We reconstruct the interaction atlas for humans using a Bayesian network-based approach utilizing existing external knowledge. By “atlas”, we mean the complete or almost complete (on the order of thousands of nodes) interactome of an organism. We have developed a construct, called the Bayesian Network Prior (BNP) that represents the acquired scientific knowledge in a distilled manner to be utilized in the Bayesian network structure learning process. The BNP serves as a proven way to incorporate external knowledge in network learning from experimental data.

We provide a divide-and-conquer approach that first identifies groups of molecules that exhibit dependency based on experimental data. Within each group, the corresponding interaction network is learned using external knowledge via the BNP framework. Each group is summarized using one representative node, and all of these nodes are used to build a meta-network that represents the interactions between the groups. The union of the nodes in pairs of interacting groups undergoes a second learning phase, and the ensemble of all of the learned edges represents the final interaction atlas.