Source-linked AI summary
Graph Representation Learning in Biomedicine
Michelle M. Li, Kexin Huang, Marinka Zitnik
TL;DR
Biomedical networks offer broad descriptors of interacting biological and medical systems, but graph representation learning must connect complex topology with useful biomedical representations. This review synthesizes methods that embed networks into compact vector spaces and relates them to systems-biology principles. Across biomedical applications, these approaches support advances including protein-function prediction, cell-type-specific analysis, disease modeling, and discovery of molecules, interactions, and functions.
Problem
Biomedical networks are broad and structurally complex, while existing reviews often cover only specific graph-learning methods, graph types, or biomedical domains.
Method
The review synthesizes graph representation learning approaches that map biomedical network elements into compact vector spaces reflecting network structure, attributes, and systems-biology principles.
Results
Graph representation learning has supported protein-function prediction, cell-type-specific disease-feature identification, disease modeling, and discovery of new molecules, interactions, and functions.
Takeaways & Limitations
The survey presents graph representation learning as a unified framework spanning molecular, genomic, therapeutic, and precision-medicine applications.
Takeaways & Limitations
Models trained solely on protein–protein interaction networks are limited because these networks cannot capture all gene-regulatory activities.
Abstract
from arXiv · showhide
Biomedical networks (or graphs) are universal descriptors for systems of interacting elements, from molecular interactions and disease co-morbidity to healthcare systems and scientific knowledge. Advances in artificial intelligence, specifically deep learning, have enabled us to model, analyze, and learn with such networked data. In this review, we put forward an observation that long-standing principles of systems biology and medicine -- while often unspoken in machine learning research -- provide the conceptual grounding for representation learning on graphs, explain its current successes and limitations, and even inform future advancements. We synthesize a spectrum of algorithmic approaches that, at their core, leverage graph topology to embed networks into compact vector spaces. We also capture the breadth of ways in which representation learning has dramatically improved the state-of-the-art in biomedical machine learning. Exemplary domains covered include identifying variants underlying complex traits, disentangling behaviors of single cells and their effects on health, assisting in diagnosis and treatment of patients, and developing safe and effective medicines.
1 Introduction
Biomedical networks represent interacting biological and medical entities, while graph representation learning converts their structure and attributes into compact embeddings for downstream analysis. The survey connects these methods to systems-biology principles and unifies applications and algorithmic paradigms across biomedical domains.
- Network foundations: Biomedical networks describe interactions ranging from molecular relationships to disease, health, and scientific knowledge.Examples include regulatory, gene–disease, and population-scale interaction networks.
- Network foundations: Network biomedicine principles link graph structure to molecular phenotypes, biological roles, disease, and health.The local hypothesis associates interacting entities with greater similarity, while shared components and disease modules explain clustering of phenotype-related entities.
- Representation learning: Graph representation learning maps nodes or larger structures into low-dimensional embeddings whose geometry reflects interaction structure and network neighborhoods.Learned transformations place nodes with similar neighborhoods near one another and support topology-related algebraic operations.
- Representation learning: Representation learning can act as a differentiable engine for key network-biomedicine principles.The survey frames learned representations as computational realizations of systems-biology concepts.
- Survey scope: The survey unifies graph representation learning across molecular, genomic, therapeutic, and precision-medicine applications.It contrasts this scope with prior reviews focused on narrower graph-learning methods, graph types, or biomedical domains.
- Algorithmic paradigms: Shallow embeddings, graph neural networks, and generative graph models form major paradigms for encoding structure and supporting prediction or graph generation.Other tasks include node, link, and graph property prediction and learning graphs that are only partially observed.
3 Application areas in biology and medicine
Graph representation learning spans biomedical applications from molecular structures and genomic networks to therapeutics and healthcare, using network structure to support prediction and personalization. Across these levels, learned graph representations address protein functions, disease mechanisms, drug candidates, and patient-specific outcomes.
- Cross-domain applications: Biomedical networks integrate heterogeneous data from molecular structures, gene interactions, diseases, drugs, and patient records across biological and healthcare levels.Graph methods can leverage structural information in multimodal datasets and connect patient records with molecular, genomic, disease, and drug networks.
- Genomics and disease: Graph embeddings capturing genome-wide interactions have enhanced disease predictions at tissue and single-cell resolution.Gene co-expression and regulatory networks can represent cellular circuitry, while network parsimony links shortest paths between disease-associated components with causal molecular pathways.
- Therapeutics: Learning drug, protein, and disease graph topology improves predictions of candidate treatments, off-target effects, and novel drug combinations.These networks model drug-drug interactions, drug-target binding, disease associations, and drug repurposing opportunities.
- Molecular biology: Graph representation learning supports protein function prediction and discovery of new molecules, interactions, and functions, including for previously unseen genes.Protein structures can be represented through residue or atom graphs, and graph convolutional models learn local and long-range dependencies for protein embeddings.
- Single-cell and spatial biology: Cell-type-aware gene embeddings enable single-cell predictions that incorporate disease or cell states and temporal or spatial dependencies.Predicted ligand-receptor interactions have helped elucidate intercellular interactions and the spatial heterogeneity of tumors after experimental validation.
- Disease modeling: Fusing noisy gene-expression data with GO annotations or PPI networks and graph convolution layers produces more interpretable disease-classification models, although incomplete PPI networks constrain coverage.Graph methods have also been developed to learn robust molecular representations despite incomplete interactomes and to infer new edges.
6 Graph representation learning for therapeutics
Graph representation learning supports therapeutics by modeling molecular structure, interactions, multimodal biomedical knowledge, and patient data for drug discovery and clinical prediction.
- Molecular representations: Molecular graphs encode atoms, bonds, spatial distances, bond angles, and rotations to represent chemical structure for learning tasks.Message passing aggregates neighboring atom and bond information to learn local chemistry, while junction-tree methods represent molecular substructures.
- Molecular representations: Graph-based molecular representations have improved quantum-chemistry predictions and helped generate candidate molecules with desirable properties.Iteratively editing molecular fragments during training has improved predictions for high-quality candidates targeting proteins of interest.
- Interaction prediction: Drug and target embeddings derived from similarity graphs are supplied to downstream machine-learning models for interaction and affinity prediction.Approaches include TDA, shallow network embedding, and graphs constructed with k-nearest neighbors.
- Interaction prediction: Fusing compound sequence, structure, and clinical information has significantly improved drug-drug and drug-target interaction predictions.Examples use attention on drug graphs, separate GNNs for proteins and molecules, and combined graph-based protein-structure and sequence representations.
- Systems and clinical applications: Heterogeneous graphs integrating drugs, targets, diseases, genomic features, and patient records support drug-action modeling and precision-medicine predictions.These systems predict drug-target interactions, drug sensitivity, diagnoses, treatments, and future prescriptions or disease codes.
- Systems and clinical applications: Cell-line-specific graph models can prioritize candidate drug combinations while accounting for anatomical differences and transfer across cell lines.In silico prioritization may reduce the cost of experimentally evaluating combinatorial therapies.
8 Outlook
The outlook extends graph representation learning toward finer-grained variant mapping and integrated biomedical knowledge graphs. It also emphasizes explainability, fairness, robustness, and attention to algorithmic bias in clinical applications.
- Future directions: Graph representation learning is proposed as a promising direction for fine-scale mapping of variants underlying disease mechanisms.Related directions include modeling GWAS and eQTL studies as networks, network propagation, chromatin interactions, and 3D chromosomal reconstruction.
- Future directions: Graph methods can analyze spatial molecular-profiling and dynamic single-cell datasets to capture expression changes over time or after perturbation.Dynamic GNNs are identified as one possible approach for these evolving datasets.
- Knowledge integration: Integrating healthcare data with molecular, genomic, disease-level, and drug-level knowledge can generate more accurate and interpretable predictions about health and disease.Knowledge graphs are being developed to synthesize multi-scaled, multimodal biomedical data.
- Responsible clinical deployment: Clinical graph-representation systems should be evaluated for explainability, fairness, and robustness in light of algorithmic bias and health disparities.The paper identifies these properties as essential as such tools enter clinical applications.
3 Further Details on Representation Learning Approaches S6
The supplementary discussion describes shallow embedding, network diffusion, topological data analysis, and manifold learning as approaches for representing graph structure.
- 3.3 Topological data analysis: Topological data analysis is presented as a supplementary representation-learning approach for extracting structure from data.The supplied passage names the topic without specifying its algorithmic steps.
- 3.4 Manifold learning: Manifold learning is included among the supplementary approaches for representing complex data in lower-dimensional spaces.The supplied passage names the topic without specifying its algorithmic steps.
1 Further Information on Graph Notation and Definitions
The notation appendix defines graph components, neighborhoods, paths, labels, and graph types used to describe biomedical network data.
- Core notation: Nodes represent biomedical entities, edges represent relations, and labels can target nodes, edges, subgraphs, or entire graphs.Examples range from atoms and patients to molecular bonds and biomedical associations.
- Core notation: A graph G = (V, E) consists of nodes V connected by edges E, commonly represented with an adjacency matrix A.The matrix can encode binary connectivity or edge weights; the graph has n = |V| nodes and m = |E| edges.
- Subgraphs and neighborhoods: A subgraph S = (V_S, E_S) is a subset of a graph and can represent structures such as disease modules or patient-doctor communities.Its node and edge sets are subsets of the corresponding sets in the parent graph.
- Subgraphs and neighborhoods: A node’s neighborhood contains connected nodes, while k-hop neighborhoods contain nodes exactly k hops away based on shortest-path distance.A walk is a node sequence, and a simple path is a walk whose nodes are distinct.
- Graph types: Biomedical graphs may be simple, weighted, attributed, heterogeneous, knowledge-based, multilayer, temporal, spatial, or combinations of these types.Heterogeneous graphs connect diverse node and edge types; knowledge graphs encode typed triplets, while multilayer, temporal, and spatial graphs represent context, change, and location.
2 Overview of Graph Machine Learning Tasks
Graph machine learning tasks include predicting labels or properties of graph elements, learning latent graph structure, and generating new graphs with desired characteristics.
- Canonical graph prediction: Canonical graph prediction assigns labels to nodes, links, edges, or entire graphs.
- Other graph prediction tasks: Additional biomedical tasks include module detection, clustering, subgraph prediction, and prediction over sequences of dynamic graphs.
- Latent graph learning: Latent graph learning infers underlying graph structure from nodes and their feature attributes.
- Latent graph learning: Learned latent graphs can support graph prediction, reveal inherent data topology, and generate low-dimensional representations of feature attributes.
- Graph generation: Graph generation learns a distribution from training graphs and samples new graphs with similar or optimized properties.
3 Further Details on Representation Learning Approaches
The review presents representation-learning approaches ranging from expert-defined graph statistics and diffusion to topology-aware, manifold, and neural methods. These approaches encode graph structure or data geometry for downstream biomedical analysis.
- Approaches: The review covers graph theoretic techniques, network propagation, topological data analysis, manifold learning, shallow embeddings, graph neural networks, and generative models.
- Network propagation: Diffusion methods summarize influence propagation and local connectivity, while GraphWave can identify similar structural roles across different graph regions.
- Topological data analysis: Topological data analysis constructs graph or vector summaries of data shape through persistent homology and Mapper.
- Manifold learning: Manifold learning maps high-dimensional data into low-dimensional representations that preserve graph-based distances or are learned end-to-end from downstream prediction signals.
- Graph theoretic techniques: Graph statistics represent node, link, and subgraph structure using expert-defined measures such as centrality, similarity, and motifs.
- Graph theoretic techniques: The protein–protein interaction network violates homophily assumptions, motivating L3, which counts length-3 paths and shows strong performance on PPI networks.
4 Further Details on Mathematical Formulations
The mathematical formulations illustrate how graph statistics encode structural roles, diffusion encodes influence profiles, and Mapper constructs topology from filtered data clusters.
- Graph theoretic techniques: Graph theoretic techniques map network components to real-valued structural statistics such as betweenness.
- Graph theoretic techniques: Betweenness measures the fraction of shortest paths between nodes s and t that pass through node u, with larger values indicating greater network influence.
- Network diffusion: Diffusion State Distance represents each node by expected random-walk visits to every destination and compares nodes using the L1 norm of their diffusion-vector difference.
- Network diffusion: DSD uses differences in node influence profiles to assess whether nodes have similar local connectivity.
- Topological data analysis: Mapper constructs a topology graph by filtering data, covering the filter space, clustering each subset, and connecting clusters that share data points.
1. Reference Map. Given a set of data D, we first define a continuous filter function
The formulations describe pipelines for embedding data, extracting persistent topology, learning manifolds and neural representations, and generating graphs from latent variables.
- Shallow network embedding: Shallow embeddings map nodes into vectors, define network and embedding similarities, and minimize a loss that preserves network distances.
- Persistent homology: Persistent homology builds Rips complexes across increasing distance scales, tracks homology classes, and records their birth and death in persistence diagrams.
- Manifold learning: Manifold learning constructs a neighborhood graph, computes shortest-path distances, and optimizes low-dimensional vectors to represent high-dimensional data.
- Graph neural networks: Graph neural networks iteratively compute messages, aggregate neighbor information, and apply nonlinear updates to produce embeddings capturing structure and node features.
- Generative modeling: Variational graph autoencoders encode adjacency and node features into latent distributions, sample embeddings, and decode them into generated networks.
- Reference map: Figure 1 organizes graph learning paradigms from graph statistics and diffusion to topology, manifold, shallow, neural, and generative approaches.