I would make one model including the interactions of SNPs and risk factors.

I would use a forward stepwise regression model which would ultimately include an interaction variable.

Stepwise is the perfect atheoretical approach, and thus the most rigorous test. 

The problem you face is that the genotype may interact with risk factors, but it also might be related to them.  To the extent that they overlap, your risk factors might bump the genotype out of the equation because the risk factors explains marginally more variance.  It would not be correct to conclude that the genotype was irrelevant, if that is the case.  

Putting all of the SNP and RF and their interactions into one huge glorious model is probably not going to give you much that could be interpreted, but you could go that route with a SEM.  It would even allow you to test mediation (first level of causality is SNP, second is RF, and have an arrow going from SNP to RF as well arrows from SNP/RF to the outcome.  It is a categorical outcome I assume (unfortunately), so use the bootstrapping/Bayesian option.

Alternatively, I'd do the stepwise, then go back and construct blocks of any significant SNPxRF that includes the whole package (SNP, RF, SNPxRF) and enter them hierarchically with the understanding that SNP is not likely to be mutable and thus deserves primacy in the causal model.  State you are doing this way due to limitations of the pure stepwise model that cannot recognize which variables are fixed and which is mutable (SEM allows you to state that, but SEM has its own weaknesses).

Thank you all for your insights. It was of great help.