From: "Tham Lai San, PHCV2"
Subject: [NMusers] Covariate models of genetic polymorphisms
Date: Thu, 6 Jul 2006 15:37:25 +0800

Dear Colleagues,

I am trying to model the variants of a few genes as covariates. On their own, when
modelled as GRPCL=POP_CL*SizeDescriptor*FGene1 where by the Fgene1(theta) has 3
values for wild type, heterozygous mutant and homozygous mutant respectively,
most have shown some functional effect on CL.

Would like to know if anybody has some prior experience putting in multiple genes
as covariates into the model and how best to approach this?
E.g. GRPCL=POP_CL*SizeDescriptor*FGene1+FGene2 or GRPCL=POP_CL*SizeDescriptor*FGene1*FGene2?
What if the genes are also known to cross-talk?

Best Regards,
Lai San Tham 

From: "Bill Bachman"
Subject: RE: [NMusers] Covariate models of genetic polymorphisms
Date: Thu, July 06, 2006 8:28 am 

The following comments apply to any covariates (as I have no
experience with genetic markers as covariates):

An argument can be made that multiplicative parameterizations are sometimes
found to be easier to get to converge than the linear counterpart (possibly
they are more descriptive of the mathematical relationships found in nature
or that they code for a fractional change in the parameter rather than a
strictly additive component to the parameter).  Many go one step further and
use multiplicative exponential parameterization that better codes for
nonlinear relationships.


If you use the linear parameterization and want to test for
interaction, add another term: something like this



Possibly the best advice I could give is: try the variations yourself.  Others experience
is nice to know, but nothing beats testing your hypotheses yourself.


From: Mark Sale - Next Level Solutions
Subject: RE: [NMusers] Covariate models of genetic polymorphisms
Date: Fri, 07 Jul 2006 09:16:49 -0700

Bill, Lai San Tham

As you seem to suspect, genetic markers (and IMHO many covariates) frequently exert their
effect only when present in combinations.  So, Bill is correct that the interaction between
the genes ought to be tested, as in Bill code
It is entirely possible that you would find that FGene1, FGene2 etc individually have
little to no effect.  However, if you put in

GRPCL=POP_CL*SizeDescriptor+FGene1*FGen2 ; if gene1 and gene2 need to be mutant
or perhaps
GRPCL=POP_CL*SizeDescriptor+FGene1*(1-FGen2) ; if gene1 is to be mutant and gene2 wild type
or perhaps
you might find a very important effect. An effect is seen only when Gene1 and Gene2 are both
mutants, but having a wild type Gene1  compensates for a mutant Gene1, with no effect on the
phenotype. That is, the effects are not independent, but very dependent.  Testing each
efffect individually is likely not adequate. 
However, the number of combination of even a modest number of genes, with only 2 or 3 polymorph
very quickly becomes prohibitively large.  Even in the model that Bill proposed, I'd be concerned
that it was over parameterized and would have convergence problems if all the effects were put
in together, rather than the usual method of one at a time.
I'm not sure if this is what you mean by cross-talk, perhaps it is this sort of interaction.
I usually think of cross talk (in biology at least) as activation/inhibition of one receptor by
some activity of another receptor, even though they aren't heterodimers.  This receptor cross
talk can be mediated by changes in gene expression, but I haven't heard the term cross talk
used when refering to polymorphisms.
Those who read this list server already know about my proposed solution, so I won't bore
them with that again.

Mark Sale MD
Next Level Solutions, LLC


From: "Tham Lai San, PHCV2"
Subject: RE: [NMusers] Covariate models of genetic polymorphisms
Date: Mon, 10 Jul 2006 08:18:53 +0800

Dear Mark, Bill and David,
Being a relatively inexperienced NM user coming from an island where few people
speak the same lingo, you advice is very much appreciated. 
I am investigating the SNP polymorphims for a bunch of transciptional regulators of
CYP450, so to be specific, Mark is right about calling them receptors rather than genes.
This is a preliminary screening exercise to see if my hypothesis holds because I am not
examining the effect of CYP polymorphims on CL but effect of receptor polymorphims on CYP
expression. However, I did notice while screening through the polymorphisms of individual
SNPs that using a binary covariate for mutants is not as good as separating heterozygous
from homzygous mutants.
Meanwhile, still working through the permutations, via the brute force method! :-)
Lai-San Tham, PharmD
Department of Hematology-Oncology
National University Hospital, Singapore