README_relation_classification.md

Relation Classification

Performance against held-out test sets of tuned BERT models used to classify sentences for particular relations between entities are shown in the table below.

Note that the predicates shown below are shortcuts for the qualified predicates used in the KG knowledge representation.

Biolink:Chemical To Disease Or Phenotypic Feature Association

relation	TP	TN	FP	FN	Precision	Recall	Fscore
causes_or_contributes_to	126	1965	37	34	0.773	0.787	0.780
treats	327	1694	90	51	0.784	0.865	0.823
no_relation	1510	466	72	114	0.954	0.930	0.942

Biolink:Chemical to Gene Association

relation	TP	TN	FP	FN	Precision	Recall	Fscore
negatively_regulates	944	8355	97	89	0.907	0.914	0.910
positively_regulates	323	9038	59	65	0.846	0.832	0.839
no_relation	7926	1285	136	138	0.983	0.983	0.983

Biolink:Gene Regulatory Relationship Association

relation	TP	TN	FP	FN	Precision	Recall	Fscore
negatively_regulates	45	991	9	13	0.833	0.776	0.804
positively_regulates	76	962	7	13	0.916	0.854	0.884
no_relation	901	127	20	10	0.978	0.989	0.984

Biolink:Gene to Disease Association

relation	TP	TN	FP	FN	Precision	Recall	Fscore
contributes_to	122	42	26	20	0.824	0.859	0.841
false	42	122	20	26	0.677	0.618	0.646

Biolink:Gene to Disease Association (Loss/Gain of Function)

relation	TP	TN	FP	FN	Precision	Recall	Fscore
contributes_to_via_gain_of_function	7	391	3	0	0.700	1.000	0.824
contributes_to_via_loss_of_function	43	305	16	37	0.729	0.537	0.619
no_relation	296	51	36	18	0.892	0.943	0.916