Computer Science Department
School of Computer Science, Carnegie Mellon University
Using Tensor Analysis to Characterize
Arvind Ramanathan*, Pratul K. Agarwal**, Christopher J. Langmead
Molecular dynamics simulations provide vast amount of information about a protein's dynamics. To interpret a protein's dynamics and how it may relate to its function, traditionally, two-way analysis techniques such as principal component analysis have been used. However, two-way analysis techniques are usually limited by the fact that they have to be done post-process, i.e., after the simulations have been run and also cannot provide insights into temporal behavior of a protein. To overcome these limitations, we are proposing to use multi-way analysis techniques to understand and interpret protein dynamics as and when the simulations are progressing i.e., online. We model MD simulations in terms of a collection of contact maps and then modeling them as tensors to capture multiple dependencies. Using two recently developed techniques to perform online analysis of streaming data, we illustrate the use of this technique to describe and interpret the behavior of a protein complex in real time. We provide both experimental evidence to support our claims and also discuss the potential advantages and disadvantages of using tensor analysis techniques.
*Joint CMU-Pitt Program in Computational Biology, Carnegie Mellon University