Improvement of Yin Yang site prediction by incorporating the interplay between phosphorylation and O-GlcNAcylation Chao Ji, Yinxing Guo, Quan Zhang What is Yin Yang site? • Yin Yang sites: The reciprocal and dynamic change between O-GlcNAcylation and phosphorylation at the same or proximal Ser/Thr • Mechanisms that might govern Yin Yang regulation – Direct competition at a single site – Competition via steric hindrance by reciprocal modification at proximal sites – Affecting the enzymatic efficiency of each other Current Yin Yang site prediction • Predict O-GlcNAcylation and phosphorylation separately – Netpohs for phosphorylation prediction – yinOyang for glycosylation prediction http://www.cbs.dtu.dk/services/YinOYang/output.php What is investigated in our project • Collect reliable Yinyang sites from MS data. – sites shown interplay between Phosphorylation and glycosylation. • Is it possible to predict Yinyang sites directly? – Phosphorylation sites are first predicted by Phosphorylation predictor. – For the sites which are predicted as Phosphorylation sites, the probability of being a Yinyang site is predicted directly. Dynamics between O-GlcNAcylation and phosphorylation • Samples treated with OA, P/N and both • Group 3 vs Group1: how phosphorylation change is response to globally elevated O-GlcNAcylation • Okadaic acid(OA): ser/thr-specific phosphatase inhibitor • PUGNAc and NAG-thiazoline(P/N): nonspecific OGlcNAcase inhibitor Measurement the dynamics • Relative Occupancy Ratio(ROR) • If the phosphorylation level drops significantly after P/N treatment for a specific site, the site is a Yin Yang site. Overview of the MS data • sites/proteins detected in MS data – 342 proteins with 573 sites from MOUSE and RAT. – 103 sites with ROR decreased significantly after treated with O-GlcNAcase inhibitor, 470 otherwise. Collecting training data • Netphos prediction of 573 sites identified in MS data. ROR decreased significantly Otherwise Total MS 103 470 573 Predicted as phosphorylation site 83 326 409 Positive examples, including 68 Ser sites and 15 Thr sites. Negative examples, including 273 Ser sites and 54 Thr sites. Evaluation of Yinyang predictor False positive False negative Error Ser 0.168 0.809 0.297 Thr 0.204 0.800 0.333 Total 0.174 0.807 0.303 Sequence context Sequence context surrounding Ser sites Positive Ser sites Negative Ser sites Sequence context surrounding Ser sites Positive Thr sites Negative Thr sites Profile Model • Select a sequence window centered at the phosphorylation for each instance in the training set. • Models are built on the sequence windows for positive set and negative set, respectively. • For a input sequence window, do the model comparison for the classification. • 4-fold cross validation is used to evaluate the model. Profile Model • Used the different window size from 3-31 bp • The minimum error rate is 0.32 when window size is 9. The max error rate is 0.344 with 29window size. Mean is 0.3350667. However, there is no obviously difference between different window size. Profile Model • Models are built for Ser sites and Thr sites separately. • For Ser set, minimum error 0.34 is achieve when window size = 29. For Thr set, minimum error 0.37 is achieved when window size = 27. Artificial Neural Network? • There’s no simple pattern for Yinyang sites, therefore the Yinyang sites and nonYinyang sites are not easily separable. • ANN is capable of classifying highly complex sequence pattern where the correlations between positions are important. • If there are fuzzy patterns in Yinyang sites, are they recognizable by ANN? Structure of ANN • Standard feed-forward artificial neural network with sigmoidal nodes and one layer of hidden unit • The input layer has n input groups for the sequence window with length n. Each group has 21 units, each of which represents 1 of the amino acids (or spacer). • The output is the probability of Yinyang site. Training and testing of ANN • 4-fold across validation is used to evaluate the performance of ANN. • During the training, 75% of the training data is used for training and 25% of the data is used to test the performance of the network. • The network is tested over window size from 3 to 31 and 3 or 5 hidden units Result of ANN False positive and False negative rate Further work • More positive instances are needed for the network to learn the patterns. • Average the results from different networks. • Incorporate the surface accessibility – O-glycan is linked post-translationally to Ser or Thr of a fully folded and assembled protein, and is thus surface exposed on the protein. – The sites predicted to be on surface are more likely to be Yinyang sites, thus the threshold for those sites could be lowered down.