Metabolomic Profiling and Machine Learning Models for Tumor Classification in Patients with Recurrent IDH-Wild-Type Glioblastoma: A Prospective Study

Abstract

Background/objectives: The recurrence of glioblastoma is an inevitable event in this disease's course. In this study, we sought to identify the metabolomic signature in patients with recurrent glioblastomas undergoing surgery and radiation therapy.

Methods: Blood samples collected prospectively from six patients with recurrent IDH-wildtype glioblastoma who underwent one surgery at diagnosis and a second surgery at relapse were analyzed using untargeted gas chromatography-time-of-flight mass spectrometry to measure metabolite abundance. The data analysis techniques included univariate analysis, correlation analysis, and a sample t-test. For predictive modeling, machine learning (ML) algorithms such as multinomial logistic regression, gradient boosting, and random forest were applied to predict the classification of samples in the correct treatment phase.

Results: Comparing samples after the first surgery and after the relapse surgeries to the pre-operative samples showed a significant decrease in sorbitol and mannitol; there was a significant increase in urea, oxoproline, glucose, and alanine. After chemoradiation, two metabolites, erythritol and 6-deoxyglucitol, showed a decrease, with a cut-off of three and a significant reduction for 6-deoxyglucitol, while 2,4-difluorotoluene and 9-myristoleate showed an increase post radiation, with a fold-change cut-off of three. The gradient-boosting ML model achieved a high performance for the prediction of tumor conditions in patients with glioblastoma who had undergone relapse surgery.

Conclusions: We developed an ML predictor for tumor phase based on the plasma metabolomic profile. Our study suggests the potential of combining metabolomics with ML as a new tool to stratify the risk of tumor progression in patients with glioblastoma.

Keywords: glioblastoma; machine learning; metabolomics; recurrence.

Figure 1

GC-TOF MS intensities of untargeted plasma metabolomics for pre-surgical and post-surgical samples. (A) Levels of decreased metabolites post surgery (PostS) compared to pre-surgery values (PreS) with a cut-off fold-change of 3. (B) Metabolites with a significant decrease post-surgery (p < 0.05). (C) Comparison of increased metabolite levels pre surgery vs. post surgery, with a cut-off fold-change of 3. (D) Metabolites with a significant increase post surgery (p < 0.05). Statistical significance was determined using an unpaired Student’s t-test, where (****) denotes a p-value < 0.0001, (***) p < 0.001, and (**) a p-value < 0.01.

Figure 2

MS intensities of plasma metabolites for pre-radiation and post-radiation samples. (A) Levels of decreased metabolites post radiation (PostRad) compared to pre radiation (PreRad), with a cut-off fold-change of 3. Statistical significance was determined using an unpaired Student’s t-test. Metabolites with a significant decrease post radiation (** p < 0.01). (B) Increased metabolites post radiation with a cut-off fold-change of 3. “ns” Not significant.

Figure 3

Heatmap of Pearson’s correlation coefficients for altered plasma metabolites with a cut-off of r > 0.90. Altered metabolites with high correlations are highlighted in black boxes. The correlation score can be tracked through the scale bar on the right side of the heatmap. Positive correlations are present between several metabolites: between sorbitol and mannitol, indoxyl sulfate and threonic acid, and gluconic acid and gluconic acid lactone.

Figure 4

Metabolomics with machine learning models for the classification of clinical stages in patients with recurrent glioblastoma undergoing repeat surgery. (A) Comparing the performance of metabolomic-based machine learning algorithms based on accuracy, precision, recall, and F1-score. (B) The learning curve on test samples as a function of the training samples. (C) ROC-AUC curve to assess the performance of the three models. (D–F) Confusion matrix for each of the three models when tested on the test dataset consisting of 12 samples. The color scales (0–5) next to each confusion matrix represent classification accuracies. The actual/prediction labels are mapped as follows: “0” for pre surgery, “1” for post surgery, “2” for pre radiation, and “3” for post radiation.