And cause it's interpretable, we can backpropagate the prediction error for the entire transcriptome to the transcription factor layer and figure out which TFs we're mispredicting (could continue fine-tuning the model this way, too). We explicitly measure that error in the context of TF activity.
