Overcoming small dataset anomalies in genetic algorithm So I am currently making my 6th version of a model designed to predict the likelihood of of a particular medical condition based on a multifactorial genetic markers and I really would appreciate some advice with regards to a method to overcome this problem. The model I have used thus far has proved very successful, but I know there is still one thing I am not accounting for statistically. 
The data I have for this medical condition is particularly limited. For some parameters I have multiple candidate genetic markers lets call this one GM1, the likelihood of developing this particular medical condition is now being shown to be related to the number of repeats of this genetic marker. However when modeling them, I run into issues with limited data for some number of repeats.
Parameter 1
For instance for 12 repeats of this genetic marker I could have 1000 people with 100 of them developing the condition. I input this into my model by checking what the % of people displaying 12 repeats was against the actual number that showed in a new test dataset so in this case 10% into my polynomial regression yields 8.7% chance of developing the condition. However in the same polynomial regression 35 repeats of the the genetic marker is also input as a 10% being 1 confirmed case from 10 people showing the same 8.7% chance of developing the condition. 
The way in which my model works is it assigns different weights to over 600 different parameters in order to best predict the actual multifactorial causation within the genetic profile. For instance parameter 2 would be a different genetic marker and a polynomial regression would be performed on the % of model data patients against the number of real observed and the same as for parameter 1, a 10% chance could input into the polynomial regression as a result of multiple different numbers of repeats of this genetic marker. 
What I need to know is how to adjust for different size datasets. Depending on what size dataset made up the original % chance of developing the condition. 
I am open to suggestion with regards to a method to achieve this.    
 A: As near as I can make out, you have 1000+ patients, one disease which the patients either have or do not have, and about 600 genetic markers, for which you have number of copies of that marker. 
What you have done with that is to run a polynomial regression of the number of repeats against the disease prevalence. You then have some kind of model (you don't specify what kind) which weights the predicted proportions from the 600 or so univariate polynomial regressions to make a unified prediction of the prevalence, which you test on a hold-out data set. You are running into problems because you may have very few patients with a particular number of repeats, meaning the prevalence could be quite biased.
I'll be basing my answer on this interpretation of your question.
OK, the answer: To begin with, I think you may not be formulating your problem in the best way. 
Right now you have: 


*

*A matrix of counts of repeats of 600 markers for each of ~1,000 patients

*A boolean vector of whether or not each patient has the condition you want to predict
The more conventional way to handle this problem would be to train a classifier on this data directly. If you particularly like regression, you could go with logistic regression, but you could also use a more sophisticated classifier like a support vector machine or random forest. All of these methods can output either absolute predictions, or probabilities, for each individual patient you feed into them.
To validate this on your test set, you make a prediction for each patient, and create a truth table against the true disease state. This will let you compute the accuracy (or most likely you want something like both sensitivity and specificity). If you want to be more sophisticated, you could take the probabilities, and compute the area under the receiver operating curve (AUROC), which gives a global measure of the tradeoff between sensitivity and specificity.
This approach is much more in line with what you will see in the literature, and avoids most of the problems (and the complexity) you have introduced into your problem, and quite likely will give you much better predictions.
