Fig. 2.

(A) Table showing the difference in final accuracy (Q₃ score) between the improved S4PRED, the AWD-GRU benchmark, and the current version of PSIPRED-Single on the CB513 test set. (B) Table of classification metrics for the S4PRED model test set predictions. These are shown for each of the three predicted class; α-helix, β-sheet and loop (or coil). The support is normalized across classes to 100 for clarity—there are a total of 84484 residue predictions in the test set. (C) Confusion matrix for the three classes in the S4PRED model test set predictions