Fragment mass prediction for phosphosite localization!

Could we improve on current phosphorylation site localization strategies? This group sure seems to think so and this new approach seems simple enough with all this deep learning peptide stuff happening to work into a lot of new workflows! 
 

If I get what they're doing, they use the deep learning model for the unmodified peptides and use the phospho and phospho loss (?) shift to score the localization. The plots vs the traditional approach suggest they're onto something solid.