This repository is built for generating a automatic pipeline to train KGML-xDTD model based on Snakemake.
- To run this pipeline, please first install conda and then run the following commands:
conda env create -f envs/graphsage_p2.7env.yml
conda env create -f envs/xDTD_training_pipeline_env.yml
## activiate the 'xDTD_training_pipeline' conda environment
conda activate xDTD_training_pipelineYou need to have permission to access the latest config_dbs.json and config_secrets.json from RTX Github Repo. If you have, run the snakemake program will automatically download these two files. Otherwise, you will get an error.
You may need to change the following parameters in the config.yaml before you run the pipeline:
RTXINFO:
GITHUB_LINK: " https://raw.githubusercontent.com/RTXteam/RTX/master" ## you might need to change this linke to specific branch that has correct config_secrets.json and config_dbs.json
KG2INFO:
BIOLINK_VERSION: "3.1.2" ## change this according to what biolink version from which the KG2 that you uses was built.
SYSTEMINFO:
NUM_CPU: 200 ## change this according to your machine configuration
TRAINING_DATA:
MOLEPRO_API_LINK: https://molepro-trapi.transltr.io/molepro/trapi/v1.3 ## please make sure this is the latest Molepro API. Check it from http://www.smart-api.info/ui/940677a65cae38c9a482e54e5c6794f7
MODELINFO:
PARAMS:
GPU: 1 ## if your machine has only one GPU, you should set this to the default value, that is 0.
PARALLEL_PRECOMPUTE:
K: 50 ## You may need to consider your machine RAM to set this parameter. We have 3T RAM to allow it to be 50.
N_drugs: 150
N_paths: 50
BATCH_SIZE: 200
DATABASE:
DATABASE_NAME: 'ExplainableDTD_v1.0_KG2.x.x.db' ## you may want to change it to something like ExplainableDTD_v1.3_KG2.8.0.1.db.You will need a drugbank acoount and request a permission from them to download that file from here.
You can run the following command to run the pipeline:
nohup snakemake --cores 16 -s Run_Pipeline.smk targets &
Please note that the last two steps (e.g., steps 24 and 25) can't be automatically executed in the pipeline since step 23 needs to be run in the background. I have commented the steps 24 and 25. Once step 23 is done, please comment out the steps 24 and 25 part in Run_Pipeline.smk and run the above command again
This step is to download the required RTX config files from the Github server and its internal server. You wil need a permission to download the config_secrets.json from its internal server.
This step is to download the training data training_data.tar.gz and from Zendo, as well as the DrugMechDB yaml file indication_paths.yaml from DrugMechDB.
This step is to download the necessary graph data from the KG2 neo4j endpoint. This step also needs config_secrets.json from its internal server. So please make sure the step1 can successufally downalod this file.
This step is to filter out some nodes with "unused" node types (as least for the drug treatment prediction) and the "SemMedDB" edges based on certain thresholds (e.g., Number of Publication Abstracts and NGD). It will take 1~2 days.
This step is to generate high-quality true positve and true negative training drug-disease pairs.
This step is to generate the ncessary input data for the downstream model training steps.
This step is to process data drugbank.xml file downloaded from DrugBank above. So please make sure you have successfully downloaded this dataset above.
This step is to extract the relationship between drug and genes from both DrugBank data and MolePro data for the downstream model training steps. It will takes a long time to run because it depends on the speed of the molepro API. To avoid calling the molepro API, we use the data molepro_df_bakup.txt collected before in default. But if you want to re-collect it, please delete this file (BUT DON'T PUSH THIS ACTION TO GITHUB).
This step is to check whether there are 3-hop reachable paths between a given drug and disease through a specific gene.
This step is to generate the input path data for the download model training steps
This step is to split data into training, validation, and test data.
This step is to calculate the attribute embedding using the PubMedBert Model.
This step is to generate the input data for runnig GraphSage model. It wil take a few hours.
This step is to generate random walk data for running GraphSage, which will take 3~4 days.
This step is to run GraphSage model to generate the input node embeddings for the Random Forest model below.
This step is tranform the file format of the input node embeddings.
This step it to train a Random Forest model for drug-disease treatment prediction.
This step is to prepare the guided path for model training and conver it to an appropriate format.
This step is pre-trained the ADAC model for drug-disease treatment path explanation.
This step is to formally traing the ADAC model.
This step is to evaluate the model in each training epoch and select the best one.
This step is to split disease into K pieces for download pre-computation.
This step is to call multiple CPUs to do pre-computation for all potential drug-disease pairs.
This step is to build the SQL database.
This step is to build the mapping tables and add them into the SQL database.
If you have any questions or need help, please contact @chunyuma or @dkoslicki.