Fig. 3

Prediction of splicing entropy with machine learning. A simplified diagram of deep learning model used to predict splicing entropy. For each event, 59 features were used as input and processed with two one-dimension convolutions, The subsequent Squeeze-and-Excitation Networks (SENet) was applied to process features. What follows is the recurrent layer which contains LSTM units that have end-to-end connection in both directions to capture dependencies between features. Recurrent outputs are the input of fully connected layer (FC) to predict the splicing entropy of events in test data. B comparison of the average performance of different methods with test data. PCC: Pearson product-moment correlation coefficient; SCC: Spearman’s rank correlation coefficient; R²: explained variation. C scatter plot shows the predictive power of xgboost and deep learning model respectively, the red line in each graph indicates the linear fit between the predicted and measured splicing entropies. D the rank of feature importance for the predictive splicing entropy (top 10) with xgboost model