1
$\begingroup$

I am analyzing the effect of one putative tumor suppressor (and its loss) on a specific cancer phenotype. This putative tumor suppressor is known in literature to have an effect on the growth of tumors in mouse models.

However, in the same region of this gene there's a second well-known tumor suppressor - actually a very powerful one. By looking at the data from cBio portal, there's a very strong co-occurrence of loss on both tumor suppressors in a pan-cancer analysis, that is they are almost always lost together.

Therefore, I would like to disentangle the effect of the two genes, and to analyze only the effect of my putative candidate on the phenotype. The simplest approach is to use only the samples that have a single loss (only the loss of my putative candidate, but not the loss of the confounding tumor suppressor). However, only ~10% of all samples shows such behavior, and this 10% is distributed across all cancer types, decreasing the statistical power of the analysis - since I would like to test each cancer type separately to avoid mixing incompatible data.

Which kind of strategy/analysis can be used in this case?

$\endgroup$
1
  • $\begingroup$ How do you model the effect of the genes? I would use all the data (or cancer specific data) and indicate which sample have lost the gene of interest, the other or both genes. That way you could disentangle the effect of each gene and the interaction of both $\endgroup$
    – llrs
    Feb 20 '19 at 8:23
2
$\begingroup$

Maybe you could fit a mixed model to predict phenotype, using the two genes as fixed effects and cancer type as a random effect, and ask the model for the separate effects of each gene, and the interaction between them.

$\endgroup$
1
$\begingroup$

Or even more simply using generalized linear modeling (GLM), where the regression weights are the genes and the accuracy of the model is measured in the residual. You also Q-Q plots to assess the fit of the model to the data.

If the residual is high and Q-Q "wobbly" transformations works wonders.


Machine learning such as a neural network is a very cool way to do this calculation, but IMO you need Python - although people on this board do this in R. The concept is very easy you are catagorising (keep in mind I'm not an oncologist) gene vs. cancer resistance and you want neural network, linear-based regression etc.. to make the call (you will see it visually). You then use the weights to assess the relative contribution of each gene.

The 10% issue is formally assessed, its not hard to obtain the power because its a common problem. The formal way to assess this, from memory, is you assign a "null model" and the accuracy of a ML model needs to a lot higher than the null model. I forget the details, I've never been in these situations. There's a specific technical name (which I should look up).


The explicit GLM / GLMM (non-ML) is the easiest approach however. If you use a Mac there is a great (relatively cheap) package for GLM, but any good software package will do. I assume GLMM is available in R (certainly Stata).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.