advertisement: compare things at compare-stuff.com!

Distinction between mainly-$\alpha$ and mixed-$\alpha \beta$ classes

The second step in the prediction of class will obviously involve the separation of the helix-containing class into the mainly-$\alpha$ and mixed-$\alpha \beta$ classes. Using exactly the same prediction method, we perform the two-class prediction on the set of 386 helix-containing domains (not including irregular domains). The overall accuracy is 74%, but with a low Matthews coefficient of 0.38 (lower than the three-class prediction: $C_\alpha=0.41$ and $C_{mixed}=0.41$). The equivalent random prediction based on prior probabilities is about 62% accurate. The reliability quartile accuracies are high: 90%, 85%, 74% and 50%, however in the top quartile only 7 of the 14 predictions for mainly-$\alpha$ domain are correct. Whilst these predictions are not as clear-cut as the mainly-$\beta$/$\alpha$-containing predictions, they may be slightly more informative than the basic three-class prediction. However we have not looked at this in detail or considered the compounding of errors in multi-stage predictions.