Quantification of biological network perturbations for mechanistic insight and diagnostics using two-layer causal models

Table 7 Prediction sensitivities and specificities for the two datasets, GSE7895 ( D ₁ ), GSE19667 ( D ₂ )

Type	Method	CV (D₁/D₂)		D₁→D₂		D₂→D₁		Mean G-perf
Type	Method	Se	Sp	Se	Sp	Se	Sp	Test sets
All Genes	tForwardLd	0.95/0.86	0.93/0.92	1.00	0.87	0.69	1.00	0.88
	NSC	0.92/0.94	0.96/0.93	1.00	0.00	0.98	0.00	0.00
	RF	0.75/0.96	1.00/0.97	1.00	0.31	0.98	0.95	0.76
	SVM	0.84/0.94	0.98/0.96	1.00	0.00	1.00	0.00	0.00
Gene in transcript layer	tForwardLd	0.95/0.89	0.92/0.93	0.97	0.91	0.61	0.85	0.83
	LDA	0.92/0.94	0.80/0.97	0.62	0.27	0.86	0.40	0.50
	NSC	0.96/0.92	0.93/0.95	1.00	0.00	1.00	0.00	0.00
	RF	0.87/0.96	1.00/0.97	0.98	0.58	0.94	1.00	0.86
	SVM	0.88/0.95	0.98/0.95	1.00	0.00	0.78	0.15	0.17
UBE downstream genes	CORG + LDA	0.97/0.94	0.90/0.94	0.83	0.36	0.61	0.55	0.56
	CORG + NSC	0.98/0.95	0.93/0.96	0.97	0.00	0.67	0.30	0.22
	Best LDA (8-Methyl-IQX)	0.96/0.95	0.88/0.96	0.9	0.80	0.80	1.00	0.89
	Best NSC (8-Methyl-IQX)	0.96/0.86	0.92/0.98	0.98	0.76	0.86	1.00	0.90
Backbone values	tForwardLd	0.97/0.90	0.95/0.97	0.8	0.82	0.84	0.90	0.86
	LDA	0.96/0.92	0.94/0.98	0.89	0.80	0.84	0.90	0.86
	NSC	0.93/0.93	0.81/0.92	0.95	0.87	0.94	0.90	0.91
	RF	0.93/0.91	0.80/0.91	0.97	0.73	0.88	0.85	0.85
	SVM	0.93/0.93	0.88/0.91	0.98	0.62	0.88	0.90	0.83

The predictions of the samples from the dataset D_j based on the model trained on the dataset D_i are reported in the columns D_i→D_j. While not systematically having the best cross-validation performance, the predictors based on the backbone values show are more robust behavior when predicting one dataset based on the model trained on the other one. The mean of the G-performances ( $= \sqrt{Sp \cdot Se}$ ) over the two independent test sets are shown in the rightmost column and is highlighted if >0.7. The best UBE-based models is chosen based on the mean cross-validation G-performance for the two datasets. The algorithms based on the backbone values leads systematically to good performances. The cross-validation standard errors are reported in Additional file 1: Table S1.

ISSN: 1471-2105