Fig. 2

ROC curves (Receiver operating characteristic curve) of model BLSTM, BLSTM + ConvNet1, ConvNet2 and BLSTM + ConvNet1 + ConvNet2, on benchmark dataset D3106. Those models were tested with fivefold cross validation. This figure plots the result of one fold validation, and all five fold testing results can be found at Additional file 1: Table S3 and Additional file 1: Figure S1. a 14 curves of model BLSTM. The best one with AUC value 0.9242 is achieved on location prediction of subcellular Lysosome, while prediction of subcellular Peroxisome is the worst one with AUC value 0.8480. The average AUC value of this model is 0.8841. The plot shows that the curves of predictions on 13 subcellular locations, except subcellular Peroxisome, are centered together. b Validation of model BLSTM + ConvNet1. With an extra convolutional neural network to extract features from encoded protein sequences, the AUC on subcellular Lysosome and Peroxisome reached to 0.9255 and 0.9134. The average AUC value of this model is 0.9064. All 14 curves are convergent together. This model is the most robust model among all the four models. c ROC curves of model ConvNet2. Different from curves in (b), those 14 curves are divergent. The max AUC value is 0.8785 and the min AUC value is 0.7058. Since it only extracts features from evolution information, this model performance isn’t as good as the other three models. d With combination of sequence information and evolution information, model BLSTM + ConvNet1 + ConvNet2 improves the best AUC value to 0.9458. However, predictions of several subcellular locations are effected by the parameters in Convnet2, so the curves of those predictions are divergent from center