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

Table 8 Prediction sensitivities and specificities for the two cohorts, A and B

Type	Method	CV (A/ B)		A→B		B→A		Mean G-perf
		Se	Sp	Se	Sp	Se	Sp	Test sets
	tForwardLd	0.38/0.63	0.80/0.65	0.50	0.82	1.00	0.88	0.79
	From Arijs, 2009	Accuracy: 0.92/0.91		0.25	1	Accuracy: 0.71		na
All Genes	RF	0.20/0.82	0.88/0.73	0.25	0.91	0.62	0.75	0.58
	SVM	0.52/0.78	0.85/0.69	0.42	0.82	0.62	0.75	0.63
	NSC	0.48/0.78	0.80/0.58	0.67	1.00	0.69	0.88	0.80
	tForwardLd	0.50/0.68	0.83/0.62	0.67	1.00	0.88	0.69	0.80
	LDA	0.43/0.90	0.84/0.65	0.42	0.82	0.88	0.62	0.66
Gene in transcript layer	NSC	0.60/0.88	0.79/0.58	0.75	0.73	0.88	0.56	0.72
	RF	0.28/0.85	0.88/0.71	0.33	0.82	0.88	0.69	0.65
	SVM	0.52/0.77	0.86/0.69	0.42	0.82	0.88	0.69	0.68
	CORG + LDA	0.33/0.73	0.68/0.62	0.25	0.64	0.75	0.69	0.56
	CORG + NSC	0.62/0.90	0.78/0.60	0.50	0.82	0.88	0.69	0.71
UBE downstream genes	Best LDA (MAP3K1)	0.45/0.98	0.78/0.69	0.58	0.91	1.00	0.75	0.80
	Best NSC (catof(TLR2))	0.85/0.92	0.80/0.56	0.75	1.00	1.00	0.69	0.85
	tForwardLd	0.53/0.82	0.75/0.75	0.50	0.64	1.00	0.62	0.68
	LDA	0.75/0.75	0.76/0.75	0.50	0.73	0.88	0.56	0.65
Backbone values	NSC	0.98/0.98	0.75/0.67	0.92	0.82	1.00	0.62	0.83
	SVM	0.78/0.78	0.40/0.77	0.91	0.50	0.69	0.88	0.73
	RF	0.75/0.82	0.62/0.78	0.82	0.75	0.75	0.88	0.80

The predictions of the samples from the cohort A (B respectively) based on the model trained on the cohorts B (A respectively) are reported in the columns A→B (B→A respectively). 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 to good performances for a majority of algorithms. The cross-validation standard errors are reported in Additional file 1: Table S2.

ISSN: 1471-2105