Table 5.
Comparison of metrics with MPI predictions from a deep-learning approach.a
| Dataset | Confidence score | Zhao et al. | SARTRE |
|---|---|---|---|
| Piazza | 0.61 | 0.68 | |
| Reznik | 0.54 | 0.73 | |
| STITCH-E.coli | 150 | 0.77 | 0.77 |
| 400 | 0.82 | 0.82 | |
| 700 | 0.84 | 0.82 | |
| 900 | 0.75 | 0.77 | |
| STITCH-Yeast | 150 | 0.63 | 0.76 |
| 400 | 0.48 | 0.84 | |
| 700 | 0.7 | 0.76 | |
| 900 | 0.58 | 0.75 |
a
The performance of SARTRE on four constructed datasets is compared to previous MPI predictions from Zhao et al. (2021) that relies on a deep-learning model with an extensive set of features. Zhao et al. use metabolite features with the size of 2325 for all datasets, and protein features with the size of 964, 328, 1365, and 1150, respectively for the four datasets. On the other hand, SARTRE uses metabolite features with the size of 168, 320, 209, and 265, and protein features with the size of 800, 333, 1000, and 864, respectively, for the four datasets. Macro AUC is calculated based on the predictions on test sets, using 10-fold cross-validation.