In article <2uvg0g$7a at tierra.santafe.ede>, stadler at SantaFe.edu (Peter
Stadler) wrote:
> Sean Eddy (sre at al.cam.ac.uk) wrote:
>> : By successfully, do you mean "the computer didn't crash in a swap
> : storm", or "I got the right answers"? I suspect the first. My advice
> : to someone wanting to fold a single molecule of 2.5 kb of RNA is
> : "don't do it". (If you have multiple RNAs, and you're using RNA
> : folding programs as a first-pass tool to help you look for a structure
> : they all share, that's a different story.) RNA folding programs
> : sometimes mis-predict 75 nt tRNA structures; I don't think I've seen
> : one ever get a 400nt group I intron right; and I don't even want to
> : think about a 2.5 kb molecule.
hear, hear !!
>> (i) Of course I "successful" means that the computer survived and
> produced a structure that minimized the free energy MODEL -
> That's all a thermodynamics folding algorithm can do by definition.
yes, but as most people think of it as a *real* structure, it can't be said
often enough. And, how well does your model agree with phylogeny/direct
structure mapping based models of rRNA?
>> (ii) The reliability of the structure prediction can - at least in my
> opinion - be enhanced by using the base pair probability matrix
> instead of THE minimum free energy structure. It tells you where
> you have well defined structural elements as opposed to weakly
> defined regions which have lots of very different structures
> with comparable energy. This type of structural information
> is much less sensitive to errors in the thermodynamic parameters
> than the minimum free energy structure itself.
>> (iii) In any case the prediction can't be better than the energy model,
> hence you'll never see pseudo-knots, triple helices etc.
>> (iv) By the way: you DO get the group one intron structure right if
> you force it to make the pseudo-knot where it belongs.
> Predicting pseudoknots is impossible at present for two reasons:
> [1] the computational effort explodes if structures at not
> tree-like any more, and [2] there by no means sufficient data on the
> thermodynamics of pseudoknotted structures.
the predicton can't be better than the energy model only if you want the
minimum free energy model. In my experience, energy minimising programs
(like Zuker's) don't work very well for longer RNAs, because they don't
account for the directionality of folding caused by synthesis (5' -> 3')
and in the case of mRNAs, translation (same). Btw, using other algorithms
to predict secondary RNA structures it is possible to predict for example
pseudoknots using "smaller than life" computers. You don't get the minimum
energy structure, but think about it: a 2.5 kb long RNA is likely to be a)
a mRNA so it gets continually sequentially unfolded by translating
ribosomes, or b) a structural RNA binding lots of proteins which will
influence the structure of the RNA. And it is likely to switch
conformations (active/inactive state), which makes predictions even harder
to make. i have succesfully used the program from the Pleij labs to predict
pseudoknots at ribosomal frameshift sites. I would use energy-minimising
programs only for shorter ( <200 bases ) sequences, preferably non-protein
binding ones. and even then as peter says:
>> I completely agree that whatever the computer throws up cannot be more
> than an educated guess, but sometimes that helps already :-)
>> Peter F Stadler
keep folding,
edwin ten dam
div. virology
cambridge university
etd at mole.bio.cam.ac.uk
