Table 2 Comparison of the PPI interface performance of the single-task model against different multi-task models.

From: Multi-task learning to leverage partially annotated data for PPI interface prediction

 

PPI dataset

PPI_extendedSFD dataset

AUC ROC

AUC PR

AUC ROC

AUC PR

IF

73.17 ± 0.36

31.71 ± 1.01

73.17 ± 0.36

31.71 ± 1.01

IFBU

74.85 ± 0.19

34.37 ± 0.32

75.15 ± 0.20

35.35 ± 0.17

IFBUSA

75.08 ± 0.24

35.62 ± 0.97

75.92 ± 0.21

36.65 ± 0.42

IFBUS3SA

75.73 ± 0.50

35.79 ± 1.44

76.32 ± 0.23

38.44 ± 0.92

IFBUS8SA

75.73 ± 0.31

36.39 ± 0.74

76.20 ± 0.24

37.95 ± 0.52

IFBUS3S8SA

75.73 ± 0.21

36.46 ± 1.13

76.06 ± 0.14

38.16 ± 0.93

  1. The mean AUC ROC and AUC PR scores and the corresponding standard deviations, on the validation set after training the models four times, are shown. Performance is measured on the validation set of both the PPI dataset and the augmented PPI_extendedSFD dataset. The multi-task models outperform the single task model (73.17 ± 0.36 AUC ROC) significantly on both dataset (P < 0.001). The overall highest AUC ROC score (76.32 ± 0.23), shown in bold, is reached when including buried residues, secondary structure in three classes and absolute solvent accessibility as related prediction tasks in addition to the PPI interface prediction on the PPI_extendedSFD dataset.
Back to article page