Analysis of variables of varying numbers

Question

I work with amino acid sequences and I want to use a self-made model to tell me something about it, lets call it $f(\text{seq})$. Now i want to know the contribution of every position in the sequence onto the model, i.e. my question is what is the importance/effect of amino acid $A$ occuring at position $I$ in the sequence with respect to the model?

How do I visualize something like that?

I want to use my model also on several sequences of differing lengths. Somehow this throws a monkeywrench into my plans of using a neural net...

My question is simple yet I did not find anything about it. Pointers would be appreciated. Or any comment you might have. This whole idea of mine is pretty unfinished and I don't really know yet what I want. So feel free to criticize, I will update the question accordingly.

Ah and if this is the wrong place to put this here please tell me also (:

cheers and thanks

note: this is a repost from th.stackexchange.

Answer 1

This is a tricky topic in general; because of the size variability you mentioned it rather boils down to motif finding. Yet I can propose four ideas:

1. Supposing that you can align sequences (ok, you always can, but let's say it yields non nonsense consensus), you can reduce the problem to the conserved region and make attributes out of aligned residues. Finding important residues will then be a simple feature selection.
2. Look for motifs using MEME or something more fancy to make PWM and just hope it will work. Sequence logo will show you the residue importance based on the predictive power it applies.
3. Make attributes out of N-gram spectra, i.e. counts of all possible N-residue words for few values of N (till your computer won't explode). Or use SVM with string kernel, it usually boils to the same thing.
4. << Can't talk about it yet, research in progress ;-) >>