Biomarker Identification with Few Samples

As expected, the performances of all the methods in terms of biomarker identification decrease with a reduction of the data set size. However, it is important to point out that even in the worst possible case (3 samples per class) early AUC for PLS-DA and the f-test are significantly greater than that obtained for completely random selection. This indicates that both methods can be used effectively in the biomarker selection phase, even with a low number of samples. In other words, features related to spiked compoxmds are consistently present in the top positions of the ordered list of experimental variables, which implies that also models constructed with very few samples can be relied upon to recognize these features. 

---