GraphPath: a graph attention model for molecular stratification with interpretability based on the pathway-pathway...

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Personalized cancer therapy necessitates understanding molecular heterogeneity.
• Identifying biological drivers of cancer is crucial for discovering therapeutic targets.
• Accurate clinical predictions require comprehensive patient characterization at molecular and pathway levels.

Purpose of the Study:

• To develop an interpretable model for classifying cancer status using multi-omics data.
• To leverage biological pathway interactions for enhanced cancer prediction.
• To identify novel therapeutic targets through interpretable model insights.

Main Methods:

• Introduction of GraphPath, a biological knowledge-driven graph neural network.
• Utilizing a multi-head self-attention mechanism within a pathway-pathway interaction network.
• Training and validation on prostate cancer multi-omics data, including external cohorts.

Main Results:

• GraphPath outperforms baseline methods like P-NET in cancer status classification.
• The model demonstrates generalizability across unseen samples in external cohorts.
• Dimensionality reduction and visualization confirm optimal model performance.
• Identification of target cancer-associated pathways contributing to predictions.

Conclusions:

• GraphPath provides a robust and interpretable framework for cancer prediction.
• The model enhances understanding of cancer's biological mechanisms.
• This approach accelerates the development of targeted cancer therapies.