I'm using tools and algorithms borrowed from recent work in deep learning to build better models for protein dynamics.
A side effect of the recent boom in deep learning has been the proliferation of well-engineered libraries for performing arbitrary symbolic differentiation. I’m using these tools to enrich our current dynamical modeling techniques.
In addition, I’m investigating the use of deep networks to find better feature representations of biomolecules.
Sodium channels initiate signaling in neurons and other cells which go on to become sensations of pleasure and pain, as well as thoughts and feelings. Potassium channels return excitable cells to normalcy.
In collaboration with the duBois lab and with the help of Folding@Home, we are conducting large scale molecular dynamics study of a voltage gated sodium channel to complement experimental work to probe sodium channel function through the use of natural toxins secreted by frogs in the Amazon rain forest.
Our lab develops and applies novel statistical techniques to understand and interpret the huge volume of data returned from a molecular dynamics study.
Particularly, we use Markov state models (MSMs) and time-structure-based independent component analysis (tICA) to draw interpretable conclusions from large timeseries data sets.
I’ve introduced a new method for including solvent degrees of freedom in MSM analysis described in this paper.
I’ve introduced a computationally tractable form of non-linear tICA that is analogous to an MSM with soft states here.