Fig. 4
From: FoldHSphere: deep hyperspherical embeddings for protein fold recognition
Cosine similarity probability histograms computed for all unique pairs within the LINDAHL test set (976 domains), grouping the negative pairs in blue color, and positive pairs in orange. To compute the cosine similarity scores, we used the embeddings extracted from the ResCNN-BGRU model trained with (a) softmax loss with sigmoid activation, (b) LMCL with tanh activation, or (c) Thomson LMCL with tanh activation