Pre-trained graph neural network and downstream tasks
- Ref: Self-supervised graph representation learning integrates multiple molecular networks and decodes gene-disease relationship
Download the following files respectively
gtex_js.csv: https://doi.org/10.6084/m9.figshare.19550818.v1 to ./fig3-b/ folder;
network.npy: https://doi.org/10.6084/m9.figshare.19550902.v1 to ./network_data/ folder；
label.npy: https://doi.org/10.6084/m9.figshare.19551259.v1 to ./pathway_member_identification/reactom/checkpoint/ folder；
test_mask.npy: https://doi.org/10.6084/m9.figshare.19551439.v1 to ./pathway_member_identification/reactom/checkpoint/ folder;
train_mask.npy: https://doi.org/10.6084/m9.figshare.19551484.v1 to ./pathway_member_identification/reactom/checkpoint/ folder;
disease_gen.npy: https://doi.org/10.6084/m9.figshare.19551697.v1 to ./RR_predict/comorbidity/ folder;
gen_feat.npy: https://doi.org/10.6084/m9.figshare.19551790.v1 to ./RR_predict/comorbidity/ folder.
We also provide STRING network we used for ablation purpose, and please download network.pk:https://doi.org/10.6084/m9.figshare.21088558 to ./disease_gene_ablation/sub_data folder.
After downloading，the following downstream tasks can be run。
- torch==1.5.1
- dgl==0.6.1
- pytorch-lightning==0.9.0
- scipy==1.5.1
run python evaluate_nci.py，and load the trained checkpoint to evaluate。
run python nci.py to start training。
results are as the following：
ValAcc 0.3080 | ValROC 0.0000 | Pathway 0.3393/0.3107/0.3070
| pathway | score |
|---|---|
| 3-10 | 0.3393 |
| 11-30 | 0.3107 |
| 31-1000 | 0.3070 |
| mean | 0.3080 |
run python evaluate_reactome.py，and load the trained checkpoint to evaluate。
run python reactome.py to start training。
results are as the following：
ValAcc 0.5008 | ValROC 0.0000 | Pathway 0.3628/0.5804/0.6907
| pathway | score |
|---|---|
| 3-10 | 0.3628 |
| 11-30 | 0.5804 |
| 31-1000 | 0.6907 |
| mean | 0.5446 |
Run python disease.py for training，After training, model will be stored in ./checkpoint/ folder.
Run python evluate.pyfor testing，the testing result will reach about Roc 0.8584 for 3000 epochs and Roc 0.8910 for 7000 epochs.
uncomment the following to save results for disease prioritization according to the checkpoint you choose.
#np.save("./result/logits203.npy", logits.t().detach().numpy()) for eye_epoch07000_valacc0.0000_val_roc0.8910_checkpoint.pt
#np.save("./result/logits202.npy", logits.t().detach().numpy()) for checkpoint.pt
The checkpoint "eye_epoch07000_valacc0.0000_val_roc0.8910_checkpoint.pt" was trained using 203 diseases with label file "gwas_cui_MAPPED_TRAIT_threshold_30_tab_2.txt", where "retinitis pigmentosa" was our recently included disease. To see the 202 diseases result, please load checkpoint "checkpoint.pt", and do the following changes:
in process.py:
(1) line 14:disease_set_path = "gwas_cui_MAPPED_TRAIT_threshold_30_tab_2.txt" change to
disease_set_path = "gwas_cui_MAPPED_TRAIT_threshold_30_tab.txt"
(2) line 33: label = [0] * 203 change to label = [0] * 202
(3) line 111: node_label_list.append([0] * 203) change to node_label_list.append([0] * 202)
in evluate.py:
line 59 n_classes = 203 change to n_classes = 202
in disease.py:
ling 78 n_classes = 203 change to n_classes = 202
- tensorflow==1.14.0
- torch==1.5.1
- networkx==2.5
- scipy==1.5.1
Run python comorbidity_train.py to start training ：set FLAGS.encoder to gat or gcn to select encoder type.
Run python comorbidity_predict.py for testing.
This pre-training code is based on the paper:
Weihua Hu*, Bowen Liu*, Joseph Gomes, Marinka Zitnik, Percy Liang, Vijay Pande, Jure Leskovec. Strategies for Pre-training Graph Neural Networks. ICLR 2020.
- pytorch==1.0.1
- torch-cluster==1.2.4
- torch-geometric==1.0.3
- torch-scatter==1.1.2
- torch-sparse==0.2.4
- torch-spline-conv==1.0.6
Run pretrain_context_predict.py to start context prediction pretraining task.