Weighted Gene Co-Expression Network Analysis and Support Vector Machine Learning in the Proteomic Profiling of Cerebrospinal Fluid from Extraventricular Drainage in Child Medulloblastoma

Abstract

Medulloblastoma (MB) is the most common pediatric malignant central nervous system tumor. Overall survival in MB depends on treatment tuning. There is aneed for biomarkers of residual disease and recurrence. We analyzed the proteome of waste cerebrospinal fluid (CSF) from extraventricular drainage (EVD) from six children bearing various subtypes of MB and six controls needing EVD insertion for unrelated causes. Samples included total CSF, microvesicles, exosomes, and proteins captured by combinatorial peptide ligand library (CPLL). Liquid chromatography-coupled tandem mass spectrometry proteomics identified 3560 proteins in CSF from control and MB patients, 2412 (67.7%) of which were overlapping, and 346 (9.7%) and 805 (22.6%) were exclusive. Multidimensional scaling analysis discriminated samples. The weighted gene co-expression network analysis (WGCNA) identified those modules functionally associated with the samples. A ranked core of 192 proteins allowed distinguishing between control and MB samples. Machine learning highlighted long-chain fatty acid transport protein 4 (SLC27A4) and laminin B-type (LMNB1) as proteins that maximized the discrimination between control and MB samples. Machine learning WGCNA and support vector machine learning were able to distinguish between MB versus non-tumor/hemorrhagic controls. The two potential protein biomarkers for the discrimination between control and MB may guide therapy and predict recurrences, improving the MB patients' quality of life.

Keywords: artificial intelligence; brain tumor; cerebral spinal fluid; extraventricular drainage; mass spectrometry; medulloblastoma; proteomics.

Figure 1

Workflow of the analysis. Cerebrospinal fluid from extraventricular drainage samples were collected, fractionated, and analyzed by mass spectrometry. The whole dataset was analyzed by the combined use of statistical and bioinformatic analyses to identify new potential biomarkers of medulloblastoma. The results of these analyses were validated by enzyme-linked immunosorbent assay (ELISA).

Figure 2

Multidimensional scaling (MDS) analysis of the proteins identified in CSF from EVD in MB or control samples. The two-dimensional scatter plot of MDS analysis shows the unsupervised cluster analysis of all CSF proteins identified from the EVD of MB (open symbol) and controls (solid symbol) in total (black square), CPLL (red circle), Mv (green triangle), and Ex (blue triangle) fractions. Ellipses show the 95% CI of the two clusters. MDS analysis identified three clusters corresponding to soluble proteins (total and CPLL fraction) and the extracellular vesicle fractions distinct in Mv and Ex fractions. No outliers were detected.

Figure 3

Heatmap of proteins that maximize the discrimination between MB and control samples. Heatmap of 192 proteins highlighted by the combined use of a t-test, Partial Least Square Discriminant Analysis (PLS-DA), and Support Vector Machine (SVM) learning analysis. In the heatmap, each row represents a protein, and each column corresponds to a sample (total, CPLL, Mv, and Ex of CSF from the EVD of MB and control patients). Normalized Z-scores of protein abundance are depicted by a pseudocolor scale (with red indicating positive expression; white, equal expression; and blue, negative expression, compared to each protein value). The dendrogram displays the outcome of unsupervised hierarchical clustering, placing similar proteome profile values next to each other. The co-expression modules identified are reported on the left of the heatmap. The statistically significant proteins identified in each comparison are highlighted in red on the right of the heatmap.

Figure 4

Partial Least Square Discriminant Analysis (PLS-DA) of all proteins identified in CSF from the EVD of MB and control samples. Two-dimensional scatter plot of PLS-DA showing the supervised cluster analysis of all proteins identified in CSF from the EVD of MB (open symbols) and control (solid symbols) in the total (black squares), CPLL (red circles), Mv, and Ex of the two clusters. PLS-DA identified eight clusters corresponding to the two clinical groups in each fraction.

Figure 5

LMNB1 evaluation of the total fraction of CSF from the EVD of MB or control patients. (a) Representative western blot analysis and (b) gel stained with Blue Silver according to Candiano G. et al. [33] (used as a loading control) of full length gel (12–16T%) for LMNB1 in the total fraction of CSF from the EVD of control and Medulloblastoma (MB); (c) Box plot showing the median and interquartile range value of the ELISA of the LMNB1 protein in the total fraction of CSF from the EVD of the control, MB, low-grade gliomas and glioneural tumors (LGGs and GN), and nine other brain tumors patients. LMNB1 proteins were more abundant in MB compared to all other patients (p < 0.0001); (d) ROC curve analysis for the LMNB1 assay.