Figure 4
From: Multi-task learning to leverage partially annotated data for PPI interface prediction
The importance of the multi-task setup and the data extension when training a PPI interface prediction model trained on limited data. The single-task model IF (pink) and the multi-task model IFBUS3SA (red and brown) are compared. The IF model and the IFBUS3SA model indicated in red are trained on a part of the PPI dataset. Differences in performance between the pink and red bars therefore presents the benefit of the multi-task learning strategy. The IFBUS3SA model in brown is trained on the PPI_extendedSFD dataset in which only a part of the PPI interface information is considered. All the brown bars are thus trained on the same number of sequences for which the related task information is available. Differences in performance between red and brown bars indicate the benefit of training the model on the augmented PPI_extendedSFD dataset. Model performance is shown by the mean AUC ROC (bars) and standard deviation (whiskers) of the PPI interface prediction on the total validation set.
