Fig. 8

The use of relevant binding mode information in the pre-training step improves model performance for TFs with small training datasets, regardless of their presence in the multi-model. Transfer learning performance for three target TFs with the same binding mode, but different number of bound regions, from each of the following five families: A HNF4A, NR2C2, and VD2R (nuclear receptors); B JUND, ATF7, and NFE2L1 (basic leucine zippers); C MAX, MNT, and SREBF2 (basic helix-loop-helix factors); D SPI1, ETV4, and ERG (tryptophan cluster factors); and E SP1, KLF9, and ZNF740 (C2H2 zinc fingers). For transfer learning, multi-models were pre-trained with five TFs with the same binding mode as the target TF (dark blue boxes), or five randomly selected TFs with a different binding mode than the target TF (red boxes), with (left) and without (right) the presence of the target TF in the pre-training step. The training dataset size of each multi-model is indicated with diamonds (secondary y-axis). The number of bound regions for each TF is shown between parenthesis. AUCPR, area under the precision-recall curve; BM, binding mode; TF, transcription factor