Table 2.
Comparison of Pscore histograms from the classification of fucosylation types between selected machine learning models of the deep neural network (DNN) and support vector machine (SVM).
| Training set (433 GSMs) | Test set (393 GSMs) | Unknown set (671 GSMs) | ||||
|---|---|---|---|---|---|---|
| DNN | SVM | DNN | SVM | DNN | SVM | |
| AUC* | 0.999 | 0.994 | 0.999 | 0.998 | 0.998 | 0.986 |
| Pscore cut <1% FDR** | 4.623 | 0.982 | 5.559 | 0.303 | 3.415 | 0.692 |
| Filtered GSMs*** | 433 | 417 | 391 | 387 | 657 | 626 |
| Union of Filtered GSMs**** (TP / FP) | 433 (433/0) | 392 (388/4) | 638 (626/12) | |||
| Sensitivity (TP /(TP TPFN)) | 100% (433/433) | 100% (388/388) | 99.21% (626/631) | |||
| Accuracy | 100% | 99.75% | 97.47% | |||
*Area under the curve (AUC) values were calculated from receiver operating characteristic curves between the target and decoy.
**Pscores were less than 1% FDR between the target and decoy, where Pscores were calculated as the natural logarithm of the difference between the first and second ranked probabilities for classification of the fucosylation types.
***Number of glycopeptide spectra matches (GSMs) was filtered with less than 1% FDR between the target and decoy.
****Union number of GSMs were classified using the DNN and SVM filtered with less than 1% FDR between the target and decoy.