262 Wuhan University Journal of Natural Sciences 2011, Vol.16 No.3
software (doc.guandang.net) is used to process the SVM classification. Here, the radial basis function (RBF) shows the best performance and is se-lected as the kernel function with the regularization pa-rameter C = 100 and the parameter g = 0.5:
k(u,v)=exp( gu v) (6)
2
Ac=∑TPi/N (10)
i=1
n
where n is the number of protein classes and N is the total number of protein sequences.
2 Results and Discussion
For the k-classification problem, we use the “one-
In statistical prediction, usually, a prediction is
against-one” strategy, which constructs k×(k 1)/2
evaluated by the re-substitution test and the k-fold
classifiers with each one trained with the data from two
cross-validation (k-CV) test and the jackknife test, re-different classes. For any protein X to be predicted, four
spectively. Among the three tests, the jackknife test is the
values of increments of diversity corresponding to the
most rigorous and objective one and there is no over-four standard measures of diversity are used as input pa-training problem in it, so, at present, its application is the
rameters to the SVM, and the expected output will be
most extensive. We calculate three indexes defined as
all-α, all-β, α/β, and α+β.
Eqs. (7), (8), and (9) under the jackknife test and use
1.2.4 Performance assessment
these indexes to examine the power of the ID-SVM ap-Three indexes are used to evaluate the prediction
proach. The test results are listed in Table 1, and the
accuracy: sensitivity (Sn), specificity (Sp), and the Mat-comparisons with other approaches for the same dataset
thew’s correlation coefficient (MCC) are defined as fol-are listed in Table 2.
lows:
Tables 1 and 2 show that ① the total success rate is
Sni=TPi/(TPi+FNi) (7)
96.6% by using the ID-SVM approach, in which the
Spi=TPi/(TPi+FPi)

(8) dipeptide amino acid composition of protein sequences
are used as the parameters. In order to look into the
MCCi=effect of parameter selection on the results, the 20 amino
acid occurrence frequencies are selected as information
parameters of sources of diversity; although the total
predictive success rate still reaches 76.5% by the ID-
(9) SVM algorithm, it is significantly less than that of using
where TPi is the number of true positives in ith structure dipeptide composition (Table 1). The data indicate that class, FPi is the number of false positives in ith class, FNi the dipeptide composition contains more information is the number of false negatives, and TNi is the number of protein structural class than that of the amino acid of true negatives in ith class. Sn represents the accuracy composition (Table 1). ② The ID-SVM approach is and Sp represents the reliability in procedure of predic-compared with the other ID-based methods. The accu-tion. MCCi is a single parameter characterizing the racy from ID-SVM is 22.6% higher than that of using matching extent between observed and predicted struc-ID method only[17] and also is 3.5% higher than that of
tural classes; its value will be in the range of from -1 to IDQD[17](Table 2). ③ Compared with other SVM-based 1, which equals to -1 for a perfectly reverse correlation,0 approaches[9-12], the total success rate is 11.3% higher than for a completely random prediction, and 1 for a perfect that is based on the PseAAC SVM by Chen et al[9] and prediction. also 21.6%, 22.1%, and 5.4% higher, respectively, than
The overall prediction accuracy (Ac) is given by the that of Vapnik SVM[10], LogitBoost SVM[11], and Binary- equation: tree SVM[12] (Table 2); ④ Compared with no- SVM-based
Table 1 The results of prediction performance for ID-SVM algorithm on 204 proteins using different
protein sample representation by the Jackknife test
Structural Using 400 dipeptides
classes S/% S/% MCC All-α 94.2 100.0 0.97 All-β 93.4 100.0 0.95
α/β 100.0 97.8 0.99 α+β 100.0 88.5 0.92
Using 20 amino acids
S/% S/% MCC 69.2 85.7 0.69 86.9 89.8 0.82 71.1 80.0 0.68 76.1 55.6 0.50
Prediction of Protein Se... 暂无评价 9页 免费...proteins using the Internet resources are described...structural diversity of proteins begins with ...
Prediction of protein protein interaction types using the decisionPrediction of protein protein interaction types using the decision隐藏>> Mathematical and ...
Prediction of Protein Topologies Using GIOHMMs and GRNNs Department of Information and Computer Science University of California, Irvine Irvine, CA 92697-3425...
Prediction of the secondary structural content might be considered the ?rst step in analyzing a new protein sequence and predicting its tertiary architecture ...
Structural Prediction of Membrane Protein: Application to Known Structures_专业资料。Joma fDogu iest(nulonhaUnvriyEg.E)V12d.o.7,No.(00)221 2 89 ...
feature selection using markov blanket induction for the prediction of protein...351 molecular structural properties that predict binding to thrombin (and thus...
Using the discovered associative feature we predict implicit protein interactions...theory based feature dimension reduction filter with fea- Prediction of ...
12页 1财富值 Prediction of Protein St... 暂无评价 5页 5财富值...更多>>We show that the rst-order theory of structural subtyping of non-...
12页 1财富值 Prediction of Protein St... 暂无评价 5页 5财富值... On the Theory of Structural Subtyping Viktor Kuncak and Martin Rinard ...
Sharan et al Box 2 Function prediction using the MRF method The Markov ...Data of a high diversity and dimensionality have been integrated using ...
热门文档