SimVec: Predicting polypharmacy side effects for new drugs

SimVec is a knowledge graph based model to predict polyphamracy side effects for new drugs by enhancing the knowledge graph structure with a chemistry-aware node initialization and weighted drug similarity edges.

Data

All required data files can be downloaded from: https://drive.google.com/drive/folders/1_6khZG4tUs1PnEh9EJLBLxD-uOQq-tyf?usp=sharing

The following data files are available to easily reproduce paper's results:

• Train/val/test split: polyphar_train_new_2.csv, polyphar_val_new_2.csv, polyphar_test_new_2.csv
• Enumeration of drugs and side effects to be used in code: ent_maps.csv, rel_maps.csv
• Single side effects: bio-decagon-mono.csv
• Precomputed molecular descriptors: mol_decsriptors_191.csv
• Precomputed morgan fingerprints: chemical_embed_morgan_fp_3_100.csv
• Precomputed nearest neighbours for drugs: weak_closest.pickle

Usage

Paper's results

To train and test SimVec_full model use the following command:

python run_simvec.py +experiment=simvec_full

Other experiments

• To run a specific experiment (SimVec version) you need to pass an experiment name
• All available experiments can be found in the folder config/experiment. You need to pass a filename (without .yaml) to cmd:

python run_simvec.py +experiment={experiment_name}

Example:

python run_simvec.py +experiment=simvec_se

Changing specific parameters of the experiments

• One can change a specific argument in the corresponding config file (just edit it) or pass a value to cmd

python run_simvec.py +experiment=simvec_se {arg1_name=value1} {arg2_name=value2}

Example:

python run_simvec.py +experiment=simvec_se params.epoch=50 run_args.gpu=False