At present, the quasi-permutation analysis relies on identification of putative 'true' DEGs, identified as those that pass some particularly stringest p-value threshold. However, the machine learning methods we wish to apply do not produce p-values, raising the question of how to transfer this analysis to such models. One option is to pick some subset of features with particularly high weighting, but that may not be appropriate: even more so than p-values, feature weights tend to defy easy interpretation.
At present, the quasi-permutation analysis relies on identification of putative 'true' DEGs, identified as those that pass some particularly stringest p-value threshold. However, the machine learning methods we wish to apply do not produce p-values, raising the question of how to transfer this analysis to such models. One option is to pick some subset of features with particularly high weighting, but that may not be appropriate: even more so than p-values, feature weights tend to defy easy interpretation.