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 have been evaluated mainly on non-biomedical networks, leaving their performance on biomedical tasks incompletely studied. This paper benchmarks 11 methods across five biomedical tasks and reports that recent methods generally perform well, with very competitive or better results than state-of-the-art methods.
Problem
Advanced graph embedding approaches have been evaluated mainly on non-biomedical networks, with few comprehensive evaluations and analyses on biomedical networks despite promising results from traditional embedding techniques.
Source: §1 · p. 2Method
Results
Recent graph embedding methods generally perform well across biomedical prediction tasks and achieve very competitive or better performance compared with state-of-the-art methods.
Source: §6 · p. 9Takeaways & Limitations
The authors present recent graph embedding methods as a starting point for future biomedical prediction tasks and provide guidelines for method selection and hyperparameter settings.
Source: §1 · p. 2, §6 · p. 9The discussion identifies incorporating external information and partial labels into advanced graph embeddings as future work that could potentially improve performance.
Source: §5 · p. 9
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
This paper systematically evaluates graph embedding methods on biomedical prediction tasks, addressing a gap in prior evaluations focused mainly on non-biomedical networks. Its results show promising, competitive performance and offer guidance for choosing methods and hyperparameters.
Introduction
- Biomedical graphs represent entities and relations for tasks including drug repositioning, lncRNA function prediction, and clinical decision support, while graph embeddings learn low-dimensional node features for downstream models.Prior graph embedding evaluations focused mainly on social, citation, and user-item networks, with few systematic comparisons on biomedical datasets.Source: §1 · pp. 1–2
- The study evaluates DDA, DDI, PPI, protein function, and medical term semantic type prediction using seven datasets and 11 methods spanning matrix factorization, random walks, and neural networks.The 11 methods comprise five matrix factorization-based, three random walk-based, and three neural network-based methods.Source: §1 · p. 2
- Across five biomedical tasks, graph embeddings achieved very competitive or better performance, remained general across graphs and tasks, and provided complementary representations for biological features.The study also compares embeddings with recent computational methods designed for DDA, DDI, and protein function prediction.Source: §1 · p. 2
- Method-selection observations distinguish task strengths: HOPE and GraRep model useful high-order proximity for link prediction, while DeepWalk and node2vec favor node classification and struc2vec favors link prediction.The paper also discusses connections to network propagation and diffusion, and identifies transfer learning as a future direction.Source: §1 · p. 2
- The authors provide BioNEV, an easy-to-use Python package with source code and datasets, to support evaluation of graph embedding methods on biomedical tasks.The package is available at https://github.com/xiangyue9607/BioNEV.Source: §1 · p. 2
2 Overview of graph embedding methods
The overview groups graph embedding methods into matrix-factorization (MF), random-walk-based, and neural-network-based approaches, each using different strategies to represent graph structure in low-dimensional node embeddings.
2.1 MF-based methods
- MF-based methods factorize graph data matrices, with newer approaches encoding higher-order proximity or network similarity measures.Traditional methods often factorize first-order matrices such as adjacency matrices; GraRep uses k-step transition matrices, while HOPE uses measures such as Katz Index and Common Neighbors.Source: §2 · p. 3
2.2 Random walk-based methods
- Random-walk methods learn node embeddings from generated node sequences, with variants designed for flexible sampling or structural similarity.DeepWalk uses truncated walks, node2vec combines breadth-first and depth-first sampling, and struc2vec applies DeepWalk to a multilayer graph encoding structural similarity.Source: §2 · p. 3
2.3 Neural network-based methods
- Neural-network methods vary in architecture and graph input to preserve node proximity or connectivity.Examples include LINE’s first- and second-order proximity, DNGR and SDNE autoencoders, GAE’s GCN encoder and inner-product decoder, and GraphGAN’s generator-discriminator framework for modeling connectivity.Source: §2 · p. 3
3 Applications of graph embedding on biomedical networks
The section reviews 11 graph-embedding methods and their use across five biomedical applications: three link-prediction tasks and two node-classification tasks.
3.3 Summary
- The authors evaluate 11 methods—five matrix-factorization, three random-walk, and three neural-network methods—across five biomedical tasks and against task-specific state-of-the-art approaches.This evaluation addresses uncertainty about newer embedding methods' performance in biomedical networks, where existing applications primarily used traditional techniques.Source: §3 · p. 3
3.1 Link prediction
- Biomedical link prediction infers unknown interactions from known links, but feature-based methods may rely on costly or unavailable biological features and inadequate biological or graph representations.The section describes matrix factorization for DDA and DDI prediction, neural methods for DDIs, and Laplacian, SVD, and autoencoder approaches for PPIs.Source: §3 · p. 3
3.2 Node classification: Protein function prediction
- Protein-function prediction uses graph-based embeddings to help address the expense of experimental annotation, with approaches spanning random walks, low-rank methods, and neural models.Some approaches combine protein sequences with PPI networks, while others learn representations from multilayer tissue networks or protein networks.Source: §3 · p. 4
3.2 Node classification: Medical term semantic type classification
- Medical-term semantic-type classification predicts unknown labels from a clinical-term co-occurrence graph, using learned embeddings as inputs to a multi-label classifier.Privacy-aware clinical datasets provide terms and co-occurrences rather than semantic information, and incomplete or mismatched UMLS knowledge leaves some term labels unknown.Source: §3 · p. 4
4 Experiments
The experiments benchmark 11 graph-embedding methods on seven biomedical datasets spanning link prediction and node classification, finding recent methods promising and their relative suitability task-dependent.
4.1 Datasets We use the following datasets for Link Prediction:
- The benchmark covers seven biomedical datasets: two DDA graphs, a DDI graph, a PPI link-prediction graph, a clinical term co-occurrence graph, and two annotated PPI graphs.Source: §4 · p. 4, §4.1 · pp. 4–5
- Link prediction uses 80/20 interaction splits, balanced random disconnected-edge negatives, and logistic-regression classifiers; node classification uses 80/20 labeled-node splits and One-vs-Rest logistic regression.The reported evaluation metrics are AUC, accuracy, and F1 for link prediction, and accuracy, Macro-F1, and Micro-F1 for node classification.Source: §4.1 · p. 6
4.3 Link prediction results
- LINE achieves 3–23% higher AUC than LEs, while struc2vec raises accuracy 3–15% over GF across the four biomedical link-prediction datasets.Source: §4.3 · p. 6, Table 3 · p. 7
- Graph embeddings match LRSSL on DDA prediction and DeepWalk embeddings improve LRSSL as complementary biological features; on DDI, they outperform or match drug-feature models across four classifiers.The comparison uses LRSSL for DDA and DeepDDI (Ryu et al., 2018) for DDI.Source: §4.3 · p. 6, Figure 3 · p. 7
- For node classification, SVD surpasses HOPE and GraRep, node2vec outperforms other random-walk methods, and LINE and GAE show promising performance.Struc2vec performs less well than DeepWalk and node2vec on these tasks.Source: §4.3 · pp. 6–7, Table 4 · p. 8
- DeepWalk, node2vec, and LINE achieve competitive or better performance than Mashup on three levels of protein-function prediction.For the comparison with Mashup, the embedding methods are run on a simple diffusion network formed by summing interaction scores across networks.Source: §4.3 · p. 7, Figure 3 · p. 7
- Around 100 dimensions is a practical choice: performance tends to saturate near that threshold, while training time rises sharply at larger dimensions.The dimensionality analysis uses the CTD DDA dataset; the paper also grid-searches selected hyperparameters for seven methods and summarizes tuning guidance.Source: §4.3 · pp. 7–8, Figure 4 · p. 8
- Method suitability varies by task: high-order proximity helps MF-based methods on link prediction, node2vec and DeepWalk suit node classification, and LINE is consistently competitive.Struc2vec is more suitable for link prediction; SDNE can perform well on link prediction but less well on node classification, while GAE may fare better on larger networks.Source: §4.3 · p. 8
5 Discussions and future directions
The discussion connects graph embedding with network propagation and identifies external graph information and pre-trained biomedical embeddings as promising directions. Fine-tuning pre-trained embeddings improved link prediction on the CTD DDA graph.
Connections of network embedding and network propagation
- Network propagation and graph embedding share the assumption that nodes with similar neighbors tend to have similar functions and interact, though they process network information differently.Propagation iteratively spreads signals, while random-walk embeddings use node sequences to learn low-dimensional representations; the same assumption also appears in LINE and SDNE.Source: §5 · pp. 8–9
- Random walk with restart, personalized PageRank, and diffusion kernels use embedding-related ideas and could be incorporated into current random-walk embedding frameworks.Source: §5 · p. 9
Modeling external information in graphs
- Incorporating external features, prior information, and partial labels into advanced graph embedding models may improve performance; benchmarking attributed methods on biomedical graphs is future work.Examples include drug and disease features in matrix factorization and joint optimization of DeepWalk with an SVM classification loss in DDRW and MMDW.Source: §5 · p. 9
Transfer learning for graph embedding
- Link prediction on the CTD DDA graph improved with pre-trained embeddings from Beam et al. (2018), which can add semantic information through fine-tuning.The authors note that entities lacking pre-trained embeddings still require random initialization, but expect pre-training followed by fine-tuning to become more promising as coverage grows.Source: §5 · p. 9
6 Conclusion
Across seven datasets and 11 representative methods, recent graph embedding methods performed well on biomedical link prediction and node classification, often matching or surpassing state-of-the-art methods. The paper also offers hyperparameter guidance and discusses connections to network diffusion and future research directions.
6 Conclusion
- Recent graph embedding methods performed well across biomedical prediction tasks and achieved competitive or better performance than state-of-the-art methods.The evaluation covered link prediction and node classification using seven datasets compiled from public databases or previous studies and 11 representative methods.Source: §6 · p. 9
- The paper positions recent graph embedding methods as a starting point for designing advanced models for future biomedical prediction tasks.Source: §6 · p. 9
- The authors tune important graph embedding hyperparameters and provide practitioners with general guidance for setting them.Source: §6 · p. 9
- The paper discusses connections between network propagation or diffusion methods and graph embedding, and identifies transfer learning for graph embedding as a potential research direction.Source: §6 · p. 9