Integrating computational pathology and multi-transcriptomics to characterize glioblastoma heterogeneity and identify prognostic biomarkers.

BACKGROUND

Glioblastoma (GBM) remains a lethal brain cancer with median survival of 12-15 months, hindered by pronounced heterogeneity. Integrating histopathological, molecular, and microenvironmental data is critical for improving outcomes.

METHODS

Pathological features from TCGA-GBM whole-slide images were combined with bulk RNA-seq (pan-apoptosis genes), scRNA-seq (monocytes), and spatial transcriptomics. A machine learning model (Lasso + plsRcox) was developed using 60 % training and 40 % validation data.

RESULTS

The prognostic model achieved C-indices of 0.75 (training) and 0.616 (validation). High-risk patients had shorter median survival (14 vs. 28 months; HR = 3.87, P = 0.001). BCL2A1 was identified as a key risk gene (r = 0.42 with risk score) and correlated with poor survival (11 vs. 26 months, P = 0.004) and abnormal nuclear morphology. scRNA-seq revealed BCL2A1+ monocytes (12.3 % of cells) with high stemness and enriched angiogenesis pathways. Spatial analysis showed these monocytes localize at the invasive front (21.4 %) near endothelial cells, promoting VEGF signaling.

CONCLUSION

Integration of computational pathology and multi-transcriptomic data identified BCL2A1+ monocytes as drivers of angiogenesis and progression. This approach offers a prognostic tool and potential therapeutic targets for GBM.