Machine learning identifies fatigue as a key symptom of fibromyalgia reflected in tyrosine, purine, pyrimidine, and glutaminergic metabolism

Abstract

Fibromyalgia patients vary in clinical phenotype and treatment can be challenging. The pathophysiology of fibromyalgia is incompletely understood but appears to involve metabolic changes at rest or in response to stress. We enrolled 54 fibromyalgia patients and 31 healthy controls to this prospective study. Symptoms were assessed using the Fibromyalgia Impact Questionnaire (FIQ) and blood samples were collected for metabolomics analysis at baseline and after an oral glucose tolerance test and a cardiopulmonary exercise test. We identified key symptoms of fibromyalgia and related them to changes in metabolic pathways with supervised and unsupervised machine learning methods. Algorithms trained with the FIQ information assigned the fibromyalgia diagnosis in new data with balanced accuracy of 88% while fatigue alone already provided the diagnosis with 86% accuracy. Supervised analyses reduced the metabolomic information from 77 to 13 key markers. With these metabolites, fibromyalgia could be identified in new cases with 79% accuracy. In addition, 5‐hydroxyindole‐3‐acetic acid and glutamine levels correlated with the severity of fatigue. Patients differed from controls at baseline in tyrosine and purine pathways, and in the pyrimidine pathway after the stress challenges. Several key markers are involved in glutaminergic neurotransmission. This data‐driven analysis highlights fatigue as a key symptom of fibromyalgia. Fibromyalgia is associated with metabolic changes which also reflect the degree of fatigue. Responses to metabolic and physical stresses result in a metabolic pattern that allows discrimination of fibromyalgia patients from controls and narrows the focus on key pathophysiological processes in fibromyalgia as treatment targets.

Study Highlights.

• WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

Among fibromyalgia patients, there is a lot of variation in symptom profiles. There are also many hypothesized pathophysiological mechanisms for fibromyalgia, including changes in metabolism. However, it remains unclear what are the contributions of individual symptoms to fibromyalgia symptom burden, what metabolic changes are relevant for fibromyalgia, and whether different metabolic anomalies could be linked to individual symptoms of fibromyalgia.

• WHAT QUESTION DID THIS STUDY ADDRESS?

Through data‐driven analysis we assessed (1) the impact of individual symptoms in fibromyalgia; (2) metabolic differences between fibromyalgia patients and healthy controls at baseline and in stress responses; and (3) the association of the most significant symptom and metabolism.

• WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE?

Our results highlight the impact of fatigue on fibromyalgia symptom burden. We also indicate tyrosine, purine, and pyrimidine pathways as well as changes in metabolites involved in glutaminergic neurotransmission as possible pathophysiological mechanisms of fibromyalgia, and 5‐hydroxyindole‐3‐acetic acid and glutamine as associating with fatigue in fibromyalgia. Importantly, most of the metabolic findings only became evident in stress responses.

• HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE?

Metabolomic research could benefit from more emphasis on dynamic assessment of metabolism under diverse metabolic conditions, such as the stress challenges employed in our study.

INTRODUCTION

Fibromyalgia affects 2%–4% of the general population, causing a long‐lasting reduction in quality of life. ¹ Chronic widespread pain, fatigue, and sleep disturbance are core symptoms of fibromyalgia, though with much variation in symptom profiles between patients. ² Fibromyalgia and the pathophysiology of its symptoms are only partly understood and its symptom profiles may be explained by different pathophysiological mechanisms. ³ Resolution of symptoms is rare with the current treatment options and symptoms respond differently to therapies. ⁴ Identifying biomarkers clarifying the pathophysiology and clinical subtypes of fibromyalgia is a potential path to developing new therapies and tailoring individual treatment strategies.

Clinical diagnostic tools, such as the Fibromyalgia Impact Questionnaire (FIQ), have been developed to capture the heterogeneous symptoms. ⁵ Possible biomarkers have been explored with metabolomics, a more comprehensive simultaneous analysis of dozens of metabolites. Metabolomics is an emerging area of research in chronic pain. Metabolomic signatures of fibromyalgia have been reported in various metabolic pathways, including energy, lipid, and amino acid metabolites, reflecting heightened oxidative stress, inflammation, and tryptophan degradation. ⁶ The consistency of these findings is limited, likely due to differences in both patient and control groups, sample type (serum, plasma, urine, or feces), genetic variability within the patient group, and statistical methods. Evolving possibilities of machine learning enable combining large datasets of metabolomic data with heterogeneous symptom profiles in fibromyalgia and clarifying their possible connections.

In this analysis, we aimed to identify the key metabolic processes associated with the core symptoms of fibromyalgia using a more controlled experimental setup. We used a data‐driven machine learning approach to identify key symptoms and associated metabolic changes at baseline and after exposure to both metabolic and physical stresses, comparing fibromyalgia patients with healthy controls. In global metabolomics analysis, the circulating metabolome has been shown to respond to acute glucose and physiological stresses by adaptations of energy, amino acid, and neurotransmitter metabolism. ⁷ , ⁸

METHODS

Study design and collection of clinical data on fibromyalgia

The study design and data acquisition have previously been described in detail. ⁹ , ¹⁰ , ¹¹ , ¹² In brief, the study cohort comprised 54 female patients with fibromyalgia, aged 18–65 years, from Helsinki University Hospital outpatient clinics, City of Vantaa primary health care, and the private clinic of one of the authors (R.M.). Thirty‐one age‐matched female controls were recruited from the staff of these health care units and a local home economics organization (Uudenmaan Martat ry). The inclusion criterion for the patients was fibromyalgia, as diagnosed according to the American College of Rheumatology 1990 criteria (ACR1990). ¹³ Exclusion criteria were atherosclerotic disease, heart disease, uncontrolled hypertension, diabetes, or neurological (not including migraine), neuromuscular, muscle, or severe psychiatric conditions, any severe musculoskeletal condition which would prevent cycle ergometry, continuous use of statins, beta‐blockers, or beta‐agonists, or poor Finnish language skills.

The study was conducted in accordance with the Declaration of Helsinki on Biomedical Research Involving Human Subjects and was approved by the Helsinki and Uusimaa Hospital District Ethics Committee (229/13/03/02/2015) and retrospectively registered in ClinicalTrials.gov (NCT03300635) on October 3, 2017. Informed written consent was obtained from each participant. The study included three visits between November 2015 and June 2018 during which: (i) a clinical interview and examination were performed, and questionnaire data acquired; (ii) after fasting, a 2‐hour oral glucose tolerance test (OGTT) was performed; and (iii) a cardiopulmonary exercise test (CPET) was performed. Serum samples for metabolomics analysis were collected at the beginning and end of both OGTT and CPET. The data analyzed included the following clinical information: FIQ score, age, and body mass index (BMI).

Assay of targeted metabolomics of blood markers

Metabolomics was performed at the Finnish Institute of Molecular Medicine (FIMM) using previously published methods. ¹⁴ The assays were performed in a similar manner to that described in Miettinen et al. ¹⁵ Of the 102 metabolites, 25 were excluded for having levels below lower limits of quantification, poor chromatography, or due to unacceptable linearity of the calibration curve, resulting in quantitation of 77 metabolites. The details of the metabolomic analyses are reported in Supplementary Materials S1. Analyses were performed on an ACQUITY UPLC system coupled to a Xevo® TQ‐S triple quadrupole mass spectrometer (Waters Corporation). MassLynx 4.1 software was used for data acquisition, data handling, and instrument control. Data processing was performed using TargetLynx software and metabolites were quantified using labeled internal standards and external calibration curves. Batches were normalized to a reference sample (quality control serum) using MetaboAnalyst (version 5.0, https://www.metaboanalyst.ca/home.xhtml). ¹⁶

Data analysis

The data analysis combined classical statistics and machine learning methods in the sense of a “mixture of experts” approach, which has repeatedly been shown to be superior to relying on a single method (such as regression analysis alone ¹⁷ ) in the analysis of “omics” ¹⁸ and other types of biomedical data. ¹⁹ Hence, the data analysis included unsupervised, semi‐supervised, and supervised methods from statistics and machine learning (Figure 1). Programming was done in the Python language ²⁰ using Python version 3.8.15 for Linux, and in the R language ²¹ using the R software package (version 4.2.2 for Linux). The detailed analyses are reported in Supplementary Materials S1. First, the key symptoms of fibromyalgia were identified from the FIQ. Because the Finnish translations of FIQ items 3 and 4 (fiq_3 and fiq_4) refer merely to “work” rather than to “work, including housework”, as in the original questionnaire, many participants who were not currently working gave no response. This led to 26.2% missing data in both items, a figure comparable to that observed by others. ²² Unsupervised analysis of the FIQ items included their correlation by calculating Spearman's rank correlation coefficient, followed by the projection of the z‐standardized data onto uncorrelated planes using principal component analysis (PCA). Supervised analyses were then performed to check how much of the full questionnaire's information was already captured by this variable. The supervised methods included three different algorithms (support vector machines [SVM], ²³ random forests,²⁴ and logistic regression ²⁵ ). Prior to classifier training, a class‐proportional random sample of 20% of the dataset was set aside as a validation sample, not touched during algorithm tuning and training. The remaining 80% of the dataset was used for classifier training. The trained classifiers were then applied to random subsets of 80% each of the validation dataset using 100‐fold nested cross‐validation. Balanced accuracy was used as the main parameter to evaluate the classification performance. ²⁶ In addition, the area under the receiver operator curve (ROC‐AUC) was calculated.

FIGURE 1.

Flowchart of the study cases and data analysis. The data analysis was designed to identify structures in the clinical and metabolomic data that supported the prior classification into fibromyalgia patients and controls (structure detection), and to extract the most relevant items from a collection of questionnaire items, clinical variables, and metabolic markers (information reduction). ANOVA, analysis of variance; FIQ, Fibromyalgia Impact Questionnaire; PCA, primary component analysis; PLS‐DA, partial least squares discriminant analysis.

The preprocessed and standardized metabolomics data were projected onto a two‐dimensional plane using partial least squares regression followed by discriminant analysis (PLS‐DA). ²⁷ Metabolic marker concentrations and their changes following stress were analyzed using classical statistics implemented as t‐test or analysis of variance with the within‐subjects factor “test time” (two levels: baseline and post‐stress) and the between‐subjects factor “group” (two levels: patients and controls). The respective effect sizes in the analyses of variance were expressed as η ^{2 28} and subjected to cABC analysis, an item categorization technique that divides a set of positive numeric data into three disjoint subsets, labeled “A” to “C”, of which subset “A” contained the “important few” and was retained. To further explore the metabolic pathways of fibromyalgia, a metabolite set enrichment analysis (MSEA) was performed. The quantitative enrichment analyses were performed using prepackaged software tools available as a web‐based comprehensive metabolomics data processing tool, MetaboAnalyst (version 5.0, https://www.metaboanalyst.ca/home.xhtml; accessed May 29, 2023). ¹⁶

RESULTS

Key symptoms of fibromyalgia

Two fibromyalgia patients had >20% missing FIQ data and were excluded from symptom analysis. Thus, the cohort analyzed consisted of fibromyalgia patients and 31 controls. They were of similar ages (mean ± standard deviation: 44.8 ± 12.4 years vs. 46 ± 11.7 years, respectively; Wilcoxon‐Mann–Whitney U test: W = 752.5, p = 0.6175), but the patients had a higher BMI than the controls (27.9 ± 5.9 kg/m² vs. 24.68 ± 3.27 kg/m²; W = 925, p = 0.01565). Being overweight or obese was more prevalent among fibromyalgia patients, of whom 33% were normal weight (BMI 18.5–25.0 kg/m2) compared with 52% of controls. Twelve missing values in the FIQ data were imputed.

Unsupervised analyses indicated that most of the FIQ items showed strong intercorrelations (Figure 2a). Fatigue (fiq_6) was the most strongly correlated with the other variables in a 100‐fold bootstrap resampling analysis and was considered a prototypical variable in the questionnaire. The PCA projection provided a clear separation of patients and controls on the first two principal components (PCs), which together explained 82% of the variance. Most of the total variance (73.7%) was accounted for by the first PC. Fatigue (fiq_6) was the most important feature for the PCA projection along the first dimension (Figure 2b).

FIGURE 2.

Fibromyalgia Impact Questionnaire (FIQ/fiq) data for the 83 participants, their correlations, and primary component analysis (PCA) projection. (a) Correlation matrix of the FIQ data, color‐coded by strength and direction of the correlation and labeled with the value of Spearman's ρ. At the right of the matrix, the variables in descending order of the strongest correlations with other variables FIQ, showing the minimum, quartiles, median (solid line inside the box), and maximum values. (b) PCA projection of the dataset instances of the eight FIQ items. Subjects are represented by their prior class membership (patients P, triangle versus controls C, dot). The projection plane (dimension 2 vs. dimension 1) consists of Voronoi cells around each datapoint, colored according to the prior class membership of the respective datapoint. (c) Raw data and statistical test results. Individual datapoints are plotted on violin plots showing the probability density distribution of the variables. The statistical significance of the differences between patients and controls for each variable was analyzed using Mann–Whitney U tests with p‐values shown above each variable. (d) Bar graph of the contributions of each variable to PC1. The dashed horizontal reference line corresponds to the expected value with uniform contributions.

Supervised analyses showed that machine learning‐based classifiers could be trained with the FIQ information to assign the fibromyalgia diagnosis for new data (i.e. the 20% validation sample not seen by the algorithms during tuning and training) with balanced accuracies of up to 88%. Fatigue alone contained enough information from the full FIQ to train algorithms to place a subject in the correct prior class, as with the full questionnaires (Table S1). This was generalizable to new subjects: in the 20% validation sample not seen during classifier training, the balanced classification accuracy was still up to 86% when only fiq_6 was used. In contrast, when permuted variables were used for training, class assignment was only at the level of guesswork with a balanced accuracy of 50%, indicating that the good classification performance was not an artifact of overfitting. In addition, the classifiers trained with only the sum score of the FIQ did not perform better than classifiers trained with the fatigue variable.

Metabolomic pattern of fibromyalgia

The data analyzed for group differences consisted of three matrices of sizes 57 × 77, 104 × 77, and 102 × 77 for baseline and before and after glucose (OGTT) or physical (CPET) stress, respectively. Thus, baseline metabolomics data were available for 57 subjects (of whom 30 were patients), glucose stress metabolomics data were available for 52 subjects (28 patients), and physical stress metabolomics data were available for 51 subjects (28 patients).

Metabolic markers distinct in patients from controls

PLS‐DA showed some separation of patients from controls in individual metabolic states (fasting and glucose and physical stress response, Figure 3a–c). PLS‐DA combining all three states provided a clear separation of patients from controls on a two‐dimensional ${ℝ}^2$ projection plane (Figure 3d). Metabolites with the largest sizes of significant effects (expressed as Cohen's η ²) comprised cAMP, cytidine, carnosine, and inosine at baseline with relatively large effect sizes (Figure 4a, Figure S1). In addition, xanthine showed different stress‐induced changes between patients and controls, glutamine and glutamate changed differently between groups only in response to CPET, whereas aspartate, hypoxanthine, spermidine, histidine, normetanephrine, and allantoin changed differently between patients and controls only after OGTT (Figure 4).

FIGURE 3.

Projection of metabolomics data on the first two dimensions obtained with partial least squares discriminant analysis (PLS‐DA). Projections (m = 77 metabolic markers) are shown separately for baseline values before glucose stress (a) and for values before and after either glucose (b) or physical stress (c). Regions of projected data, according to patients/controls or before/after stress, are indicated by ellipses covering the 96% confidence limits of the groupwise projections. Datapoints are labeled according to the probe code in the data. The suffix “G” indicates glucose stress, “E” indicates exercise stress, and the following numbers indicate the time of probe sampling (1 = baseline, 2 = after stress). (d) In addition, the metabolic markers were rearranged into d = 231 metabolomic variables including baseline (before glucose stress) and differences from baseline after glucose stress and after exercise stress. The standardized data were the projected using PLS‐DA.

FIGURE 4.

Statistical analyses of differences in metabolic markers. Differences shown between patients and controls with largest effect sizes at baseline or in their changes after glucose stress (G) and exercise (E). (a) Bar graph of effect sizes of interest, expressed as η ², including differences in baseline concentrations before glucose stress (univariate analysis of variance or t‐tests) and changes induced by stress (interaction “group” by “test time” in analysis of variance for repeated measures). (b) ABC analysis plot (blue line) showing the cumulative distribution function of effect sizes. The red lines show the boundaries between the ABC subsets “A”, “B”, and “C”. (c) Venn diagram of the metabolites that produced the largest effects as assigned to ABC set “A” or “B”, separately for baseline and glucose/physical stress.

The task of the supervised analyses of discriminating between patients and controls in a validation sample of 20% of the cases, separated prior to classifier training, was unsuccessful when using the statistically significant metabolomics variables identified in the above‐reported analysis (Table S2). The balanced classification accuracy in the independent validation sample reached 79% and the ROC‐AUC 84%, while the 95% confidence intervals obtained from 100‐fold nested cross‐validation runs were well above the 50% mark. Specifically, the included informative metabolomic variables were cAMP, cytidine, carnosine, and inosine at baseline before OGTT, differences from baseline in cytidine, carnosine, inosine, aspartate, hypoxanthine, spermidine, histidine, normetanephrine, allantoin, and xanthine after OGTT, and differences from baseline in xanthine, glutamine, and glutamate after CPET. When the PLS‐DA was trained with permuted metabolomics information, it regressed to mere guessing of class assignment, indicating that the results were not an overfitting artifact.

Metabolic markers reflecting the degree of fatigue

The strongest 100‐fold cross‐validated correlations with the degree of fatigue, with absolute values of Pearson's r > 0.4, were 5‐hydroxyindole‐3‐acetic acid (5‐HIAA) at baseline and exercise‐induced changes in glutamine and kynurenine concentrations (Figure S2). The same metabolites were also identified as significant variables by linear regression analysis (Table S3). The levels of 5‐HIAA were negatively associated with fatigue, and greater increases in glutamine after exercise were associated with higher levels of fatigue (Figure S2). It is noteworthy that the same metabolites were not correlated with the subjects’ BMIs (5‐HIAA: r = −0.094, p = 0.5593; glutamine: r = 0.21, p = 0.188; L‐kynurenine: r = 0.095, p = 0.5531).

Metabolic purine and pyrimidine pathways deregulated in fibromyalgia

The largest differences in metabolic pathways between fibromyalgia patients and controls were seen in the purine and tyrosine pathways in the fasting state, in the purine and pyrimidine pathways in response to OGTT, and in the pyrimidine pathway in response to CPET (Figure 5).

FIGURE 5.

Metabolic pathway enrichment analysis. (a) Fasting state pathway enrichment between patient and control groups. (b) Oral glucose tolerance test (OGTT)‐induced difference pathway enrichment between patient and control groups. (c) Cardiopulmonary exercise test (CPET)‐induced difference pathway enrichment between patient and control groups.

DISCUSSION

Based on differences in metabolomic patterns reflecting the purine, pyrimidine, and tyrosine pathways, fibromyalgia patients could be distinguished from healthy controls with 79% accuracy. The present results therefore support the existence of a metabolic signature for fibromyalgia. The signature was narrowed down to the most relevant pathways for this syndrome using information reduction, a typical procedure in machine learning. On the clinical side, information reduction limited the fibromyalgia symptoms asked about in the FIQ to fatigue, which alone provided enough information to discriminate patients from controls with 86% accuracy. As mentioned in the introduction, the primary goal of this analysis was to identify critical metabolic processes associated with the primary symptoms of fibromyalgia. The approach was to use machine learning techniques for knowledge discovery, ²⁹ as opposed to constructing a diagnostic tool from metabolic markers alone. While these markers did not surpass the efficacy of traditional clinical diagnostic tools, they proved to be instrumental in narrowing the scope of potential pathophysiological processes associated with fibromyalgia, as discussed below.

When using more comprehensive measures such as the Multidimensional Fatigue Inventory, fatigue has been shown to correlate negatively with physical activity, walking ability and motivation, while correlating positively with anxiety and depression. ³⁰ We observed that glutamine increase in response to physical exercise correlated with greater fatigue: similar associations have previously been reported, as have conflicting results, with glutamine even being seen as an “antifatigue amino acid” in the diet. ³¹ A neurochemical basis for fibromyalgia involving fatigue and glutamine is supported by reports of increased brain glutamine and glutamate levels in fibromyalgia patients assessed by magnetic resonance spectroscopy. ³² As a further metabolic marker, 5‐HIAA levels were observed to correlate negatively with fatigue, in line with previous reports ³³ and consistent with its involvement in depression ³⁴ as another possible symptom in fibromyalgia.

Sustained intensive physical activity would quickly deplete cell energy stores of ATP which could not be replenished by glycolysis, while glucose loading, as in the glucose tolerance test, would lead to detrimental hyperglycaemia if not controlled. Therefore, metabolomic adaptivity between catabolism and anabolism is crucial to maintain homeostasis and steady energy metabolism. Hormones, such as catecholamines, cortisol, glucagon, insulin, and leptin, control many of these adaptations. Decreased adaptivity to these stressors can be used to diagnose many conditions, such as cardiopulmonary or muscle diseases, and diabetes. ³⁵

A major finding of the present work was that capturing metabolomic markers in fibromyalgia may be difficult at baseline, therefore necessitating the induction of stress, such as by the OGTT or the CPET. For example, while metabolic changes in tyrosine and purine pathways were observed at baseline, differences in pyrimidine pathways became manifest only after stress. The biological roles of the 13 metabolites that passed the feature selection process in the present analysis are summarized in Table S4 according to the pathways involved.

At baseline, fibromyalgia patients differed from controls by higher blood levels of cyclic adenosine monophosphate (cAMP). Overactivation or chronic cAMP signaling has been shown to be involved in pain chronification. ³⁶ , ³⁷ Cyclic‐AMP‐activated pathways affect mitochondrial biogenesis, anti‐inflammatory and antioxidant responses, and are also involved in purine metabolism. ³⁷

Purine and pyrimidine metabolism, de novo synthesis, salvage, and degradation are tightly controlled. Disturbances in these pathways can lead to a variety of diseases, like primary gout (increased de novo synthesis of purine nucleotides resulting in hyperuricaemia), Lesch–Nyhan syndrome (defective salvage), or immunodeficient conditions associated with adenosine deaminase defects (reduced degradation). Baseline differences in carnosine, cytidine, and inosine link fibromyalgia to tyrosine and purine pathways, which have also been previously linked to fibromyalgia. ³⁸ Altering purine metabolism, such as with dietary replacement of S‐adenosylmethionine, which inhibits some purine and pyrimidine metabolizing systems, ³⁹ has been suggested to have some beneficial effects on fibromyalgia. ⁴⁰

In response to stress, a partially different metabolic pattern emerged. The major differences shifted toward glutamine and pyrimidine pathways. Glutamine, which is also a precursor of spermidine, is a non‐essential amino acid necessary for immune system function. Glutamine is an important nitrogen deposit with glucose availability affecting nitrogen balance through its effect on glutamine metabolism. Glutamine is also transformed into glutamate, a key component in cellular metabolism and also an excitatory neurotransmitter mainly affecting AMPA and NMDA receptors. The latter are involved in hyperexcitability in nociception and are thought to play a part in fibromyalgia pathophysiology. ⁴¹ Glutamine and glutamate are also involved in purine metabolism. ⁴² Glutamate levels are expected to rise in response to exercise ⁴³ whereas glutamine levels remain stable. ⁴⁴ This was seen in controls while fibromyalgia patients showed an increase in glutamine levels but no increase in glutamate. Systemic glutamate and glutamine have been reported to be elevated in fibromyalgia, and this has been suggested to be due to reduced bacterial transformation of glutamate to GABA. ⁴⁵ Several of these metabolites are known to be involved in glutaminergic neurotransmission, mainly involving AMPA and NMDA receptors. This is in line with evidence about altered NMDA receptor activity or metabolic traces in fibromyalgia patients, ⁴⁶ and supports a role for (low‐dose) NMDA antagonists as a treatment of fibromyalgia, ⁴⁷ as well as a dietary therapy component via magnesium substitution which acts as an NMDA antagonist. ⁴⁸ Cytidine is a pyrimidine component of RNA and controls neuronal–glial glutamate cycling, cerebral phospholipid metabolism, catecholamine synthesis, and mitochondrial function. ⁴⁹ Cytidine supplementation lowers brain glutamate and glutamine levels, with elevated glutamate levels being shown in pain‐controlling brain areas in fibromyalgia patients. ³ Cytidine supplementation may have analgesic, anti‐inflammatory, neuroprotective, and cognitive enhancing properties. ⁵⁰

Our study identified or verified several key pathways involved in fibromyalgia, particularly those related to fatigue, that can be targeted by pharmacological or dietary interventions. While higher BMI is known to play a role in the prevalence and severity of fibromyalgia, and likely affects fatigue severity as well, it is noteworthy that key metabolite concentrations correlated with fatigue but not with BMI in the present cohort. Therefore, a general weight loss diet may not be as effective as more targeted interventions on the identified pathways to effectively address fibromyalgia symptoms.

The validated targeted metabolomics method used in this study was designed for screening serum biomarkers that have important roles in many biological processes and are biomarkers for several diseases involving major metabolic pathways from 24 classes in a single assay. ¹⁴ The targeted analysis method used was performed with a combined liquid chromatography (LC) separation technique with tandem mass spectrometry (MS/MS). This method was chosen for its high accuracy and better reproducibility with matrix effect validation and higher sensitivity for the detection of metabolites that are low in abundance. The method applied also enabled quantification with high throughput. Untargeted relative metabolomics could also have been a valid approach for new biomarker discovery. However, this approach might have created data for unannotated compounds requiring confirmation with quantitative analysis methods.

In a collaborative effort between clinicians, biochemists, and data scientists, complex clinical and biochemical information collected from fibromyalgia patients under various conditions was narrowed down to a set of relevant metabolic pathways that play key roles in the disease. These metabolic patterns contained enough information to train an algorithm that could determine whether a person was a fibromyalgia patient or a healthy control almost as accurately as a clinical diagnosis. The data therefore likely suggest a metabolic basis for fibromyalgia, supporting the idea that fibromyalgia is, at least in part, a metabolic syndrome.

Importantly, our findings support the involvement of microbiome–gut–brain axis in fibromyalgia pathogenesis. We found reduced 5‐HIAA levels in fibromyalgia patients, indicating reduced serotonin levels. Tryptophan degradation into kynurenine instead of metabolism to serotonin in fibromyalgia has also been reported before and this explains the reduced serotonin, as has been found in fibromyalgia previously. ⁶ Kynurenine has pro‐inflammatory and pro‐nociceptive properties. Other products of serotonin include melatonin. Reduced melatonin levels may be involved in sleep disturbances in fibromyalgia patients, but this has not been demonstrated in fibromyalgia to our knowledge. In humans, serotonin is mainly produced in the gut and systemic serotonin affects the brain through the vagal nerve. The ANS also plays an important role in gastrointestinal tract function through the vagal nerve. Consistent with previous findings, we found elevated glutamate levels in the fibromyalgia patients. This has been proposed to be due to reduced transformation of glutamate into GABA by gut bacteria. ⁵¹ We also found elevated levels of purines, both inosine and its metabolites hypoxanthine and xanthine, in the fibromyalgia patients. This has also been reported previously, together with reduced adenosine levels in fibromyalgia, with the proposed cause being reduced conversion of inosine into adenosine, possibly involving gut bacteria. ³⁸ , ⁵¹ The existence of gut dysbiosis in fibromyalgia is supported by studies showing changes in serum bile acid levels explainable by dysbiosis, and, importantly, by studies of the microbiome itself showing dysbiosis in fibromyalgia. ⁵¹ These interconnections are outlined in Figure 6. This may provide clinical applications in the future, as all relevant metabolic pathways are possible targets for dietary interventions, which have already been associated with beneficial effects in fibromyalgia. Further, the effect of sleep interventions or interventions to enhance vagal tone, such as vagal nerve stimulation, should also be studied in fibromyalgia patients.

FIGURE 6.

Possible connections of gut microbiota, metabolisms, and fibromyalgia symptoms.

Our study has some limitations. Because of the small population size, the metabolic differences seen in our data cannot be generalized to all fibromyalgia patients. Fibromyalgia is likely a multifactorial condition with great variability in both genetic and epigenetic factors which cause downstream effects on metabolism. Also, our study population consisted only of females. The symptom profiles and metabolism of male patients are likely to differ somewhat from those of females. However, our results lend support to the involvement of metabolism in fibromyalgia. While some sex differences are likely to exist, we would expect similar mechanisms in male patients as, to our knowledge, the metabolites indicated are not particularly sex‐dependent.

To conclude, this data‐driven analysis highlights fatigue as a driving symptom of fibromyalgia associated with metabolic changes. In particular, the response to stress results in a metabolic pattern that allows fibromyalgia patients to be discriminated from healthy controls with good accuracy and narrows the focus to key pathophysiological processes in fibromyalgia.

AUTHOR CONTRIBUTIONS

T.Z., A.I.N., R.M., E.K., and J.L. wrote the manuscript. T.Z., A.I.N., R.M., E.K., and J.L. designed the research. T.Z. and R.M. performed the research. A.I.N. and J.L. analyzed the data. J.L. contributed new analytical tools.

FUNDING INFORMATION

This work was supported by Finnish State Research Funding (TYH2017215), the Signe and Ane Gyllenberg Foundation, the Emil Aaltonen Foundation, and the Juhani Aho Foundation for Medical Research. JL was supported by the Deutsche Forschungsgemeinschaft (DFG LO 612/16‐1).

CONFLICT OF INTEREST STATEMENT

The authors declared no competing interests for this work.

Supporting information

Figure S1.

Table S1.

Data S1.

Data S2.

Zetterman T, Nieminen AI, Markkula R, Kalso E, Lötsch J. Machine learning identifies fatigue as a key symptom of fibromyalgia reflected in tyrosine, purine, pyrimidine, and glutaminergic metabolism. Clin Transl Sci. 2024;17:e13740. doi: 10.1111/cts.13740