Graph Representation Learning, Geometric Deep Learning, and Foundation Models for Biological Network Analysis
WIREs Computational Molecular Science · 2026-11 · 16 (6)
DOI: 10.1002/wcms.70092 ↗Kurz gefasst
Fachartikel in der Zeitschrift WIREs Computational Molecular Science, online erschienen am 16.11.2026. Verfasst von Gleb Svinin und 2 weiteren Autor:innen. Die Arbeit nennt 108 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Fonds National de la Recherche Luxembourg.
Abstract (Originalsprache)
Biological systems are inherently relational. Genes, proteins, metabolites, and cells interact through complex networks that encode the molecular logic of development, homeostasis, and disease. However, conventional machine learning methods that operate on tabular or sequence‐based representations are not well‐suited to capture this graph‐structured organization. Graph neural networks (GNNs) and their modern descendants, such as graph transformers, equivariant architectures, and foundation models, provide a framework for learning directly from biological interaction networks. This review provides a broad, methodologically focused overview of graph representation learning and geometric deep learning approaches applied to biological network analysis. It covers graph convolutional networks, graph attention networks, message‐passing neural networks, graph transformers, equivariant networks for three‐dimensional molecular structures, and large‐scale pre‐trained foundation models. We introduce a taxonomy that distinguishes graph‐native, knowledge‐graph, graph‐informed, and non‐graph biological foundation models, and we organize the discussion of architectures around the design decisions that matter for biological networks, such as heterophily, edge direction and sign, edge uncertainty, and oversquashing. We evaluate how these architectures can be applied to key network biology tasks, such as gene regulatory network inference, protein–protein interaction prediction, multi‐omics network integration, and disease gene prioritization, and we compare the problem formulations underlying these tasks. We also compare representational strategies, such as heterogeneous graphs and hypergraphs, to capture the multilayer complexity of biological systems. Benchmarking resources, open software tools, and key datasets used across these tasks are cataloged. We discuss challenges, such as data sparsity, network incompleteness, uncertainty in graph construction, data leakage, negative sampling, biological interpretability, and limited ground‐truth labels, and we provide practical guidance for method selection and validation. This review is intended for computational biologists, bioinformaticians, and machine learning researchers who are seeking an overview and introduction to this field with an emphasis on its applicability to disease‐relevant omics data analysis.
Förderung
- Fonds National de la Recherche Luxembourg
Zitieren
APA
Svinin, G., Loo, R. T. J., & Glaab, E. (2026). Graph Representation Learning, Geometric Deep Learning, and Foundation Models for Biological Network Analysis. WIREs Computational Molecular Science, 16(6). https://doi.org/10.1002/wcms.70092
Vancouver
Svinin G, Loo RTJ, Glaab E. Graph Representation Learning, Geometric Deep Learning, and Foundation Models for Biological Network Analysis. WIREs Computational Molecular Science. 2026;16(6). doi:10.1002/wcms.70092
BibTeX
@article{Svinin2026,
author = {Svinin, Gleb and Loo, Rebecca Ting Jiin and Glaab, Enrico},
title = {Graph Representation Learning, Geometric Deep Learning, and Foundation Models for Biological Network Analysis},
journal = {WIREs Computational Molecular Science},
year = {2026},
volume = {16},
number = {6},
doi = {10.1002/wcms.70092}
}
EUVIMED