[bip] Fwd: online sites for doing binding site search

[Mark Robinson]

Hi Andy

On 01/12/2007, Andrew Cameron <acameron at caltech.edu> wrote:
> On a more practical plane I agree with both Titus and Erich.  I've used
> JASPAR and Cluster-buster to search but I haven't been satisfied with the
> results.  It's been difficult to design Cluster-buster parameters to find
> known well characterized modules.  The known ones are usually embedded in a
> forest of false positives, probably due to the laxity of the PWMs used.

One thing I have been working on for a while now is integrating the
results of a range of search algorithms (and more recently other
sources of evidence) using a *suitably trained* machine learning
algorithm such as an SVM. In this way it is possible to reduce the
false positive rate without losing too many false positives in the
process. Seems to me, the implication is that the false positives are
not just a result of "lax PWMs", perhaps they stem additionally from
dependencies that the individual algorithms are unable to entirely
capture?

Mark