DGP-PGTN:

End-to-End Interpretable Disease-Gene Association Prediction with Parallel Graph Transformer Networks

Quick start

• Run DGP-PGTN:

$ python dg_main_sparse.py --epoch 25 --node_dim 32 --lr 0.01 --weight_decay 0.1

The essential python packages were listed in requirements.txt.

Code and data

• dg_model_sparse.py: PGTN model
• gcn.py: GCN model
• dg_main_sparse.py: use the dataset to run DGP-PGTN

data/ directory.

• edges_g.pkl: gene heterogeneous network, which contains five edge types.
• gene_feature.npy: node feature of gene heterogeneous networks, which contains 32-dimensional features of 68061 nodes in the gene heterogeneous network
• edges_d.pkl: disease heterogeneous network, which contains three edge types.
• disease_feature.npy: node feature of disease heterogeneous networks, which contains 32-dimensional features of 19182 nodes in the gene heterogeneous network
• result.npy: disease gene dataset(10000 disease-gene assoications) contains three columns: gene_id, disease_id and their true label.

The file format of the input gene and disease heterogeneous networks is “.pickle”. The data type in the file is a list and the sparse matrices of all association relationships are stored in the list. For examlpe in the gene heterogeneous networks edges_g.pkl:

five types of edges in the gene heterogeneous:	
gene-gene	<with 1051038 stored elements in Compressed Sparse Row format> [gene-gene binary(0 or 1) matrix]
gene-GO		<with 290214 stored elements in Compressed Sparse Row format>  [gene-GO binary(0 or 1) matrix]
GO-gene		<with 290214 stored elements in Compressed Sparse Row format>  [GO-gene binary(0 or 1) matrix]
gene-HPO	<with 182144 stored elements in Compressed Sparse Row format>  [gene-HPO binary(0 or 1) matrix]
HPO-gene	<with 182144 stored elements in Compressed Sparse Row format>  [HPO-gene binary(0 or 1) matrix]

Data statistics

Disease and gene heterogeneous networks are constructed from five data resources, including HumanNet, Human Phenotype Ontology (HPO), Gene Ontology (GO), Disease Ontology (DO) and DisGeNet

Network	Type	Name	Number
gene network	node	gene	21,354
		GO	18,330
		HPO	8,153
	relation	gene-gene	1,051,038
		gene-GO	290,214
		gene-HPO	182,144
disease network	node	disease	6,453
		gene	21,354
	relation	disease-disease	13,444
	relation	disease-gene	86,297

Runtime environment

• Linux terminal and Intel(R) Xeon(R) Silver 4208 CPU @ 2.10GHz, 8 Core(s), 32 Logical Processor(s)