Lyme neuroborreliosis, the neurological manifestation of Lyme disease, presents a formidable diagnostic challenge in children. Its symptoms—facial nerve palsy, severe headache, meningeal signs—are non-specific and overlap with numerous other inflammatory and infectious conditions of the central nervous system. Conventional diagnosis relies on detecting anti-Borrelia antibodies in serum and cerebrospinal fluid alongside pleocytosis, an elevated white cell count in the spinal fluid. Yet serological assays have limited sensitivity early in infection, and children appear to be disproportionately affected, likely because of greater outdoor exposure and the particular characteristics of their developing immune systems. Metabolomics, the systematic measurement of small molecules, offers something serology cannot: a direct, real-time snapshot of the biochemical alterations occurring inside the central nervous system as the infection unfolds.

The Polish team enrolled 20 children with confirmed Lyme neuroborreliosis and 20 healthy age-matched controls, all recruited from the Lower Silesian Voivodeship. The patients, aged 6 to 17, had a median symptom duration of ten days before admission, and crucially, all were treatment-naïve—none had received antibiotics before sampling. This detail matters enormously for metabolomics, since even a short course of antimicrobial therapy can distort the metabolic landscape. From each patient, the researchers collected paired serum and cerebrospinal fluid samples simultaneously, before any treatment began, allowing a rare within-patient comparison between the systemic circulation and the central compartment. The study was approved by the institutional ethics committee and conducted in accordance with the Declaration of Helsinki, with informed consent obtained from all legal guardians.

The analytical approach was deliberately comprehensive. The team employed both nuclear magnetic resonance spectroscopy on a 600 MHz instrument and untargeted liquid chromatography-tandem mass spectrometry coupled to a quadrupole time-of-flight mass spectrometer. To maximize coverage of the CSF metabolome, every spinal fluid sample was run through two complementary chromatographic columns—one based on BEH Amide chemistry and the other on a zwitterionic ZIC-pHILIC stationary phase—generating two distinct datasets from the same twenty patients. Quality control samples, prepared by pooling aliquots from all extracts, were injected periodically throughout the analytical sequences to monitor instrument stability. In the NMR analysis, chemical shift reproducibility was exceptional, with coefficients of variation below 1% for 25 of the 27 annotated biomolecules, a result that validates the peak assignment strategy and confirms that the observed differences reflect genuine pathophysiology rather than instrumental noise.

Handling the data required considerable statistical care. Missing values in metabolomics datasets are typically “missing not at random,” arising because metabolite concentrations fall below the limit of detection rather than through random technical failure. The researchers compared several imputation approaches—including half-minimum, k-nearest neighbors, and quantile regression imputation of left-censored data—and found that a Random Forest algorithm consistently performed best, yielding the lowest normalized root mean square error and the optimal sum-of-ranks metric. After imputation and log2 transformation, the data were normalized using probabilistic quotient normalization, with quality control samples incorporated into the construction of the reference spectrum, then mean-centered and Pareto-scaled before statistical modeling.

The results were striking. In the serum, univariate analysis identified 17 significantly altered metabolites, and the most consistent finding was profound dysregulation of the purine pathway. Hypoxanthine, xanthine, and uric acid were all markedly elevated in the children with neuroborreliosis, with uric acid showing one of the largest fold changes in the entire dataset. This pattern admits at least two interpretations. On one hand, elevated purines may simply reflect the non-specific consequences of infection: increased cellular turnover, tissue injury, inflammatory activation, and oxidative stress. On the other hand, the researchers point to a more intriguing possibility grounded in Borrelia biology. The spirochete lacks a de novo purine synthesis pathway entirely and must salvage purines from its host to build its own nucleic acids. Mouse studies have shown that Borrelia burgdorferi depends on host-derived hypoxanthine for survival and infection, and evidence links purine regulation to bacterial resistance against reactive oxygen species. The authors are careful to note that their untargeted data cannot establish causality, and the purine elevation may combine both host and pathogen-driven contributions.

Equally significant was the elevation of L-tyrosine in the patients’ serum, which displayed the highest statistical significance among all metabolites measured, with a false discovery rate far below conventional thresholds. The researchers hypothesize that this mirrors the activation of tyrosine-dependent signaling cascades during acute inflammation, a mechanism previously implicated in the inflammatory damage that Borrelia inflicts on oligodendrocytes, the myelin-producing cells of the nervous system. Pathway enrichment analysis, mapping significant metabolites onto the KEGG database, repeatedly flagged the biosynthesis of phenylalanine, tyrosine, and tryptophan, alongside purine metabolism and glycerophospholipid metabolism. Notably, these enriched pathways appeared consistently across both blood and cerebrospinal fluid samples and across both statistical approaches, indicating a reproducible signal independent of analytical method.

The cerebrospinal fluid analysis told its own compelling story. Unsupervised hierarchical clustering of the 68 metabolites consistently detected across all 20 CSF samples revealed that most patients formed a broad, heterogeneous cluster, but three individuals branched off at the earliest node of the dendrogram, exhibiting a pronounced, coordinated up-accumulation of a large block of co-regulated metabolites. This group included amino acids such as glutamate, tyrosine, and phenylalanine; methylated osmolytes including choline, betaine, and asymmetric dimethylarginine; and purine intermediates including xanthine and uric acid. Several of these compounds overlapped with the metabolites altered in serum, suggesting that these three children may harbor a more pronounced systemic-to-central metabolic disturbance, and pointing toward greater permeability of the blood-CSF barrier in these individuals. Changes in glycerophospholipid metabolism—the chemical family that includes the membrane components phospholipids—further hint at transient disruption of the blood-brain barrier during active neuroinfection, although the authors caution that confirming structural damage requires complementary imaging or biophysical evidence.

Within the CSF, choline and glutamate emerged as candidate signatures of particular interest. Choline contributed strongly to group discrimination in the supervised PLS-DA models, with a variable importance in projection score exceeding 1.75. Elevated choline has previously been associated with active neuroinflammation and cellular membrane turnover, and it participates in epigenetic regulation as a methyl-group donor, mechanisms that may modulate inflammatory and oxidative stress responses. Supporting this interpretation, magnetic resonance spectroscopy studies of neuroborreliosis patients have independently identified choline abnormalities. Glutamate, the brain’s principal excitatory neurotransmitter, presents a different concern: its accumulation in the spinal fluid could theoretically precipitate excitotoxic cascades, a process in which excessive neurotransmitter signaling damages neurons. Both findings, the researchers emphasize, remain hypotheses requiring validation, particularly because the study lacked a non-Lyme neuroinflammatory control group that would establish whether these shifts are specific to Borrelia infection or represent a generalized central nervous system response to inflammation.

The study also connects to the broader tryptophan-kynurenine story in neuroborreliosis. Although tryptophan and kynurenine were not directly quantified in the CSF in this cohort, the enriched pathways indirectly suggest a metabolic shift consistent with activation of this pathway, which is driven by interferon-gamma-induced indoleamine 2,3-dioxygenase. Previous work has shown that children with Lyme disease display elevated CSF kynurenine and kynurenic acid, and the kynurenine-to-tryptophan ratio has been proposed as a tool for distinguishing bacterial from viral central nervous system infections. The authors also speculate that elevated uric acid, beyond being a purine catabolite, may function as a damage-associated molecular pattern, activating NLRP3 inflammasome complexes and amplifying local inflammation—a speculative metabolic synergy in which purine and tryptophan metabolites co-modulate the immunometabolic landscape.

As a pilot study, the work carries clear limitations, most notably the absence of matched CSF controls, which forced the researchers to draw conclusions about the central nervous system primarily through multivariate statistics and the identification of internal subgroups within the patient population. Yet the simultaneous collection of paired serum and CSF samples is a genuine methodological strength, enabling direct within-patient comparison between the systemic and central compartments. The robustness of the supervised models was verified through permutation testing with 1,000 random label permutations, achieving balanced error rates of zero with empirical permutation p-values below 0.001, indicating that the observed separation between patients and controls was unlikely to have arisen by chance.

What emerges from this study is a coherent immunometabolic narrative of pediatric neuroborreliosis: a pathogen that scavenges host purines, an inflammatory response generating oxidative stress and purine catabolites, aromatic amino acid shifts signaling receptor-level immune activation, membrane lipid changes suggesting barrier compromise, and choline and glutamate elevations pointing toward neuroinflammation and potential excitotoxicity. The identified signatures—particularly those governing purine metabolism and the inferred kynurenine pathway—now constitute a promising foundation for larger, controlled biomarker discovery studies. For a disease whose pediatric presentation is frequently nonspecific and whose standard tests often fall short, the prospect of a metabolomics-based diagnostic that reads the biochemical conversation between Borrelia and the developing brain is a compelling one, and this pilot study provides the detailed map from which such a test could eventually be built.

The study spearheaded by Nielsen, Fjordside, Drici, and their colleagues dives into the proteomic landscape—essentially the full complement of proteins present in cerebrospinal fluid (CSF)—to identify unique biomarkers that distinguish Lyme neuroborreliosis from other neurological disorders and healthy controls. This exploration into proteomics is crucial because proteins serve as both effectors and indicators of disease processes, offering a much richer and more dynamic snapshot of pathophysiology than genetic material alone. By profiling CSF with high-resolution mass spectrometry and subsequently analyzing the data through sophisticated machine learning models, the researchers have pushed the boundaries of diagnostic precision.

One of the paramount obstacles in Lyme neuroborreliosis diagnosis lies in its symptom overlap with other neurological diseases such as multiple sclerosis or viral meningitis. Traditional diagnostic methods rely heavily on serology, often yielding false negatives or inconclusive results due to immune evasion tactics employed by Borrelia. The innovative proteomic approach, however, overcomes these limitations by detecting subtle changes in protein expression and signaling pathways that are uniquely perturbed during infection. This method offers clinicians a powerful, unbiased window into the host-pathogen interaction, which could dramatically enhance early and accurate detection.

This landmark investigation involved collecting cerebrospinal fluid samples from a large cohort encompassing patients diagnosed with Lyme neuroborreliosis, individuals with other neurological conditions, and healthy controls. Employing next-generation mass spectrometry, the team cataloged thousands of proteins, analyzing quantitative shifts in abundance that correlated strongly with disease status. The dataset was then fed into machine learning algorithms designed to train on patterns within the proteomic data, enabling them to classify samples with remarkable accuracy. The marriage of cutting-edge proteomics and machine learning created a diagnostic tool that surpasses conventional methods both in sensitivity and specificity.

The machine learning model at the heart of this study embodies state-of-the-art artificial intelligence techniques, leveraging supervised learning paradigms such as random forests and support vector machines. These algorithms excel at detecting complex, nonlinear relationships within high-dimensional data—precisely the challenge posed by proteomic datasets that can include thousands of protein measurements per sample. By iteratively refining decision boundaries, the models distilled the proteomic signatures into diagnostic outputs, effectively giving clinicians a molecular fingerprint indicative of Lyme neuroborreliosis.

What sets this study apart is not just the use of proteomics or machine learning individually, but their strategic integration. The researchers demonstrated that combining these approaches allows for detection of disease-specific protein alterations that might be invisible to standard statistical analyses. Proteomics unearths a vast trove of biological signals, but without advanced computation, much of that wealth remains unexploited. Artificial intelligence serves not only as a pattern recognition tool but also enhances interpretability by highlighting key biomarker candidates that drive diagnostic predictions.

Beyond diagnosis, this work opens new avenues for exploring disease mechanisms and potential therapeutic targets. The proteins identified as critical markers often belong to pathways involved in immune response, inflammation, and neural tissue integrity. Understanding how Borrelia infection perturbs these pathways at a molecular level may spur development of novel interventions aimed at halting or reversing neurological damage. By providing a molecular roadmap, this integrated approach holds promise not just for Lyme disease but for a spectrum of neuroinfectious disorders.

The clinical implications are profound. Current diagnostic delays in Lyme neuroborreliosis frequently result in progression to severe neurological impairment, reduced treatment efficacy, and chronic symptoms. An objective, rapid, and reliable test based on proteomic signatures and machine learning classification could transform patient outcomes by enabling earlier intervention. Furthermore, this strategy could reduce unnecessary treatments in patients mistakenly diagnosed with Lyme neuroborreliosis, sparing them from potential side effects and healthcare costs.

This study also underscores the transformative potential of applying systems biology and artificial intelligence to infectious disease diagnostics. It exemplifies how cross-disciplinary collaboration among clinicians, bioinformaticians, and proteomics experts can yield tools capable of tackling conditions that have long evaded precise diagnosis. The broader research community stands to benefit from these methodologies as they are adapted to other pathogens and clinical contexts where diagnostic challenges prevail.

Notably, the integration of proteomics and machine learning in this work navigates around several common pitfalls in biomarker discovery, such as overfitting and batch effects. The researchers implemented rigorous validation protocols including independent test sets to ensure that the diagnostic models generalize well to new patient samples. This commitment to robustness buttresses confidence that the findings can be translated into clinically actionable assays.

Continued research will focus on refining the sensitivity thresholds of these proteomic markers, expanding patient cohorts for broader validation, and developing user-friendly platforms for clinical implementation. Portable mass spectrometers and automated data pipelines portend the feasibility of bringing these high-tech diagnostics directly to healthcare settings. Additionally, integrating these proteomic classifiers with other modalities such as neuroimaging and genomic data could enhance diagnostic comprehensiveness.

Ultimately, this synergistic blend of proteomics and machine learning heralds a new era in infectious disease diagnostics—one where the invisible molecular signatures of disease can be harnessed algorithmically to provide definitive answers. As Lyme neuroborreliosis exemplifies the challenges of diagnosing elusive infections, this pioneering study serves as a beacon, illuminating how advanced technologies can be leveraged to overcome diagnostic uncertainty and improve patient care worldwide.

The implications reverberate beyond Lyme disease, inspiring optimism that similar multi-omics and AI strategies might soon revolutionize diagnostics across a gamut of neurological, infectious, and autoimmune disorders. As we stand on the cusp of personalized medicine, this work exemplifies the promise of integrating biological complexity with computational power to unravel and accurately identify the molecular fingerprints of human disease.

Subject of Research:
Lyme neuroborreliosis diagnosis through proteomics and machine learning.

Article Title:
The diagnostic potential of proteomics and machine learning in Lyme neuroborreliosis.

Article References:
Nielsen, A.B., Fjordside, L., Drici, L. et al. The diagnostic potential of proteomics and machine learning in Lyme neuroborreliosis. Nat Commun 16, 9322 (2025). https://doi.org/10.1038/s41467-025-64903-z

Image Credits: AI Generated