Pan-cancer Distant Metastasis Prediction Based on Graph Neural Network.

Area of Science:

• Computational pathology
• Biomarker discovery
• Graph representation learning

Background:

• Distant metastasis (DM) is the primary cause of cancer mortality.
• Predicting DM is difficult due to a lack of effective biomarkers.
• Histopathological whole slide images (WSIs) contain rich morphological data.

Purpose of the Study:

• To develop a novel graph representation for identifying discriminative morphological features from WSIs.
• To leverage graph neural networks (GNNs) for capturing spatial dependencies in metastatic progression.
• To improve the clinical prediction of distant metastasis for personalized cancer intervention.

Main Methods:

• Transforming high-resolution WSIs into topological graphs.
• Utilizing graph neural networks (GNNs) to analyze cellular organization and spatial relationships.
• Evaluating the method on a large-scale pan-cancer dataset and independent cohorts.

Main Results:

• The proposed graph representation effectively identifies discriminative morphological features.
• The GNN-based approach captures complex spatial dependencies critical for metastasis.
• Superior performance was demonstrated in distilling shared metastatic patterns across diverse cancers.
• Cross-dataset validation confirmed the robustness of the representation.

Conclusions:

• Computational pathology offers a scalable and objective approach for cancer risk stratification.
• The novel graph representation shows potential for high-accuracy prediction of distant metastasis.
• This method can aid in personalized clinical decision-making for cancer patients.