This repository is a fork of CBR-SUBG from Das el at 2022 work. Here we implement CBR-SUBG on Mechanistic Repositioning Network with Indications (MIND), a biomedical knowledge graph that integrates two biomedical resources: MechRepoNet and DrugCentral. Results of our implementation are reported here: https://ai4d3.github.io/papers/39.pdf
- MechRepoNet with DrugCentral Indications Knowledge Graph and Train/Test/Dev data can de dowloaded from here: here
- You can download our collected train/test/dev subgraphs here: here
Run your own subgraph collection procedure running:
1-Collect chains around train drug queries, joining the drug query entitiy to the disease answer.
python src/01_collect_subgraphs/01_find_paths.py --data_name <dataset_name>
--data_dir_name <path_to_train_graph_files>
--output_dir <path_to_save_output_file>
--cutoff 3
--paths_to_collect 1000
2-For a given query in train/dev/test set, we retrieve its K-nearest neighbor (KNN) queries from the training set. We then gather the collected paths from step 1 and traverse the KG. The following code snippet specifies that we consider 5 KNN queries and explore up to 100 nodes at each traversal step.
python src/01_collect_subgraphs/02_graph_collection.py --data_name <dataset_name>
--data_dir_name <path_to_train_test_dev_files>
--knn 5
--collected_chains_name <path_to_collected_chains_file_in_step_1>
--branch_size 100
--output_dir <path_to_save_output_file>
The runner.py
file is the main file that is needed to run the code. The commands to reproduce CBR-SUBG on MIND dataset are listed below:
python runner.py --output_dir 01_results/
--data_dir rc/00_data/
--data_name MIND
--paths_file_dir MIND_cbr_subgraph_knn-5_branch-200.pkl
--train_batch_size 1
--num_neighbors_train 5
--eval_batch_size 1
--num_neighbors_eval 10
--gcn_dim_init 64
--hidden_channels_gcn 128
--drop_gcn 0.796562802302663
--conv_layers 1
--transform_input 0
--use_wandb 1
--dist_metric l2
--dist_aggr1 mean
--dist_aggr2 mean
--sampling_loss 0.33912131320071265
--temperature 0.10531080354774307
--learning_rate 0.1
--warmup_step 300
--weight_decay 0.01
--num_train_epochs 100
--gradient_accumulation_steps 2
--check_steps 10
Note: Because PyTorch does not ensure perfect reproducibility, even when using the same random seed (as explained in the PyTorch documentation at https://pytorch.org/docs/stable/notes/randomness.html), there may be slight deviations in the results compared to those reported in the paper.
This project builds upon work from Das el at 2022
Please open an issue or contact agonzalez@scripps.edu if you have any questions.