Finding transmembrane domains

[A_Morozov]

I need to locate transmembrane domains in a protein, so I've run it trough webservers linked at expasy in "Topology prediction" section (SOSUI, TMHMM, HMMTOP, etc.) And you know what? Every single one of them gives me different positions!
What parameters should I estimate to correctly predict these domains? Perhaps some advice on which servers behave better on which taxa?

[JamieHeather]

Are the positions similar or overlapping at all?

You could also consider using a positive control - pick a protein (or some) for which you know the TMD, run them through, and see which is most accurate.

[A_Morozov]

They are more or less overlapping, though sometimes only by a few AA. The amount of transmembrane domains is also the same in all estimates. This is actually all I need (how many domains there are and whether they are before or after a certain site).
But the general question still remains: if I want to reliably estimate domains having only some protein sequences, say, from genome we are annotating, what tool should I use and what parameters and how should be optimised?

[winsettz]

Is this a single pass transmembrane protein? Is it a homodimer, each with a single TM domain? Is it a GPCR which is seven-pass?