Graph-based multi-modality integration for prediction of cancer subtype and severity.

Area of Science:

• Computational biology
• Bioinformatics
• Cancer research

Background:

• Personalized cancer screening requires integrating diverse data types for improved accuracy.
• Current methods often analyze single data sources, neglecting valuable interactions and complementary insights.

Purpose of the Study:

• To develop and evaluate a graph-theory-based approach for integrating multi-omics data (RNA-Seq, histopathology) for personalized cancer screening.
• To assess the impact of different data fusion strategies (early, intermediate, late) on predicting cancer subtypes and severity.

Main Methods:

• Construction of individual-specific networks using RNA-Seq and whole-slide imaging data.
• Computation of inter-individual similarity matrices based on graph structures.
• Training and evaluation of classification models using similarity matrices and macro F1 score.

Main Results:

• Graph-based methods demonstrate improved performance over non-interaction-aware approaches.
• Multi-modal data integration, particularly through intermediate fusion, enhances classification accuracy compared to single-data models.
• The proposed workflow shows adaptability and potential for personalized healthcare applications.

Conclusions:

• Graph theory provides a powerful framework for integrating multi-modal data in personalized cancer screening.
• Intermediate fusion of RNA-Seq and histopathology data offers significant advantages for cancer subtype and severity prediction.
• The developed methodology holds promise for advancing personalized medicine across various disease contexts.