Skip to main content
. Author manuscript; available in PMC: 2010 Jul 1.
Published in final edited form as: Proteins. 2009 Jul;76(1):176–183. doi: 10.1002/prot.22329

Table 2.

Ten-fold cross-validated accuracy and coverage given by the two-hidden-layer (ANN2) neural networks for predicting full-contact (|i - j| ≥ 6) and nonlocal contact (|i - j| ≥ 24) maps at different number of top predictions (L, L/2, L/5) are shown for proteins of different structural classes (all α, all β and others) and different lengths (L < 120, 120 ≤ L ≤ 240, and L > 240).

α β Other L < 120 120 ≤ L ≤ 240 L > 240
Number of Proteins 110 52 338 154 214 132
Full Contact Mapsa
(L) Accu.b 16.9 33.7 32.4 27.6 30.1 29.3
Cov.c 17.1 14.4 15.9 20.1 15.2 12.5
(L/2) Accu.b 22.0 41.8 41.8 35.1 38.4 38.2
Cov.c 10.9 8.8 10.3 13.0 9.7 8.2
(L/5) Accu.b 29.8 51.4 52.4 44.5 48.2 48.9
Cov.c 5.9 4.3 5.3 7.0 4.9 4.2

Nonlocal C ontact Mapsd
(L) Accu.b 9.4 14.1 17.1 13.5 16.3 15.0
Cov.c 16.0 10.7 13.1 18.2 12.5 9.3
(L/2) Accu.b 11.3 16.4 21.9 16.2 20.4 19.8
Cov.c 10.3 6.3 8.3 11.5 7.8 6.1
(L/5) Accu.b 13.8 20.4 28.2 19.7 25.5 27.1
Cov.c 5.5 3.1 4.1 5.9 3.9 3.3
a

All residue-residue contacts with the separation of the sequence positions greater than or equal to 6.

b

Accuracy: fraction of true positives in number of top predictions.

c

Coverage: fraction of true positives in total number of native contacts.

d

Nonlocal residue-residue contacts with the separation of the sequence positions greater than or equal to 24.