Network inference for analyzing protein dynamics

Proteins and many other systems are often conceptualized as networks to access analysis methods. Several approaches use molecular dynamics simulations of proteins to construct networks using correlational statistics. However, in the field of network science, it is well-established to solve the inverse problem for a network that can produce the observed correlations. We apply this inverse approach to two adhesion proteins, FimH and Siglec-8, to identify networks that are distinct from correlation networks and instead resemble a contact map. In the inverse networks, covalent interactions are stronger than hydrogen-bonds and non-bonding interactions. This pattern is not present in correlation networks. Moreover, interactions within the backbone dominate the inverse networks, while interactions between sidechains dominate the correlation networks. Due to the differences in the networks constructed by correlation and by solving the inverse problem, there are also differences in topological properties, community detection, and comparing connectivity. While more computationally expensive, solving the inverse problem can remove transitive correlations to produce networks with physically interpretable interactions.