Biology Reference
In-Depth Information
MHC binders and 63 MHC non-binders. ANN and SVM outper-
formed the QM method. For training of both the ANN and SVM,
peptide sequences were changed to numerical vectors by encoding
sequences with 20-bin representation. The ANN performance was
72.2% accuracy, 73.2% sensitivity, and 71.2% specificity. The SVM
performance was 75.4%, 73.8%, and 77.0%, respectively, while the
QM method performance was 70.0%, 65.2%, and 74.9%, respectively.
Also, the ANN and SVM were used to perform consensus and com-
bined prediction of CTL epitopes. In consensus prediction, epitopes
predicted by both methods were considered “epitopes,” otherwise
they were considered as “non-epitopes.” In combined prediction, epi-
topes predicted by either method were considered as “epitopes.” In
addition, two models of such implementation can be used, i.e. in one
model the ANN can be used as the base method, while in the second
model the SVM is used as the base method. The CTLPred server
(available at http://www.imtech.res.in/raghava/ctlpred/) is based
on such an approach. They also developed the ComPred method
169
based on a hybrid method using ANN and QM filtering for the com-
bined prediction of MHC binding peptides or CTL epitopes. Using
this combined prediction approach, they achieved a sensitivity of
79.4% and a specificity of 88.4% for discriminating between T-cell epi-
topes and non-epitope MHC binders. Another of their programs,
ANNPred, uses a feed-forward back-propagation ANN for the pre-
diction of MHC binders to 30 MHC alleles. Both ComPred and
ANNPred are available at http://bioinformatics.uams.edu/mir-
ror/nhlapred/index.html.
Data encoding for T-cell epitope and MHC binding predic-
tions with ANN.
Representation of peptide sequences using 20-bin
numerical representation (0s and 1s), appears to provide an ANN
with sufficient discriminative power to predict (or classify) CTL epi-
topes
169,170
and MHC binding activities of peptides.
119,163,214
However,
this sparse encoding of the amino acids ignores their physicochemical
similarities or dissimilarities. Furthermore, in some cases, selection of
numerical representation schemes for signal processing is reported to
have statistically significant effects on ANN performance.
201,202,215
As
previously discussed, Milik
et al
. (1998) reported that an ANN derived
