Table 3.
Summary of studies with deep learning based methods for splice site detection with the HS3D dataset. (D) Donor, (A) Acceptor, Acc. (Accuracy), Sn. (Sensitivity), Sp. Specificity.
| Reference | Base | Architecture | Data | Sequence numbers/lengths | Measure | Performance | ||
|---|---|---|---|---|---|---|---|---|
| DeepSplice43 | CNN |
Input layer 2 Conv layers 1 dense layer Output layer |
HS3D |
2796 (true)(D) 271937 (false)(D) 2880 (true)(A) 329,374 (false)(A) |
140 nt | Donor | Acceptor | |
| Acc. | 0.946 | 0.923 | ||||||
| Sn. | 0.957 | 0.934 | ||||||
| Sp. | 0.938 | 0.914 | ||||||
| DeepSS44 | CNN |
Input layer 2 Conv layers + max pooling layer 2 dense layer Output layer |
HS3D |
2796 (true)(D) 90,953 (false)(D) 2880 (true)(A) 90,353 (false)(A) |
140 nt | Acc. | 0.97a | 0.98a |
| Sn. | 0.96a | 0.97a | ||||||
| Sp. | 0.97a | 0.98a | ||||||
| Pr. | 0.88a | 0.92a | ||||||
| MCC | 0.90a | 0.93a | ||||||
| AUC ROC | 99.02 | 98.79 | ||||||
| AUC PR | 95.93 | 94.28 | ||||||
| DeepDSSR34 | Hybrid (CNN + BLSTM) |
2 inception like layers A convolutional layer A bidirectional layer Dense layer |
HS3D |
2796 (true)(D) 90,924 (false)(D) |
140 nt | Sn. | 0.988 | – |
| Sp. | 0.891 | – | ||||||
| MCC | 0.914 | – | ||||||
aThe approximate values were taken from the graphs since exact values were not given in the paper. Only imbalanced dataset results were mentioned for DeepSS.