Fig. 9

Model interpretation elucidates the transfer learning mechanism. A TF motif representations learnt in the first convolutional layer of the original multi-model of 50 TFs. Filters (dots) are plotted with respect to their information content (x-axis) and overall influence (y-axis). The color-scale reflects the Tomtom similarity of a given filter to a TF profile from the JASPAR database. When applicable, the name of the most significant TF is shown. Filters 60 and 92 correspond to the JASPAR profiles for ESRRA and NR2C2 (shown in bold), respectively, and are highlighted with dotted line circles. B Transfer learning fine-tunes the filters learned in the pre-training step to resemble the binding motif of the target TF. Specific positions (black triangles) from filters 60 and 92 (highlighted above) are refined in the fine-tuning step to resemble the JASPAR profiles for HNF4A and SPI1, respectively. C Learning the motifs of cofactors contributes to the predictive capacity of the model. Using DeepLIFT importance scores from a JUND model initialized with the weights of a multi-model pre-trained with five cofactors, TF-MoDISco recovers the motifs of two of the cofactors, CEBPB and SP1, as well as the JUND motif