Learning Probabilistic Models of Hydrogen Bond Stability from Molecular Dynamics Simulation Trajectories

Hydrogen bonds (H-bonds) play a key role in both the formation and stabilization of protein structures. H-bonds involving atoms from residues that are close to each other in the main-chain sequence stabilize secondary structure elements. H-bonds between atoms from distant residues stabilize a protein’s tertiary structure. However, H-bonds greatly vary in stability. They form and break while a protein deforms. For instance, the transition of a protein from a non-functional to a functional state may require some H-bonds to break and others to form. The intrinsic strength of an individual H-bond has been studied from an energetic viewpoint, but energy alone may not be a very good predictor. Other local interactions may reinforce (or weaken) an H-bond. This paper describes inductive learning methods to train a protein-independent probabilistic model of H-bond stability from molecular dynamics (MD) simulation trajectories. The training data describes H-bond occurrences at successive times along these trajectories by the values of attributes called predictors. A trained model is constructed in the form of a regression tree in which each non-leaf node is a Boolean test (split) on a predictor. Each occurrence of an H-bond maps to a path in this tree from the root to a leaf node. Its predicted stability is associated with the leaf node. Experimental results demonstrate that such models can predict H-bond stability quite well. In particular, their performance is roughly 20% better than that of models based on H-bond energy alone. In addition, they can accurately identify a large fraction of the least stable H-bonds in a given conformation. The paper discusses several extensions that may yield further improvements.