Inferring signaling pathway abnormalities from histopathological images via logicconstrained gene-pathway heterogeneous knowledge graph

Jianing Xi

Abstract:

Conventional histopathological analysis focuses on single-gene mutations and struggles to capture
pathway-level dysregulation driving cancer. To address this, we propose LCG-HGNN, a Logic-
Constrained Gene-Pathway Heterogeneous Graph Neural Network that enables collaborative
recognition of gene groups and infers signaling pathway alterations from whole-slide images. By
integrating a gene-pathway graph structure, dynamic edge weighting within our proposed
KePathGraph framework, which also incorporates logical clauses, our framework achieves superior
prediction accuracy and clinical interpretability over single-gene and multi-label baselines. This work
establishes a pathway-oriented paradigm for histopathological interpretation, providing deeper
insights into the mechanisms underlying cancer initiation and progression.