Particle Filtering and Protein Image Recognition
Computer Science Department
University of Wisconsin, Madison
Friday, February 16, 2007, 12:00-1:00
Our work aims to automatically place a complete 3D model of a protein
into an electron density map, a 3D protein image produced by X-ray
crystallography. Our previous work computes, for each amino acid in the protein, an approximate marginal distribution of the location of each
amino acid's center atom. These distributions are computed over the
entire image on a course grid. However, it is more useful to biologists
to have a single physically-feasible protein structure containing the
location of every non-hydrogen atom (or perhaps a small set of such
structures). We investigate the use of particle filtering
(or sequential Monte Carlo) to automatically construct an all-atom model of a protein, using the approximate marginal distributions to successively guide placement of each amino acid in a protein sequence into the 3D electron-density map.
Particle-filtering methods model some probability distribution as the
sum of a finite number of point estimates. In our modeling of proteins,
a "point estimate" refers to an all-atom configuration of some portion
of the protein chain (e.g., from amino acid 20 to 45). At each
iteration, we want to add one more amino acid's atoms to each trace (say amino acid 46), given all the atoms in the partial structure.
We show sequential Monte Carlo produces accurate all-atom models,
resulting in reduced RMS error compared to taking the location for each
amino acid that maximizes the marginal distribution, then placing the
best matching-matching sidechain. Our combined approach produces a more accurate protein model than two other commonly used techniques.
Return to seminar list