Source-linked AI summary
Graph Embedding on Biomedical Networks: Methods, Applications, and Evaluations
Xiang Yue, Zhen Wang, Jingong Huang, Srinivasan Parthasarathy, Soheil Moosavinasab, Yungui Huang, Simon M. Lin, Wen Zhang, Ping Zhang, Huan Sun
TL;DR
Recent graph embedding methods had not been systematically evaluated on biomedical networks, despite promising traditional techniques. This paper benchmarks 11 methods across seven datasets and five biomedical tasks, finding competitive or better performance and offering selection and hyper-parameter guidelines.
Problem
Recent graph embedding methods were mainly evaluated on non-biomedical networks, leaving their effectiveness on biomedical tasks insufficiently studied.
Method
The paper benchmarks 11 representative graph embedding methods on seven datasets covering three link prediction and two node classification tasks.
Results
Recent graph embedding methods generally perform well across biomedical prediction tasks and achieve competitive or better performance than state-of-the-art methods without biological features.
Takeaways & Limitations
Recent graph embedding methods can serve as a starting point for future biomedical prediction models, with method and hyper-parameter choices guided by the summarized experiments.
Takeaways & Limitations
GAE performance is omitted for the large-scale Clini COOC graph because the authors' source code does not support graphs with more than 40k nodes.
Abstract
from arXiv · showhide
Graph embedding learning that aims to automatically learn low-dimensional node representations, has drawn increasing attention in recent years. To date, most recent graph embedding methods are evaluated on social and information networks and are not comprehensively studied on biomedical networks under systematic experiments and analyses. On the other hand, for a variety of biomedical network analysis tasks, traditional techniques such as matrix factorization (which can be seen as a type of graph embedding methods) have shown promising results, and hence there is a need to systematically evaluate the more recent graph embedding methods (e.g. random walk-based and neural network-based) in terms of their usability and potential to further the state-of-the-art. We select 11 representative graph embedding methods and conduct a systematic comparison on 3 important biomedical link prediction tasks: drug-disease association (DDA) prediction, drug-drug interaction (DDI) prediction, protein-protein interaction (PPI) prediction; and 2 node classification tasks: medical term semantic type classification, protein function prediction. Our experimental results demonstrate that the recent graph embedding methods achieve promising results and deserve more attention in the future biomedical graph analysis. Compared with three state-of-the-art methods for DDAs, DDIs and protein function predictions, the recent graph embedding methods achieve competitive performance without using any biological features and the learned embeddings can be treated as complementary representations for the biological features. By summarizing the experimental results, we provide general guidelines for properly selecting graph embedding methods and setting their hyper-parameters for different biomedical tasks.
1 Introduction
Biomedical graphs support diverse prediction and classification tasks, but advanced graph embedding methods have been insufficiently evaluated on biomedical networks. This paper systematically benchmarks representative methods and derives task-specific guidance.
- Biomedical graphs represent entities and relations for tasks including drug repositioning, interaction prediction, and clinical decision support.
- Graph embedding methods learn low-dimensional node representations that preserve graph structure for downstream prediction, classification, clustering, and community detection.
- Advanced graph embedding approaches have mainly been evaluated on non-biomedical networks, leaving biomedical performance insufficiently studied.
- Traditional methods such as Laplacian eigenmap and matrix factorization have shown promising biomedical results, motivating evaluation of newer approaches.
- The study benchmarks 11 graph embedding methods on seven datasets spanning three link prediction and two node classification tasks.
- Results indicate competitive graph embedding performance, complementary value for biological features, and practical guidance for method and hyper-parameter selection.
2 Overview of graph embedding methods
The paper organizes graph embedding methods into matrix factorization-based, random walk-based, and neural network-based approaches. These methods differ in how they encode graph structure and learn node representations.
- Graph embedding methods are categorized into MF-based, random walk-based, and neural network-based groups.
- MF-based methods: MF-based methods factorize data matrices into lower-dimensional representations while preserving manifold structure and topological properties.
- MF-based methods: GraRep and HOPE model high-order graph proximity through designed matrices before factorization.
- Random walk-based methods: Random walk-based methods generate node sequences by traversing graph neighbors and learn representations with word2vec-inspired models.
- Random walk-based methods: DeepWalk uses truncated random walks, node2vec uses biased walks combining breadth-first and depth-first sampling, and struc2vec models structural identity.
- Neural network-based methods: Neural network-based methods include MLPs, autoencoders, GANs, and GCNs with differing architectures and inputs.
3 Applications of graph embedding on biomedical networks
The paper applies graph embedding to biomedical link prediction and node classification, using learned representations as classifier features. It addresses limitations of hand-engineered biological features and evaluates methods across diverse biomedical applications.
- The study covers DDA, DDI, PPI, protein function, and medical term semantic type prediction or classification tasks.
- Link prediction: Biomedical link prediction seeks potential interactions from known links across networks such as DDA, DDI, and PPI graphs.
- Link prediction: Biological and hand-crafted graph features may be unavailable, costly, or insufficiently precise for robust modeling.
- Link prediction: Graph embeddings automatically learn low-dimensional representations and have been used, for example, by factorizing DDA matrices into drug and disease latent spaces.
- Node classification: Protein function prediction addresses costly experimental annotation, while medical term classification uses a term co-occurrence graph derived from clinical text.
- The pipeline learns embeddings first, then feeds them into binary or multi-label classifiers for unobserved links or node labels.
4 Experiments
The experiments systematically evaluate 11 graph embedding methods across biomedical link prediction and node classification datasets, comparing method families, baselines, and hyper-parameter effects. Recent methods generally improve or match traditional techniques and state-of-the-art systems, with method suitability varying by task.
- Experimental setup: The study compiles seven biomedical datasets and evaluates 11 methods spanning matrix factorization, random walk, and neural network approaches.The datasets cover four link prediction networks and three node classification datasets.
- Datasets: 92 813 CTD DDA edges connect 12 765 nodes, while STRING PPI contains 359 776 interactions among 15 131 proteins after filtering for confidence above 0.7.The DDA graph includes 9580 chemicals and 3185 diseases.
- Link prediction results: 3–23% AUC improvement over LEs was achieved by LINE across four link prediction datasets, while Struc2vec improved accuracy over GF by 3–15%.These results indicate that recent graph embedding methods can improve link prediction over traditional techniques.
- State-of-the-art comparisons: Graph embeddings achieved competitive performance against LRSSL for DDA prediction and outperformed or matched DeepDDI under four classifiers for DDI prediction.Adding DeepWalk embeddings as a fourth LRSSL feature improved LRSSL performance, indicating complementarity with biological features.
- Node classification results: For node classification, node2vec generally outperformed other random-walk methods, LINE surpassed SDNE, and GAE performed promisingly.Struc2vec was less suitable when biomedical graphs lacked clear structural roles; GAE worked better on relatively large networks than on small datasets.
- Hyper-parameter analysis: Performance generally saturated near embedding dimensionality 100, while training time increased sharply beyond that threshold.The experiments therefore identify around 100 dimensions as a practical balance between prediction performance and time efficiency.
5 Discussions and future directions
The discussion connects graph embedding with propagation, external information, transfer learning, and hyper-parameter selection. It identifies attributed biomedical graph embedding and broader biomedical pretraining as promising directions, while noting current coverage limits.
- Connections of network embedding and network propagation: Network propagation and several embedding methods share the assumption that nodes with similar neighbors tend to interact.Propagation iteratively spreads node information through edges; related assumptions appear in random walk-based methods, LINE, and SDNE.
- Hyper-parameter selection: Table 5 summarizes the meanings of major hyper-parameters and provides general guidelines for setting them across embedding methods.The table is presented as a practical reference for method configuration.
- Connections of network embedding and network propagation: Random-walk variants such as random walk with restart, personalized PageRank, and diffusion kernels can be incorporated into graph embedding frameworks.These variants use embedding-related constructions including normalized Laplacian matrices and inverse-Laplacian factorization.
- Modeling external information in graphs: External features and partial labels can improve biomedical graph models, but benchmarking attributed network embedding methods is left for future work.Examples include drug and disease features, prior gene-network information, and label-aware objectives.
- Transfer learning for graph embedding: Pre-trained biomedical entity embeddings improved link prediction on the CTD DDA graph, supporting further investigation of transfer learning.The paper notes that entities without released pre-trained embeddings must currently be initialized randomly.
6 Conclusion
The paper surveys graph embedding techniques and benchmarks them across biomedical link prediction and node classification tasks. Using seven datasets and eleven representative methods, it finds that recent methods generally perform well across these tasks.
- 6 Conclusion: The study compiles seven public or previously studied datasets to benchmark eleven representative graph embedding methods.The evaluation covers biomedical link prediction and node classification.
- 6 Conclusion: Recent graph embedding methods generally perform well across various biomedical prediction tasks.This conclusion follows from the paper’s extensive experiments across the compiled benchmarks.
Funding
The work was supported by the Patient-Centered Outcomes Research Institute under grant ME-2017C1-6413, and no conflicts of interest were declared.
- Funding: The study was supported by the Patient-Centered Outcomes Research Institute under grant ME-2017C1-6413.
- Funding: The authors declared no conflicts of interest.