Fig. 4
From: Molecular complex detection in protein interaction networks through reinforcement learning
Example trajectory for finding a complex with the RL pipeline using a learned value function. A This network comprises 7 nodes and 11 edges, each with corresponding with an edge weight. B In this network, a known complex consists of the nodes A, B, C, and E. The goal is to predict this known complex from the network using the learned value function. C A seed edge is identified to begin the walk (edge AB). At this seed edge, the complex is at state (density) S1 = 0.8. D Then, we evaluate all possible neighbors of nodes A and B, i.e., C and D. Adding node C gives a temporary complex {A, B, C} with S2 = 0.57, and a learned value V({A, B, C}) = 0.35, while adding node D gives a temporary complex {A, B, D} with S2 = 0.33, V({A, B, D}) = 0.25. E The neighbor with the highest value function is node C and hence, node C is added resulting in S2 = 0.57, V({A, B, C}) = 0.35. F The next neighbors are evaluated, i.e., D, E, and G. Adding node D leads to S3 = 0.35, V({A, B, C, D}) = 0.2, node E results in S3 = 0.38, V({A, B, C, E}) = 0.36, and node G leads to S3 = 0.35, V({A, B, C, G}) = 0.2. G Since the neighbor yielding the highest value function is node E, this node is added to the complex resulting in S3 = 0.38, V ({A, B, C, E}) = V(0.38) = 0.36. H Each neighbor (D, F, and G) results in a value function less than the current complex {A, B, C, E}. Thus, no neighbor is added, and the predicted community is complete