Table 3. The top hit@10 results (partial and free settings) for relation prediction of our biological knowledge graph as feature generators models.

Partial and free settings are two evaluation strategies correspond to: partially removing 80% of the relation edges and completely removing all relation edges, respectively.

Relation	Walking RDF/OWL	TransE	Poincaré	Rescal	SimplE	R-GCN
	partial
has function	12.97	10.51	7.36	5.62	19.20	9.35
has interaction	11.51	13.18	3.75	8.62	8.60	11.40
has disease annotation	31.10	26.31	21.76	19.40	24.02	20.42
has sideeffect	22.42	25.70	5.70	17.07	17.26	17.21
has indication	26.09	18.40	9.77	13.09	14.66	11.83
has target	15.32	13.70	7.68	5.50	20.28	10.93
has gene phenotype	15.23	14.89	8.67	8.59	8.97	10.07
has disease phenotype	34.69	19.86	11.36	10.32	8.61	7.22
	free
has function	3.16	2.54	0.43	0.75	2.31	2.56
has interaction	4.14	2.53	1.84	0.74	2.81	0.89
has disease annotation	22.51	21.90	8.02	14.24	10.96	15.75
has sideeffect	18.14	21.20	0.05	13.94	5.82	6.77
has indication	14.78	11.95	5.25	9.61	9.09	13.40
has target	0.24	0.26	0.06	0.12	1.73	0.63
has gene phenotype	9.08	9.61	0.94	6.80	1.94	4.65
has disease phenotype	10.83	4.73	1.10	3.05	1.59	1.89

Note:

The highest performing method is indicated in bold.