[miso-users] eQTL advice requested

[Singh, Larry (NIH/NHGRI) [E]]

Hi Yarden,

Thanks very much for your e-mail.  I've responded below.

On 8/15/13 4:01 PM, "Yarden Katz" <yarden at mit.edu> wrote:

>Hi Larry,
>
>See below for some thoughts on this topic.
>
>On Aug 8, 2013, at 2:45 PM, Singh, Larry (NIH/NHGRI) [E] wrote:

Š

>There's are generally two subtly different ways of defining isoform
>expression change when dealing with more than two isoforms (in your
>example, four.)  You can ask whether any of the isoforms change; e.g. is
>isoform 1 different between sample A and sample B? Is isoform 2 different
>between sample A and sample B? And view each isoform as a potential unit
>that can be used as an eQTL.

How would you represent a given isoform numerically though?  You can't use
PSI.  I guess that's why I was proposing a Bayes factor relative to a
"reference" sample.  Is there another way?

>
>Another way of defining it is to look for changes in the distribution of
>isoforms. You can imagine a 4-valued vector of isoform expression in
>sample A being different as a distribution from that in sample B, without
>very large magnitude changes occurring in any one particular isoform
>(i.e. one particular value in those vectors.)  Difference in expression
>in this scenario would mean some sort of difference between vectors,
>rather than a Delta Psi and Bayes factor on each of the isoforms in the
>vector.  This two ways are of course related (and will be correlated)
>since the isoform expression values sum to 1.

If I understand correctly, the hope would be that the vectors that are
most similar, by some measurement (Euclidean distance perhaps) should be
closer to each other if they have the same genotype.

>
>I would generally try several variants of this and see how they behave in
>your data.  What you propose makes sense, and you can also do it without
>the Bayes factor if you'd like by comparing relative delta psi values and
>filtering on isoform expression estimates that have narrow confidence
>intervals.  

So you'd just keep the isoforms that have narrow confidence intervals?

>
>These choices will depend on how you expect isoform-specific eQTLs to
>behave.  I can imagine scenarios where among a set of samples, there's a
>shift toward the dominant isoform, and the distribution of Psi values
>(e.g. in a four valued Psi-vector) shifts from something roughly uniform
>to something peaked on one isoform (e.g. [0.25, 0.25, 0.25, 0.25] to
>[0.8, 0.05, 0.05, 0.1]) depend on a genetic variant, but I'm not sure how
>common that is.
>
>You can also imagine different ways of discretizing the multi-isoform
>MISO estimates to something simpler.  In many tissues and for many genes,
>there's only one or a handful of "major" isoforms that account for most
>of the expression.  You can define the "dominant" isoform in your
>reference sample as the one with the maximum Psi-value, e.g. isoform 1 in
>a vector like [0.8, 0.05, 0.05, 0.1] and then see how that one isoform
>changes across your samples.  That's one simple way to do it, which would
>not be sensitive to more subtle variations in the Psi vectors as
>distributions, but could be simpler to interpret biologically.


I guess there are many options here.  I can simplify things to "dominant"
isoform and then cluster all others to "non-dominant".  I'll probably do
as you suggest and just try different possibilities.  Or as you say, just
consider the dominant isoform to try to get a more coarse representation.

Thanks so much for your help and advice.  I'll play with a bit and see
what happens.

Kind regards,
-Larry.