Fig. 4: Functional Hierarchies in Model Input.
From: Deep neural network prediction of genome-wide transcriptome signatures – beyond the Black-box
a We found the explanatory power on test data to decrease as the number of TFs were removed from the model. Particularly, we found around 500 TFs to carry a greater predictive power, as the explanatory power after that point more rapidly decreased. b We measured the median number of scientific studies at each iteration of our backward-selection algorithm, and found that as the number of input TFs decreased, the more studied were those that were left. c The TFs that were identified as more important also had more recorded interactions in the TRRUST database.
