NSF4SL: Negative-Sample-Free Contrastive Learning for Ranking Synthetic Lethal Partner Genes in Human Cancers
- This is the code for our paper ``NSF4SL: Negative-Sample-Free Contrastive Learning for Ranking Synthetic Lethal Partner Genes in Human Cancers''.
- NSF4SL is a negative-sample-free model for prediction of synthetic lethality (SL) based on a self-supervised contrastive learning framework.
git clone git@github.com:JieZheng-ShanghaiTech/NSF4SL.git
cd NSF4SL
pip install -r requirements.txt
SL Data
We test the model performance under 3 kinds of 5-fold cross-validation:
- CV1: we split the train/test data by SL pairs, and both genes of a testing pair would be likely to appear
in the training set.
- data path: .data/pair_split_cv1
- CV2: the data is split by genes, where only one gene of a testing pair is present in the training set.
- data path: .data/gene_split_cv2
- CV3: the data is split by genes, where neither genes of a testing pair is present in the training set.
- data path: .data/gene_split_cv3
The format of one input sample is (gene1, gene2). Please modify the data path in main.py if you would like to test different CVs.
Knowledge Graph Embedding
We train the SynLethKG embedding based on DGL-KE. The format of one input sample is (entity1, relation, entity2). The pre-trained KG embedding is under the data path: .data/kg_embed.
python main.py --aug_ratio=0.1 --batch_size=256 --gpu=0 --latent_size=256 --train_ratio=0.8
Illustration of the above hyper-parameters:
-------------------------------------------
aug_ratio # ratio of augmentraion for each gene feature
batch_size # batch size for training
gpu=0 # ID of GPU
latent_size # latent size for encoders and predictor
train_ratio # ratio of training data
The code is inspired by BUIR.
Bootstrapping user and item representations for one-class collaborative filtering
Lee, Dongha, et al. Bootstrapping user and item representations for one-class collaborative filtering, Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2021.
SL data and SynLethKG are constructed based on SynLethDB 2.0.
SynLethDB 2.0: A web-based knowledge graph database on synthetic lethality for novel anticancer drug discovery
Wang, Jie, et al. SynLethDB 2.0: a web-based knowledge graph database on synthetic lethality for novel anticancer drug discovery, Database, Volume 2022, 2022.