Figure 4

Performance of PPPs on genome-wide prediction. A, B, C: PPPs' performances on CpG-rich sequence at A) high (50 bp), B) medium (200 bp) and C) low (2 kbp) resolution, respectively. D, E, F: Performances on CpG-poor sequence at D) high (50 bp), E) medium (200 bp) and F) low (2 kbp) resolution, respectively. The performance is measured by sensitivity, which is the proportion of all true promoters that are predicted correctly, and specificity, which is the proportion of total prediction that are correct. The evaluation was based on 37,793 CpG-rich promoters and 4,743 CpG-poor promoters on 1.06 × 10⁹ bps of the human genome. For baseline prediction, we repeated the randomization (see text) 10 times and the means are reported. The standard deviations (sensitivity, specificity) are (± 0.2%, ± 0.3%), (± 0.2%, ± 0.2%), (± 0.0%, ± 0.1%) for the high, medium and low resolutions, respectively.