How to connect groups if they are not large enough? I am new at statistics and I have problem that I have no idea how to solve.
I have biological data, which includes expression levels (130 patients) for different alleles (I'm analysing SNP's) of different genes (over 500 000 genes).
First thing I did was finding those genes, for whom I have at least 3 patients with each allel (AA, AB, BB). The thing is that I have to analyse those which have less than 3 patients. 
I have to connect this groups somehow, but I'm not sure how. The only thing I figured out is that I can't connect AA and BB, which from biological point of view is pretty obvious. Only possibilities are to create groups AA+AB or BB+AB, but I have no idea how to choose proper group.
Is there some statistical rule that would allow me to fit one (or two) value into group?
Any help will be appreciated
 A: What you're talking about isn't really statistical, it's more about inheritance types in genetics. Your inheritance model has to stay constant over your whole analysis, otherwise it doesn't really make sense.  You can analyze different inheritance models, but only one at once.
Going further, there are four basic types of inheritance and coding of genetic alleles.
An additive model assumes that the effect of two risk alleles is double that of one risk allele.  Following this logic, AA is coded as 0 (0 risk alleles), AB is coded as 1 (1 risk allele) and BB is coded as 2 (2 risk alleles). This is the default in programs like plink or snptest from oxstatgen. 
A dominant model assumes that effect of one risk allele is the same as having both, and thus AA is coded 0, while both AB and BB are coded 1. 
A recessive model assumes that the effect of 1 risk allele is masked when a ref allele is present. Thus, AA is coded 1, while AB and BB are coded 0.
The general inheritance model is somewhere in between, with the effect of one allele not being exactly half of two.  So, this could be 0 for AA, 0.75 for AB, and 2 for BB, as a rough example.
There's no general statistical rule about how to do this because it's purely based on biology. If you believe the gene has a recessive inheritance, then you would code it thusly. It's interesting to analyze your 500,000 SNPs using different inheritance models and get different results.
Generally in statistical genetics, we exclude SNPs which have a minor allele frequency less than 1%, unless you are specifically analyzing rare variants. I suspect some of the SNPs that you are looking at with only 3 patients in the AA or BB group might be cut out through standard quality control procedures. 
As for what you should do, if your supervisor does not wish you to cut out these low MAF SNPS, or if they don't actually have that low MAF, you can look specifically at which SNPs have these small groups. The small group will either be the AA or the BB (usually the BB unless someone reverse coded your data). I would graph phenotype versus the three genotypes, and see if you observe one of the above patterns. If you do, then you can go ahead and code it as either dominant or recessive.  Make sure to mention the coding when you present the results. 
Do not forget to do multiple testing correction for the 500,000 SNPs that you are analyzing. You will get a lot of false positives just by chance. 
