Table 2.
The prediction performance of SVM models with combinations of three kinds of regulatory features such as over-represented hexamer nucleotides (OR), nucleotide composition (NC), and DNA stability (DS), is evaluated by benchmark "Cross-validation" based on the specified window size -200 to +100 of TSS(+1).
| Training set | Window size | Features | Precision | Sensitivity | Specificity | Accuracy |
|---|---|---|---|---|---|---|
| All (6,452) |
-200 ~+100 | OR+NC | 77% | 71% | 79% | 75% |
| -200 ~+100 | OR+DS | 76% | 69% | 78% | 74% | |
| -200 ~+100 | NC+DS | 75% | 74% | 76% | 75% | |
| -200 ~+100 | OR+NC+DS | 79% | 76% | 79% | 78% | |
| With CpG (4,898) |
-200 ~+100 | OR+NC | 79% | 81% | 79% | 80% |
| -200 ~+100 | OR+DS | 77% | 80% | 76% | 78% | |
| -200 ~+100 | NC+DS | 77% | 82% | 75% | 78% | |
| -200 ~+100 | OR+NC+DS | 80% | 84% | 79% | 82% | |
| Without CpG (1,554) | -200 ~+100 | OR+NC | 68% | 70% | 67% | 68% |
| -200 ~+100 | OR+DS | 68% | 71% | 66% | 68% | |
| -200 ~+100 | NC+DS | 66% | 67% | 66% | 66% | |
| -200 ~+100 | OR+NC+DS | 69% | 69% | 71% | 70% | |
The number of training sequences used to construct the SVM models is shown in parenthesis of the column "Training set".