Table 1.

LEM outperforms existing network inference algorithms on both in silico and biological data

Network	# Nodes	LEM (AUC)	Inferelator (AUC)	Granger Causality (AUC)	Hill-DBN (AUC)	Jump3 (AUC)
In silico 1	3	1.0000	0.9000	0.7000	0.5000	0.9000
In silico 2	3	1.0000	0.5667	0.8111	0.3667	0.7222
In silico 3	5	0.9900	0.7857	0.7791	0.4003	0.6794
In silico 4	10	0.8884	0.5541	0.5949	0.5131	0.7727
In silico 5	20	0.8781	0.6789	0.7441	0.6770	0.7540
Yeast cell-cycle 1	17	0.8693	0.6705	0.6893	0.6253	0.6481
Network	# Nodes	LEM (MCC)	TD-ARACNE (MCC)	Banjo DBN (MCC)
In silico 1	3	1.0000	0.0000	−0.5000
In silico 2	3	1.0000	0.0000	−0.5000
In silico 3	5	0.7379	0.4528	−0.0624
In silico 4	10	0.7463	0.0636	0.0294
In silico 5	20	0.5908	0.2147	0.0086
Yeast cell-cycle 1	17	0.0478	0.0292	−0.0380

Using in silico networks 1–2 (Fig. 2) and 3–5 (Fig. 3), as well as a yeast cell-cycle network (Fig. 4), we compared LEM performance to existing algorithms. AUC-ROC scores labeled (AUC) were used to compare the performance of LEM to Inferelator, Granger Causality, Hill-DBN and Jump3. Matthew’s correlation coefficient (MCC) was used to compare LEM to TD-ARACNE and BANJO, which are binary classifiers and do not output numerical scores for network edges. No biological prior information was used for this comparison. Using dynamics data from each network, LEM better approximates the underlying network model than the other algorithms. See Additional file 1: Section 5 for a complete explanation of AUC-ROC and MCC scoring

AUC area under the curve, LEM Local Edge Machine, MCC Matthew’s correlation coefficient, ROC receiver-operating characteristic