Extracting a simplified dependency structure to build a tractably trainable GPM. To overcome the intractability shown in Figure 2B, we need to simplify the edge structure of the resulting complete multi/metagraph such that it contains only the “most important” edges representing dependencies in the training data. Here we have shown a subset of the most important dependencies present in the data shown in Figure 1. While edge weights are not shown here, it is important in a working interface to provide the user with edge-weight information, and to avoid arbitrarily filtering edges based on their magnitude. To a biological end-user, small edges between infrequently occurring subnodes can be as important as larger edges between common subnodes, depending on the features they connect. Edges are colored based on disjoint subnetworks of dependencies.