Figure 2

Networks inferred by MI3 or control methods from a 350-sample synthetic dataset using the following 4 scoring metrics: (a) MI3, (b) dMI3, (c) BN (log conditional probability) and (d) MI2. The best two parent model for each target gene was selected by using different methods and compared to true models. Here our interesting nodes are all the dependent nodes, u1–u6. Local regulatory networks are learned on these nodes and then assembled. When there is no information on dependent versus independent nodes, local networks are learned for all nodes including x1–x3. Conflicting local structures can be resolved in step (2) of Figure 1. For instance, the best two parents for x1 are u3 and u5, which conflicts with the local model for u5 whose parents are x1 and u3. Such conflicts were solved easily based on MI3 score, u3+u5->x1 scores 1.07 while x1+u3->u5 scores 1.49; hence the latter is the true model. The results remained essentially the same for MI3, BN and dMI3, but not for MI2.