Information theoretic approaches for inference of biological networks from continuous-valued data

BMC Systems Biology

Table 1 Performance of MI and TE-based methods of GRN inference, presented as the mean AUC (and standard deviation) across a variety of random [46], small-world [47] and scale-free [48] networks from the Mendes ‘Century’ and ‘Jumbo’ collections [42]

Collection	Networks	Nodes	Edges	Topology	AUC (Mutual Information)		AUC (Transfer Entropy)
					Kernel (ARACNE [30])	KSG	Kernel	KSG
CenturyRND	50	100	200	Random	0.514	0.478	0.589	0.603
					(0.030)	(0.028)	(0.024)	(0.027)
CenturySF	50	100	200	Scale-free	0.475	0.505	0.526	0.561
					(0.036)	(0.033)	(0.030)	(0.030)
CenturySW	50	100	200	Small-world	0.477	0.471	0.602	0.598
					(0.035)	(0.035)	(0.028)	(0.030)
JumboRND	5	1000	1000	Random	0.473	0.439	0.540	0.564
					(0.014)	(0.013)	(0.006)	(0.009)
JumboSF	5	1000	1000	Scale-free	0.526	0.577	0.606	0.649
					(0.007)	(0.010)	(0.007)	(0.012)

Kernel-based methods apply the uniform kernel (see (2)) with bandwidth h=0.1. For KSG-based methods, KSG algorithm 1 (better suited to small networks, see (3)) was applied to ‘Century’ data and algorithm 2 (see (4)) to ‘Jumbo’ data, both with K=4 [33] and assuming length-1 Markovian processes. Gene expression time-series were simulated until convergence (d x/d t=0) using Gepasi with default parameters [45]

ISSN: 1752-0509