MODAPro: Explainable Heterogeneous Networks with Variational Graph Autoencoder for Mining Disease-Specific Functional Molecules and Pathways from Omics Data

The rapid growth of multiomics data has revolutionized
our ability
to investigate disease mechanisms, yet significant challenges persist
in achieving meaningful integration due to inherent data heterogeneity,
characteristic sparsity patterns, and the persistent interpretability
gap in current analytical approaches. To address these critical limitations,
we introduce MODAPro, a biologically informed deep learning framework
that synergistically integrates variational graph autoencoders (VAE)
with graph convolutional networks (GCN). This novel architecture enables
MODAPro to capture and meaningfully interpret complex, nonlinear molecular
relationships across different omics layers with unprecedented resolution.
Through systematic benchmarking across diverse disease-related data
sets, MODAPro consistently outperforms existing approaches in identifying
disease-associated biomarkers and functionally coherent modules. Importantly,
MODAPro reveals latent biomolecular information that is often missed
by conventional methods. In realistic and challenging scenarios, MODAPro
effectively captures intricate across–omic interactions, enhancing
functional annotation and offering new insights into disease from
a systems biology perspective, bridging a critical gap between computational
analysis and biological understanding. Furthermore, MODAPro retains
robust performance on single-omics data sets by leveraging the multiomics
context, facilitating discovery even from sparse or incomplete data.
The framework’s adaptability to various data types and conditions
makes it particularly valuable for precision medicine applications,
where it can uncover actionable disease signatures and regulatory
networks that would otherwise remain undetected. By effectively addressing
the major challenges in multiomics integration, MODAPro provides a
useful approach for systems biology research and translational medicine.
The source code of MODAPro is publicly available in our GitHub repository: https://github.com/zhaoxiaoqi0714/MODAPro.