Speaker
Prof.
Erik Aurell
(KTH)
Description
Inverse statistical mechanics means to determine model
parameters (couplings, external fields etc) from
observations (one- and two-point correlations, or other
data). In the course of an ongoing investigation into
methods to do the "inverse Ising" problem we recently found
that a pseudo-likelihood method works better than many other
alternatives; it is particularly good in the parameter range
of strong interactions and few samples [1].
We have recently tried to extend this method to determine
amino acid contacts from protein sequences in the same
protein family, following the approach of Morcos et al [2].
The relevant model is then a Potts model with 20 or 21
states. We find that here also the pseudo-likelihood
provides somewhat better reconstruction of known protein
structures [3].
This is joint work with Magnus Ekeberg and Martin Weigt.
[1] Erik Aurell, Magnus Ekeberg "Inverse Ising inference
using all the data", Physical Review Letters (2012, in
press) [arXiv:1107.3536]
[2] Faruck Morcos et al, "Direct-coupling analysis of
residue coevolution captures native contacts across many
protein families", PNAS November 21, 2011
[3] Magnus Ekeberg, Erik Aurell, Martin Weigt (2012, in
preparation)
