Table 4.

Fold classification accuracy by 3DZD and the deep neural network.

Method	3DZD Type	Accuracy	Precision	Recall	F-Measure
Fold
3DZD-NN	Full Atom	0.954	0.945	0.964	0.954
	Main Chain	0.977	0.974	0.979	0.977
3DZD	Full Atom	0.508	0.504	0.998	0.670
	Main Chain	0.616	0.571	0.939	0.710

This benchmark is computed using the test set from the SCOPe dataset. Balanced positive and negative test pairs were constructed from the set of 2521 protein structures in SCOPe. There were 167,872 test pairs in total. 3DZD is the original method where the 3DZD of two structures are compared with a score that uses Euclidean distance of 3DZDs of two proteins, which is defined as 1/(1+Euclidean distance). Thus, the score ranges from 0 to 1. 3DZD-NN is the deep network that outputs predicted probability that input two structures are in the same SCOPe fold. Probability values output by 3DZD-NN range from 0 to 1. We used the best threshold that maximized F-measure. The threshold values of 3DZD-NN full atom, 3DZD-NN main-chain, and 3DZD were 0.5, 0.6, and 0.1, respectively. See Table 1 in Supplementary Information for results of all different thresholds. See Methods for definitions of accuracy, precision, recall, and F-measure.