This article is part of the supplement: Selected articles from the Ninth Asia Pacific Bioinformatics Conference (APBC 2011)
Learning probabilistic models of hydrogen bond stability from molecular dynamics simulation trajectories
1 Mathematical and Computer Sciences & Engineering Division, King Abdullah University of Science and Technology, Thuwal 23955-6900, Saudi Arabia
2 Biomedical Informatics, Stanford University, Stanford, CA 94305, USA
3 Computer Science Department, Stanford University, Stanford, CA 94305, USA
BMC Bioinformatics 2011, 12(Suppl 1):S34 doi:10.1186/1471-2105-12-S1-S34Published: 15 February 2011
Hydrogen bonds (H-bonds) play a key role in both the formation and stabilization of protein structures. They form and break while a protein deforms, for instance during the transition from a non-functional to a functional state. 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.
This paper describes inductive learning methods to train protein-independent probabilistic models of H-bond stability from molecular dynamics (MD) simulation trajectories of various proteins. The training data contains 32 input attributes (predictors) that describe an H-bond and its local environment in a conformation c and the output attribute is the probability that the H-bond will be present in an arbitrary conformation of this protein achievable from c within a time duration Δ. We model dependence of the output variable on the predictors by a regression tree.
Several models are built using 6 MD simulation trajectories containing over 4000 distinct H-bonds (millions of occurrences). Experimental results demonstrate that such models can predict H-bond stability quite well. They perform roughly 20% better than models based on H-bond energy alone. In addition, they can accurately identify a large fraction of the least stable H-bonds in a conformation. In most tests, about 80% of the 10% H-bonds predicted as the least stable are actually among the 10% truly least stable. The important attributes identified during the tree construction are consistent with previous findings.
We use inductive learning methods to build protein-independent probabilistic models to study H-bond stability, and demonstrate that the models perform better than H-bond energy alone.