We extracted HE-stained WSI features of GBM and used machine learning algorithms to predict MMP9 expression levels in patients. This paper presents the first report of such findings on GBM based on available information. These findings indicate that MMP9 expression is an independent prognostic factor for GBM patients (HR = 1.917, 95% CI: 1.211–3.033). Among all extracted features, we constructed a classifier of the most informative six features that could clearly classify GBM patients with high MMP expression from those with low MMP9 expression (AUC = 0.754 in the test set). Furthermore, in GSEA, the PS obtained from the SVM model showed significant correlations with angiogenesis, apoptosis, and tumor-related pathways, including JAK-STAT, P53, and MAPK signaling pathways. Notably, PS was significantly associated with OS of GBM patients (p= 0.002).
Previous studies have demonstrated that overexpression of MMP9 correlates with poor prognosis and treatment outcomes in various malignant tumors, including colorectal cancer.19kidney20breast cancertwenty oneAn association between high MMP9 expression and poor prognosis has also been demonstrated in patients with glioma.twenty twoOn the other hand, patients with low MMP9 expression experienced an increased OS of 5.2 months when treated with bevacizumab.twenty threeThis is consistent with the results of our study, which demonstrated that patients with high MMP9 expression levels had a poor prognosis as evidenced by Kaplan-Meier curve analysis. Furthermore, multivariate analysis revealed that MMP9 was an independent risk factor for OS in GBM patients.
GBM exhibits significant heterogeneity, and prognostic insights gained from key molecular marker expression would be beneficial for clinical decision making. Pathological examination plays a key role in the diagnosis and staging of GBM patients, but its accuracy can depend on the pathologist's expertise. Obtaining sufficient information from subjective evaluation of tissue sections poses both challenges and opportunities. Recently, computer systems and medical image processing algorithms have been developed to aid in the extraction of image features associated with tumor characteristics and survival outcomes.24,25Automated methods also offer the benefits of increased efficiency and reduced labor.26.
Changes in tissue architecture and nuclear morphology often reflect changes in molecular expression within tumors and can be predicted by deep learning algorithms from breast pathology images.27,prostate28gastrointestinal cancer26In lung cancer, specific genetic alterations, including STK11, EGFR, FAT1, SETBP1, KRAS, and TP53, have been predicted with an AUC range of 0.733–0.856 and have been externally validated.29Furthermore, Hollon et al.30 We use stimulated Raman histology and genomic data to predict molecular alterations (IDH mutation, 1p19q codeletion, and ATRX mutation) in diffuse gliomas. Here, we showed that pathological image features showed good ability to classify high and low MMP9 expression in GBM patients. The two pathomics models showed strong predictive performance in both training and validation subsets, with AUC values of 0.778 and 0.754, respectively, for the SVM model and 0.776 and 0.722, respectively, for the LR model. It was speculated that pathomics was effective in predicting MMP9 expression.
We conducted additional investigations to pair WSI with clinical data to explore the relationship between PS and OS in pathology cohorts.The integration of pathology images and deep learning has been used to predict prognosis in several types of cancer.31,32,33,34demonstrated a level of accuracy in predicting outcomes for patients diagnosed with glioma that exceeds existing clinical frameworks.35Another glioma study classified patient survival into four classes based on pathology images.36In this study, high PS was significantly associated with poor prognosis of GBM, which supported previous studies. Taken together, our pathogenesis model demonstrated considerable potential for risk stratification based on OS.
Correlations between pathology, imaging, molecular and genetic data have important value in precision oncology. Given the high heterogeneity of tumors, quantitative depictions of pathological and imaging phenotypes can more effectively compensate for microscopic molecular or genetic defects. Correlations between pathological imaging information and patient gene expression have been observed in several studies.37,38In this study, functional annotation was performed using GSEA to reveal the underlying biological processes associated with PS. In the KEGG and Hallmark gene sets, a significant increase in apoptosis, angiogenesis, and signaling pathways involved in cancer development, including P53, JAK-STAT, and MAPK signaling pathways, was observed. There is strong evidence in the literature that these signaling pathways and phenotypes are intricately linked to the initiation and progression of GBM. The tumor P53 pathway is a classic therapeutic target with a complex regulatory network, the deregulation of which mediates GBM cell invasion, migration, and proliferation.39The JAK-STAT pathway is a key oncogenic hub in the GBM microenvironment that includes reactive astrocytes, gliomas and immune cells, promoting tumor growth and therapy-resistant invasion.40The MAPK signaling pathway responds to various stresses and plays a key role in multiple cellular processes, including apoptosis, and alterations in MAPK signaling promote the malignant phenotype.41This finding may provide valuable knowledge about the molecular mechanisms underlying the morphological characteristics of GBM.
This study has several limitations. First, the study design was retrospective, which inherently introduced potential confounding factors. In addition, this study was conducted at a single center with a relatively small sample size, and further validation by large multicenter studies was needed. Third, the patients included in this study were obtained from a public dataset, which may have led to heterogeneity in the data. In future investigations, we plan to collect data for validation purposes.
In conclusion, MMP9 expression was significantly associated with the prognosis of GBM patients. The prediction model based on H&E stained WSI features showed good stability and diagnostic efficiency, making it a promising tool for predicting MMP9 expression in GBM patients. With the advent of the big data era, pathology integrating genomics and proteomics has emerged as a new research direction to meet the demands of precision medicine.
