Study on Metabolic Differences between IDH Mutant and Wild-type Glioma Cells Using Raman Spectroscopy and Machine Learning Platform

Background Introduction

In the diagnosis and treatment of gliomas, formalin-fixed, paraffin-embedded (FFPE) tissue sections are commonly used. However, due to background noise interference from the embedding medium, the application of FFPE tissues in Raman spectroscopy-based studies is limited. To overcome this issue and identify tumor subtypes, this study developed a novel Raman spectroscopy-based machine learning platform named APOLLO (Raman Spectroscopy Pathology of Malignant Gliomas), which can predict glioma subtypes from FFPE tissue sections.

Paper Source

This paper was written by scholars Adrian Lita, Joel Sjöberg, David Păcioianu, among others, from the National Cancer Institute (USA), University of Turku (Finland), University of Bucharest (Romania), and Henry Ford Health System. The paper was published in the June 2024 issue of the journal “Neuro-Oncology.”

Research Process

Research Subjects and Samples

The study obtained FFPE samples from 46 patients. These samples had known methylation subtypes and were confirmed to contain tumor cells using H&E staining. Subsequently, molecular fingerprinting was performed on these samples using spontaneous Raman spectroscopy, and classifiers for tumor/non-tumor, IDH1 mutant/wild-type, and methylation subtypes were constructed using Support Vector Machine and Random Forest algorithms.

Data Collection and Preprocessing

Raman spectral data for each sample were collected using ThermoFisher DXR2xi Raman Microscope and Leica Stellaris 8 CRS Microscope. During data preprocessing, silent regions were removed, and baseline correction and normalization were performed to ensure the reliability of the Raman spectral data.

Algorithm and Model Training

The DBSCAN clustering algorithm was used to identify tumor and non-tumor tissues, and machine learning clustering models were trained for each scanned area. Multiple classifiers, including Random Forest models and Support Vector Machines, were trained using 5-fold cross-validation to optimize the accuracy, precision, and recall of the models. Finally, the most important Raman frequencies required to distinguish between tumor/non-tumor and IDH mutant/wild-type were extracted through Random Forest feature ranking and statistical tests.

Experimental Results Verification

Stimulated Raman Scattering (SRS) was used to verify the Raman spectral frequencies. The results showed that the APOLLO platform could efficiently distinguish between tumor and non-tumor tissues and between IDH1 mutant and wild-type tumors. Specifically, the APOLLO platform identified a significant increase in cholesterol ester levels in IDH1 mutant gliomas, indicating metabolic specificity in IDH mutant tumors.

Research Results

Distinction between Tumor and Non-Tumor Tissues

When ranking the importance of Raman frequencies using ANOVA, chi2, and Random Forest models, it was found that the Raman frequency 2850 cm^-1 (rich in lipid CH2 bonds) is of extreme importance in distinguishing between tumor and non-tumor tissues. Other newly identified Raman frequencies such as 2883 cm^-1, 1690 cm^-1, 1607 cm^-1, 1401 cm^-1, and 1335 cm^-1 also showed high intensity in tumor tissues.

Distinction between IDH1 Mutant and Wild-type Gliomas

By training multiple classifier models, APOLLO accurately distinguished between IDH1 mutant and wild-type gliomas, with an average ROC AUC of 0.82. Notably, cholesterol ester signals at Raman spectral frequencies 2883 cm^-1, 1440 cm^-1, and 532 cm^-1 were crucial in distinguishing between IDH1 mutant tumors.

More Detailed Subtype Distinction

APOLLO effectively differentiated between high CpG island methylator phenotype (g-CIMP-high) and low methylation phenotype (g-CIMP-low) gliomas within the IDH1 mutant class. This classification is clinically significant as the g-CIMP-high subtype is associated with better prognosis.

Significance of the Study

The APOLLO platform demonstrates the potential of label-free Raman spectroscopy in extracting meaningful biological information from FFPE sections, opening up prospects for the application of FFPE samples in other cancer research. This study not only shows the effectiveness of machine learning-based automated classification methods but also identifies specific metabolic differences in cholesterol metabolism in IDH1 mutant gliomas, providing new insights for drug targeting.

Highlight Introduction

1. Innovative Platform: APOLLO is a novel Raman spectroscopy and machine learning platform that can accurately predict glioma subtypes from FFPE tissue sections.
2. Discovery of High Cholesterol Metabolism: The study found significantly increased levels of cholesterol esters in IDH1 mutant gliomas, offering new insights into tumor biology.
3. Fully Automated Workflow: The APOLLO platform achieves fully automated data processing and classification without human intervention, greatly improving analysis efficiency.

This study provides important scientific basis and clinical application value for the molecular classification and treatment of gliomas. In the future, the APOLLO platform could also be extended to other types of FFPE tissues and cancer research, promoting the development of oncology and molecular pathology.