Drug repurposing through joint learning on knowledge graphs and literature

Here, we developed a novel method that combines information in literature and structured databases, and applies feature learning to generate vector space embeddings. We apply our method to the identification of drug targets and indications for known drugs based on heterogeneous information about drugs, target proteins, and diseases. We demonstrate that our method is able to combine complementary information from both structured databases and from literature.

Below are the steps for the drugs repurposing pipleine

Requirements

• python 2.7.6
• numpy
• keras
• boost libraries for running multithreaded implementations of randomwalk.

Running

1. Build the graph as described in link

2. The output graph is in the data folder in this repository

3. Before generating the corpus, remove the has-target edges for (Drug target interactions) prediction, and has-indication edges for Drug indications prediction.

python remove_relation_links.py

4. Generate the knowledge graph corpus from the edgelist after removing edges, run

./deepwalk ../data/edgelist_WalkingRDFOWL_has_indication_free.txt ../data/corpus_WalkingRDFOWL_has_indication_free.txt

5. Run word2vec on the generated corpus

python word2vec_gensim.py

6. Normalize the knowledge graph entities with the PubMed abstracts corpus by running

python normalize_text.py

7. Use the the generated corpus from step 5 with Word2Vec to create independent Pubmed abstracts embeddings.

8. Combine the generated corpus from step 5 with the knowledge graph corpus similar to the following and run Word2Vec on the combined corpus.

cat ../data/corpus_WalkingRDFOWL_has_indication_free.txt ../data/medline_abstracts_mapped_drugsrepo.txt > ../data/combined_corpus.txt

9. Run word2Vec on the combined corpus.
10. Run Ind_ann_graph_common.py and other scripts to train the Artificaial Neural Networks with different embeddings from the knowledge graph and PubMed abstracts available in the data folder.

Data

Knowledge graph and literature

The PubMed abstarcts used in this project was downloaded from [Pubtator] (ftp://ftp.ncbi.nlm.nih.gov/pub/lu/PubTator/), The normalization script can be used to normalize the knowledge graph and literature. The normalized corpus used in this study is available upon request. The knowledge graph edgelist is edgelist_WalkingRDFOWL.txt and the mapping to knowledge graph node is mapping_WalkingRDFOWL.txt

Embeddings

embeddings_WalkingRDFOWL_has_indication_free.txt knowledge graph embeddings for predicting drug indications embeddings_WalkingRDFOWL_has_targets_free.txt knowledge graph embeddings for predicting drugs targets drugs_text_embeddings.txt, diseases_text_embeddings.txt and genes_text_embeddings.txt are Medline abstracts embeddings. drugs_embeddings_combined_has_indication.txt, diseases_embeddings_combined_has_indication.txt and genes_embeddings_combined_has_indication.txt are knowledge graph and Medline abstracts jointly trained.

Evaluations and Mapping

All generated embeddings and mapping data used to normalize Literature information to knowledge graph used in this project is available as python dictionary in the data folder. All drug indications drugs2ind_doid.dict and drug targets drugs2tars_stitch.dict evaluations are available as well. The drug indications is from SIDER database. The drug target is from STITCH database. Chemicals alias from STITCH was used to convert drugs mentions in text to STITCH ID available in chemical_map.dict.

Disease ontology was used to extract MESH to DOID mapping in mesh2doid.dict and OMIM to DOID in omim2doid.dict

Predictions

We make drug indications predictions for approved drugs from SIDER available predicted_indications_approved_processed.tsv in the data folder. The first column is the drug ID and drug name, indications disease ontology ID and name, and the prediction score. The full list of the tested drugs and the predicted ranks for indications and targets are included as indications_ranked_graph.txt, indications_ranked_concat_embeddings.txt and indications_ranked_concat_corpus.txt, etc. The first is the drug PubChem ID followed by the diseases and their ranks.

For the complete data including the mapping files, embeddings and normalized PubMed corpus, please download from here

Sample results

The tables below illustrates few examples of the method's ablility to combine complemnetary information betwene the knowledge graph and the literature which result in improved predictions ranks for drugs indications and targets

Drug	Indication	Knowledge graph	Pubmed abstracts	Concatenated embeddings	Concatenated corpora
CID00002678 (Cetirizine)	allergic hypersensitivity disease (DOID:1205)	ranked 34	ranked 4	ranked 1	ranked 10
CID05464096 (Ramiprilat)	cerebrovascular disease (DOID:6713)	ranked 76	ranked 1	ranked 1	ranked 3
CID00002786 (Clindamycin)	impetigo (DOID:8504)	ranked 16	ranked 11	ranked 1	ranked 1
CID00002658 (Cefuroxime)	pneumonia (DOID:552)	ranked 46	ranked 7	ranked 3	ranked 1
CID00004091 (Metformin)	diabetes mellitus (DOID:9351)	ranked 3	ranked 6	ranked 1	ranked 3
CID00003310 (Etoposide)	leukemia (DOID:1240)	ranked 177	ranked 3	ranked 11	ranked 1

Drug	Target (gene Entrez)	Knowledge graph	Pubmed abstracts	Concatenated embeddings	Concatenated corpora
CID00004048 (Megestrol acetate)	2908	ranked 13	ranked 10	ranked 6	ranked 4
CID00004934 (Propantheline)	1131	ranked 91	ranked 13	ranked 1	ranked 1
CID00003155 (Dothiepin)	1129	ranked 62	ranked 26	ranked 19	ranked 1
CID00004666 (Paclitaxel)	7157	ranked 5	ranked 3	ranked 5	ranked 2
CID00003640 (Cortisol)	1551	ranked 13	ranked 20	ranked 3	ranked 10
CID00004594 (Omeprazole)	1544	ranked 53	ranked 18	ranked 7	ranked 2

Citation