Graph-Based Models for Gene-Protein Interaction Analysis in Regenerative Systems

Abstract

The molecular machinery of tissue regeneration operates through highly interconnected networks of gene regulatory interactions and protein-protein associations that collectively orchestrate stem cell activation, proliferation, differentiation, and tissue remodelling. Understanding the topology and dynamics of these gene-protein interaction (GPI) networks is essential for identifying key regulatory nodes that can be targeted to enhance regenerative capacity. While individual protein-protein interaction (PPI) databases and gene regulatory network (GRN) tools provide partial views of this molecular landscape, integrated graph-based analysis that combines PPI topology, GRN regulatory structure, and tissue-specific expression context within a unified computational framework has been lacking. This paper proposes the Regenerative Gene-Protein Interaction Graph (RGPIG) framework, a graph neural network methodology for integrated analysis of GPI networks in regenerative biology contexts, comprising four graph analytical components: a heterogeneous GPI graph constructor (HGPIC) that builds integrated gene-protein interaction graphs from multi-source biological databases; a graph neural network link predictor (GNN-LP) that predicts novel gene-protein interactions from graph topology and node features; a regenerative module detector (RMD) that identifies functionally coherent regeneration-associated subgraphs; and a network perturbation simulator (NPS) that models the effect of gene knockdown or drug perturbation on the regenerative GPI network. RGPIG is applied to tissue-specific GPI networks from three regeneration contexts -- skeletal muscle (satellite cell-mediated repair), liver (hepatocyte regeneration), and neural tissue (neurogenesis) -- constructed from 284,000 experimentally confirmed interactions across 22,847 human genes and proteins. GNN-LP predicts novel GPI links with AUC-ROC = 0.934 (SD = 0.018), outperforming network topology baselines (0.784) and sequence-only predictors (0.812). RMD identifies 84 regenerative modules across three tissues, of which 72 are enriched for known regeneration pathway members (FDR < 0.01). NPS simulates 1,284 perturbation conditions, identifying 48 high-impact network hubs whose silencing reduces regenerative module connectivity by > 40%. The study contributes the RGPIG specification, the RegenInteractome database of tissue-specific GPI networks, and 384 computationally predicted novel GPI interactions with experimental priority scores.