Skip to main content
Fig. 3 | BMC Bioinformatics

Fig. 3

From: FoldHSphere: deep hyperspherical embeddings for protein fold recognition

Fig. 3

Cross-validation fold classification accuracy (%) results for different LMCL margins and scales \(s=\{30, 50\}\), using the SCOPe 2.06 training set. The results are provided separately for each neural network model: CNN-GRU, CNN-BGRU, ResCNN-GRU and ResCNN-BGRU, trained using different combinations of activation function (in the embedding layer) and loss function. These are: softmax loss with sigmoid activation (dash-dotted horizontal line), LMCL with sigmoid activation (blue lines), LMCL with tanh activation (magenta lines) and Thomson LMCL with tanh activation (green lines). For the LMCL and Thomson LMCL results, solid lines and dashed lines correspond to scale values 30 and 50, respectively

Back to article page