The difficulty we encountered is that there were too few available datasets with linked data that can directly be used for concept learning using our setting.
However, when too few sample datasets are chosen, noise inherent in the microarray data will also affect the results obtained from 'condition-dependent' GCA.
If the preliminary TRN is too small or of poor quality, or if there are too few expression datasets, the training versus random set probability distributions are difficult to distinguish.
If the former, we would expect most trees to have at least some clades resolved, so we counted the number of clades resolved in each tree of the genome dataset (the other datasets had too few trees to be informative).
17 families with fewer than 6 sequences were excluded, since too few taxa in the dataset results in poor power to identify amino acid sites under diversifying selection [ 12].
Since, the numbers of FNs in Arabidopsis dataset were too few to deduce any meaningful statistics, we did not use Arabidopsis FNs dataset for further comparisons.
This is also what makes the TIES data unique, as most existing European datasets have too few young adults of migrant descent to fully capture their life course decisions.
The results on the real datasets in the orthologous space (Table S8 (B)) were rather poor and similar to what was observed for the synthetic data: when the dataset contains too few closely related orthologs (less than six for bacterial genes and four for the yeast genes) the algorithms failed in detecting the motif (data not shown).
This fourth dataset had too few samples to perform survival analyses and NPV calculations, but using 48 samples with > 90% tumor cellularity[ 14], we applied 5%and10%0% normal correction.
For dataset of pathway, like result from Chia et al. [ 33], ISA algorithm returned no bicluster, which was attributed to the fact that this dataset contains too few conditions.
