Effect of size and heterogeneity of samples on biomarker discovery: synthetic and real data assessment