PS8-Net: A Deep Convolutional Neural Network to Predict the Eight-State Protein Secondary Structure

Ensemble of Template-Free and Template-Based Classifiers for Protein Secondary Structure Prediction

“…With the result obtained by our ensemble of template-free and template-based classifiers, we can compare the Q8 accuracy obtained by our method with the classifiers in the literature. Our proposed method was able to outperform the state-of-the-art approaches [ 33 ] by 8.2 p.p., as reported in Table 16 . The basic ensembles (template-free and template-based) also surpassed the state-of-the-art approaches.…”

“…Among the methods in the literature, our methods surpassed the state-of-the-art approaches [ 33 ]. According to Table 20 , our final ensemble surpassed the state-of-the-art approaches by 17.6 percentage points.…”

“…The recent results of inception-based [ 57 , 58 ] protein secondary structure prediction methods [ 33 , 49 , 51 ] showed that this kind of architecture can aggregate local information, generated by the first layers, and global information, which is generated by deep layers. Driven by this success, we evaluated the newest inception architecture (Inception-v4 [ 58 ]) for the protein secondary structure prediction task.…”

“…CNN classifiers [ 33 , 34 , 35 ] have filters that can analyze several different windows of a sequence. Deep convolutional networks can obtain long-distance information, but local information, gained in the early layers of the network, is lost.…”

“…With that, several methods have started to use inception blocks, where information from the shallower layer is kept, concatenating the information with the deeper layers. This approach has enabled an advance in the results of the prediction of secondary structures, such as the method developed by Ratul et al [ 33 ], which reached the state-of-the-art in CB6133 [ 7 ] and CB513 [ 20 ] datasets, with a classifier with inception blocks.…”