TABLE 1.

Model performance based on different tissue-specific GIANT networks and PPI networks

Network type	5-fold cross-validationa		Independent testa
	AUC	AUPRC	AUC	AUPRC
GIANT networks
T cells	0.751 ± 0.003	0.554 ± 0.010	0.703 ± 0.013	0.483 ± 0.011
Adipose tissue	0.747 ± 0.002	0.546 ± 0.008	0.703 ± 0.015	0.476 ± 0.014
Epidermis tissue	0.747 ± 0.002	0.552 ± 0.006	0.701 ± 0.025	0.468 ± 0.022
PPI networks
PPI network in this studyb	0.643 ± 0.004	0.502 ± 0.011	0.552 ± 0.020	0.368 ± 0.025
InWeb_InBioMapb	0.669 ± 0.006	0.501 ± 0.010	0.590 ± 0.014	0.390 ± 0.012

^aThe results are based on five different repeats of negative sample selections, which are expressed as average ± SD.

^bWe used the same encoding strategy as the GIANT network to infer the compiled PPI network- or InWeb_InBioMap-based predictive model. Since there are a total of 16,745 proteins in the compiled PPI network, each sample can be converted into a 16,745-dimensional feature vector. Regarding the InWeb_InBioMap PPI network, the number of proteins is 16,948, and thus each sample can be represented as a 16,948-dimensional vector. To train and assess the compiled PPI network- or InWeb_InBioMap-based model, note that some HDFs in the original training and independent test sets were removed since they were not included in these two PPI networks.