Computational Systems Biology Methods and Protocols.7z

MCC¼

TPTNFPFN

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

ðÞTPþFPðÞTPþFNðÞTNþFPðÞTNþFN

p

where TP, TN, FP, and FN are the number of true positives (cor-

rectly predicted O-GlcNAcylated sites), true negatives (correctly

predicted non-O-GlcNAcylated sites), false positives (falsely pre-

dicted O-GlcNAcylated sites), and false negatives (falsely predicted

non-O-GlcNAcylated sites), respectively.

2.4 Classify
Algorithms

All the prediction algorithms shown in Table1 for identifying

protein O-GlcNAcylated sites use machine learning methods,

including support vector machines (SVM), artificial neural net-

works, and hidden Markov model (HMM). SVM is a set of related

supervised learning methods used for classification and regression

based on statistical learning theory, which has been shown to be a

powerful tool in many fields of bioinformatics [8, 49–52]. Five of

the six predictors (Table1) were designed using a SVM classifier.

The predictor designed by Kao et al. [10] is a two-layered machine

learning method, incorporating a HMM profiles and SVM.

3 Performance of Online Tools for O-GlcNAcylation Sites Prediction

Each of the computational predictors for O-GlcNAcylation sites

identification shown in Table1 has self-reported their sensitivity

and specificity values. However, these values can often deviate

significantly from the actual sensitivity and specificity of the predic-

tor for two reasons. First, developers can mistakenly include testing

data when refining the predictor, leading to an “over-fit.” Second,

when new O-GlcNAcylation modification data are published, the

overall characteristics of the available data may change [53]. Three

of the six predictors do not have an available web server. Therefore,

to assess the performances of the six predictors, we selected the

independent dataset used in O-GlcNAcPRED and PGlcS

[8, 10]. The dataset is composed of 67 O-GlcNAcylation sites

and 7244 non-O-GlcNAcylation sites extracted from 38 experi-

mentally identified O-GlcNAcylated proteins that were not

included in the original dbOGAP. The sequences of each of the

38 proteins were uploaded into each predictor, and the prediction

results are shown in Table2.

Among the six predictors, PGlcS achieved the best sensitivity of

64.62%, which obviously does not meet the requirements of exper-

imental biologists. The predictor O-GlcNAcscan gave the best

specificity of 92.45%, but also gave the worst sensitivity of

31.34%. The differences between sensitivity and specificity in

O-GlcNAcscan were mainly due to the predictor being trained on

Computational Prediction of Protein O-GlcNAc Modification 243