Pathology-genomic fusion via biologically informed cross-modality graph learning for survival analysis

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Cancer survival prediction is complex due to tumor heterogeneity.
• Integrating histopathology images and genomics shows promise but faces challenges in fusion, interpretability, and gene detection variability.

Purpose of the Study:

• To develop SurvPAGE, a model for cancer survival prediction using fused pathology images and genomic data.
• To address challenges in multi-modal data fusion and enhance cross-modal interpretability.

Main Methods:

• SurvPAGE utilizes a biologically informed heterogeneous graph learning framework.
• It constructs modality-specific graphs for histology and genomics, linked by genotype-phenotype edges.
• A masked graph autoencoder (genGraphMAE) enhances robustness against gene detection variability.

Main Results:

• SurvPAGE achieved state-of-the-art performance in survival prediction C-index on glioma and renal carcinoma datasets.
• The model identified prognostic histology regions and potential driver genes.
• Interpretable methods provided multi-scale insights into prognostic factors.

Conclusions:

• SurvPAGE effectively integrates pathology and genomic data for improved cancer prognosis.
• Biologically informed graph learning offers a robust and interpretable framework for multi-modal cancer data analysis.
• This approach advances personalized cancer treatment strategies through enhanced survival prediction.