Causality-aware graph neural networks for functional stratification and phenotype prediction at scale

Area of Science:

• Computational biology
• Systems biology
• Genomics

Background:

• Functional phenotypic heterogeneity in disease is complex.
• Understanding genotype-to-phenotype translation is crucial for disease research.
• Existing methods may not fully capture pathway-level functional differences.

Purpose of the Study:

• To develop a computational framework for classifying entire biological pathways.
• To integrate mathematical programming and Graph Neural Networks (GNNs) for disease pathway analysis.
• To enable genotype-to-phenotype translation by linking genetic variations to functional profiles.

Main Methods:

• Reconstruction of gene networks using Prior Knowledge Networks (PKNs) via mathematical programming.
• Classification of reconstructed gene networks using Graph Neural Networks (GNNs) at the graph-level.
• Application to the human DNA damage and repair pathway, focusing on the TP53 regulon in a pancancer study.

Main Results:

• Successful classification of Gene Regulatory Networks (GRNs) across different TP53 mutation types.
• Identification of mutations with distinct functional profiles linked to specific phenotypes.
• Demonstration of a data-driven pipeline for genotype-to-phenotype translation.

Conclusions:

• The integrated framework effectively elucidates functional phenotypic heterogeneity in disease.
• This approach provides a scalable method for analyzing diverse disease conditions and polygenic complexity.
• The study reveals new functional patterns by disentangling disease complexity through causal representation.