PMC

Venn diagram of experimental phosphorylation sites retrieved from the PhosPhAt database for different nutrient stress experiments. The overlap between different experiments comprises mainly plasma membrane proton ATPases and aquaporins, the proteins unique for each condition include transporters and kinases among others. Nitrate: nitrate starvation and resupply; Phosphate: phosphate starvation and resupply; Carbon: carbon starvation and sucrose re-supply.

Receiver operating characteristics curves of the prediction by pSer predictor in comparison to NetPhos 2.0 () (see Supplementary material for details). In the diagram, improved classification performance is indicated for predictors with increased area under the ROC. The area under the ROC curve was A1 = 0.81 ± 0.01 for the pSer predictor and A2 = 0.67 ± 0.01 for NetPhos and was significantly better with a z-score = (A1−A2)/SE(A1−A2) of 24.1 corresponding to a P-value of 3.3E−128 in the limiting case of a normal distribution according to the algorithm proposed in ().

Schematic diagram outlining the structure of the PhosPhAt service illustrating the two main query entry points to query experimental data and pSer prediction information. Both services merge into a common output at the ‘Summary Page’ on which the prediction results are displayed on top of the page and all experimental phosphopeptides for the given AGI code are listed below. In instances where no experimental phosphopeptides are available, only the prediction result will be displayed. External links to published references at PubMed and MS/MS data at the ProMex mass spectral library () are also shown.

Negative log(P-values) from Fisher exact test on the occurrences of GO: function terms associated with predicted phosphoproteins. P-values were corrected for multiple testing by using the False Discovery Rate (FDR) formalism (). Overrepresented GO: terms are colored red, underrepresented blue. GO: terms were included if p_FDR <0.001. GO annotations were taken from TAIR (). To avoid training bias, phosphorylation sites used during the training of the classifier have been removed in the Fisher exact test. Only GO assignments with evidence categories: direct assay, mutant phenotype, physical and genetic interaction as well as sequence of structural similarity have been considered.

Prediction performance of the pSer predictor in comparison to NetPhos 2.0 (). The recall rate versus the associated precision is plotted. The curved lines indicate lines of equal correlation coefficient. In the diagram, improved classification performance is indicated for predictors falling into the upper right corner. Performance results for our classifier correspond to results obtained in the 10-fold cross-validation test (see Supplementary material for details.) The classifier NetPhos 2.0 was applied to our dataset without training; i.e. NetPhos 2.0 was applied to an independent dataset as it was technically not possible to perform a cross-validation for NetPhos 2.0. While the testing protocols differed, the results still suggest that a plant-specific predictor may yield better performance when applied to plant proteins than a generic predictor.