Application of Machine Learning to Metabolomic Profile Characterization in Glioblastoma Patients Undergoing Concurrent Chemoradiation

Metabolites. 2023 Feb 17;13(2):299. doi: 10.3390/metabo13020299.

Abstract

We here characterize changes in metabolite patterns in glioblastoma patients undergoing surgery and concurrent chemoradiation using machine learning (ML) algorithms to characterize metabolic changes during different stages of the treatment protocol. We examined 105 plasma specimens (before surgery, 2 days after surgical resection, before starting concurrent chemoradiation, and immediately after chemoradiation) from 36 patients with isocitrate dehydrogenase (IDH) wildtype glioblastoma. Untargeted GC-TOF mass spectrometry-based metabolomics was used given its superiority in identifying and quantitating small metabolites; this yielded 157 structurally identified metabolites. Using Multinomial Logistic Regression (MLR) and GradientBoostingClassifier (GB Classifier), ML models classified specimens based on metabolic changes. The classification performance of these models was evaluated using performance metrics and area under the curve (AUC) scores. Comparing post-radiation to pre-radiation showed increased levels of 15 metabolites: glycine, serine, threonine, oxoproline, 6-deoxyglucose, gluconic acid, glycerol-alpha-phosphate, ethanolamine, propyleneglycol, triethanolamine, xylitol, succinic acid, arachidonic acid, linoleic acid, and fumaric acid. After chemoradiation, a significant decrease was detected in 3-aminopiperidine 2,6-dione. An MLR classification of the treatment phases was performed with 78% accuracy and 75% precision (AUC = 0.89). The alternative GB Classifier algorithm achieved 75% accuracy and 77% precision (AUC = 0.91). Finally, we investigated specific patterns for metabolite changes in highly correlated metabolites. We identified metabolites with characteristic changing patterns between pre-surgery and post-surgery and post-radiation samples. To the best of our knowledge, this is the first study to describe blood metabolic signatures using ML algorithms during different treatment phases in patients with glioblastoma. A larger study is needed to validate the results and the potential application of this algorithm for the characterization of treatment responses.

Keywords: concurrent chemoradiation; glioblastoma; machine learning; metabolomic profiling; treatment response.