Fig. 3

Neural networks can classify Smad-bound sites as being Smad2- or Smad3-bound.A. Precision recall curve of CNN (blue) and CNN-LSTM (black) models, taking the average of 10 models for final classification. An average precision of 0.95 was observed for the CNN model, as compared to the slightly higher average precision of 0.96 of the CNN-LSTM model. The model was better able to classify Smad3 (0.87) as compared to Smad2 (0.7) B. Confusion matrix of CNN model in classifying Smad2 and Smad3 sites. The model was able to better classify Smad3 (0.7 vs 0.87). C Confusion matrix of CNN-LSTM. Similar to the CNN model, the CNN-LSTM model was also better at classifying Smad3 (0.84) as compared to Smad2 (0.78), but performed better than the CNN model (as shown in A). D. The effect of ensemble learning on model performance evaluated using AUCPR. We evaluated the performance of increasing the number of models used from one to ten, with increase in AUCPR observed as the number of models increased. The standard deviation, indicative of stability, also decreased as more models were included in the final ensemble. E. Confusion matrix of Smad2/3 binding in hESC, showing model performance in a novel cell type was not included in the training dataset