Signed versus isomorphic graph autoencoders for microRNA–single–nucleotide polymorphisms network reconstruction in periodontal osteo-genomic landscapes

Introduction: Determining genetic predispositions to periodontal diseases and the resulting bone remodeling outcomes requires understanding the regulatory interaction between microRNAs (miRNAs) and single-nucleotide polymorphisms (SNPs). Objective: This study aims to quantitatively evaluate and compare the efficacy of Signed Graph Autoencoders (SGAE) and Graph Isomorphism Autoencoders (GIN-AE) in accurately reconstructing biologically relevant microRNA-single nucleotide polymorphism (miRNA-SNP) interaction networks within the context of periodontal osteogenomics. Methods: Using a carefully selected dataset of miRNA–SNP interactions linked to bone disease from HMDD v4.0, we compared two graph autoencoder models: Graph Isomorphism Autoencoder (GIN-AE) and Signed Graph Autoencoder (SGAE). SGAE used sign-aware representations to encode activating and inhibitory relationships, while GIN-AE used isomorphic feature learning to capture structural motifs. Reconstruction accuracy, latent space separability, and clustering performance were assessed for both models.