Fig. 1

The ExplaiNN model and its application to TFBS prediction in OCRs. A ExplaiNN takes as input one-hot encoded DNA sequences. Architecturally, it is composed of multiple independent CNNs (i.e., units), each of which comprising one convolutional layer with a single filter, batch normalization, exponential activation, and max pooling, and two fully connected layers with batch normalization, ReLU activation, and dropout. The final linear layer of ExplaiNN (i.e., the output) combines the outputs from each unit (denoted here using Xs). B Performances (AUPRC; y-axis) of ExplaiNN models trained using increasing numbers of units (x-axis) on predicting the binding of 50 TFs in OCRs (green line). The performances of DanQ [7], a deep CNN with 3 convolutional layers (i.e., DeepCNN), and two shallow CNNs with 1 convolutional layer (i.e., CNN₁ and CNN₁Exp featuring an exponential activation function instead of ReLU) are provided as baselines (gray lines). C Pairwise comparison of the individual performances (AUPRC) of the 50 TFs from the previous dataset between the ExplaiNN model trained using 100 units and either DanQ (green dots) or the DeepCNN (gray dots). D Number of binding modes (y-axis) detected with ExplaiNN using increasing numbers of units (x-axis). E Number of binding modes detected with DanQ, the DeepCNN, and ExplaiNN trained using 100 and 300 units on the previous dataset using either filter visualization (no pinstripes) or TF-MoDISco [23] clustering on DeepLIFT [22] attribution scores (pinstripes). The 50 TFs in the dataset are represented by 33 unique binding modes (dashed line), some of which are detected using different combinations of models and interpretive approaches (green); other detected binding modes (i.e., different from those 33) are shown in gray. F Heatmap of the final linear layer weights of the ExplaiNN model trained using 100 units, with rows representing units with assigned biological annotations based on their Tomtom [27] similarity to known TF profiles from the JASPAR database [18] and columns representing the 50 TFs predicted by the model. More than one filter can learn the same TF motif representation, but some may not contribute to the model’s predictions (black arrows). G (top) Visualization of importance scores for a unit annotated as FOXA1 from the ExplaiNN model trained using 100 units. This unit contributes the most to the prediction of FOXA1 binding. (bottom) Filter nullification analysis for the seven DanQ filters annotated as FOX TFs. The results are consistent with the unit importance scores. AUPRC, area under the precision-recall curve; CNN, convolutional neural network; OCR, open chromatin region; ReLU, rectified linear unit; TF, transcription factor; TFBS, TF binding site