This repository contains source code and datasets for paper "Graph Embedding on Biomedical Networks: Methods, Applications, and Evaluations" (accepted by Bioinformatics). This work aims to systematically evaluate recent advanced graph embedding techniques on biomedical tasks. We compile 5 benchmark datasets for 4 biomedical prediction tasks (see paper for details) and use them to evaluate 11 representative graph embedding methods selected from different categories:

• 5 matrix factorization-based: Laplacian Eigenmap, SVD, Graph Factorization, HOPE, GraRep
• 3 random walk-based: DeepWalk, node2vec, struc2vec
• 3 neural network-based: LINE, SDNE, GAE

The code can also be applied to graphs in other domains (e.g., social networks, citation networks). More experimental details can be found in Supplementary Materials.

Please kindly cite the paper if you use the code, datasets or any results in this repo or in the paper:

@article{yue2020graph,
  title={Graph embedding on biomedical networks: methods, applications and evaluations},
  author={Yue, Xiang and Wang, Zhen and Huang, Jingong and Parthasarathy, Srinivasan and Moosavinasab, Soheil and Huang, Yungui and Lin, Simon M and Zhang, Wen and Zhang, Ping and Sun, Huan},
  journal={Bioinformatics},
  volume={36},
  number={4},
  pages={1241--1251},
  year={2020},
  publisher={Oxford University Press}
}

Fig. 1: Pipeline for applying graph embedding methods to biomedical tasks. Low-dimensional node representations are first learned from biomedical networks by graph embedding methods and then used as features to build specific classifiers for different tasks. For (a) matrix factorization-based methods, they use a data matrix (e.g., adjacency matrix) as the input to learn embeddings through matrix factorization. For (b) random walk-based methods, they first generate sequences of nodes through random walks and then feed the sequences into the word2vec model to learn node representations. For (c) neural network-based methods, their architectures and inputs vary from different models.

Datasets used in the paper:

• CTD DDA : a drug-disease association graph extracted from Comparative Toxicogenomics Database
• NDFRT DDA : a drug-disease association graph extracted from UMLS National Drug File
• DrugBank DDi : a drug-drug interaction graph extracted from DrugBank database
• STRING PPI : a protein-protein interaction graph extracted from STRING database

• Clin Term COOC : a medical term-term co-occurrence graph from (Finlayson et al., 2014) [source data], [paper]
• node2vec PPI: a PPI graph with functional annotations used in node2vec (Grover and Leskovec, 2016)
• Mashup PPI: a experimental PPI graph with functional annotations used in Mashup (Cho et al., 2016)

Statistics:

We also release the best-performing pre-trained representations of nodes (e.g., drugs, diseases, proteins, UMLS concepts) on each dataset. These pre-trained vectors can be used as:

• External representations to complement the biological features. In the paper, we showed that by adding the network embedding feature into an existing computational model for predicting drug-disease associations, the performance is further improved (Section 4.3 in the paper).

• Initialized values of the embedding vectors before training. We can initialize the embedding vector for each node on a graph with its pre-trained embedding (e.g., by looking for the corresponding entity in pre-trained vocab look-up table rather than by random initialization, and then continue training various graph embedding methods as before (which is often referred to as “fine-tuning”). We conducted experiment with this "transfer learning" idea on the "CTD DDA" graph and showed the improvement (Section 5 in the paper).

All the pretrained vectors can be downloaded here. The files are formatted as:

node_num, embedding_dimension
index_1, embedding vector 1
index_2, embedding vector 2
...

The corresponding index to node name (or their original ID) can be found in the each dataset directory.

The graph embedding learning for Laplician Eigenmap, Graph Factorization, HOPE, GraRep, DeepWalk, node2vec, LINE, SDNE uses the code from OpenNE The code of struc2vec and GAE is from their authors. To ensure different source code could run successfully in our framework, we modify part of their source code.

Use the following command to install directly from GitHub;

$ pip install git+https://github.com/xiangyue9607/BioNEV.git

Alternatively, use the following commands to install the latest code in development mode (using -e):

$ git clone https://github.com/xiangyue9607/BioNEV.git
$ cd BioNEV
$ pip install -e .

• --input, input graph file. Only accepted edgelist format.
• --output, output graph embedding file.
• --task, choose to evaluate the embedding quality based on a specific prediction task (i.e., link-prediction, node-classification, none (no eval), default is none)
• --testing-ratio, testing set ratio for prediction tasks. Only applied when --task is not none. The default is 0.2
• --dimensions, the dimensions of embedding for each node. The default is 100.
• --method, the name of embedding method
• --label-file, the label file for node classification.
• --weighted, true if the input graph is weighted. The default is False.
• --eval-result-file, the filename of eval result (save the evaluation result into a file). Skip it if there is no need.
• --seed, random seed. The default is 0.

• Matrix Factorization-based methods:

  • --kstep, k-step transition probability matrix for GraRep. The default is 4. It must divide the --dimension.
  • --weight-decay, coefficient for L2 regularization for Graph Factorization. The default is 5e-4.
  • --lr, learning rate for gradient descent in Graph Factorization. The default is 0.01.

• Random Walk-based methods:

  • --number-walks, the number of random walks to start at each node.
  • --walk-length, the length of the random walk started at each node.
  • --window-size, window size of node sequence.
  • --p, --q, two parameters that control how fast the walk explores and leaves the neighborhood of starting node. The default values of p, q are 1.0.
  • --OPT1, --OPT2, --OPT3, three running time efficiency optimization strategies for struc2vec. The default values are True.
  • --until-layer, calculation until the layer. A hyper-parameter for struc2vec. The default is 6.

• Neural Network-based methods:

  • --lr, learning rate for gradient descent. The default is 0.01.
  • --epochs, training epochs. The default is 5. Suggest to set a small value for LINE and SDNE (e.g., 5), and a large value for GAE (e.g., 500).
  • --bs, batch size. Only applied for SDNE. The default is 200.
  • --negative-ratio, the negative sampling ratio for LINE. The default is 5.
  • --order, the order of LINE, 1 means first order, 2 means second order, 3 means first order + second order. The default is 2.
  • --alpha, a hyperparameter in SDNE that balances the weight of 1st-order and 2nd-order proximities. The default is 0.3.
  • --beta', a hyperparameter in SDNE that controls the reconstruction weight of the nonzero elementsin the training graph. The default is 0.
  • --dropout, dropout rate. Only applied for GAE. The default is 0.
  • --hidden, number of units in hidden layer. Only applied for GAE. The default is 32.
  • --gae_model_selection, GAE model variants: gcn_ae or gcn_vae. The default is gcn_ae.

bionev --input ./data/DrugBank_DDI/DrugBank_DDI.edgelist \
       --output ./embeddings/DeepWalk_DrugBank_DDI.txt \
       --method DeepWalk \
       --task link-prediction \
       --eval-result-file eval_result.txt

bionev --input ./data/Clin_Term_COOC/Clin_Term_COOC.edgelist \
       --label-file ./data/Clin_Term_COOC/Clin_Term_COOC_labels.txt \
       --output ./embeddings/LINE_COOC.txt \
       --method LINE \
       --task node-classification \
       --weighted True

Feel free to contact Xiang Yue <yue.149 AT osu DOT edu> or Huan Sun <sun.397 AT osu DOT edu> for any questions about the paper, datsaets, code and results.