Table 1. Reported prediction results from different studies.
| Studya | Number of TFs | Number of unique TFBSs | Dataset size (Total TF-TFBS links) | True TF-TFBS links | False TF-TFBS links | Testing method | Highest accuracy |
|---|---|---|---|---|---|---|---|
| Qian Z. et al. (reported in (33)) | 480 | 2,341 | 10,206 | 3,356 | 6,850 | Leave one out | 76.6% |
| Qian Z. et al. (reported in (34)) | 143 | 571 | 10,430 | 3,430 | 7,000 | Leave one out | 87.9% |
| Cai, Y. et al. (reported in (35)) | 599 | 2,402 | 35,410 | 3,541 | 31,869 | Leave one out | 91.1% |
| DRAF models (on the datasets from this study) | 232 | 44,710 | 1,214,389 | 110,399 | 1,103,990 | 30% holdout | 99.16% |
aThis table shows the prediction accuracy of the DRAF models on the holdout dataset (30% of the total), and the other models as reported in the original references (33–35) that used different TF-TFBS test datasets. Our holdout dataset is 34-, 116- and 119-fold larger than the datasets from (35), (34) and (33), respectively. The test dataset for DRAF has 364 317 (positive and negative) TF-TFBS links which is more than 10 times larger than the next largest dataset used in (35).