This repository aims to simplify the drug-target interaction prediction process which is based on protein residue interaction networks (RINs)
The repository aims to go from a simple collections of inputs - structures of proteins, interactions data on drugs to a fully-function GNN model
- clone the repository with
git clone https://github.com/ilsenatorov/rindti - change in the root directory with
cd rindti - (Optional) install mamba with
conda install -n base -c conda-forge mamba - create the conda environment with
mamba env create -f workflow/envs/main.yml(might take some time) - activate the environment with
conda activate rindti - clone rinerator repository into your home dir with
git clone --branch perf https://wibi-git.helmholtz-hzi.de/ske18/rinerator/ ~/rinerator - install rinerator with
pip install ~/rinerator
Snakemake pipeline (located in workflow directory) is responsible for parsing all the data.
Usually the result is a single pickle file containing all the necessary information for training, testing etc.
The pipeline relies on the resources folder which should contain resources/structures/ folder with pdb files for acquiring protein networks.
For DTI prediction the model also needs csv files in resources/drugs, more info will be provided later.
The pipeline is controlled by a config file config/snakemake.yaml which can be modified to change the output of the pipeline
In order to obtain the pickle file necessary for training the DTI model please ensure the only_proteins entry in the yaml config file is set to false.
Once pipeline is run with snakemake -j <num_cores> --use-conda, it will generate a pickle file in the results/prepare_all/ directory.
In order to omit the drug data, one can set the only_proteins to true in the config file.
Running the pipeline with snakemake -j <num_cores> --use-conda will generate a pickle file in results/prepare_proteins/ directory.
Since snakemake sometimes struggles with calculating DAG for really large pipelines, steps of this pipeline could be ran manually
- For rinerator-based networks please run the rinerator on all the necessary pdb files, then use the
workflow/scripts/prepare_proteins.pyscript to generate the necessary pickle file. - For distance-based networks please run the
workflow/scripts/distance_based.pyscript to generate the necessary pickle file.
Information on both scripts can be found through workflow/scripts/distance_based.py --help or workflow/scripts/prepare_proteins.py --help
The GNN model is shipped as a package, located in the rindti directory.
Additionally, scripts that give an example of model training are provided (train.py and pretrain.py), which can be used as plug-and-play, or as inspiration to create custom training approaches.
A model can be trained with a simple train.py results/prepare_all/<pickled_data>.pkl, assuming pipeline has been ran in advance.
Three types of models are currently available - class, reg and noisy nodes model.
Both class and noisy nodes require the final label to be a bool value (0 or 1), however noisy nodes introduces corruption of nodes and auxiliary loss of predicting the original labels of such nodes.
In order to pretrain the model one can use the pretrain.py script, which has three types of self-supervised pretraining models available
GraphLog and Infograph can be used directly with the result of prepare_proteins step of the workflow (located in the results/prepare_proteins directory).
Pfam siamese network will require adding the information on the protein family to that dataframe in order to function.