The protein P is represented by a vector of 343 numbers, all of which are zero except for f276 (356), f89 (562), f13 (621), f149 (214), f71 (142), f158 (424), f23 (241), and f4 (411). Because the AC coding method was superior to the CT coding method for this task, we used AC in the subsequent model construction.

The AC method, which describes how variables at different positions are correlated and interact, has been widely used for coding proteins [ 12, 34].We built our final model with the architecture and parameters of the best trained AC model trained on pre-training dataset. We screened out the PPIs with ‘confidence score’ equal 1 as the ‘high quality’ (HQ) data and treated the rest as the ‘low quality’(LQ) data. Though various computational methods for predicting PPI have been developed, their robustness for prediction with external datasets is unknown. An SAE consists of multiple layers of autoencoders, which are layer-wise trained in turn, and the output of the former layer is wired to inputs of the successive layer.

which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. was slightly higher than ours, we compared prediction abilities of the two models on external test sets. High performance image conversion IR viewers based on high-grade image converter are designed to observe indirect radiation of infrared laser, light emitting diodes (LED), dye and other IR-sources in 350 – 1700 nm spectral range. Then AC and CT models with the best performance with 10-CV were recruited to predict the hold-out test set.

For machine-learning algorithms, support vector machine (SVM) and its derivatives [ 9, 10], random forest [ 11] and neural networks [ 12], have been applied. First, all 20 amino acids are clustered into seven groups according to their dipole and side chain volumes (Additional file 4: Table S2).

Other groups have tried different methods to build negative data, but did not achieve promising results [ 6].

It is noteworthy that our model gave a satisfying prediction accuracy for a large number of newly verified PPIs.