Rapid visualization of PD-L1 expression level in glioblastoma immune microenvironment via machine learning cascade-based Raman histopathology

Abstract

Introduction: Combination immunotherapy holds promise for improving survival in responsive glioblastoma (GBM) patients. Programmed death-ligand 1 (PD-L1) expression in immune microenvironment (IME) is the most important predictive biomarker for immunotherapy. Due to the heterogeneous distribution of PD-L1, post-operative histopathology fails to accurately capture its expression in residual tumors, making intra-operative diagnosis crucial for GBM treatment strategies. However, the current methods for evaluating the expression of PD-L1 are still time-consuming. Objective: To overcome the PD-L1 heterogeneity and enable rapid, accurate, and label-free imaging of PD-L1 expression level in GBM IME at the tissue level. Methods: We proposed a novel intra-operative diagnostic method, Machine Learning Cascade (MLC)-based Raman histopathology, which uses a coordinate localization system (CLS), hierarchical clustering analysis (HCA), support vector machine (SVM), and similarity analysis (SA). This method enables visualization of PD-L1 expression in glioma cells, CD8+ T cells, macrophages, and normal cells in addition to the tumor/normal boundary. The study quantified PD-L1 expression levels using the tumor proportion, combined positive, and cellular composition scores (TPS, CPS, and CCS, respectively) based on Raman data. Furthermore, the association between Raman spectral features and biomolecules was examined biochemically. Results: The entire process from signal collection to visualization could be completed within 30 min. In an orthotopic glioma mouse model, the MLC-based Raman histopathology demonstrated a high average accuracy (0.990) for identifying different cells and exhibited strong concordance with multiplex immunofluorescence (84.31 %) and traditional pathologists' scoring (R2 ≥ 0.9). Moreover, the peak intensities at 837 and 874 cm−1 showed a positive linear correlation with PD-L1 expression level. Conclusions: This study introduced a new and extendable diagnostic method to achieve rapid and accurate visualization of PD-L1 expression in GBM IMB at the tissular level, leading to great potential in GBM intraoperative diagnosis for guiding surgery and post-operative immunotherapy.

Keywords

• Glioblastoma
• PD-L1
• Immune microenvironment
• Raman spectroscopy
• Machine learning