Table 2.
Performance comparison of various deep learning models on the training dataset P.ELM, 10-fold cross validation was used.
| Methods | Residue = S | ||||||
|---|---|---|---|---|---|---|---|
| AUC (%) | Acc (%) | Sn (%) | Sp (%) | Pre (%) | F1 (%) | MCC | |
| TransPhos | 78.67 | 71.53 | 67.16 | 75.89 | 73.59 | 70.23 | 0.432 |
| CNN | 74.34 | 68.40 | 61.14 | 75.65 | 71.52 | 65.93 | 0.372 |
| LSTM | 77.04 | 70.48 | 65.01 | 75.95 | 72.99 | 68.77 | 0.412 |
| RNN | 75.53 | 68.84 | 61.44 | 76.24 | 72.11 | 66.35 | 0.381 |
| FCNN | 75.30 | 69.14 | 60.68 | 77.61 | 73.04 | 66.29 | 0.388 |
| Methods | Residue = T | ||||||
| AUC | Acc | Sn (%) | Sp (%) | Pre (%) | F1 (%) | MCC | |
| TransPhos | 67.19 | 61.77 | 47.32 | 76.22 | 66.56 | 55.32 | 0.246 |
| CNN | 64.44 | 59.19 | 42.03 | 76.34 | 63.98 | 50.74 | 0.196 |
| LSTM | 66.59 | 60.64 | 41.85 | 79.43 | 67.05 | 51.54 | 0.230 |
| RNN | 66.03 | 61.21 | 48.57 | 73.84 | 65.00 | 55.60 | 0.232 |
| FCNN | 63.94 | 59.63 | 45.30 | 73.96 | 63.50 | 52.88 | 0.201 |
| Methods | Residue = Y | ||||||
| AUC | Acc | Sn (%) | Sp (%) | Pre (%) | F1 (%) | MCC | |
| TransPhos | 60.09 | 55.41 | 38.52 | 72.30 | 58.17 | 46.35 | 0.115 |
| CNN | 59.11 | 54.59 | 34.81 | 74.37 | 57.60 | 43.40 | 0.100 |
| LSTM | 59.49 | 55.56 | 40.74 | 70.37 | 57.89 | 47.83 | 0.116 |
| RNN | 61.71 | 59.48 | 58.96 | 60.00 | 59.58 | 59.27 | 0.190 |
| FCNN | 59.30 | 56.44 | 43.26 | 69.63 | 58.75 | 49.83 | 0.134 |
Accuracy (Acc), Sensitivity (Sn), Specificity (Sp), Precision (Pre), F1 Score (F1) and Matthews correlation coefficient (MCC) were calculated to measure the performance of models. Data in bold indicates that the model performs best for that evaluation metric.