Predictive integration of gene functional similarity and co-expression defines treatment response of endothelial progenitor cells

Table 1 EPCs biosignatures of Ado-treatment response.

Name	Biosignature gene composition	BS	AUC	k
EDQ	Expert-driven queries: CXCR4, CXCL2, CXCL5, CXCL12, CCL7, CCL2, CCL23	7	0.75	-
EDQ+15NN	Expert-driven queries (CXCR4, CXCL2, CXCL5, CXCL12, CCL7, CCL2) together with their most functionally similar genes from integrated k NN strategy*	6	0.92	15
DE	Top-4 data-driven queries: EFNA1, SH3BP5, PEA15, B2M	4	0.75	-
DE+4NN	Top-4 data-driven queries together with their most functionally similar gene from integrated k NN strategy*	4	0.83	4
EDQ+PPI	Expert-driven queries (CXCR4, CXCL2, CXCL5, CXCL12, CCL7, CCL2) together with their interacting partners in the PPI network	6	0.67	-

Top classification performances (summarized with AUC values) obtained from different query-driven schemes and the integrated k NN. AUC values estimated through LOOCV and correspond to SVM classification models (see Methods). Prediction models based on biosignatures identified by standard statistical techniques or query-driven inputs only are also included (models indicated with " - "). "*": In EDQ+15NN, GO-based similarity in integrated k NN model was estimated with BP terms, in DE+4NN with MF terms. BS: Biosignature size defined as the number of inputs to prediction models based on two schemes: individual gene (EDQ and DE models) and integrated gene neighborhood input representations.

ISSN: 1752-0509