DG-MSGAT: A Biologically-informed Differential Gene Multi-Scale Graph Attention Network for predicting neoadjuvant...

Area of Science:

• Oncology
• Genomics
• Bioinformatics

Background:

• Accurate prediction of neoadjuvant therapy efficacy is crucial for rectal cancer patient outcomes.
• Personalized treatment strategies and reduced interventions depend on predicting pathological complete response (pCR).
• Existing predictive models often neglect gene-gene interactions and struggle with high-dimensional gene expression data.

Purpose of the Study:

• To develop a novel, biologically informed model for predicting neoadjuvant therapy response in rectal cancer.
• To address limitations of existing models by integrating differential gene expression and multi-scale gene interactions.
• To enhance the accuracy of pathological complete response (pCR) prediction.

Main Methods:

• Constructed patient-specific gene graphs using gene expression profiles and differential expression signals, with edges based on protein-protein interactions.
• Employed the Differential Gene Multi-Scale Graph Attention Network (DG-MSGAT), a multi-scale graph attention network with stacked attention layers and residual connections.
• Modeled hierarchical gene dependencies and preserved feature integrity to estimate pCR probability.

Main Results:

• DG-MSGAT significantly outperformed conventional algorithms (SVM, decision trees, random forests) in predicting neoadjuvant therapy efficacy for locally advanced rectal cancer.
• Network analysis identified key genes (e.g., TP53, EGFR, CTNNB1) and immune-related pathways linked to therapeutic response.
• The model demonstrated robust performance in a clinical context.

Conclusions:

• The DG-MSGAT model represents a significant advancement in predicting neoadjuvant therapy outcomes for rectal cancer.
• It effectively models complex gene interactions and overcomes challenges of high-dimensional gene expression data.
• DG-MSGAT offers a clinically relevant tool to guide personalized treatment decisions.