Fig. 2

The proposed ResCNN-BGRU neural network model for fold-related embedding learning through protein fold classification. The model architecture contains three differentiated parts. The residual-convolutional network a processes the input \(L\times 45\) residue-level features and consists of two residual blocks with two 1D-convolutional layers each. Its output is passed through a bidirectional layer of gated recurrent units (b) to obtain a fixed size representation of the input domain, which is further processed by two fully-connected layers (c). The first FC layer learns a 512-dimensional embedding vector for each input, while the second one learns a class weight matrix \(\mathbf {W}\) to perform the classification into K fold classes. The ResCNN-GRU model is identical but using a unidirectional GRU layer instead