Fig. 7

Pre-training with a few biologically relevant TFs only offers a slight advantage over a multi-model pre-trained with a large number of TFs. Transfer learning performance for the target TFs HNF4A (A), JUND (B), MAX (C), SPI1 (D), and SP1 (E): on the right, from multi-models pre-trained with five TFs with the same binding mode as the target TF (dark blue boxes), five cofactors of the target TF with a different binding mode than the target TF (light blue boxes), five functional partners of the target TF from STRING with a different binding mode than the target TF (yellow boxes), and the best cofactors and functional partners from STRING of the target TF regardless of their binding mode (dotted light blue and yellow boxes, respectively); on the left, from “burying” each group of five biologically relevant TFs among 45 TFs drawn from the original multi-model, each with a different binding mode than the target TF. The performance of de novo PWMs (black stars) is provided as a baseline for each TF. AUCPR, area under the precision-recall curve; BM, binding mode; PWM, position weight matrix; TF, transcription factor