Duchenne Muscular Dystrophy Drug Reourposing using Knowledge Graphs and eXplainable AI

In this repository you will find the code necessary to run the pipeline shown in: ThesisFinal.pdf. This pipeline is part of a Master Thesis project and it is expected to be improved throughout time.

The objective of this project is that, given a number of seeds related to a target (rare) disease, several drug candidates and explanations will be obtained for the symptoms of the targeted disease.

The environment used to run this project can be created using requirements.txt the In total, three Jupyter Notebooks are used to obtain the predictions and the explanations. The first one, Notebook_1.ipynb, is used to create a Knowledge Graph containing information of the disease. In this first notebook, only information from Monarch, is used. To incorporate drug information (Drug-Target and Drug-Disease) the second notebook, Notebook_2.ipynb, is used. This notebook incorporates information from DrugCentral and Therapeutic Terget Database. The final notebook, Notebook_3.ipynb, is the one that provides the predictions and the explanations. Each notebook contains information about its use, but in the next sections there is a small summary showing its details.

List of files

The description of the files and folders used in this project is the following:

• requirements.txt: list of requirements necessary to load the work environment used in the project
• DMD_KG: Folder containing the files to necessary to create the Knowledge Graphs
• monarch.py: Python file containing the necessary modules to run Bioknowledge Reviewer
• Notebook_1.ipynb: Jupyter Notebook that is used to build the Knowledge Graph
• Notebook_2.ipynb: Jupyter Notebook used to include drug information to the Knowledge Graph
• Notebook_3.ipynb: Jupyter Notebook used to obtain the predictions and the explanations
• edge2vec3.py & transition3.py: Python files containing the modules to run Edge2Vec
• graph_nodes_v2022-01-11.csv: CSV file containing the nodes of the Knowledge Graph created using two seeds: Duchenne Muscular Dystrophy (MONDO:0010679) and DMD (HGNC:2928)
• graph_edges_v2022-01-11.csv: CSV file containing the edges of the Knowledge Graph created using two seeds: Duchenne Muscular Dystrophy (MONDO:0010679) and DMD (HGNC:2928)
• drug.target.interaction.tsv: TSV file downloaded from DrugCentral, contains Drug-Target information
• drug_to_disease.txt: TXT file downloaded from TTD, contains Drug-Disease information
• matched.csv: CSV file containing the TTD diseases matched to the Human Phenotype Ontology (HPO), this was done using SORTA.
• drug_nodes.csv: CSV file containing the drug nodes of the Knowledge Graph
• drug_target_edges.csv: CSV file containing Drug-Target edges of the Knowledge Graph
• drug_disease_edges.csv: CSV file containing Drug-Disease edges of the Knowledge Graph

Graph building (Monarch)

In this first part, the Knowledge Graph is created using information from Monarch, this is done in Notebook_1.ipynb. To extract information from Monarch, Bioknowledge Reviewer is used. As input you will need a number of seeds of the targeted diseases. The seeds are the IDs used by Monarch to identify its elements. You can use the Monarch interface to search for the seeds. In our example, we have used seeds of Duchenne Muscular Dystrophy (MONDO:0010679) and the gene affected innthe disease: DMD (HGNC:2928). As output, two .csv files are obtained, one containing the nodes of the Knowledge Graph and one containing the edges of the Knowledge Graph.

Graph building (DrugCentral and TTD)

The second part is to incorporate drug information from DrugCentral and TTD, this is done using Notebook_2.ipynb. As input, this notebook receives the two csv files generated in the previous step, a csv file containing information from DrugCentral, a txt file containing information from Therapeutic Terget Database and a file matching TTD IDs to Monarch IDs (this file corresponds to drug_to_disease_final (1).csv, and was obtained using SORTA. As output this notebook produces two csv files, one containing all the nodes of the Knowledge Graph and one containing all the edges of the Knowledge Graph.

Predictions and Explanations

In this final part predictions and explanations are obtained, this is done using Notebook_3.ipynb. To obtain predictions a Graph Neural Network (GNN) model was build. This model was created using Pytorch Geometric and DeepSnap. To obtain the explanations a modified version of GNNExplainer is used. This modified version is allows for explanations on a link prediction task. As input, this notebook receives the two csv files obtained in the previous step, and as output in provides a list of drug candidates for each symptom of the disease and a list of explanations supporting these predictions.