The methods showed large differences in prediction accuracy but the supervised method was found to perform best, despite the parameters of the unsupervised methods having been optimized.
As said before, all methods showed large bias if events were rare, particularly if accompanied by other problems, such as strong selection and also a large treatment effect.
The accuracy of MinCut is improved by ModMinCut (54%) and MaxCut (31.5%), but all cut methods show larger distances than the average complete gene tree (the baseline distance, 18.5%).
When k=5, all methods show larger mean square error values when there is a a high percentage of zeros in the data compared with when there is only a small amount of zeros.
For the dual-channel datasets, (Spellman, PramilaL, PramilaS), LoessOnly methods show large average correlation, in contrast to the low number of highly correlated pairs, noted for the same methods (previous paragraph).
Based on this rationale, data from yeast two hybrid methods are sufficiently consistent to allow for intra species comparisons (between different experiments) and inter species comparisons, while data from affinity purification mass spectrometry methods show large differences even within intra species comparisons.
When the underlying distribution is bimodal (Fig. 3d), all three methods show large underestimation, although ABC continues performing best for n >40, showing smaller AREs.
However, undiluted samples analyzed by the DCC method showed large standard deviations of repeatability and SCC values lower than those by the DMSCC or FSCC methods, probably because of the high solids content in ovine milk.
Also the Trim and Fill method and the Peters method showed large bias, particularly if the event was very rare (5%).
15 18 Moderate inter-rater agreement has been shown, 17 19 and in the study according to oncological care a Bland-Altman analysis of the GTT method showed large random errors when comparing review teams.
