Fig. 5
From: Biologically relevant transfer learning improves transcription factor binding prediction
Transfer learning is more beneficial for TFs with the same binding mode as a TF from the pre-training step. A Performance difference (i.e., ΔAUCPR) of individual models trained with and without transfer learning (x-axis) is plotted with respect to the size of the training dataset (y-axis) for TFs used to train the multi-model (yellow dots), TFs with the same binding mode as a multi-model TF (blue dots), and TFs with different binding modes than multi-model TFs (red dots). B ΔAUCPRs for TFs used to train the multi-model (yellow boxes), TFs with the same binding mode as a multi-model TF (blue boxes), and TFs with different binding modes than multimodel TFs (red boxes). AUCPR, area under the precision-recall curve; ΔAUCPR, AUCPR from transfer learning - AUCPR from training from scratch; BM, binding mode; TF, transcription factor