Supplementary Material: Incorporating Prior Knowledge into Network Inference in a Robust, Data Driven Manner

For a given gene, the total number of potential predictors p in BBSR is determined by the size of the union of the 10 highest scoring predictors based on tlCLR and all those predictors that have previously been reported as regulators (see Method section in main article). If p is large (> 10) it becomes infeasible to compute all 2 possible regression models during the model selection step. To further reduce the number of predictors, we look at a subset of of all possible models, and employ an averaging method to discover the 10 most promising predictors. We first build all models containing one or two predictors, and compute the expected BIC for each one. For every predictor we compute the average expected BIC of all models containing that particular predictor. These averages allow us to rank the predictors, and to reduce the set to the 10 best predictors as defined by the BIC in this small subspace of all models.