Biologically relevant transfer learning improves transcription factor binding prediction

Deep learning has proven to be a powerful technique for transcription factor (TF) binding prediction but requires large training datasets. Transfer learning can reduce the amount of data required for deep learning, while improving overall model performance, compared to training a separate model for each new task.

We assess a transfer learning strategy for TF binding prediction consisting of a pre-training step, wherein we train a multi-task model with multiple TFs, and a fine-tuning step, wherein we initialize single-task models for individual TFs with the weights learned by the multi-task model, after which the single-task models are trained at a lower learning rate. We corroborate that transfer learning improves model performance, especially if in the pre-training step the multi-task model is trained with biologically relevant TFs. We show the effectiveness of transfer learning for TFs with ~ 500 ChIP-seq peak regions. Using model interpretation techniques, we demonstrate that the features learned in the pre-training step are refined in the fine-tuning step to resemble the binding motif of the target TF (i.e., the recipient of transfer learning in the fine-tuning step). Moreover, pre-training with biologically relevant TFs allows single-task models in the fine-tuning step to learn useful features other than the motif of the target TF.

Our results confirm that transfer learning is a powerful technique for TF binding prediction.