least squares discriminant analysis (PLS-DA) using MetaboAnalyst 6.0. Both linear dimensionality-reduction methods failed to cleanly separate patients from controls, reflecting the heterogeneity of the STS group. The analysis then advanced to a random forest (RF) classifier, built from 2,000 decision trees using bagging and random feature selection, with an out-of-bag (OOB) error rate serving as an internal cross-validation metric alongside five-fold cross-validation. The RF model also computed permutation-based variable importance measures (VIM) to rank each metabolite's contribution to class discrimination, enabling unbiased biomarker discovery without requiring predefined hypotheses.
When linear multivariate methods (PCA and PLS-DA) were applied to the NMR metabolite dataset, the score plots showed significant overlap between the STS patient group and healthy controls. This failure is attributable to the profound intra-group variation within the STS cohort, driven by its histological heterogeneity across eight distinct histotypes. In contrast to the minimal variation seen among healthy controls, patients showed broad metabolic scatter, consistent with the diverse molecular biology of different STS subtypes. The inability of linear methods to capture nonlinear relationships between metabolite features and class labels justified the move to a machine learning approach.
Random forest performance: The RF model with 2,000 decision trees achieved an OOB error rate that stabilised at 4% as tree count increased, indicating strong and stable classification performance. The five-fold cross-validation error converged to a similar value, confirming that the model was not overfitting the training data. Variable importance was ranked using the mean decrease in classification accuracy upon permutation of each feature, a measure that directly quantifies each metabolite's contribution to correct patient-vs-control classification. This approach identified eleven metabolites with statistically significant deregulation (p less than 0.01 by t-test), which were confirmed as the key contributors to the RF model's discriminant performance.
The eleven-metabolite panel: The significantly deregulated metabolites identified were acetone, hypoxanthine, inosine, acetate, histidine, choline, 3-hydroxybutyrate, citrate, acetoacetate, lactate, and pyruvate. These metabolites span several distinct biochemical pathways, including ketone body metabolism, fatty acid synthesis, purine salvage, one-carbon metabolism, and aerobic glycolysis. This breadth reflects the metabolic reprogramming that cancer cells undergo across multiple fronts simultaneously. The directionality of dysregulation varied: ketone bodies (acetone, 3-hydroxybutyrate, acetoacetate), acetate, citrate, hypoxanthine, inosine, and lactate were all significantly upregulated in STS patients, while choline, histidine, and pyruvate were significantly downregulated.
Attempts to correlate individual metabolite levels with specific clinical and demographic variables, including histotype, grade, age, and gender, did not yield statistically significant associations. This absence of correlation likely reflects the intrinsic heterogeneity of STS, which blurs metabolic signatures that might otherwise be histotype-specific. The authors note that future subtype-stratified analyses are essential to refine the biomarker panel for precision medicine applications. At the level of the overall STS dataset, however, the eleven-metabolite panel provides a reproducible and statistically robust signal distinguishing tumour-bearing patients from healthy controls.
The Warburg effect, also called aerobic glycolysis, describes a well-established phenomenon in which cancer cells preferentially convert glucose to lactate even in the presence of abundant oxygen, rather than pursuing the more energy-efficient route of oxidative phosphorylation in the mitochondria. This shift supports rapid cell proliferation by generating not only ATP but also the metabolic intermediates needed for biosynthesis. In the serum of STS patients, the authors observed clear evidence of Warburg activation: lactate was significantly upregulated (p less than 0.001) while pyruvate was significantly downregulated (p less than 0.01). The resulting lactate-to-pyruvate ratio exceeded that of healthy controls by three-fold on average, a shift so pronounced it could be visually estimated directly from the raw NMR signals.
Ketone body accumulation and metabolic inflexibility: All three major ketone bodies, acetoacetate, 3-hydroxybutyrate, and acetone, were significantly elevated in STS patient serum (acetone and 3-hydroxybutyrate at p less than 0.00001; acetoacetate at p less than 0.001). Healthy cells efficiently catabolise ketone bodies to acetyl-CoA via the enzyme succinyl-CoA:3-ketoacid CoA transferase (SCOT), which feeds the TCA cycle to generate ATP. Tumour cells, however, frequently exhibit deficient ketolytic enzyme activity (particularly SCOT) and mitochondrial dysfunction, a state described as acquired metabolic inflexibility. As a result, ketone bodies accumulate in the circulation rather than being utilised for energy by the tumour. Acetone is particularly notable because, unlike acetoacetate and 3-hydroxybutyrate, it cannot be recycled to acetyl-CoA and is instead exhaled or excreted. It was the most upregulated ketone body in the STS panel.
Acetate, citrate, and fatty acid synthesis: Acetate (p less than 0.00001) and citrate (p less than 0.0001) were both significantly elevated in STS patient serum. Both molecules feed into fatty acid synthesis, a pathway that tumours upregulate to meet the heightened demand for membrane lipids and other biomass components during rapid proliferation. Under hypoxic or nutrient-limited conditions, cancer cells increasingly rely on acetate as an alternative carbon source for acetyl-CoA synthesis via acetyl-CoA synthetase 2 (ACSS2), bypassing the TCA-derived citrate pathway. Elevated extracellular citrate may also be taken up directly by tumour cells to sustain cytosolic fatty acid synthesis without relying on de novo citrate generation from the Krebs cycle. The simultaneous elevation of both acetate and citrate in the STS circulation is consistent with this dual-substrate model of lipogenic support for tumour growth.
Inosine and hypoxanthine were among the most significantly elevated metabolites in the STS serum panel, both at p less than 0.00001. In healthy controls, these purines were below the NMR detection limit (approximately 1 micromolar). In STS patients, inosine and hypoxanthine were detectable up to concentrations in the tens of micromolar range (with representative values of 29 micromolar for inosine and 25 micromolar for hypoxanthine shown in the NMR spectra). Crucially, xanthine, the further catabolic product of hypoxanthine oxidation, was not elevated, indicating that tumour cells are intercepting and recycling these purines rather than degrading them further.
Purine salvage pathway mechanics: The purine salvage pathway recycles purine bases to synthesise nucleotides without the energetically expensive de novo synthesis route. Under conditions of hypoxia or cell stress such as apoptosis, cells release extracellular ATP and related adenosine nucleotides, which are progressively dephosphorylated by ectonucleotidases to produce adenosine, inosine, and ultimately hypoxanthine. Tumour cells in hypoxic microenvironments generate and release these intermediates in large quantities. Inosine can be transported into cells via equilibrative nucleoside transporters (ENT1 and ENT2), then converted back to hypoxanthine by purine nucleoside phosphorylase (PNP). Hypoxanthine can be salvaged to inosine monophosphate (IMP) by hypoxanthine-guanine phosphoribosyltransferase (HGPRT), with IMP then converted to AMP and GMP for DNA and RNA synthesis. By capturing circulating inosine and hypoxanthine from the bloodstream and re-routing them through HGPRT, rapidly dividing tumour cells avoid the cost of de novo purine synthesis while maintaining high nucleotide availability.
Histidine depletion: Histidine was significantly downregulated in STS patient serum (p less than 0.00001), with concentrations averaging approximately 35 micromolar compared to higher levels in healthy controls. This aligns with a value of approximately 30 micromolar previously associated with lung cancer progression. Histidine serves as a donor of one-carbon units to tetrahydrofolate through the enzyme formiminotransferase-cyclodeaminase, contributing to one-carbon metabolism and the synthesis of purines and pyrimidines. The hypothesis is that tumours with overactive nucleic acid metabolism excessively absorb histidine from surrounding tissue and bloodstream, depleting circulating levels. This downregulation of histidine was previously suggested in a smaller PCA analysis of STS patients by Miolo et al., and this study's machine learning approach provides a more rigorous validation of it as a circulating STS biomarker.
The simultaneous upregulation of circulating purines and downregulation of histidine paints a coherent picture of nucleotide metabolism in STS: tumour cells drain histidine to support de novo pyrimidine synthesis while also scavenging circulating inosine and hypoxanthine to supplement purine nucleotide pools via the salvage route, both strategies driven by the high proliferative demand for nucleic acid precursors.
Among all eleven dysregulated metabolites, choline emerged as the most clinically significant finding in this study. Choline was significantly downregulated in STS patient serum (p less than 0.00001), and it was the only metabolite to survive correction for multiple comparisons in the survival analysis, achieving a false discovery rate q-value of 0.0092, compared to acetate (q-value 0.1236), inosine (q-value 0.2683), and hypoxanthine (q-value 0.2770).
Why choline decreases in sarcoma: Choline is a key precursor of phosphatidylcholine (PtdCho), the dominant phospholipid in cell membranes. Cancer cells undergoing rapid growth require substantially increased phosphatidylcholine production to build new membrane bilayers. The altered signalling pathways characteristic of cancer cells, including upregulated phospholipase D and choline kinase, further amplify choline uptake and utilisation. The observed decrease in circulating choline reflects this heightened consumption by the developing tumour mass, with serum levels dropping as the tumour captures more of the available choline supply from the bloodstream. This mechanism is well established in other solid tumours and in MRS-based imaging studies, but this study provides the first untargeted metabolomics evidence in a large STS cohort.
Survival analysis: Overall survival data were available for 60 of the 75 patients, with 18 deaths recorded at the checkpoint analysis. The median overall survival was 29 months (95% CI: 24-33 months). When stratified by grade, low-grade cases fared better with a median of 35 months (95% CI: 24-39 months), while gender and age at diagnosis did not contribute significant differences. Univariate Cox proportional hazards regression linked choline levels to overall survival with a raw p-value of 0.0008 and a Benjamini-Krieger-Yekutieli corrected q-value of 0.0092, making it the only metabolite passing FDR correction. No clinical features were found to be significantly associated with choline levels themselves, which may indicate that choline depletion represents a pan-histotype metabolic response to STS rather than a feature of specific subtypes or grades.
Therapeutic implications: The choline metabolism pathway has been recognised as a potential therapeutic target in oncology. Choline transporter-like proteins (CTLs/SLC44 family) and choline kinase alpha represent potential drug targets. Interference with choline supply or its conversion to phosphatidylcholine could impair membrane synthesis in rapidly dividing tumour cells. Additionally, choline-based imaging using PET tracers such as 11C-choline and 18F-fluorocholine is already in clinical use for prostate cancer; adapting these approaches to STS monitoring could provide a non-invasive imaging biomarker that reflects the same metabolic axis identified in this serum study.
Histological heterogeneity and lack of subtype stratification: The most fundamental limitation of this study is the impossibility of performing histotype-specific subgroup analysis within a cohort of 75 patients covering eight distinct STS subtypes. The authors explicitly acknowledge that the high degree of STS heterogeneity prevented meaningful stratification of patients by histotype. The metabolic findings therefore reflect a pan-STS signature rather than subtype-specific biomarkers. Given that different STS subtypes have distinct molecular drivers, differential chemotherapy responses, and varied prognoses, it is highly likely that each histotype has its own characteristic metabolic signature that is obscured in the aggregated analysis. For example, liposarcoma (25% of cohort) involves aberrant lipid metabolism driven by MDM2/CDK4 amplification, while synovial sarcoma (7%) is driven by SS18-SSX fusion proteins, and these distinct biologies may generate different metabolic reprogramming patterns.
Cohort size and external validation: The total cohort of 75 STS patients and 85 controls, while larger than most prior STS metabolomics studies, remains modest for a machine learning biomarker discovery study. The RF model was evaluated using OOB error and five-fold cross-validation, both internal approaches. No external validation on an independent institutional cohort was performed. Performance metrics derived from internal cross-validation are known to overestimate generalisation accuracy, and external validation on STS patients treated at different centres with different blood processing and storage protocols would be essential before these biomarkers could be considered for clinical deployment.
Single time-point sampling: All serum samples were collected at a single point in time prior to or at diagnosis. The study therefore cannot address questions about longitudinal dynamics: how do metabolite levels change during chemotherapy, after surgical resection, or at disease relapse? The authors propose that serial monitoring of metabolites such as choline over time could track tumour progression and treatment response, but this hypothesis remains untested. Longitudinal NMR metabolomics studies with pre-treatment, during-treatment, and post-treatment sampling would be needed to establish the utility of these biomarkers for monitoring.
Absence of targeted molecular validation: While NMR is inherently quantitative and reproducible, it has lower sensitivity than mass spectrometry-based approaches and cannot detect metabolites present at sub-micromolar concentrations. Some potentially relevant lipid metabolites and low-abundance signalling molecules may have been missed. Additionally, the study did not validate the identified metabolite changes at the tissue level through complementary approaches such as immunohistochemistry for choline transporters or metabolic enzyme expression. Mechanistic validation linking serum choline decrease to tumour choline uptake and membrane synthesis would strengthen the biological interpretation.
Subtype-specific metabolomic studies: The most immediately actionable next step identified by the authors is the design of larger, histotype-stratified metabolomics studies. For this to be feasible, multi-institutional collaboration will be required to assemble adequate sample sizes for each individual STS subtype. Rare subtypes such as synovial sarcoma, rhabdomyosarcoma, and fibrosarcoma individually account for 6-7% of this cohort, which translates to fewer than six patients per subtype. Pan-European or global sarcoma consortia, which already exist for clinical trials (such as the European Organisation for Research and Treatment of Cancer, EORTC, sarcoma group), represent natural infrastructure for assembling the biobank material needed to power subtype-specific metabolomics analyses. Such studies could identify unique metabolic vulnerabilities for each histotype, with direct implications for targeted therapy development.
Longitudinal liquid biopsy monitoring: The authors envision a paradigm in which serial serum NMR metabolomics is integrated into the clinical follow-up of STS patients, enabling the monitoring of treatment response, disease recurrence, and pharmacodynamic effects of systemic therapy. Given that choline levels correlated significantly with overall survival, tracking choline over the treatment course could provide a dynamic indicator of disease burden that is faster, less invasive, and less expensive than CT or MRI surveillance. The NMR protocol used in this study requires minimal sample manipulation and preserves the serum for future re-analysis, both properties that favour its integration into longitudinal biobanking workflows.
Integration with other liquid biopsy modalities: NMR metabolomics can be combined with other circulating biomarker modalities to create multimodal liquid biopsy panels. Circulating tumour DNA (ctDNA) and cell-free DNA (cfDNA) provide information about tumour-specific somatic mutations, copy number alterations, and methylation patterns. Circulating tumour cells (CTCs) offer information about tumour cell viability and phenotype. Combining metabolomic data with genomic and proteomic liquid biopsy readouts through multi-omics machine learning classifiers could substantially improve the sensitivity and specificity of non-invasive STS monitoring compared to any single modality alone.
Targeting choline metabolism therapeutically: The consistent and statistically robust downregulation of circulating choline, together with its association with overall survival, positions the choline metabolic axis as a candidate therapeutic target in STS. Choline kinase alpha inhibitors (such as MN58b and TCD-717) have shown preclinical activity in other solid tumours by depleting the phosphatidylcholine supply needed for membrane synthesis. Exploring these agents in STS models, and using serum choline as a pharmacodynamic biomarker in early-phase trials, would translate the metabolomic discovery in this study into a clinical development programme. The potential for choline-based PET imaging to serve as a companion diagnostic biomarker in such trials provides an additional translational pathway.