Table 4.
The internal consistency of groups over all datasets before and after correction
| Datasets | Original | First pass | Corrected |
|---|---|---|---|
| Histone | 91.7% | 95.3% | 98.8% |
| Histone rare | 77.7% | 83.4% | 91.3% |
| Transposons(median) | 93.4% | 95.8% | 99.2% |
| Transposons(mean) | 92.7% | 95.6% | 98.8% |
| Transposons(best) | 94.7% | 98.1% | 99.8% |
| Trans. rare(median) | 85.4% | 90.2% | 96.8% |
| Trans. rare(mean) | 85.7% | 89.6% | 95.4% |
| Trans. rare(best) | 92.7% | 95.5% | 99.7% |
| Simulated | 91.1% | 98.5% | 99.9% |
| Simulated rare | 85.4% | 95.8% | 99.7% |
‘Original’ denotes the uncorrected signatures, ‘corrected’ the corrected signatures, while ‘first pass’ describes the signatures after a single application of the neural network corrector. In the ‘corrected’ and ‘first pass’ cases the internal consistency is the similarity of corrected signatures of a ground truth group (i.e. manual assembly copy group) to the consensus of the corrected group, instead of to the consensus of the uncorrected group. In the uncorrected case, internal consistency is just accuracy, that means 100% - error rate.