Identification of potential biomarkers for Lyme disease using bioinformatics and machine learning

Abstract

Lyme disease (LD) presents significant diagnostic challenges due to the absence of a reliable screening method for initial detection. This study aimed to identify potential biomarkers using bioinformatics and machine learning algorithms, which may contribute to future biomarker-based research for Lyme disease diagnostics. The gene expression profile datasets GSE145974 and GSE63085 were analyzed using machine learning to identify hub genes among differentially expressed genes. High-throughput data and receiver operating characteristic curves were used to validate these genes. The molecular mechanisms underlying LD were explored using functional enrichment analysis. The correlation between immune cell counts and insomnia in LD was further validated using clinical data from the GEO database. Gene set enrichment analysis indicated that hub genes were enriched in circadian rhythms. The integration of machine learning revealed FCGR1B, MPP1, and HSPA6 as potential central genes involved in immune response and diagnostic biomarkers for Lyme disease. Immune infiltration analysis showed that LD is frequently associated with the monocyte–macrophage system and humoral immunity. This study provides novel insights into the targeted treatment of LD by revealing novel diagnostic biomarkers.

Introduction

Lyme disease (LD) is a prevalent vector-borne infectious disease caused by spirochetes and transmitted by ticks [1–3]. It progresses through three stages, namely, early localized, early disseminated, and late disseminated [4, 5]. The characteristic early sign of LD is an erythema migrans rash, which is painless and typically > 2 inches in diameter and presents in 60–70% of the affected patients [6]. Advanced stages can lead to severe complications, including arthritis, carditis, and central nervous system diseases such as meningitis and encephalitis [7–9]. The pathogenesis of LD involves evasion of the host’s immune system by Borrelia spirochetes through alterations in membrane proteins and surface proteins [10, 11]. This evasion facilitates their survival and proliferation within different host environments. Annually, more than 470,000 individuals are diagnosed and treated for Lyme disease annually in the USA alone [12]; however, the actual number of cases may be higher, which remain undetected due to diagnostic challenges.

Currently, clinical diagnosis of LD requires a comprehensive evaluation of epidemiological history, clinical signs, and laboratory tests. Standard detection methods include serological tests and polymerase chain reaction (PCR) [13, 14]. Nevertheless, these methods have limitations because serological tests can detect antibodies only 2–4 weeks post-infection and lack early-stage sensitivity [15, 16], and PCR requires samples from the tick-infested area and can yield false results [17].

Despite advancements in LD research, there is a crucial need for reliable early diagnostic methods. Most studies have focused on immunological and therapeutic strategies, with less emphasis on molecular mechanisms and biomarkers [17]. Investigating the impact of LD on the expression of human genes could provide novel insights into its diagnosis and treatment [18, 19]. For instance, a 31-gene LD classifier developed using machine learning demonstrated 90% sensitivity and 100% specificity for detecting early-stage LD [20].

This study aimed to identify biomarkers for the early diagnosis of LD using bioinformatics and machine learning. We integrated three machine learning techniques—least absolute shrinkage and selection operator (LASSO), random forest (RF), and support vector machine with recursive feature elimination (SVM–RFE)—to improve the identification accuracy of biomarkers. This approach focuses on molecular data analysis and may facilitate the development of specific diagnostic and therapeutic strategies.

Materials and methods

Data sources and search criteria

Two gene expression datasets were obtained from the Gene Expression Omnibus (GEO) database: GSE145974 and GSE63085. The GSE145974 dataset contains expression profiles of peripheral blood mononuclear cells (PBMCs) from patients with acute Lyme disease and healthy controls, based on the GPL11154 platform (Illumina HiSeq 2000). The GSE63085 dataset includes PBMC samples collected at three time points (pre-treatment, post-treatment, and 6 months follow-up) from patients with early disseminated Lyme disease, based on the GPL13667 platform (Affymetrix Human Genome U219 Array).

These datasets represent transcriptomic profiles of early-stage Lyme disease patients, and both contain well-annotated case and control groups. GSE145974 was used as the discovery dataset for differential gene expression analysis, WGCNA, immune infiltration analysis, and machine learning modeling. GSE63085 was used exclusively as an external validation dataset to confirm hub gene performance.

Identification of differentially expressed genes (DEGs)

Gene expression data from GSE145974 were preprocessed using R software version 3.6.3. Samples were grouped into control and Lyme disease-infected groups. Probes with expression values < 50 were filtered and log2 transformed if not already normalized. The normalize between arrays function from the limma package was used for normalization. Differentially expressed genes (DEGs) were identified using the limma package, applying thresholds of |log2 fold change (logFC)|≥ 1 and adjusted p-value (adj.P) < 0.05 [21]. A volcano plot and hierarchical clustering heatmap were generated to visualize DEG distribution.

Weighted gene coexpression network analysis (WGCNA) and module gene selection

WGCNA was performed on the top 5000 most variable genes from the GSE145974 dataset to identify gene modules associated with Lyme disease. Soft-thresholding power was determined using scale-free topology criteria [22–24]. WGCNA uses biological markers to identify tightly connected gene modules, which can be instrumental in determining candidate genes for studies on diseases [22]. Modules with a minimum of 30 genes and module merging threshold of 0.4 were constructed [25]. The module most significantly correlated with clinical Lyme disease status was subjected to Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses using the cluster Profiler package.

Machine learning approaches for identifying hub genes

Three machine learning algorithms were employed to identify robust diagnostic biomarkers (hub genes): LASSO Regression was implemented using the glmnet package, with tenfold cross-validation to select the optimal penalty parameter (lambda) [26]. Random Forest (RF) classification was conducted using the randomForest package, ranking genes by importance score [27]. Support Vector Machine Recursive Feature Elimination (SVM-RFE) was applied using the e1071 package, with tenfold cross-validation to identify the top-performing features [28]. Genes identified by all three methods were considered hub genes.

Validation of diagnostic performance

To assess the diagnostic utility of the selected hub genes, their expression profiles were validated using the GSE63085 dataset. Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) values were computed using SPSS 26.0 to evaluate diagnostic accuracy.

Gene set enrichment analysis (GSEA)

GSEA was performed using the Broad Institute’s GSEA software. Gene expression matrices from GSE145974 were used to identify enriched pathways at the gene set level. Single-gene GSEA was also conducted for each hub gene to examine their functional roles. Only the top three significantly enriched pathways per gene were reported.

Immune infiltration analysis

Immune cell composition was estimated using the CIBERSORT algorithm. The runCIBERSORT pipeline was applied to the GSE145974 expression matrix, utilizing the LM22 signature matrix. The relative proportions of 22 immune cell types were computed. The Spearman correlation coefficient was used to assess associations between hub gene expression and immune cell abundance [29].

Statistical analysis and data validation

All statistical analyses were conducted using R version 3.6.3 and GraphPad Prism 8.3.0. Welch’s t-test was applied to compare immune cell fractions between groups. A p-value < 0.05 was considered statistically significant.

Results

Overview of study workflow

A comprehensive workflow of the analytical process is summarized (Fig. 1). The schematic outlines the full data processing pipeline, including data acquisition from GEO, DEG identification, WGCNA, machine learning-based biomarker screening, enrichment analyses, and immune cell deconvolution. It illustrates how the GSE145974 dataset was primarily used for discovery analyses, while GSE63085 served as a validation cohort. The integration of multiple analytical strategies enhances the robustness and interpretability of the study.

Fig. 1.

Schematic of the study workflow. ROC: Receiver Operating Characteristic; GSEA: Gene Set Enrichment Analysis; GO: Genetic Ontology; KEGG: Kyoto Encyclopedia of Genes and Genomes; AUC: Area Under the Curve; and FDR: False Discovery Rate

Weighted gene coexpression network analysis

To investigate transcriptomic modules associated with Lyme disease, we conducted WGCNA on the top 5000 most variable genes from the GSE145974 dataset. Three major modules showed significant correlations with disease status: The blue module was positively associated (r = 0.56, p = 1e–4), while the red and black modules showed negative correlations (r = − 0.57 and − 0.56, respectively) (Fig. 2a). These findings highlight discrete gene coexpression patterns linked to LD pathology. The soft-thresholding power used to ensure scale-free topology was selected as 19 (Fig. 2b).

Fig. 2.

Construction of weighted gene coexpression network analysis (WGCNA) module. a Associations between gene modules and LD. Each row corresponds to a module eigengene, and each column corresponds to a clinical status. Each cell contains the correlations and p values between each module and clinical status, b the cluster dendrogram of the top 25% of genes with the highest variance in the GSE63085 dataset. Each color represents a specific gene module, wherein genes within the same module share significant biological functions, c, d, and e gene Ontology (GO)/Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses of the significant modules identified via WGCNA. WGCNA: Weighted Gene Coexpression Network Analysis and LD: Lyme disease

We then performed functional enrichment analysis of each key module. The black module was significantly enriched in GO terms related to “ubiquitin protein ligase binding,” indicating altered proteostasis mechanisms (Fig. 2c). The red module was associated with mitochondrial dysfunction and neurodegeneration pathways, such as “amyotrophic lateral sclerosis” and “respiratory chain” components (Fig. 2d). The blue module, most positively associated with LD, was enriched in processes related to neutrophil activation and degranulation, underscoring a strong innate immune activation signature (Fig. 2e).

Identification of differentially expressed genes and hub genes through machine learning algorithms

To identify specific transcriptional differences between LD and healthy controls, we analyzed DEGs in the GSE145974 dataset. A total of 91 DEGs were identified, with 65 upregulated and 26 downregulated in LD samples (adjusted p < 0.05; |logFC|≥ 1). A heatmap of the top DEGs demonstrates clear separation between LD and control profiles (Fig. 3a). A volcano plot illustrates the global differential expression landscape, with significant genes highlighted (Fig. 3b). The upregulation of immune-related and inflammatory genes suggests that LD is associated with a pronounced pro-inflammatory transcriptional profile. These DEGs may reflect the host immune response to Borrelia infection, particularly the activation of innate immune signaling and stress response pathways. This differential expression signature provides a foundational dataset for the subsequent identification of robust biomarkers.

Fig. 3.

a Heatmaps illustrating the fold change between patients with LD and controls, b a volcano plot representing differentially expressed genes. Candidate biomarkers were identified using machine learning algorithms, c and d The LASSO regression algorithm identified 15 genes as biomarkers with minimal binomial deviation, e the top 10 genes were selected and ranked based on the importance scores obtained using the random forest algorithm, f and g the support vector machine recursive feature elimination (SVM–RFE) algorithm was used for screening candidate genes. The highlighted point indicates minimal error rate and optimal accuracy, with the corresponding genes representing the best signature selected by SVM–RFE, h three hub genes (FCGR1B, MPP1, and HSPA6) were selected. A Venn diagram was used to visualize the intersection between the three machine learning algorithms. Lasso: least absolute shrinkage and selection operator and SVM-RFE: support vector machine recursive feature elimination

To identify core diagnostic markers, we applied three machine learning approaches: LASSO regression, random forest (RF), and SVM–RFE. LASSO selected 15 genes with nonzero coefficients, including FCGR1B and MPP1 (Fig. 3c–d). RF ranked the top genes by importance score, with FCGR1B and MPP1 again ranking highly (Fig. 3e). SVM–RFE identified a 36-gene optimal set (Fig. 3F–G). Intersection analysis using a Venn diagram revealed three overlapping hub genes: FCGR1B, MPP1, and HSPA6 (Fig. 3h).

The consistent identification of these three genes across independent algorithms suggests that they are not only statistically robust but also biologically meaningful. FCGR1B is known to be involved in immunoglobulin binding and monocyte signaling, MPP1 plays a role in membrane–cytoskeletal interactions and immune cell polarity, and HSPA6 is part of the heat shock protein family, implicated in cellular stress responses. Their emergence as central nodes may reflect their involvement in the host’s immune activation and cellular response to bacterial infection in Lyme disease.

Validation of diagnostic performance

To validate the diagnostic utility of the hub genes, we evaluated their performance using ROC curve analysis. In the GSE145974 dataset, all three genes achieved strong classification performance with AUC values exceeding 0.90, indicating high sensitivity and specificity (Fig. 4a). External validation using the GSE63085 dataset confirmed robust diagnostic performance for FCGR1B and MPP1, with AUC values above 0.80 (Fig. 4b). HSPA6 could not be assessed due to missing expression data in GSE63085.

Fig. 4.

Visualization of gene expression and external data verification of biomarkers. a ROC curve analysis of three hub genes in the GSE145974 dataset, b ROC curve analysis of two hub genes in the GSE63085 dataset, c box plot showing the expression levels of FCGR1B, MPP1, and HSPA6. Significance levels are indicated as follows: *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001

In both datasets, box plot analysis demonstrated significantly elevated expression of the three hub genes in LD samples compared to healthy controls (Fig. 4c). These consistent results across independent datasets support the robustness and generalizability of the selected biomarkers.

Gene set enrichment analysis (GSEA)

To gain insight into the biological roles of the hub genes, we performed single-gene GSEA for FCGR1B, HSPA6, and MPP1. Only the top three significantly enriched pathways per gene were reported and categorized into overarching biological themes (e.g., immune signaling, metabolism, and circadian regulation) to facilitate interpretation. FCGR1B was associated with upregulation of pathways such as “Proteasome,” “Collecting Duct Acid Secretion,” and “One Carbon Pool by Folate” (Fig. 5a). These pathways are involved in protein degradation, renal ion transport, and folate metabolism, suggesting that FCGR1B may play a role in metabolic–immune crosstalk. HSPA6 was enriched in neutrophil extracellular trap formation and 2-oxocarboxylic acid metabolism while being negatively associated with circadian rhythm regulation and taste transduction (Fig. 5b). MPP1 was positively associated with antifolate resistance and glycan degradation, and negatively associated with circadian signaling and hedgehog pathways (Fig. 5c). These findings imply that circadian dysregulation and altered metabolic responses are potential features of the LD transcriptome. Global GSEA of all DEGs revealed that immune-related and metabolic processes, including “proteasome,” “aldosterone-regulated sodium reabsorption,” and “circadian rhythm,” were significantly enriched, supporting the notion of systemic metabolic and inflammatory activation in LD (Fig. 5d).

Fig. 5.

Functional enrichment analysis. a The main upregulated and downregulated signaling pathways significantly enriched in FCGR1B, b the main upregulated and downregulated signaling pathways significantly enriched in HSPA6, c the main upregulated and downregulated signaling pathways significantly enriched in MPP1, d the key signaling pathways of DEGs. Abbreviations: DEGs: differentially expressed genes

Immune cell infiltration profiling

To assess the immune microenvironment changes in Lyme disease, we performed immune cell deconvolution analysis using CIBERSORT on the GSE145974 dataset. The relative proportions of 22 immune cell types were quantified across all samples. A correlation heatmap revealed distinct relationships among immune cell populations in LD (Fig. 6a).

Fig. 6.

Immune infiltration analysis. a Correlation heatmap of the immune microenvironment in patients with LD, b bar plot comparing the expression of immune cells between patients with LD and controls, c box plot comparing the expression of immune cells between patients with LD and controls, d correlation between infiltrating immune cells and FCGR1B/MPP1 expression. Significance levels are indicated as follows: *p < 0.05; **p < 0.01; ***p < 0.001; and ****p < 0.0001

Monocytes and M0 macrophages were significantly elevated in LD samples compared to controls, while resting CD4 memory T cells were markedly decreased (Fig. 6b–c). These changes suggest an enhanced innate immune state and a possible suppression or redistribution of adaptive T-cell populations. Such an immune profile is consistent with early host responses to bacterial infection.

Moreover, correlation analyses demonstrated that expression levels of FCGR1B and MPP1 were positively correlated with the abundance of monocytes (Fig. 6d). This further supports the functional role of these genes in shaping the innate immune response to Borrelia infection. Collectively, these findings suggest that Lyme disease is characterized by a restructured immune landscape, featuring monocyte/macrophage activation and adaptive immune modulation.

Discussion

In this study, we employed a multi-dimensional bioinformatics framework to identify key biomarkers and immune signatures associated with Lyme disease (LD). By integrating differential gene expression analysis, WGCNA, machine learning, and immune infiltration analysis, we identified FCGR1B, MPP1, and HSPA6 as potential diagnostic markers with biologically plausible functions in the LD immune response.

Our results support and extend existing study on LD-related immune activation. The finding that the blue WGCNA module—positively associated with LD—is enriched in neutrophil activation and degranulation pathways (Fig. 2E) aligns with reports that innate immune responses, particularly involving neutrophils and monocytes, are critical in the early defense against Borrelia burgdorferi infection [30, 31]. Prior studies have shown that TLR1/2 signaling and Fc gamma receptor (FcγR) pathways are engaged upon pathogen exposure [32, 33], providing a mechanistic context for our observation that FCGR1B is upregulated and correlated with monocyte infiltration (Fig. 6d). FCGR1B, although lacking an intracellular signaling domain, may still influence immune cell engagement and serve as a surrogate marker of monocyte activation.

Our machine learning models converged on three hub genes—FCGR1B, MPP1, and HSPA6—selected consistently by LASSO, RF, and SVM–RFE (Fig. 3h). The previous studies identified a 31-gene classifier for early LD with high accuracy [20]; our findings demonstrate that comparable diagnostic performance (AUC > 0.90 in GSE145974) can be achieved using only three genes (Fig. 4a), thus offering a more parsimonious biomarker panel. Importantly, two of the three hub genes (FCGR1B and MPP1) were validated in an independent dataset (Fig. 4b), strengthening their potential translational utility.

Notably, HSPA6 could not be validated in the GSE63085 dataset due to missing expression data. However, it was consistently identified by all three machine learning algorithms and enriched in pathways related to immune activation and circadian rhythm disruption, suggesting biological relevance. Future validation using other independent datasets with full gene coverage, as well as functional experiments, will be necessary to confirm the diagnostic potential of HSPA6. It is part of the HSP70 family, known to be involved in cellular stress and inflammation [34, 35], and its downregulation was associated with circadian rhythm disruption (Fig. 5b), a phenomenon increasingly linked to chronic infection and inflammatory disease [36]. The association of all three hub genes with proteasome function and immune–metabolic pathways (Fig. 5a–c) suggests broader involvement in cellular homeostasis and stress adaptation.

Moreover, our immune cell deconvolution findings corroborate recent transcriptome-based LD studies showing that LD alters monocyte populations and suppresses adaptive immunity [37, 38]. The observed increase in monocytes and M0 macrophages, along with reduced resting CD4 memory T cells (Fig. 6C), reflects a shift toward a myeloid-dominated immune profile, possibly at the expense of T-cell memory generation.

While this study provides a comprehensive transcriptomic profile of LD, several limitations remain. Although two independent datasets, GSE145974 and GSE63085, were used for biomarker identification and validation, both cohorts are of moderate sample size. The consistency of results across datasets supports the reliability of the identified biomarkers; however, their generalizability still requires confirmation in larger-scale studies. Future research should include broader clinical populations, particularly from diverse geographic regions and clinical backgrounds, to ensure the wide applicability of these biomarkers. Second, although hub gene expression correlates with immune infiltration, functional validation in vitro and in vivo will be necessary to establish causality. Finally, the absence of precise clinical metadata (e.g., time since infection and symptom severity) in the datasets limits stratified analysis.

In conclusion, our integrative approach identifies FCGR1B, MPP1, and HSPA6 as potential diagnostic biomarkers for Lyme disease and highlights innate immune activation, circadian disruption, and proteasome activity as key features of its molecular pathology. These findings offer new insights into LD diagnosis and pathogenesis, with implications for early detection and targeted intervention.

Conclusions

Three key biomarkers—FCGR1B, MPP1, and HSPA6—were identified and validated as potential biomarkers associated with LD based on transcriptomic data analysis using advanced bioinformatics and machine learning techniques. Our multi-algorithm approach and comprehensive analyses provide robust and novel insights, improving the diagnostic precision and providing potential targets for therapeutic intervention. These findings represent a significant advancement in LD diagnostics, paving the way for developing more effective therapeutic strategies.

Author’s contributions

Q.W.L.: Conceptualization, Formal analysis, Writing—original draft; Y.Z.W.: Writing—original draft, Writing—review & editing; Z.S. W.: Formal analysis, Writing—original draft; X.L.H.: Writing—original draft; Y.L.H.: Writing—review & editing; X.J.C.: Writing—original draft; Y.F.L.: Writing—review & editing; X.G.G.: Conceptualization; J.J.W.: Conceptualization.

Funding

This work was supported by Guangdong Medical Research Fund project [grant number: A2022340]; 2023 Guangzhou Higher Education Teaching Quality and Teaching Reform Engineering First-Class Course Project [grant number: 2023YLKC024]; and First-Class Undergraduate Major Construction Funding Project of High-Level University [grant number: 02–408-2304-02085XM].

Data availability

The datasets used and analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Conflict of interest

The authors declare no competing interests.

Ethics approval

Not applicable.

Consent to participate

Not applicable.

Consent to publish

Not applicable.