Plasma immune profiling combined with machine learning contributes to diagnosis and prognosis of active pulmonary tuberculosis

Abstract

Tuberculosis (TB) remains one of the deadliest chronic infectious diseases globally. Early diagnosis not only prevents the spread of TB but also ensures effective treatment. However, the absence of non-sputum-based diagnostic tests often leads to delayed TB diagnoses. Inflammation is a hallmark of TB, we aimed to identify biomarkers associated with TB based on immune profiling. We collected 222 plasma samples from healthy controls (HCs), disease controls (non-TB pneumonia; PN), patients with TB (TB), and cured TB cases (RxTB). A high-throughput protein detection technology, multiplex proximity extension assays (PEA), was applied to measure the levels of 92 immune proteins. Based on differential analysis and the correlation with TB severity, we selected 9 biomarkers (CXCL9, PDL1, CDCP1, CCL28, CCL23, CCL19, MMP1, IFNγ and TRANCE) and explored their diagnostic capabilities through 7 machine learning methods. We identified combination of these 9 biomarkers that distinguish TB cases from controls with an area under the receiver operating characteristic curve (AUROC) of 0.89-0.99, with a sensitivity of 82-93% at a specificity of 88-92%. Moreover, the model excels in distinguishing severe TB cases, achieving AUROC exceeding 0.95, sensitivities and specificities exceeding 93.3%. In summary, utilizing targeted proteomics and machine learning, we identified a 9 plasma proteins signature that demonstrates significant potential for accurate TB diagnosis and clinical outcome prediction.

Keywords: PEA; Tuberculosis; biomarker; diagnosis; machine learning.

Figure 1.

Comparison of proinflammatory immune signals among the HC, PN, and TB plasma samples. (A) Principal component analysis (PCA) of proinflammatory immune signals among HC, PN, and TB plasma samples. Each plot represents one sample. (B) Venn diagram illustrating the differentially expressed inflammation-related proteins for each pairwise comparison, both specific and commonly shared proteins. (C) Violin plot displaying the common differentially expressed inflammation-related proteins in (B). *p < 0.05, **p < 0.01, ***p < 0.001. (D) Volcano plot depicting the differentially expressed proteins for each pairwise comparison. (E) Protein–protein interaction (PPI) network analysis of the differentially expressed inflammation-related proteins for each comparison.

Figure 2.

Functional annotations of differentially expressed inflammation-related proteins. (A and B) Gene Ontology (GO) (A) and Kyoto Encyclopedia of Genes and Genomes (KEGG) (B) enrichment analysis of the differentially expressed inflammation-related proteins for each comparison based on the background of all annotated proteins. (C) Schematic diagram of the reconstructed interaction network of Toll-like receptor signalling, NF-κB signalling, JAK-STAT signalling, and TB signalling, along with their functionally associated differentially expressed immune proteins between HC and TB samples. (D) Schematic diagram of the reconstructed interaction network of Toll-like receptor signalling, IL-17 signalling, TNF signalling, and lipid and atherosclerosis signalling, along with their functionally associated differentially expressed immune proteins between HC and PN samples. (E) Scatter plot showing the IL-17 signalling-related differentially expressed immune proteins between PN and TB samples. *p < 0.05, **p < 0.01. (F) Gene set enrichment analysis (GSEA) demonstrating the enrichment of four pathways between HC and TB cohorts using published RNA-seq datasets.

Figure 3.

Immune signatures associated with TB treatment. (A) Expression pattern and functional analysis of inflammation-related proteins among HC, TB, and RxTB cohorts. (B) Principal component analysis (PCA) displaying proinflammatory immune signals among HC, TB, and RxTB plasma samples. Each plot represents one sample. (C) Venn diagram illustrating the differentially expressed inflammation-related proteins for each comparison, both specific and commonly shared proteins. (D) Dot plot depicting the differentially expressed inflammation-related proteins derived from RxTB vs. HC and RxTB vs. TB pairwise comparisons. (E) Violin plot displaying the common differentially expressed inflammation-related proteins in (C). *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. (F) PPI network analysis of the differentially expressed inflammation-related proteins between RxTB individuals and patients with TB.

Figure 4.

Diagnostic capacity of the identified 9 potential biomarkers for TB based on machine learning. (A) Pie chart showing the correlations (r) between HRCT score and inflammation proteins. (B) Heatmap displaying inflammatory proteins in (A) that are associated with HRCT score with a |correlation|≥ 0.3. (C) Venn diagram representing the intersection of differentially expressed proteins with proteins in (A) that are associated with HRCT with a |correlation|≥ 0.3. (D) Heatmap indicating the expression profiles of the 9 potential diagnostic biomarkers in the HC, TB, and PN cohorts. (E and F) Performance of various machine learning algorithms for pairwise prediction in the train (left) and test (right) datasets for TB vs. HC (E) and TB vs. PN (F). (G) The specificity of SVM model in distinguishing TB from HC under a fixed sensitivity level. (H) The sensitivity of SVM model in distinguishing TB from PN under a fixed specificity level.

Figure 5.

Diagnostic capacity of the identified 9 potential biomarkers for TB clinical outcomes. (A) Table showing the sample size and the range of HRCT score based on TB development. (B) Receiver operating characteristic (ROC) curves demonstrating the performance of combination of 9 biomarkers for pairwise prediction of mild, moderate, and severe TB patients. (C) Scatter plot showing the expression levels of 9 biomarkers among mild, moderate and severe TB patients. *p < 0.05, **p < 0.01, ***p < 0.001. (D) Bar graph illustrating the gender distribution in mild, moderate, and severe TB patients. (E–G) Bar chart representing the expression levels of 9 biomarkers in male and female samples across mild (E), moderate (F), and severe (G) TB patients.