MUFFIN

Overview

This repository contains source code for our paper "MUFFIN: Multi-Scale Feature Fusion for Drug–Drug Interaction Prediction".

Dataset Preparation

You need to provide datasets defined as below:

1. Approved Drug SMILES file:

ALl we want to consider is the FDA-approved drug, the form is look like:

Compound::DB00119 CC(=O)C(O)=O

In our work, we have 2322 drugs. Just 2322 lines in this file.

2. DDI dataset file:

'DDI_pos_neg.txt': store the DDI dataset, the form is "drug1 \t drug2 \t type". For binary data: type is in {0,1}, for the multi-class DDI dataset "multi_ddi_sift.txt", type ranges from 0 to 80, and for the multi-label dataset, it is in {0-200}.

For the TWOSIDES dataset, you can obtained from http://tatonettilab.org/offsides/

you can also use following command to get the total multi-label DDI dataset.

wget http://tatonettilab.org/resources/nsides/TWOSIDES.csv.xz

For the DrugBank dataset, you can obtained from https://go.drugbank.com/releases/latest

3. Knowledge Graph file:

DRKG : 'train.tsv' which is defined as "h \t r \t t" id form

you can get DRKG dataset from https://github.com/gnn4dr/DRKG or just download files using command below:

wget https://dgl-data.s3-us-west-2.amazonaws.com/dataset/DRKG/drkg.tar.gz

Now, you get "drkg.tsv" file, put it into the "./data/DRKG" directory. Next you need to change the entity ID inconsitent with your drug smiles file (in first step, you get it).

just run the code below:

python process_raw_DRKG.py

After that, you can get "entities.tsv","relation.tsv" and "train.tsv".

4. Customize your dataset:

you can use your dataset follow the form as ours.

you can define your KG dataset file name as "kg2id.txt" and DDI dataset name as "DDI_pos_neg.txt".

Pretrain your data Preparation

1. Graph-based drug embedding:

you can use "pretrain_smiles_embedding.py" file to generate your drug embedding, which is at last shown at ".npy" form in the data/DRKG directory.

try this code:

python pretrained_smiles_embedding.py -fi ./data/DRKG/your_smiles_file.csv -m gin_supervised_masking -fo csv -sc smiles

-fi : your smiles file position

-fo : your file's type: txt or csv

-sc : the smiles "column name" in your smiles file.

when you run this code, you can then get the final Graph-based drug embedding.

tips: For convenient, the drug smiles order(in the "npy" file) is consistent with the KG entity, which means if you have 2322 drugs, and the graph-based embedding [ID：0-2321] is the same as the 2322 former KG entity embedding [ID:0-2321]

2. KG-based drug embedding:

We use DGL-KG tools to train our KG entities. if you want to generate KG-embedding, just run the following code according to your needs:

python pretrain_kg_embedding.py --model_name TransE_l2 --dataset DRKG --data_path data/DRKG/ --data_files entities.tsv relations.tsv train.tsv --format udd_hrt --batch_size 2048 --neg_sample_size 128 --hidden_dim 100 --gamma 12.0 --lr 0.1 --max_step 100000 --log_interval 1000 --batch_size_eval 16 -adv --regularization_coef 1.00E-07 --test --num_thread 1 --gpu 1 2 --num_proc 2 --neg_sample_size_eval 10000 --async_update

Now, you can get the KG-based embedding named "DRKG_TransE_l2_entity.npy", just moved it into your dataset directory!

Environment Setting

This code is based on Pytorch 3.6.5. You need prepare your virtual enviroment early.

Running the code

You can run the following command to re-implement our work:

python main.py

what you need to do according to your customized file:

1. change “—graph_embedding_file” name with “your smiles file position”
2. change “—entity_embedding_file” name with “your entity file position”
3. change “—relation_embedding_file” name with “your relation file position”
4. change “—out_dim” name with “1 or 81”, which is depends on your task “binary or multi-class”
5. change “—multi_type” name with “False or True”, which is depends on your task “binary or multi-class”

Cite

@article{chen2021muffin,
  title={MUFFIN: multi-scale feature fusion for drug--drug interaction prediction},
  author={Chen, Yujie and Ma, Tengfei and Yang, Xixi and Wang, Jianmin and Song, Bosheng and Zeng, Xiangxiang},
  journal={Bioinformatics},
  year={2021}
}

Contact

If you have any question, just contact us.

Chinese version is provided here.