Deep learning meets ontologies: experiments to anchor the cardiovascular disease ontology in the biomedical literature

Table 5 Exemplifying results for Experiment II: Top twelve ranked candidate terms (highest cosine similarity) from the word embeddings created with CBOW and Skip-gram using two terms as target: “OLR1” from the abstract of the PubMed article with ID = 22,738,689; and “oxidized_low-density_lipoprotein receptor_ 1” that is the CVDO protein class name (rdfs:label) for the CVDO class gene with name (rdfs:label) OLR1. Hence, the target term exploits the protein class expressions within the CVDO

	CBOW		Skip-gram
Rank	Candidate terms from word embeddings	Cosine	Candidate terms from word embeddings	Cosine
1	atherogenesis	0.469405	lectin-like_oxidized_low-density_lipoprotein	0.688603
2	atherosclerosis	0.465861	(LOX-1)_is	0.672042
3	CD36	0.439280	atherosclerosis_we_investigated	0.669050
4	LOX-1	0.424173	receptor-1	0.664891
5	atherosclerotic_lesion_formation	0.416537	lectin-like_oxidized_LDL_receptor-1	0.663988
6	vascular_inflammation	0.414620	lOX-1_is	0.660110
7	inflammatory_genes	0.411186	human_atherosclerotic_lesions	0.657075
8	atherosclerotic_lesions	0.405906	oxidized_low-density_lipoprotein_(ox-LDL)	0.655515
9	monocyte_chemoattractant_protein-1	0.398739	oxidized_low-density_lipoprotein_(oxLDL)	0.654965
10	plaque_destabilization	0.398201	(LOX-1)	0.652099
11	oxidized_low-density_lipoprotein_(oxLDL)	0.397967	proatherosclerotic	0.651571
12	atherosclerosis_atherosclerosis	0.396677	receptor-1_(LOX-1)_is	0.649000

ISSN: 2041-1480