Table 2.
Prediction accuracies of the specialized predictors. The per-chain accuracies and standard errors were estimated via a 10-fold cross-validation procedure (see Performance evaluation). Using all 54 features, the two predictors were built as linear SVMs with parameter C, which represents the trade-off between training error and margin, set to 0.5 and 1, respectively. Default decision threshold of 0.5 was used for both predictors. SNS and SNL are sensitivities, or true positive rates, on short and long disordered regions, respectively. SP is specificity, or true negative rate, on ordered regions. ACCS is the overall accuracy for VSL2-S calculated as (SNS + SP)/2, while ACCL is for VSL2-L as (SNL + SP)/2.
| W in | W out | SN S | SN L | SP | ACCS/ACCL | |
| 15 | 5 | 82.0 ± 1.1 | 70.7 ± 1.9 | 81.5 ± 0.3 | 81.7 ± 0.6 | |
| VSL2-S | 15 | 31 | 56.5 ± 1.7 | 71.3 ± 2.3 | 89.1 ± 0.4 | 73.0 ± 0.8 |
| 41 | 5 | 79.8 ± 1.3 | 72.8 ± 1.9 | 81.2 ± 0.4 | 80.5 ± 0.7 | |
| 41 | 31 | 68.8 ± 1.6 | 74.3 ± 2.2 | 85.3 ± 0.4 | 77.1 ± 0.8 | |
| 15 | 5 | 70.8 ± 1.4 | 78.6 ± 1.9 | 80.1 ± 0.6 | 79.4 ± 1.0 | |
| VSL2-L | 15 | 31 | 53.4 ± 1.7 | 79.9 ± 2.2 | 84.9 ± 0.6 | 82.4 ± 1.1 |
| 41 | 5 | 50.1 ± 1.7 | 80.9 ± 2.1 | 85.7 ± 0.5 | 83.3 ± 1.1 | |
| 41 | 31 | 44.1 ± 1.8 | 82.1 ± 2.2 | 87.3 ± 0.5 | 84.7 ± 1.2 | |