Table 6.
Comparison of effects of models using different pre-training loss
| Loss | Pre-train | Train | ||||
|---|---|---|---|---|---|---|
| Precision | Recall | F1-score | Precision | Recall | F1-score | |
| PRAUC-loss | 0.0351 | 0.6177 | 0.0665 | 0.8599 | 0.7897 | 0.8233 |
| Negative-F1 | 0.1564 | 0.0987 | 0.1210 | 0.3587 | 0.2107 | 0.2655 |
| Weighed-logistic | 0.0283 | 0.7559 | 0.0546 | 0.8963 | 0.8015 | 0.8462 |
| No-pretraining | – | – | – | 0.8907 | 0.5861 | 0.7070 |
In the pre-training phase, the hard constraint layer is removed from the network and different loss functions are used for training. In the training phase, we used different pre-trained models for transfer learning to obtain prediction models on the Rfam dataset. The loss function of training phase is Neagtive-F1 function