HTINet2: Herb Target Prediction via Knowledge Graph Embedding and Residual-like Graph Neural Network

Overview

We proposed a deep learning-based target prediction framework termed HTINet2, which designed three key modules, namely, KG Construction and Embedding Learning, Graph Representation Learning, and Target Prediction.

Environment Requirement

If you want to use env of anaconda:

conda create -n HTINet2
conda activate HTINet2
pip install -r requirements.txt

Usage

Knowledge Embedding

The TMKG knowledge graph we constructed comes frome multiple sources. These databases include SymMap, soFDA, STRING v11, KEGG, GO of gene ontology, the TCM books consist of “ChP 2020”, “100 Classic Famous Formulas”, “TCM Diagnostics”, “Treatise on Febrile Diseases”, “Formulas of TCM”, and “Internal and External Women and Children’s Diseases”.

Using OpenNE methods to learn the deep knowledge embedding of TMKG for herbs and targets. The output using LINE method is located at /Knowledge_embedding.

Graph Representation

Dataset

• training dataset training_set.npy, testing dataset testing_set.npy, validation dataset val_set.npy, they all contain three parts, the genes corresponding to herbs, the herbs corresponding to genes, and the total number of genes corresponding to herbs. All dataset set_all.npy is all genes corresponding to herbs.
• in datasets, each id represents a herb or target. The relationship between the id and their names can be found in name2id.
• graph representation component inspired by LR-GCCF, and the initialized embeddings of GNN are the result of pre-training of the Knowledge Embedding module /Knowledge_embedding/herb_LINE.txt and /Knowledge_embedding/gene_LINE.txt.

Running

• train

cd code
python train.py

• test

cd code
python test.py

To adjust the dimension of the model, it is necessary to also modify the dimension of the pre-training embeddings in the Knowledge Embedding stage. This step ensures consistency of embedding dimensions within the framework, especially since the framework is not end-to-end.