Protein Fold Recognition From Sequences Using Convolutional and Recurrent Neural Networks

Abstract

The identification of a protein fold type from its amino acid sequence provides important insights about the protein 3D structure. In this paper, we propose a deep learning architecture that can process protein residue-level features to address the protein fold recognition task. Our neural network model combines 1D-convolutional layers with gated recurrent unit (GRU) layers. The GRU cells, as recurrent layers, cope with the processing issues associated to the highly variable protein sequence lengths and so extract a fold-related embedding of fixed size for each protein domain. These embeddings are then used to perform the pairwise fold recognition task, which is based on transferring the fold type of the most similar template structure. We compare our model with several template-based and deep learning-based methods from the state-of-theart. The evaluation results over the well-known LINDAHL and SCOP_TEST sets, along with a proposed LINDAHL test set updated to SCOP 1.75, show that our embeddings perform significantly better than these methods, specially at the fold level. Supplementary material, source code and trained models are available at http://sigmat.ugr.es/~amelia/CNN-GRU-RF+/.

FoldHSphere: deep hyperspherical embeddings for protein fold recognition

“…We tested the effectiveness of our hyperspherical embeddings using both the well-known LINDAHL dataset [ 3 ] and the updated LINDAHL_1.75 dataset we recently proposed in [ 48 ]. The original LINDAHL dataset includes 976 domains from SCOP 1.37 covering 330 folds.…”

“…In order to represent the protein amino acid sequence with variable length L , we considered 45 features for each amino acid as in previous works [ 42 , 48 ]. These 45 residue-level features contain the following information: Amino acid encoding: one-hot vector of size 20 representing the amino acid type.…”

“…Furthermore, deep learning-based methods have been recently proposed to identify the protein fold, such as DeepSF [ 42 ], DeepFR [ 43 ], DeepSVM-Fold [ 44 ], MotifCNN-fold [ 45 ], SelfAT-Fold [ 46 ], VGGfold [ 47 ], and CNN-BGRU [ 48 ]. In these methods, a supervised neural network model is trained to classify the input protein domain into one of the possible fold classes.…”

“…From the trained model, a fold-related embedding representation is extracted, which is then used to measure the similarity between each two protein domains. In this context, the learned embeddings constitute a d -dimensional space in which we can map high-dimensional protein representations such as evolutionary profiles [ 48 ] ( , where L is the protein sequence length) or contact maps [ 43 ] ( ). Moreover, these embeddings capture the fold information during training by placing inputs from the same fold close together in the embedding space.…”

“…Here, the suffix ‘BGRU’ refers to bidirectional GRU, while ‘GRU’ indicates the use of a unidirectional GRU. These two architectures are derived from our previous CNN-GRU and CNN-BGRU models [ 48 ], and can also process arbitrary length protein sequences represented by residue-level features.…”