Awesome
DeepAEG
DeepAEG: A model for predicting cancer drug response based on data enhancement and edge-collaborative update strategies
Requirements
Installation
DeepAEG can be downloaded by
git clone git@github.com:zhejiangzhuque/DeepAEG.git
Installation has been tested in a Linux/MacOS platform.
Environment
conda env create -f DeepAEG.yml
conda activate DeepAEG
Model implementation
Step 1: gene data Preparing
Four types of raw data are required to generate genomic mutation matrix(the order is: copy number, Gene expression, Gene methlation, Gene mutation).
Data of the same class were saved as after normalization（data/CCLE/Result_StandardScaler.csv）, Four items of data were combined into a four-element group.
The Four types of raw data files can be downloaded from CCLE database.
Step 2: drug data Preparing and Feature extraction
Each drug in our study will be represented as a graph with nodes and edges, and we collected a total of 221 drugs. Here, we use the deepchem library to extract node features and graphs of drugs.
drugfeature=[node_features adj_np, adj_np_01,smiles_feature]
node_features：features of all atoms within a drug with size 50
adj_np：adjacent list of all atoms within a drug， The elements therein are replaced by the eigenvector of that edge with size 11.
adj_np_01: adjacent list of all atoms within a drug. It denotes the all the neighboring atoms indexs
smiles_feature: The resulting features are extracted directly from SMILES by Transformer
Selection of feature extraction for data enhancement or not
python extract_drug_features_auc.py -use_aug True -aug_num 2
python extract_drug_features_auc.py -use_aug False
[-use_aug] Whether to conduct data augmentation (default: Fasle)
[-aug_num] The number of each SMILES recombined into new virtual SMILES (default: 2)
Step 3: Conduct training
python run_DeepAEG_newaug2.py