Fig. 3

Transfer learning improves TF binding prediction. A The transfer learning strategy used in this study consists of two steps: pre-training a multi-task CNN with multiple TFs (top) and fine-tuning a single-task CNN, initialized with the weights of the pre-trained multi-task CNN, for one TF (bottom). The CNN architecture is similar to those of Basset [57] and AI-TAC [58]: three convolutional layers, each with ReLU activation, batch normalization, and max-pooling, followed by two fully connected layers and one output layer. Models were trained with one-hot encoded, 200-bp long DNA sequences, and their reverse complements. B Performance of individual models trained with (blue boxes) and without (red boxes) transfer learning for the 50 TFs from the pre-training step. The performance of each TF on the multi-model is provided as baseline (yellow boxes). C Quantification of the effect of training dataset size on model performance. TFs (dots) are plotted with respect to the size of their training dataset (x-axis) and performance of their individual models trained with (blue) and without (red) transfer learning (y-axis). The performance of de novo PWMs (black stars) is provided as a baseline for each TF. AUCPR, area under the precision-recall curve; CNN, convolutional neural network; PWM, position weight matrix; ReLU, rectified linear units; TF, transcription factor; TL, transfer learning