I'm working with a data set similar to horseracing problem (it's
actually a healthcare problem) in which I've got a groups (sets) of 8
proteins.  I trying to predict the final order of the proteins, so in
many ways it's a horserace type problem.

My question is a design one.  Is there any way to input the set of
protein (vectors) - as a 8x8 matrix but in a way that would enable me
to look across the 300+ (8x8) matrixes.  I want to be able to identify
those that I can successfully predict using non-neural net methods.
I've got about a 30% hit rate but if I can go up to around 45%+ it
becomes clinically useful.  If I turn it all into a string vector (64
items + correct/incorrect) I don't have enough data to be meaningful.
I think.

The additional compounding problem is that it's both time
consuming and expensive to get more data.