Since our knowledge of the positive cases is very very limited, I would say the same for the negative ones. However, since it is expected that the regulation matrix miRNA-transcripts is quite sparse (..by the way, is it true?), I would not mind for the negative cases. Random sampling should work, and the small number of true positives you are putting in your negative dataset simply by chance should be very low and should not affect your accuracies.

 Thank you Matteo,

but depending on the algorithm used i found the number of false positives to be quite high in some cases! Random sampling does work in cases of moderate hit probability, though.